feat(pi052): condition low-level prompt on state + fix eval slowdown

- Inject discretized proprioceptive state (256 bins, pi05 format) into
  low-level action-conditioning prompts in both training
  (PI052TextTokenizerStep) and eval (_with_low_level_subtask_prompt),
  matching the recipe's documented "[images, subtask, state]" intent.
  Higher-level subtask/memory text streams stay state-free.
- Cache the loc-token tokenizer (_get_loc_tokenizer) instead of reloading
  it from disk on every _build_text_batch/select_message call (it ran
  twice per env per replan and dominated eval runtime).
- Add a KV cache to select_message decode (bit-identical output to the
  recompute path) to avoid O(n^2) generation.

Net: pi052 eval ~2.9 s/it -> ~0.1 s/it (~25x).
Co-authored-by: Cursor <cursoragent@cursor.com>

Makefile

@@ -178,3 +178,9 @@ test-smolvla-ete-eval:
 		--env.episode_length=5 \
 		--eval.n_episodes=1 \
 		--eval.batch_size=1
+
+# E2E annotation pipeline smoke test against a tiny in-memory fixture
+# dataset. Opt-in (not part of `make test-end-to-end`) and uses a stub VLM
+# backend, so it does not require a real model checkpoint or GPU.
+annotation-e2e:
+	uv run python -m tests.annotations.run_e2e_smoke

docs/source/_toctree.yml

@@ -9,6 +9,8 @@
 - sections:
   - local: il_robots
     title: Imitation Learning for Robots
+  - local: lelab
+    title: LeLab - Lerobot GUI
   - local: bring_your_own_policies
     title: Adding a Policy
   - local: integrate_hardware
@@ -43,6 +45,8 @@
     title: Language Columns and Recipes
   - local: tools
     title: Tools
+  - local: annotation_pipeline
+    title: Annotation Pipeline
   - local: video_encoding_parameters
     title: Video encoding parameters
   - local: streaming_video_encoding
@@ -59,6 +63,10 @@
     title: π₀-FAST (Pi0Fast)
   - local: pi05
     title: π₀.₅ (Pi05)
+  - local: molmoact2
+    title: MolmoAct2
+  - local: vla_jepa
+    title: VLA-JEPA
   - local: eo1
     title: EO-1
   - local: groot
@@ -73,6 +81,10 @@
 - sections:
   - local: sarm
     title: SARM
+  - local: robometer
+    title: ROBOMETER
+  - local: topreward
+    title: TOPReward
   title: "Reward Models"
 - sections:
   - local: inference

docs/source/act.mdx

@@ -79,17 +79,13 @@ If your local computer doesn't have a powerful GPU, you can utilize Google Colab
 Once training is complete, you can evaluate your ACT policy using the `lerobot-record` command with your trained policy. This will run inference and record evaluation episodes:
 
 ```bash
-lerobot-record \
-  --robot.type=so100_follower \
+lerobot-rollout \
+  --strategy.type=base \
+  --policy.path=${HF_USER}/act_policy \
+  --robot.type=so101_follower \
   --robot.port=/dev/ttyACM0 \
-  --robot.id=my_robot \
   --robot.cameras="{ front: {type: opencv, index_or_path: 0, width: 640, height: 480, fps: 30}}" \
   --display_data=true \
-  --dataset.repo_id=${HF_USER}/eval_act_your_dataset \
-  --dataset.num_episodes=10 \
-  --dataset.single_task="Your task description" \
-  --dataset.streaming_encoding=true \
-  --dataset.encoder_threads=2 \
-  # --dataset.camera_encoder.vcodec=auto \
-  --policy.path=${HF_USER}/act_policy
+  --task="Your task description" \ # can be skipped for ACT
+  --duration=60
 ```

docs/source/annotation_pipeline.mdx

@@ -0,0 +1,281 @@
+# Annotation Pipeline
+
+`lerobot-annotate` watches each episode's video with a vision-language
+model (VLM) and writes natural-language annotations back into your
+dataset. It fills the two language columns from the
+[Language Columns and Recipes](./language_and_recipes) page —
+`language_persistent` and `language_events` — straight into
+`data/chunk-*/file-*.parquet`.
+
+In short: point it at a LeRobot dataset, and it adds subtasks, plans,
+memory, interjections, speech, and visual Q&A that a policy can be
+trained on.
+
+## How it fits together
+
+```text
+  your dataset                  lerobot-annotate
+  (LeRobot v3.1)
+        │
+        ▼
+  ┌─────────────────────────────────────────────────────┐
+  │                    read episodes                     │
+  └──────────────────────────┬──────────────────────────┘
+                             │
+        ┌────────────────────┼────────────────────┐
+        ▼                    ▼                     ▼
+  ┌──────────┐      ┌───────────────┐        ┌──────────┐       one shared Qwen-VL
+  │   plan   │      │ interjections │        │   vqa    │  ◀──   server (vLLM, OpenAI
+  └────┬─────┘      └───────┬───────┘        └────┬─────┘        API) drives all three
+       └────────────────────┼─────────────────────┘
+                            │   each module stages raw JSONL
+                            ▼   into .annotate_staging/
+                  ┌─────────────────┐
+                  │    validator    │  ◀──  checks everything
+                  └────────┬────────┘
+                           ▼
+                  ┌─────────────────┐
+                  │     writer      │
+                  └────────┬────────┘
+                           ▼
+              data/chunk-*/file-*.parquet
+              (+ meta/info.json tools)
+```
+
+Three modules (`plan`, `interjections`, `vqa`) all talk to **one** shared
+VLM. Each module stages its output to disk, a validator checks it, and a
+single writer rewrites the dataset shards in place.
+
+## What the pipeline produces
+
+Each module emits a few kinds of annotation ("styles"), routed to one of
+the two language columns:
+
+| Style / atom                                | Column                | Module          |
+| ------------------------------------------- | --------------------- | --------------- |
+| `subtask` (Pi0.7-style "how, not what")     | `language_persistent` | `plan`          |
+| `plan` (initial + refresh on interjection)  | `language_persistent` | `plan`          |
+| `memory` (MEM-style compression)            | `language_persistent` | `plan`          |
+| `task_aug` (rephrasings of the task)        | `language_persistent` | `plan`          |
+| `interjection`                              | `language_events`     | `interjections` |
+| speech tool-call atom (`style=null`, `say`) | `language_events`     | `interjections` |
+| `vqa` (user / assistant pair)               | `language_events`     | `vqa`           |
+
+### How subtasks are generated
+
+The `plan` module doesn't ask the VLM for subtasks in one shot. Instead
+it uses a two-step **describe → segment** flow:
+
+1. **Describe** — the VLM narrates only what it actually sees in the
+   chosen camera (no guessing about the task).
+2. **Segment** — that description is fed back in, and the VLM splits the
+   episode into consecutive atomic subtasks.
+
+The resulting spans are then stitched into a gap-free, full-episode
+cover, so **every frame has exactly one active subtask**. See
+[`run_hf_job.py`](https://github.com/huggingface/lerobot/blob/main/examples/annotations/run_hf_job.py)
+for the production settings (single camera, embedded frames, windowed
+subtask generation).
+
+### Tools
+
+The writer does **not** add a `tools` column to the parquet. The tool
+catalog lives in `meta/info.json["tools"]` instead (see [Tools](./tools)).
+After every run, the pipeline makes sure the canonical `say` schema is in
+that list, keeping any tools you declared beforehand.
+
+Want to add your own tool? Edit `meta/info.json["tools"]` directly — the
+pipeline preserves whatever is already there. That makes the tool visible
+to the chat template, so the model can learn to _generate_ the call. The
+runtime layer that actually _executes_ a generated call (the `Tool`
+protocol / `TOOL_REGISTRY` under `src/lerobot/tools/`) is not part of
+this PR — the [Tools](./tools) doc marks those pieces as
+not-yet-implemented.
+
+## Running on Hugging Face Jobs
+
+Annotation runs on [Hugging Face Jobs](https://huggingface.co/docs/hub/en/jobs).
+The repo ships a launcher script you copy and tweak for your dataset:
+
+```bash
+HF_TOKEN=hf_... uv run python examples/annotations/run_hf_job.py
+```
+
+[`run_hf_job.py`](https://github.com/huggingface/lerobot/blob/main/examples/annotations/run_hf_job.py)
+starts a single-GPU `h200` job (bump it to `h200x4` for big datasets)
+that:
+
+1. installs `lerobot` (from `main`) plus the annotation extras,
+2. boots one vLLM server per GPU (using the `vllm/vllm-openai` image) and
+   drives it over the OpenAI-compatible API,
+3. runs the `plan` / `interjections` / `vqa` modules across the dataset
+   with `lerobot-annotate`,
+4. with `--push_to_hub=true`, uploads the result to `--new_repo_id` (or
+   back to `--repo_id` in place if you leave that unset).
+
+To use a different dataset, model, or hub repo, edit the `CMD` block in
+the script. Every flag there maps directly to a `lerobot-annotate` flag
+(run `lerobot-annotate --help` for the full list).
+
+## Key options
+
+These are the flags you'll reach for most often. Run
+`lerobot-annotate --help` for everything else; the defaults are tuned for
+short manipulation episodes.
+
+### Dataset in / out
+
+| Flag              | Default | What it does                                                            |
+| ----------------- | ------- | ----------------------------------------------------------------------- |
+| `--repo_id`       | —       | Hub dataset to annotate (downloaded if `--root` unset).                 |
+| `--root`          | —       | Annotate a local dataset directory instead.                             |
+| `--new_repo_id`   | —       | Push the result to a new repo (leaves the source repo untouched).       |
+| `--push_to_hub`   | `false` | Upload after annotating (to `--new_repo_id`, else back to `--repo_id`). |
+| `--only_episodes` | all     | Annotate just these episode indices (handy for a test run).             |
+| `--seed`          | `1729`  | Seeds the RNGs that pick interjection timestamps + VQA question types.  |
+
+### Which modules run
+
+Every module is on by default and can be toggled independently (set to
+`false` to skip it, e.g. to iterate on one module at a time):
+
+| Flag                      | Default | Turns off                           |
+| ------------------------- | ------- | ----------------------------------- |
+| `--plan.enabled`          | `true`  | subtasks + plan + memory + task_aug |
+| `--interjections.enabled` | `true`  | interjections + speech atoms        |
+| `--vqa.enabled`           | `true`  | the VQA pairs                       |
+
+### The VLM (`--vlm.*`)
+
+| Flag                       | Default            | What it does                                                                        |
+| -------------------------- | ------------------ | ----------------------------------------------------------------------------------- |
+| `--vlm.model_id`           | `Qwen/Qwen3.6-27B` | The model to serve and prompt.                                                      |
+| `--vlm.camera_key`         | first `images.*`   | Which camera every prompt is grounded on.                                           |
+| `--vlm.serve_command`      | auto               | The exact `vllm serve …` command (set TP size, GPU memory, `--max-model-len` here). |
+| `--vlm.parallel_servers`   | `1`                | Independent servers for round-robin routing (one per GPU).                          |
+| `--vlm.num_gpus`           | `0`                | GPUs per server (`0` = one each).                                                   |
+| `--vlm.client_concurrency` | `16`               | In-flight requests across all servers.                                              |
+| `--vlm.max_new_tokens`     | `512`              | Generation cap per call.                                                            |
+| `--vlm.temperature`        | `0.2`              | Sampling temperature.                                                               |
+
+### Subtasks / plan / memory (`--plan.*`)
+
+| Flag                            | Default    | What it does                                                                                                              |
+| ------------------------------- | ---------- | ------------------------------------------------------------------------------------------------------------------------- |
+| `--plan.frames_per_second`      | `1.0`      | How densely the episode video is sampled.                                                                                 |
+| `--plan.max_video_frames`       | `32`       | Hard cap on frames per call (context-budget guard — don't exceed ~32 for a 32k context).                                  |
+| `--plan.subtask_window_seconds` | `0`        | Split long episodes into fixed windows for constant frame density (`0` = whole episode).                                  |
+| `--plan.plan_max_steps`         | `8`        | Upper bound on subtasks per episode.                                                                                      |
+| `--plan.subtask_describe_first` | `true`     | Run the describe→segment grounding pass (best subtask quality; +1 call/episode).                                          |
+| `--plan.emit_plan`              | `true`     | Emit the numbered `plan` rows (`false` = subtasks + memory only).                                                         |
+| `--plan.n_task_rephrasings`     | `10`       | How many `task_aug` rephrasings to emit (`0` disables).                                                                   |
+| `--plan.derive_task_from_video` | `if_short` | Use the dataset task as-is (`off`), only when it's missing/short (`if_short`), or always re-derive from video (`always`). |
+| `--plan.use_video_url`          | `false`    | Send a server-side video clip instead of embedded frames.                                                                 |
+
+### Interjections + VQA
+
+| Flag                                            | Default | What it does                                               |
+| ----------------------------------------------- | ------- | ---------------------------------------------------------- |
+| `--interjections.max_interjections_per_episode` | `3`     | Cap on interjection/speech pairs per episode.              |
+| `--vqa.vqa_emission_hz`                         | `1.0`   | How often VQA pairs are emitted.                           |
+| `--vqa.restrict_to_default_camera`              | `false` | Ground VQA only on `--vlm.camera_key` (else every camera). |
+| `--executor.episode_parallelism`                | `16`    | Episodes processed concurrently within each phase.         |
+
+## Contributing new modules
+
+The pipeline is built to grow, and **contributions are very welcome** —
+a brand-new module (say, trajectory traces or affordances), a new prompt
+template, a smarter grounding flow, or quality fixes to the existing
+`plan` / `interjections` / `vqa` modules.
+
+Every module lives under
+`src/lerobot/annotations/steerable_pipeline/modules/`, shares the VLM
+client and the keyframe cache, writes its raw output to the staging
+tree, and plugs into the executor as its own phase. Got an idea? Open an
+issue or PR on [the repo](https://github.com/huggingface/lerobot).
+
+## How recipes consume the output
+
+The annotations are meant to be read by recipes (see
+[Language Columns and Recipes](./language_and_recipes)). Typically:
+
+- low-level / high-level / memory-update branches read
+  `subtask` / `plan` / `memory` from `language_persistent`.
+- an interjection-response branch reads `interjection` events plus the
+  paired speech atom (merged into one assistant turn via `tool_calls_from`)
+  and the matching `plan` refresh at the same timestamp.
+- a VQA branch reads the `(vqa, user)` and `(vqa, assistant)` pairs from
+  `language_events`.
+
+## Why state and events are split
+
+Two ideas shape the design:
+
+1. **Persistent state vs. exact events.** Persistent rows (`subtask`,
+   `plan`, `memory`) apply to the whole episode and answer "what's true
+   right now?". Event rows (`interjection`, `vqa`, speech) appear only on
+   the one frame whose timestamp matches. Timestamps are copied straight
+   from the source parquet — never recomputed in floating point.
+2. **One VLM pass.** All three modules share a single VLM client (the
+   OpenAI-compatible client talking to the job's vLLM server), so you pay
+   for one model load per dataset, not three.
+
+## Re-running a single module
+
+Each module stages its raw output to
+`<root>/.annotate_staging/episode_{N:06d}/<module>.jsonl`. This makes
+prompt iteration cheap: re-running one module overwrites only its own
+JSONL, then the writer recomposes the final parquet. Disable modules you
+don't want with `--plan.enabled=false` (and likewise
+`--interjections.enabled` / `--vqa.enabled`) to test one at a time.
+
+## What the validator checks
+
+Before the writer runs, `StagingValidator` confirms:
+
+- every event row lands exactly on a real frame timestamp;
+- no speech / interjection pairs are left orphaned;
+- `plan` is refreshed at every interjection timestamp;
+- `memory` rows fall on subtask boundaries (a warning, not an error);
+- each VQA assistant `content` is valid JSON in one of the
+  bbox / keypoint / count / attribute / spatial shapes;
+- every row goes to the column chosen by `column_for_style(style)`.
+
+Any error aborts the writer. Pass `--skip_validation=true` to override
+while debugging.
+
+## Where each module's ideas come from
+
+- **`plan` — subtasks.** Hi Robot ([Shi 2025](https://arxiv.org/abs/2502.19417))
+  for atom granularity ("pick up one piece of lettuce", "place bowl to
+  box"); Pi0.7 ([Physical Intelligence 2025](https://pi.website/pi07))
+  for "how, not what" detail.
+- **`plan` — memory.** MEM ([Torne 2026](https://arxiv.org/abs/2603.03596)):
+  keep only the minimal relevant information — preserve outcomes, drop
+  specific attributes.
+- **`interjections`.** Hi Robot's scenario taxonomy: negative task,
+  situated correction, specific constraint, preference. Speech is a
+  tool-call-only atom
+  (`tool_calls=[{type:function, function:{name:"say", arguments:{text:...}}}]`).
+- **`vqa`.** ECoT ([Zawalski 2024](https://arxiv.org/abs/2407.08693)) for
+  grounded features (pixel bounding boxes `[x_min, y_min, x_max, y_max]`,
+  keypoints) and Steerable VLA Policies
+  ([Zhao 2025](https://arxiv.org/abs/2509.07626)) for multi-abstraction
+  grounding. Pi0.7 also grounds answers across abstraction levels.
+
+When improving a module, tweak its prompt template in
+`src/lerobot/annotations/steerable_pipeline/prompts/` rather than
+rewriting from scratch.
+
+## Roughly how much it costs
+
+Per episode, the pipeline makes about `max_steps` plan calls,
+`max_interjections_per_episode` interjection calls, and
+`vqa_emission_hz × episode_seconds` VQA calls. With the defaults (8
+subtasks, 1 interjection, 1 Hz × 3 pairs) on a 30-second episode, that's
+~50 VLM calls.
+
+Storage stays small: `language_persistent` is at most tens of KB per
+episode (parquet dictionary-encodes the one entry that repeats across
+frames), and `language_events` is empty on most frames — its size scales
+with the number of emissions, not `num_frames × num_emissions`.

docs/source/groot.mdx

@@ -105,10 +105,12 @@ These results demonstrate GR00T's strong generalization capabilities across dive
 
 ### Evaluate in your hardware setup
 
-Once you have trained your model using your parameters you can run inference in your downstream task. Follow the instructions in [Imitation Learning for Robots](./il_robots). For example:
+Once you have trained your model using your parameters you can run inference in your downstream task. Follow the instructions in [Policy Deployment (lerobot-rollout)](./inference). For example:
 
 ```bash
-lerobot-record \
+lerobot-rollout\
+  --strategy.type=sentry \
+  --strategy.upload_every_n_episodes=5 \
   --robot.type=bi_so_follower \
   --robot.left_arm_port=/dev/ttyACM1 \
   --robot.right_arm_port=/dev/ttyACM0 \
@@ -119,14 +121,12 @@ lerobot-record \
   }' \
   --display_data=true \
   --dataset.repo_id=<user>/eval_groot-bimanual  \
-  --dataset.num_episodes=10 \
   --dataset.single_task="Grab and handover the red cube to the other arm" \
   --dataset.streaming_encoding=true \
   --dataset.encoder_threads=2 \
   # --dataset.camera_encoder.vcodec=auto \
   --policy.path=<user>/groot-bimanual \ # your trained model
-  --dataset.episode_time_s=30 \
-  --dataset.reset_time_s=10
+  --duration=600
 ```
 
 ## License

docs/source/il_robots.mdx

@@ -68,13 +68,13 @@ from lerobot.teleoperators.so_leader import SO101Leader, SO101LeaderConfig
 from lerobot.robots.so_follower import SO101Follower, SO101FollowerConfig
 
 robot_config = SO101FollowerConfig(
-    port="/dev/tty.usbmodem58760431541",
-    id="my_red_robot_arm",
+    port="/dev/tty.usbmodem5AB90687491",
+    id="my_follower_arm",
 )
 
 teleop_config = SO101LeaderConfig(
-    port="/dev/tty.usbmodem58760431551",
-    id="my_blue_leader_arm",
+    port="/dev/tty.usbmodem5AB90689011",
+    id="my_leader_arm",
 )
 
 robot = SO101Follower(robot_config)
@@ -108,13 +108,13 @@ With `rerun`, you can teleoperate again while simultaneously visualizing the cam
 <hfoption id="Command">
 ```bash
 lerobot-teleoperate \
-    --robot.type=koch_follower \
-    --robot.port=/dev/tty.usbmodem58760431541 \
-    --robot.id=my_awesome_follower_arm \
-    --robot.cameras="{ front: {type: opencv, index_or_path: 0, width: 1920, height: 1080, fps: 30}}" \
-    --teleop.type=koch_leader \
-    --teleop.port=/dev/tty.usbmodem58760431551 \
-    --teleop.id=my_awesome_leader_arm \
+    --robot.type=so101_follower \
+    --robot.port=/dev/tty.usbmodem5AB90687491 \
+    --robot.id=my_follower_arm \
+    --robot.cameras="{front: {type: opencv, index_or_path: 0, width: 640, height: 480, fps: 30}}" \
+    --teleop.type=so101_leader \
+    --teleop.port=/dev/tty.usbmodem5AB90689011 \
+    --teleop.id=my_leader_arm \
     --display_data=true
 ```
 </hfoption>
@@ -122,34 +122,48 @@ lerobot-teleoperate \
 
 <!-- prettier-ignore-start -->
 ```python
+import time
+from lerobot.teleoperators.so_leader import SO101Leader, SO101LeaderConfig
+from lerobot.robots.so_follower import SO101Follower, SO101FollowerConfig
 from lerobot.cameras.opencv import OpenCVCameraConfig
-from lerobot.teleoperators.koch_leader import KochLeader, KochLeaderConfig
-from lerobot.robots.koch_follower import KochFollower, KochFollowerConfig
+from lerobot.utils.visualization_utils import init_rerun, log_rerun_data, shutdown_rerun
 
-camera_config = {
-    "front": OpenCVCameraConfig(index_or_path=0, width=1920, height=1080, fps=30)
-}
-
-robot_config = KochFollowerConfig(
-    port="/dev/tty.usbmodem585A0076841",
-    id="my_red_robot_arm",
-    cameras=camera_config
+robot_config = SO101FollowerConfig(
+    port="/dev/tty.usbmodem5AB90687491",
+    id="my_follower_arm",
+    cameras={
+        "wrist": OpenCVCameraConfig(index_or_path=0, width=640, height=480, fps=30),
+        "top": OpenCVCameraConfig(index_or_path=1, width=640, height=480, fps=30)
+    }
 )
 
-teleop_config = KochLeaderConfig(
-    port="/dev/tty.usbmodem58760431551",
-    id="my_blue_leader_arm",
+teleop_config = SO101LeaderConfig(
+    port="/dev/tty.usbmodem5AB90689011",
+    id="my_leader_arm",
 )
 
-robot = KochFollower(robot_config)
-teleop_device = KochLeader(teleop_config)
+init_rerun(session_name="teleoperation")
+
+robot = SO101Follower(robot_config)
+teleop_device = SO101Leader(teleop_config)
 robot.connect()
 teleop_device.connect()
 
+TARGET_HZ = 30
+TIME_PER_FRAME = 1.0 / TARGET_HZ
+
 while True:
+    start_time = time.perf_counter()
+
     observation = robot.get_observation()
     action = teleop_device.get_action()
     robot.send_action(action)
+    log_rerun_data(observation=observation, action=action)
+
+    elapsed_time = time.perf_counter() - start_time
+    sleep_time = TIME_PER_FRAME - elapsed_time
+    if sleep_time > 0:
+        time.sleep(sleep_time)
 ```
 <!-- prettier-ignore-end -->
 
@@ -202,10 +216,11 @@ lerobot-record \
 <!-- prettier-ignore-start -->
 ```python
 from lerobot.cameras.opencv import OpenCVCameraConfig
-from lerobot.datasets import LeRobotDataset
+from lerobot.datasets.lerobot_dataset import LeRobotDataset
 from lerobot.utils.feature_utils import hw_to_dataset_features
-from lerobot.robots.so_follower import SO100Follower, SO100FollowerConfig
-from lerobot.teleoperators.so_leader import SO100Leader, SO100LeaderConfig
+from lerobot.robots.so_follower import SO101Follower, SO101FollowerConfig
+from lerobot.teleoperators.so_leader.config_so_leader import SO101LeaderConfig
+from lerobot.teleoperators.so_leader.so_leader import SO101Leader
 from lerobot.common.control_utils import init_keyboard_listener
 from lerobot.utils.utils import log_say
 from lerobot.utils.visualization_utils import init_rerun
@@ -218,71 +233,56 @@ EPISODE_TIME_SEC = 60
 RESET_TIME_SEC = 10
 TASK_DESCRIPTION = "My task description"
 
-# Create robot configuration
-robot_config = SO100FollowerConfig(
-    id="my_awesome_follower_arm",
-    cameras={
-        "front": OpenCVCameraConfig(index_or_path=0, width=640, height=480, fps=FPS) # Optional: fourcc="MJPG" for troubleshooting OpenCV async error.
-    },
-    port="/dev/tty.usbmodem58760434471",
-)
-
-teleop_config = SO100LeaderConfig(
-    id="my_awesome_leader_arm",
-    port="/dev/tty.usbmodem585A0077581",
-)
-
-# Initialize the robot and teleoperator
-robot = SO100Follower(robot_config)
-teleop = SO100Leader(teleop_config)
-
-# Configure the dataset features
-action_features = hw_to_dataset_features(robot.action_features, "action")
-obs_features = hw_to_dataset_features(robot.observation_features, "observation")
-dataset_features = {**action_features, **obs_features}
-
-# Create the dataset
-dataset = LeRobotDataset.create(
-    repo_id="<hf_username>/<dataset_repo_id>",
-    fps=FPS,
-    features=dataset_features,
-    robot_type=robot.name,
-    use_videos=True,
-    image_writer_threads=4,
-)
-
-# Initialize the keyboard listener and rerun visualization
-_, events = init_keyboard_listener()
-init_rerun(session_name="recording")
-
-# Connect the robot and teleoperator
-robot.connect()
-teleop.connect()
-
-# Create the required processors
-teleop_action_processor, robot_action_processor, robot_observation_processor = make_default_processors()
-
-episode_idx = 0
-while episode_idx < NUM_EPISODES and not events["stop_recording"]:
-    log_say(f"Recording episode {episode_idx + 1} of {NUM_EPISODES}")
-
-    record_loop(
-        robot=robot,
-        events=events,
-        fps=FPS,
-        teleop_action_processor=teleop_action_processor,
-        robot_action_processor=robot_action_processor,
-        robot_observation_processor=robot_observation_processor,
-        teleop=teleop,
-        dataset=dataset,
-        control_time_s=EPISODE_TIME_SEC,
-        single_task=TASK_DESCRIPTION,
-        display_data=True,
+def main():
+    # Create robot configuration
+    robot_config = SO101FollowerConfig(
+        port="/dev/tty.usbmodem5AB90687491",
+        id="my_follower_arm",
+        cameras={
+            "wrist": OpenCVCameraConfig(index_or_path=0, width=640, height=480, fps=30),
+            "top": OpenCVCameraConfig(index_or_path=1, width=640, height=480, fps=30)
+        }
     )
 
-    # Reset the environment if not stopping or re-recording
-    if not events["stop_recording"] and (episode_idx < NUM_EPISODES - 1 or events["rerecord_episode"]):
-        log_say("Reset the environment")
+    teleop_config = SO101LeaderConfig(
+        port="/dev/tty.usbmodem5AB90689011",
+        id="my_leader_arm",
+    )
+
+    # Initialize the robot and teleoperator
+    robot = SO101Follower(robot_config)
+    teleop = SO101Leader(teleop_config)
+
+    # Configure the dataset features
+    action_features = hw_to_dataset_features(robot.action_features, "action")
+    obs_features = hw_to_dataset_features(robot.observation_features, "observation")
+    dataset_features = {**action_features, **obs_features}
+
+    # Create the dataset
+    dataset = LeRobotDataset.create(
+        repo_id="<hf_username>/<dataset_repo_id>",
+        fps=FPS,
+        features=dataset_features,
+        robot_type=robot.name,
+        use_videos=True,
+        image_writer_threads=4,
+    )
+
+    # Initialize the keyboard listener and rerun visualization
+    _, events = init_keyboard_listener()
+    init_rerun(session_name="recording")
+
+    # Connect the robot and teleoperator
+    robot.connect()
+    teleop.connect()
+
+    # Create the required processors
+    teleop_action_processor, robot_action_processor, robot_observation_processor = make_default_processors()
+
+    episode_idx = 0
+    while episode_idx < NUM_EPISODES and not events["stop_recording"]:
+        log_say(f"Recording episode {episode_idx + 1} of {NUM_EPISODES}")
+
         record_loop(
             robot=robot,
             events=events,
@@ -291,26 +291,50 @@ while episode_idx < NUM_EPISODES and not events["stop_recording"]:
             robot_action_processor=robot_action_processor,
             robot_observation_processor=robot_observation_processor,
             teleop=teleop,
-            control_time_s=RESET_TIME_SEC,
+            dataset=dataset,
+            control_time_s=EPISODE_TIME_SEC,
             single_task=TASK_DESCRIPTION,
             display_data=True,
         )
 
-    if events["rerecord_episode"]:
-        log_say("Re-recording episode")
-        events["rerecord_episode"] = False
-        events["exit_early"] = False
-        dataset.clear_episode_buffer()
-        continue
+        # Reset the environment if not stopping or re-recording
+        if not events["stop_recording"] and (episode_idx < NUM_EPISODES - 1 or events["rerecord_episode"]):
+            log_say("Reset the environment")
+            record_loop(
+                robot=robot,
+                events=events,
+                fps=FPS,
+                teleop_action_processor=teleop_action_processor,
+                robot_action_processor=robot_action_processor,
+                robot_observation_processor=robot_observation_processor,
+                teleop=teleop,
+                control_time_s=RESET_TIME_SEC,
+                single_task=TASK_DESCRIPTION,
+                display_data=True,
+            )
 
-    dataset.save_episode()
-    episode_idx += 1
+        if events["rerecord_episode"]:
+            log_say("Re-recording episode")
+            events["rerecord_episode"] = False
+            events["exit_early"] = False
+            dataset.clear_episode_buffer()
+            continue
 
-# Clean up
-log_say("Stop recording")
-robot.disconnect()
-teleop.disconnect()
-dataset.push_to_hub()
+        dataset.save_episode()
+        episode_idx += 1
+
+    # finalize dataset
+    log_say("Finalizing dataset...")
+    dataset.finalize()
+    # Clean up
+    log_say("Stop recording")
+    robot.disconnect()
+    teleop.disconnect()
+    dataset.push_to_hub()
+
+
+if __name__ == "__main__":
+    main()
 ```
 <!-- prettier-ignore-end -->
 
@@ -348,7 +372,7 @@ The `record` function provides a suite of tools for capturing and managing data
 ##### 2. Checkpointing and Resuming
 
 - Checkpoints are automatically created during recording.
-- If an issue occurs, you can resume by re-running the same command with `--resume=true`. When resuming a recording, `--dataset.num_episodes` must be set to the **number of additional episodes to be recorded**, and not to the targeted total number of episodes in the dataset !
+- If an issue occurs or you want to record additional episodes in the same dataset, you can resume by re-running the same command with `--resume=true`. When resuming a recording, `--dataset.num_episodes` must be set to the **number of additional episodes to be recorded**, and not to the targeted total number of episodes in the dataset! Make sure that you also set `--dataset.root="local_path"`, it's a local path to save the new part of the dataset and is required to resume.
 - To start recording from scratch, **manually delete** the dataset directory.
 
 ##### 3. Recording Parameters
@@ -422,7 +446,7 @@ from lerobot.utils.utils import log_say
 
 episode_idx = 0
 
-robot_config = SO100FollowerConfig(port="/dev/tty.usbmodem58760434471", id="my_awesome_follower_arm")
+robot_config = SO100FollowerConfig(port="/dev/tty.usbmodem5AB90687491", id="my_follower_arm")
 
 robot = SO100Follower(robot_config)
 robot.connect()
@@ -490,6 +514,83 @@ Additionally you can provide extra `tags` or specify a `license` for your model
 
 If your local computer doesn't have a powerful GPU you could utilize Google Colab to train your model by following the [ACT training notebook](./notebooks#training-act).
 
+#### Train using Hugging Face Jobs
+
+Hugging Face jobs let's you easily select hardware and run the training in the cloud. So if you don't have a powerful GPU or you need more VRAM or just want to train a model much faster use HF Jobs! It's pay as you go and you simply pay for each second of use, you can see the pricing and additional information [here](https://huggingface.co/docs/hub/jobs).
+
+To run the training use this command:
+
+<hfoptions id="train_with_hf_jobs">
+<hfoption id="Command">
+```bash
+hf jobs run \
+  --flavor a10g-small \
+  --timeout 4h \
+  --secrets HF_TOKEN \
+  huggingface/lerobot-gpu:latest \
+  -- \
+  python -m lerobot.scripts.lerobot_train \
+    --dataset.repo_id=username/dataset \
+    --policy.type=act \
+    --steps=5000 \
+    --batch_size=16 \
+    --policy.device=cuda \
+    --policy.repo_id=username/your_policy \
+    --log_freq=100
+```
+</hfoption>
+<hfoption id="API example">
+
+<!-- prettier-ignore-start -->
+```python
+from huggingface_hub import run_job, get_token
+
+run_name = "act_so101_hf_jobs"
+dataset_id = "username/dataset"
+user_hub_id = "username"
+
+command_args = [
+    "python", "-m", "lerobot.scripts.lerobot_train",
+    "--dataset.repo_id", dataset_id,
+    "--policy.type", "act",
+    "--steps", "5000",
+    "--batch_size", "16",
+    "--num_workers", "4",
+    "--policy.device", "cuda",
+    "--log_freq", "100",
+    "--save_freq", "1000",
+    "--save_checkpoint", "true",
+    "--wandb.enable", "false",
+    "--policy.repo_id", f"{user_hub_id}/{run_name}"
+]
+
+print(f"Submitting job '{run_name}' to Hugging Face Infrastructure...")
+
+job_info = run_job(
+    image="huggingface/lerobot-gpu:latest",
+    command=command_args,
+    flavor="a10g-small",
+    timeout="4h",
+    secrets={"HF_TOKEN": get_token()}
+)
+
+print("\n🚀 Job successfully launched!")
+print(f"🔹 Job ID: {job_info.id}")
+print(f"🔗 Live UI Dashboard & Logs: {job_info.url}")
+```
+<!-- prettier-ignore-end -->
+
+</hfoption>
+</hfoptions>
+
+You can modify the `--flavor` to use different hardware, for example: `t4-small`, `a100-large`, `h200`. Use `hf jobs hardware` to see the full list with pricing.
+Depending on the model you want to train and the hardware you selected you can also modify the `--batch_size` and `--number_of_workers`.
+For longer training sessions increase the timeout.
+
+Once the training is started you can go to [Jobs](https://huggingface.co/settings/jobs) and see if your jobs is running as well as all the outputs. Sometimes it takes a few minutes to schedule your job so be patient.
+
+After training the model will be pushed to hub and you can use it as any other model with LeRobot.
+
 #### Upload policy checkpoints
 
 Once training is done, upload the latest checkpoint with:
@@ -546,5 +647,6 @@ The `--strategy.type` flag selects the execution mode:
 - `sentry`: Continuous recording with auto-upload (useful for large-scale evaluation)
 - `highlight`: Ring buffer recording with keystroke save (useful for capturing interesting events)
 - `dagger`: Human-in-the-loop data collection (see [HIL Data Collection](./hil_data_collection))
+- `episodic`: Episode-oriented policy recording with reset phases between episodes
 
 All strategies support `--inference.type=rtc` for smooth execution with slow VLA models (Pi0, Pi0.5, SmolVLA).

docs/source/inference.mdx

@@ -157,6 +157,44 @@ Foot pedal input is also supported via `--strategy.input_device=pedal`. Configur
 | `--strategy.input_device`            | Input device: `keyboard` or `pedal` (default: keyboard) |
 | `--teleop.type`                      | **Required.** Teleoperator type                         |
 
+### Episodic (`--strategy.type=episodic`)
+
+Episode-oriented recording that mirrors the behavior of `lerobot-record`. The policy drives the robot for each episode; an optional teleoperator can drive the robot during the reset phase between episodes.
+
+```bash
+lerobot-rollout \
+    --strategy.type=episodic \
+    --policy.path=${HF_USER}/my_policy \
+    --robot.type=so100_follower \
+    --robot.port=/dev/ttyACM0 \
+    --teleop.type=so100_leader \
+    --teleop.port=/dev/ttyACM1 \
+    --dataset.repo_id=${HF_USER}/my_eval_data \
+    --dataset.num_episodes=20 \
+    --dataset.episode_time_s=30 \
+    --dataset.reset_time_s=10 \
+    --dataset.single_task="Pick up the red cube"
+```
+
+Teleop is optional — if omitted the robot holds its position during the reset phase.
+
+**Keyboard controls:**
+
+| Key         | Action                           |
+| ----------- | -------------------------------- |
+| `→` (right) | End the current episode early    |
+| `←` (left)  | Discard episode and re-record it |
+| `ESC`       | Stop the recording session       |
+
+| Flag                                            | Description                                                                |
+| ----------------------------------------------- | -------------------------------------------------------------------------- |
+| `--dataset.num_episodes`                        | Number of episodes to record                                               |
+| `--dataset.episode_time_s`                      | Duration of each recording episode in seconds                              |
+| `--dataset.reset_time_s`                        | Duration of the reset phase between episodes in seconds                    |
+| `--teleop.type`                                 | Optional. Teleoperator to drive the robot during resets                    |
+| `--strategy.reset_to_initial_position`          | Whether to reset the robot to its initial position between episodes        |
+| `--strategy.smooth_leader_to_follower_handover` | Whether to turn on or off the leader -> follower smooth handover behavior. |
+
 ---
 
 ## Inference Backends

docs/source/language_and_recipes.mdx

@@ -141,6 +141,11 @@ sample["target_message_indices"]
 
 The renderer does not apply a tokenizer chat template. Policy processors decide how to serialize the messages for their backbone, which keeps the same dataset usable across SmolVLA, Pi0.5, and any future VLM that expects OpenAI-style chat messages.
 
+## Blends
+
+Blend recipes select one weighted sub-recipe deterministically from the sample index.
+`recipes/subtasks_vqa.yaml` trains the core blend — high-level subtask prediction, low-level execution, and VQA. `recipes/subtask_mem_vqa_speech.yaml` is the fuller variant that also adds memory updates and spoken interjection responses.
+
 ## Graceful absence
 
 If both language columns are missing, `None`, or empty, `RenderMessagesStep` is a no-op.

docs/source/lelab.mdx

@@ -0,0 +1,29 @@
+# LeLab - LeRobot Guide
+
+LeLab is a graphical user interface built on top of the LeRobot library, designed to make robotics accessible without needing to memorize CLI commands. From a single app you can configure your robot, teleoperate it, collect datasets, train policies locally or on cloud GPUs via HF Jobs, and deploy trained models back onto your robot. It's the easiest way to go from an unboxed SO-101 to a working policy, and a great companion for anyone learning the LeRobot workflow. Source code and issues live on GitHub: [huggingface/leLab](https://github.com/huggingface/leLab).
+
+> [!TIP]
+> For now LeLab is compatible only with SO-ARM101
+
+<Youtube id="VqyKUuW9V1g" />
+
+### Installation
+
+Requires [`uv`](https://docs.astral.sh/uv/getting-started/installation/). Install and launch in one command:
+
+```
+uv tool install git+https://github.com/huggingface/leLab.git && lelab
+```
+
+After install, run `lelab` from your terminal anytime to start the app.
+
+### Features
+
+- **Add robots** — Select arm type (leader/follower), calibrate each joint from the middle position, and attach cameras.
+- **Teleoperation** — Control the follower arm with the leader and see a live 3D visualization of the arms.
+- **Dataset recording** — Define a task description, number of episodes, and episode/reset durations. Press spacebar to advance between episodes. 30+ episodes recommended.
+- **Local training** — Train a policy directly on your own machine with a selected dataset, policy type, batch size, and step count.
+- **Cloud training with HF Jobs** — Train on powerful GPUs via [HF Jobs](https://huggingface.co/docs/huggingface_hub/en/guides/jobs) with transparent pricing. Run `hf auth login` first. See the [Compute HW Guide](hardware_guide) for hardware/batch size tips.
+- **Training visualization** — Watch progress live in the app, with checkpoints saved automatically.
+- **Run trained policies** — Pick any model from your jobs list and run inference on your robot with one click.
+- **Use community datasets** — Provide any Hugging Face dataset ID to train on datasets you didn't record yourself.

docs/source/lerobot-dataset-v3.mdx

@@ -275,7 +275,7 @@ A converter aggregates per‑episode files into larger shards and writes episode
 pip install "https://github.com/huggingface/lerobot/archive/33cad37054c2b594ceba57463e8f11ee374fa93c.zip"
 
 # Convert an existing v2.1 dataset hosted on the Hub:
-python -m lerobot.datasets.v30.convert_dataset_v21_to_v30 --repo-id=<HF_USER/DATASET_ID>
+python -m lerobot.scripts.convert_dataset_v21_to_v30 --repo-id=<HF_USER/DATASET_ID>
 ```
 
 **What it does**

docs/source/molmoact2.mdx

@@ -0,0 +1,433 @@
+# MolmoAct2 Policy
+
+MolmoAct2 is the LeRobot policy implementation of
+[MolmoAct2](https://allenai.org/blog/molmoact2), ported into the LeRobot
+training, evaluation, checkpointing, and dataset interfaces for easier use with
+LeRobot datasets.
+
+This implementation currently supports training and evaluation for the regular
+MolmoAct2 model. MolmoAct2-Think, which supports adaptive depth reasoning, is
+not included in this LeRobot policy yet and is coming soon.
+
+For the original MolmoAct2 training code used for the experiments reported in
+the paper, see [allenai/molmoact2](https://github.com/allenai/molmoact2).
+
+## Installation Requirements
+
+Install LeRobot with the MolmoAct2 optional dependencies:
+
+```bash
+pip install -e ".[molmoact2]"
+```
+
+To run the models in this repository, you need an NVIDIA GPU. The measurements
+below were taken on a single NVIDIA H100 80GB with bf16 model loading, LIBERO with two RGB cameras. MolmoAct2 rows use `chunk_size=10`, action dim 7
+padded to `expected_max_action_dim=32`, and `num_flow_timesteps=8`. Training measurements use
+`gradient_checkpointing=true` and include the forward pass, backward pass,
+gradient clipping, optimizer step, and optimizer state allocation. Values are
+peak GPU memory sampled with `nvidia-smi`. Leave a few GiB of headroom for
+dataloader workers, CUDA context, and fragmentation.
+
+Multi-GPU training through `accelerate` increases throughput and global batch
+size, but this LeRobot port does not currently expose the original MolmoAct2
+`fsdp_devices` model-parallel training path. The current training script has
+not been tested for multi-node training.
+
+| Mode                                             | Peak Memory, bs=8 | Peak Memory, bs=16 | Peak Memory, bs=32 |
+| ------------------------------------------------ | ----------------: | -----------------: | -----------------: |
+| Inference, continuous, CUDA graph enabled (bs=1) |          12.1 GiB |                  - |                  - |
+| Fine-tuning, action expert only, continuous      |          16.5 GiB |           18.3 GiB |           21.4 GiB |
+| Fine-tuning, LoRA VLM, both action modes         |          20.2 GiB |           26.8 GiB |           41.3 GiB |
+| Fine-tuning, full model, both action modes       |          48.3 GiB |           49.8 GiB |           60.1 GiB |
+
+The repo has been tested with Ubuntu 22.04.
+
+## Usage
+
+To use MolmoAct2 in a LeRobot training config, set:
+
+```python
+policy.type=molmoact2
+```
+
+## Training
+
+MolmoAct2 can be fine-tuned from either the released MolmoAct2 Hugging Face
+checkpoint format or from a checkpoint already saved by LeRobot. Both routes use
+the same LeRobot training loop, dataset transforms, checkpoint saving, and
+logging. The difference is only how the initial policy weights and processor
+state are loaded.
+
+### Training With Original MolmoAct2 Weight
+
+Use `policy.checkpoint_path` when starting from a released MolmoAct2 checkpoint,
+for example `allenai/MolmoAct2` or `allenai/MolmoAct2-LIBERO`. LeRobot will load
+the original HF model files, then build its own policy processor from the
+dataset metadata and the policy options below.
+
+The command below shows full fine-tuning on the merged LIBERO dataset. It uses
+bf16 model loading, 8 flow timesteps, LeRobot dataset statistics, image
+augmentation, and LeRobot's checkpointing/logging path.
+
+```bash
+accelerate launch \
+  --num_processes=8 \
+  --mixed_precision=bf16 \
+  -m lerobot.scripts.lerobot_train \
+  --dataset.repo_id=allenai/MolmoAct2-LIBERO-Dataset \
+  --dataset.root=/path/to/lerobot/data/allenai/MolmoAct2-LIBERO-Dataset \
+  --dataset.video_backend=pyav \
+  --dataset.image_transforms.enable=true \
+  --policy.type=molmoact2 \
+  --policy.checkpoint_path=allenai/MolmoAct2-LIBERO \
+  --policy.device=cuda \
+  --policy.action_mode=both \
+  --policy.chunk_size=10 \
+  --policy.n_action_steps=10 \
+  --policy.setup_type="single franka robotic arm in libero" \
+  --policy.control_mode="delta end-effector pose" \
+  --policy.image_keys='["observation.images.image","observation.images.wrist_image"]' \
+  --policy.model_dtype=bfloat16 \
+  --policy.num_flow_timesteps=8 \
+  --policy.gradient_checkpointing=true \
+  --policy.freeze_embedding=true \
+  --policy.normalize_gripper=false \
+  --policy.enable_knowledge_insulation=false \
+  --policy.push_to_hub=false \
+  --wandb.enable=true \
+  --wandb.entity=<wandb_entity> \
+  --wandb.project=<wandb_project> \
+  --job_name=<job_name> \
+  --output_dir=outputs/<job_name> \
+  --steps=10000 \
+  --batch_size=32 \
+  --num_workers=4 \
+  --log_freq=20 \
+  --eval_freq=-1 \
+  --save_checkpoint=true \
+  --save_freq=2000
+```
+
+### Training With LeRobot MolmoAct2 Weight
+
+Use `policy.path` when starting from a MolmoAct2 checkpoint that was saved by
+LeRobot, either from a local `pretrained_model` directory or from the Hub. This
+restores the saved LeRobot policy config, model weights, processor, and
+normalization statistics. You can still override training-time options such as
+`batch_size`, `steps`, LoRA flags, or `policy.action_mode`.
+
+```bash
+accelerate launch \
+  --num_processes=8 \
+  --mixed_precision=bf16 \
+  -m lerobot.scripts.lerobot_train \
+  --dataset.repo_id=allenai/MolmoAct2-LIBERO-Dataset \
+  --dataset.root=/path/to/lerobot/data/allenai/MolmoAct2-LIBERO-Dataset \
+  --dataset.video_backend=pyav \
+  --dataset.image_transforms.enable=true \
+  --policy.path=/path/to/pretrained_model \
+  --policy.device=cuda \
+  --policy.action_mode=both \
+  --policy.chunk_size=10 \
+  --policy.n_action_steps=10 \
+  --policy.model_dtype=bfloat16 \
+  --policy.num_flow_timesteps=8 \
+  --policy.gradient_checkpointing=true \
+  --wandb.enable=true \
+  --wandb.entity=<wandb_entity> \
+  --wandb.project=<wandb_project> \
+  --job_name=<job_name> \
+  --output_dir=outputs/<job_name> \
+  --steps=10000 \
+  --batch_size=32 \
+  --num_workers=4 \
+  --log_freq=20 \
+  --eval_freq=-1 \
+  --save_checkpoint=true \
+  --save_freq=2000
+```
+
+### Common Practices
+
+For fine-tuning on a comparatively small dataset, such as a single LIBERO suite
+or a real-world dataset with less than 200 demonstrations, a global batch size of
+16 to 32 is a good starting point. In these settings, `policy.enable_lora_vlm=true` or `policy.train_action_expert_only=true` is also a practical choice. In both
+cases, we intentionally keep the action expert fully trainable, which we found
+to be crucial for model performance. For larger fine-tuning datasets, larger
+global batch sizes and full fine-tuning are usually preferred.
+
+### Common Policy Options
+
+- `policy.checkpoint_path`: original MolmoAct2 HF checkpoint to initialize from.
+  Use this for released MolmoAct2 weights.
+- `policy.path`: LeRobot checkpoint to initialize from. Use this for checkpoints
+  created by LeRobot training.
+- `policy.action_mode`: training target, one of `continuous`, `discrete`, or
+  `both`. `both` trains the flow-matching action expert and the discrete
+  action-token loss.
+- `policy.train_action_expert_only`: trains only parameters whose names contain
+  `action_expert`. It requires `policy.action_mode=continuous`.
+- `policy.enable_lora_vlm`: enables LoRA on VLM linear layers. Use
+  `policy.enable_lora_action_expert=true` only if LoRA should also cover action
+  expert linear layers. When `policy.enable_lora_action_expert=false`, the
+  action expert base weights remain fully trainable while the VLM is trained
+  through LoRA adapters. When `policy.enable_lora_action_expert=true`, the
+  action expert is also adapter-tuned instead of fully fine-tuned.
+- `policy.enable_knowledge_insulation`: when `true`, detaches action-expert
+  context K/V states before the action loss. The default is `false`.
+- `policy.chunk_size`: action horizon used by the policy. For LIBERO we use
+  `10`. This LeRobot port overrides the loaded checkpoint's
+  `max_action_horizon` with this value.
+- `policy.n_action_steps`: number of actions consumed from each predicted
+  chunk before querying the policy again. For LIBERO, set it to `chunk_size`.
+- `policy.setup_type`: text inserted into the prompt to describe the robot and
+  scene, e.g. `single franka robotic arm in libero`. More examples are listed
+  in the `metadata_by_tag` entries of
+  [`norm_stats.json`](https://huggingface.co/allenai/MolmoAct2/blob/main/norm_stats.json).
+- `policy.control_mode`: text inserted into the prompt to describe the action
+  space, e.g. `delta end-effector pose` or `absolute joint pose`.
+- `policy.image_keys`: ordered LeRobot image observation keys passed to the
+  processor.
+- `policy.model_dtype`: checkpoint/forward dtype, one of `float32`,
+  `bfloat16`, or `float16`. Use `bfloat16` for normal training.
+- `policy.num_flow_timesteps`: number of flow-matching timesteps sampled per
+  example during training. We use `8` for fine-tuning.
+- `policy.num_inference_steps`: optional override for continuous action
+  generation steps at inference time.
+- `policy.gradient_checkpointing`: enables checkpointing in the VLM/action path
+  to reduce activation memory.
+- `policy.freeze_embedding`: freezes input embeddings. The default is `true`.
+- `policy.normalize_gripper`: controls whether gripper dimensions are included
+  in state/action quantile normalization. The default is `false`.
+- `policy.normalize_language`: normalizes task strings before prompt
+  construction. The default is `true`.
+- `policy.mask_action_dim_padding`: masks padded dimensions in the flow loss.
+  Released checkpoints use `policy.expected_max_action_dim=32`.
+- `policy.max_sequence_length`: optional manual sequence cap. Leave unset to
+  infer it from images, state dimension, action dimension, action horizon, and
+  discrete-action mode.
+
+### Learning Rates
+
+MolmoAct2 uses parameter-group learning rates to match the original MolmoAct2
+fine-tuning experiments.
+
+- Full fine-tuning uses `policy.optimizer_lr=1e-5` for the VLM,
+  `policy.optimizer_vit_lr=5e-6` for the vision tower,
+  `policy.optimizer_connector_lr=5e-6` for image connector layers, and
+  `policy.optimizer_action_expert_lr=5e-5` for the action expert.
+- LoRA VLM fine-tuning sets the VLM, vision, and connector LoRA parameter
+  groups to `5e-5` when `policy.enable_lora_vlm=true`. By default,
+  `policy.enable_lora_action_expert=false`, so the action expert is still fully
+  fine-tuned with `policy.optimizer_action_expert_lr`. If
+  `policy.enable_lora_action_expert=true`, the action expert is trained through
+  LoRA adapters instead.
+- Action-expert-only fine-tuning trains only the action expert and uses
+  `policy.optimizer_action_expert_lr=5e-5`.
+
+You can override the full fine-tuning and action-expert learning rates with
+`policy.optimizer_lr`, `policy.optimizer_vit_lr`,
+`policy.optimizer_connector_lr`, and `policy.optimizer_action_expert_lr`.
+Scheduler settings can be changed with `policy.scheduler_warmup_steps`,
+`policy.scheduler_decay_steps`, and `policy.scheduler_decay_lr`.
+
+### Dataset Quantile Statistics
+
+MolmoAct2 defaults to quantile normalization for state and action features. If
+your dataset has not been converted with quantile statistics, you can add them
+with:
+
+```bash
+python src/lerobot/scripts/augment_dataset_quantile_stats.py \
+  --repo-id=your_dataset
+```
+
+Alternatively, train MolmoAct2 with mean/std normalization:
+
+```bash
+--policy.normalization_mapping='{"ACTION": "MEAN_STD", "STATE": "MEAN_STD", "VISUAL": "IDENTITY"}'
+```
+
+## Evaluation
+
+Evaluation also supports both LeRobot-saved checkpoints and original MolmoAct2
+HF checkpoints. For LIBERO replication, keep the EGL rendering environment
+fixed and use `policy.per_episode_seed=true`.
+
+**Important:** We found that `num_steps_wait=10` does not reliably let the
+LIBERO scene stabilize and can degrade measured success. All LIBERO evaluation
+results reported here use `num_steps_wait=50`.
+
+### Evaluation With LeRobot MolmoAct2 Weight
+
+Use `policy.path` for a checkpoint saved by LeRobot. The saved processor and
+normalization statistics are restored together with the model.
+
+```bash
+export MUJOCO_GL=egl
+export PYOPENGL_PLATFORM=egl
+export OMP_NUM_THREADS=1
+export MKL_NUM_THREADS=1
+
+lerobot-eval \
+  --policy.path=allenai/MolmoAct2-LIBERO-LeRobot \
+  --policy.inference_action_mode=continuous \
+  --policy.model_dtype=bfloat16 \
+  --policy.use_amp=true \
+  --policy.enable_inference_cuda_graph=true \
+  --policy.device=cuda \
+  --policy.per_episode_seed=true \
+  --policy.eval_seed=1000 \
+  --env.type=libero \
+  --env.task=libero_10,libero_goal,libero_object,libero_spatial \
+  --env.camera_name_mapping='{"agentview_image":"image","robot0_eye_in_hand_image":"wrist_image"}' \
+  --eval.batch_size=1 \
+  --eval.n_episodes=50 \
+  --seed=1000
+```
+
+### Evaluation With Original MolmoAct2 Weight
+
+You can evaluate a released Hugging Face checkpoint directly without first
+converting it to a LeRobot checkpoint. In this case, set
+`policy.checkpoint_path` to the HF model repo and provide `policy.norm_tag`.
+For LIBERO, `policy.norm_tag=libero` loads the LIBERO action/state
+normalization statistics, action horizon, prompt metadata, and image-key order
+from the checkpoint's `norm_stats.json`.
+
+To fully replicate the MolmoAct2 paper results with released Hugging Face
+checkpoints, we recommend using the v0.5.1-pinned
+[`allenai/lerobot` `molmoact2-hf-inference`](https://github.com/allenai/lerobot/tree/molmoact2-hf-inference)
+branch. That branch matches the original evaluation settings used for the
+reported numbers.
+
+```bash
+export MUJOCO_GL=egl
+export PYOPENGL_PLATFORM=egl
+export OMP_NUM_THREADS=1
+export MKL_NUM_THREADS=1
+
+lerobot-eval \
+  --policy.type=molmoact2 \
+  --policy.checkpoint_path=allenai/MolmoAct2-LIBERO \
+  --policy.norm_tag=libero \
+  --policy.inference_action_mode=continuous \
+  --policy.model_dtype=float32 \
+  --policy.use_amp=false \
+  --policy.enable_inference_cuda_graph=true \
+  --policy.device=cuda \
+  --policy.per_episode_seed=true \
+  --policy.eval_seed=1000 \
+  --env.type=libero \
+  --env.task=libero_goal \
+  --env.camera_name_mapping='{"agentview_image":"image","robot0_eye_in_hand_image":"wrist_image"}' \
+  --eval.batch_size=1 \
+  --eval.n_episodes=50 \
+  --seed=1000
+```
+
+Use `--env.task=libero_10,libero_goal,libero_object,libero_spatial` to run the
+full LIBERO suite. The same command works for other released MolmoAct2
+checkpoints as long as the requested `policy.norm_tag` exists in that
+checkpoint's `norm_stats.json`.
+
+### Common Evaluation Options
+
+- `policy.inference_action_mode`: required for rollout. Use `continuous` for
+  flow-matching inference or `discrete` for action-token inference. It must be
+  compatible with the training-time `policy.action_mode` saved in the
+  checkpoint.
+- `policy.path`: LeRobot checkpoint path or Hub repo. Use this for checkpoints
+  saved by LeRobot.
+- `policy.checkpoint_path`: original MolmoAct2 HF checkpoint path or Hub repo.
+  Use this with `policy.type=molmoact2` and `policy.norm_tag`.
+- `policy.norm_tag`: selects normalization statistics, prompt metadata,
+  image-key order, and action horizon from the original checkpoint's
+  `norm_stats.json`. It is required for direct original-HF checkpoint
+  evaluation.
+- `policy.model_dtype`: model load/forward dtype. Use `bfloat16` for normal
+  GPU evaluation. Use `float32` only when you explicitly want fp32 inference.
+- `policy.use_amp`: runs the policy forward under autocast during eval. For
+  `model_dtype=bfloat16`, keep this enabled.
+- `policy.enable_inference_cuda_graph`: enables the MolmoAct2 inference CUDA
+  graph path for faster repeated continuous-action rollout.
+- `policy.per_episode_seed` and `policy.eval_seed`: make stochastic continuous
+  action generation deterministic per episode for replication.
+- `env.task`: comma-separated LIBERO suites or a single suite. Use
+  `libero_10,libero_goal,libero_object,libero_spatial` for the full benchmark.
+- `env.camera_name_mapping`: maps LIBERO camera names to the image keys expected
+  by the policy processor.
+
+## Performance Results
+
+### LIBERO Benchmark Results
+
+MolmoAct2 has demonstrated strong performance on the LIBERO benchmark suite. To
+compare and test its LeRobot implementation, we fine-tuned
+[`allenai/MolmoAct2-LIBERO`](https://huggingface.co/allenai/MolmoAct2-LIBERO)
+for an additional 10k steps on the LIBERO dataset with per-GPU batch size 32 on
+8 H100 GPUs, then compared the results to the original MolmoAct2 reference
+results.
+
+The LeRobot fine-tuned checkpoint reported here is available at
+[`allenai/MolmoAct2-LIBERO-LeRobot`](https://huggingface.co/allenai/MolmoAct2-LIBERO-LeRobot)
+and was trained on
+[`allenai/MolmoAct2-LIBERO-Dataset`](https://huggingface.co/datasets/allenai/MolmoAct2-LIBERO-Dataset).
+
+| Benchmark      | LeRobot Implementation | MolmoAct2 Original |
+| -------------- | ---------------------: | -----------------: |
+| LIBERO Spatial |                  98.4% |              97.8% |
+| LIBERO Object  |                 100.0% |             100.0% |
+| LIBERO Goal    |                  98.0% |              97.8% |
+| LIBERO 10      |                  96.6% |              93.2% |
+| Average        |                 98.25% |             97.20% |
+
+These results demonstrate MolmoAct2's strong performance across diverse robotic
+manipulation tasks. To reproduce them, follow the instructions in the LIBERO
+evaluation section.
+
+## Differences From the Original Implementation
+
+This LeRobot port is intended to match MolmoAct2 behavior while using LeRobot's
+dataset, training, evaluation, checkpoint, and logging infrastructure. The main
+differences from the original training repository are:
+
+- The original paper training stack loads the model in fp32 and trains under
+  mixed precision. This LeRobot port usually loads the checkpoint directly in
+  `policy.model_dtype=bfloat16` for lower memory use.
+- The original repository uses its own FSDP/model-parallel training path. The
+  LeRobot port uses the standard LeRobot/Accelerate training path and has not
+  been tested for multi-node training.
+- The original repository supports sequence packing. The LeRobot port trains on
+  one LeRobot sample per item and pads to an inferred fixed sequence budget.
+- The LeRobot port follows LeRobot's optimizer, scheduler, checkpoint saving,
+  dataset transforms, image augmentation, and Weights & Biases logging
+  conventions.
+- The original training path supports mixed action horizons by padding to
+  `max_action_horizon` and masking padded horizon slots in the action expert
+  self-attention. This is useful when training across datasets with different
+  control frequencies. The LeRobot port currently targets single-dataset
+  fine-tuning, so `policy.chunk_size` overrides the checkpoint
+  `max_action_horizon` and horizon masking is not implemented yet. Support for
+  this mixed-horizon path is planned.
+
+## Citation
+
+```bibtex
+@misc{fang2026molmoact2actionreasoningmodels,
+      title={MolmoAct2: Action Reasoning Models for Real-world Deployment},
+      author={Haoquan Fang and Jiafei Duan and Donovan Clay and Sam Wang and Shuo Liu and Weikai Huang and Xiang Fan and Wei-Chuan Tsai and Shirui Chen and Yi Ru Wang and Shanli Xing and Jaemin Cho and Jae Sung Park and Ainaz Eftekhar and Peter Sushko and Karen Farley and Angad Wadhwa and Cole Harrison and Winson Han and Ying-Chun Lee and Eli VanderBilt and Rose Hendrix and Suveen Ellawela and Lucas Ngoo and Joyce Chai and Zhongzheng Ren and Ali Farhadi and Dieter Fox and Ranjay Krishna},
+      year={2026},
+      eprint={2605.02881},
+      archivePrefix={arXiv},
+      primaryClass={cs.RO},
+      url={https://arxiv.org/abs/2605.02881},
+}
+```
+
+## License
+
+This model is licensed under Apache 2.0. It is intended for research and
+educational use in accordance with
+[Ai2's Responsible Use Guidelines](https://allenai.org/responsible-use),
+consistent with [allenai/molmoact2](https://github.com/allenai/molmoact2).

docs/source/pi05.mdx

@@ -91,7 +91,7 @@ lerobot-train \
 If your dataset is not converted with `quantiles`, you can convert it with the following command:
 
 ```bash
-python src/lerobot/datasets/v30/augment_dataset_quantile_stats.py \
+python src/lerobot/scripts/augment_dataset_quantile_stats.py \
     --repo-id=your_dataset \
 ```
 

docs/source/policy_molmoact2_README.md

@@ -0,0 +1,39 @@
+# MolmoAct2
+
+This repository contains the LeRobot policy implementation of
+[MolmoAct2](https://allenai.org/blog/molmoact2), ported into LeRobot for
+training, evaluation, checkpointing, and dataset compatibility.
+
+This implementation currently supports training and evaluation for the regular
+MolmoAct2 model. MolmoAct2-Think, which supports adaptive depth reasoning, is
+not included in this LeRobot policy yet and is coming soon.
+
+For the original MolmoAct2 training code used for the experiments reported in
+the paper, see [allenai/molmoact2](https://github.com/allenai/molmoact2).
+
+## LIBERO Evaluation
+
+Important: we found that `num_steps_wait=10` does not reliably let the LIBERO
+scene stabilize and can degrade measured success. All LIBERO evaluation results
+reported for this LeRobot implementation use `num_steps_wait=50`.
+
+## Citation
+
+```bibtex
+@misc{fang2026molmoact2actionreasoningmodels,
+      title={MolmoAct2: Action Reasoning Models for Real-world Deployment},
+      author={Haoquan Fang and Jiafei Duan and Donovan Clay and Sam Wang and Shuo Liu and Weikai Huang and Xiang Fan and Wei-Chuan Tsai and Shirui Chen and Yi Ru Wang and Shanli Xing and Jaemin Cho and Jae Sung Park and Ainaz Eftekhar and Peter Sushko and Karen Farley and Angad Wadhwa and Cole Harrison and Winson Han and Ying-Chun Lee and Eli VanderBilt and Rose Hendrix and Suveen Ellawela and Lucas Ngoo and Joyce Chai and Zhongzheng Ren and Ali Farhadi and Dieter Fox and Ranjay Krishna},
+      year={2026},
+      eprint={2605.02881},
+      archivePrefix={arXiv},
+      primaryClass={cs.RO},
+      url={https://arxiv.org/abs/2605.02881},
+}
+```
+
+## License
+
+This model is licensed under Apache 2.0. It is intended for research and
+educational use in accordance with
+[Ai2's Responsible Use Guidelines](https://allenai.org/responsible-use),
+consistent with [allenai/molmoact2](https://github.com/allenai/molmoact2).

docs/source/policy_vla_jepa_README.md

@@ -0,0 +1,39 @@
+# VLA-JEPA
+
+This repository contains the LeRobot port of **VLA-JEPA**, a Vision-Language-Action model that combines a Qwen3-VL language backbone with a self-supervised video world model (V-JEPA2) and a flow-matching DiT action head.
+
+Converted from [ginwind/VLA-JEPA](https://huggingface.co/ginwind/VLA-JEPA).
+
+---
+
+## Architecture Overview
+
+| Component               | Module                            | Role                                                    |
+| ----------------------- | --------------------------------- | ------------------------------------------------------- |
+| **Qwen3-VL backbone**   | `Qwen3VLInterface`                | Fuses images + language instruction into context tokens |
+| **DiT-B action head**   | `VLAJEPAActionHead`               | Flow-matching diffusion over the action chunk           |
+| **V-JEPA2 world model** | `ActionConditionedVideoPredictor` | Self-supervised video prediction loss (training only)   |
+
+At inference time only the Qwen backbone and action head are used; the world model is not needed.
+
+---
+
+## Citation
+
+```bibtex
+@misc{sun2026vlajepaenhancingvisionlanguageactionmodel,
+  title         = {VLA-JEPA: Enhancing Vision-Language-Action Model with Latent World Model},
+  author        = {Jingwen Sun and Wenyao Zhang and Zekun Qi and Shaojie Ren and Zezhi Liu and Hanxin Zhu and Guangzhong Sun and Xin Jin and Zhibo Chen},
+  year          = {2026},
+  eprint        = {2602.10098},
+  archivePrefix = {arXiv},
+  primaryClass  = {cs.RO},
+  url           = {https://arxiv.org/abs/2602.10098},
+}
+```
+
+---
+
+## License
+
+Weights are distributed under the license terms of the original [ginwind/VLA-JEPA](https://huggingface.co/ginwind/VLA-JEPA) repository (**Apache 2.0 License**). The LeRobot integration code follows the **Apache 2.0 License**.

docs/source/porting_datasets_v3.mdx

@@ -300,7 +300,7 @@ This replaces the old episode-per-file structure with efficient, optimally-sized
 If you have existing datasets in v2.1 format, use the migration tool:
 
 ```bash
-python src/lerobot/datasets/v30/convert_dataset_v21_to_v30.py \
+python src/lerobot/scripts/convert_dataset_v21_to_v30.py \
     --repo-id your_id/existing_dataset
 ```
 

docs/source/robometer.mdx

@@ -0,0 +1,185 @@
+# ROBOMETER
+
+ROBOMETER is a **general-purpose video-language robotic reward model**. It predicts dense, frame-level task progress and frame-level success from a trajectory video and a task description.
+
+**Paper**: [ROBOMETER: Scaling General-Purpose Robotic Reward Models via Trajectory Comparisons](https://arxiv.org/abs/2603.02115)
+**Project**: [robometer.github.io](https://robometer.github.io/)
+**Original code**: [github.com/robometer/robometer](https://github.com/robometer/robometer)
+**Checkpoint**: [lerobot/Robometer-4B](https://huggingface.co/lerobot/Robometer-4B)
+
+## Overview
+
+ROBOMETER builds on `Qwen/Qwen3-VL-4B-Instruct` and adds three lightweight prediction heads:
+
+- **Progress head**: predicts per-frame task progress in `[0, 1]`.
+- **Success head**: predicts per-frame task success probability.
+- **Preference head**: predicts which of two trajectories better completes the task during training.
+
+The paper trains ROBOMETER with a composite objective:
+
+```text
+L = L_pref + L_prog + L_succ
+```
+
+The LeRobot integration is currently **inference-only**. It preserves the preference head so that the published `Robometer-4B` checkpoint loads without remapping, but `compute_reward()` queries the progress or success head only.
+
+## What the LeRobot Integration Covers
+
+- Standard `reward_model.type=robometer` configuration through LeRobot.
+- Qwen3-VL image and text preprocessing through `RobometerEncoderProcessorStep`.
+- LeRobot reward-model save/load APIs through `PreTrainedRewardModel`.
+- Dense, frame-level progress and success predictions internally.
+- A scalar reward through `compute_reward()` for downstream LeRobot reward-model usage.
+
+This page focuses on using the published ROBOMETER checkpoint as a zero-shot reward model. Training ROBOMETER from scratch is outside the current LeRobot integration.
+
+## Installation Requirements
+
+1. Install LeRobot by following the [Installation Guide](./installation).
+2. Install the ROBOMETER dependencies:
+
+```bash
+pip install -e ".[robometer]"
+```
+
+If you use `uv` directly from a source checkout:
+
+```bash
+uv sync --extra robometer
+```
+
+ROBOMETER uses a Qwen3-VL-4B backbone, so GPU inference is strongly recommended.
+
+## Model Inputs and Outputs
+
+ROBOMETER expects:
+
+- A trajectory video or sequence of frames.
+- A natural-language task description.
+
+In LeRobot datasets, the preprocessor reads:
+
+| Config field              | Default                  | Meaning                                               |
+| ------------------------- | ------------------------ | ----------------------------------------------------- |
+| `reward_model.image_key`  | `observation.images.top` | Camera/video observation used by ROBOMETER            |
+| `reward_model.task_key`   | `task`                   | Key in complementary data that stores the task string |
+| `reward_model.max_frames` | `8`                      | Maximum number of frames passed to ROBOMETER          |
+
+The model predicts per-frame progress and success internally. The LeRobot reward API returns a scalar per sample:
+
+- `reward_output="progress"` (default): return the last-frame progress, clamped to `[0, 1]`.
+- `reward_output="success"`: return `1.0` if the last-frame success probability is above `success_threshold`, otherwise `0.0`.
+
+## Usage
+
+### Load the Reward Model Directly
+
+```python
+from lerobot.rewards.robometer import RobometerConfig, RobometerRewardModel
+
+cfg = RobometerConfig(
+    pretrained_path="lerobot/Robometer-4B",
+    device="cuda",
+    reward_output="progress",
+)
+reward_model = RobometerRewardModel.from_pretrained(cfg.pretrained_path, config=cfg)
+```
+
+### Encode Frames and Compute a Reward
+
+For a direct Python call, provide frames as `uint8` arrays with shape `(T, H, W, C)` and a task string:
+
+```python
+from lerobot.rewards.robometer.modeling_robometer import ROBOMETER_FEATURE_PREFIX
+from lerobot.rewards.robometer.processor_robometer import RobometerEncoderProcessorStep
+
+# frames: np.ndarray, shape (T, H, W, C), dtype uint8
+# task: str
+encoder = RobometerEncoderProcessorStep(
+    base_model_id=cfg.base_model_id,
+    use_multi_image=cfg.use_multi_image,
+    use_per_frame_progress_token=cfg.use_per_frame_progress_token,
+    max_frames=cfg.max_frames,
+)
+
+encoded = encoder.encode_samples([(frames, task)])
+batch = {f"{ROBOMETER_FEATURE_PREFIX}{key}": value for key, value in encoded.items()}
+
+reward = reward_model.compute_reward(batch)
+```
+
+`reward` is a tensor of shape `(batch_size,)`.
+
+### Use the Reward Factory
+
+You can also instantiate ROBOMETER through the reward factory:
+
+```python
+from lerobot.rewards import make_reward_model, make_reward_model_config, make_reward_pre_post_processors
+
+cfg = make_reward_model_config(
+    "robometer",
+    pretrained_path="lerobot/Robometer-4B",
+    device="cuda",
+    image_key="observation.images.top",
+)
+reward_model = make_reward_model(cfg)
+preprocessor, postprocessor = make_reward_pre_post_processors(cfg)
+```
+
+The preprocessor writes Qwen-VL tensors under the `observation.robometer.*` namespace, and `compute_reward()` reads those encoded tensors.
+
+## Configuration Notes
+
+### Backbone and Vocabulary
+
+The published checkpoint uses a Qwen3-VL-4B backbone. ROBOMETER adds five special tokens to the tokenizer in a fixed order:
+
+```text
+<|split_token|>
+<|reward_token|>
+<|pref_token|>
+<|sim_token|>
+<|prog_token|>
+```
+
+`<|prog_token|>` is inserted after each frame and is the hidden-state position used for per-frame progress and success prediction. `<|split_token|>` and `<|pref_token|>` are used by the paper's pairwise trajectory preference objective. `<|reward_token|>` and `<|sim_token|>` are preserved for checkpoint compatibility.
+
+The LeRobot config stores a serialized `vlm_config` with the post-resize vocabulary so the model can reload from `config.json` without downloading the base Qwen weights first. For `Qwen/Qwen3-VL-4B-Instruct`, the tokenizer length is `151669`, and the five ROBOMETER tokens produce the checkpoint vocabulary size `151674`.
+
+### Progress Prediction
+
+In the published checkpoint, progress is discrete. The progress head outputs logits over `progress_discrete_bins=10` uniformly spaced bin centers in `[0, 1]`. LeRobot converts these logits into a continuous value by applying a softmax and taking the expectation over bin centers, matching the upstream ROBOMETER implementation.
+
+### Success Prediction
+
+The success head outputs raw logits per frame. LeRobot converts them to probabilities with `sigmoid`. When `reward_output="success"`, `compute_reward()` thresholds the last-frame success probability using `success_threshold`.
+
+## Limitations
+
+- The current LeRobot integration is inference-only; it does not implement ROBOMETER training or preference-pair training.
+- `compute_reward()` returns a scalar per sample for the LeRobot reward-model API, even though ROBOMETER predicts per-frame progress and success internally.
+- ROBOMETER is video-language based; it does not use privileged robot state such as contact forces or object poses.
+
+## References
+
+- [ROBOMETER project](https://robometer.github.io/)
+- [ROBOMETER paper](https://arxiv.org/abs/2603.02115)
+- [Original ROBOMETER code](https://github.com/robometer/robometer)
+- [Published ROBOMETER-4B checkpoint](https://huggingface.co/lerobot/Robometer-4B)
+- [Qwen3-VL-4B-Instruct](https://huggingface.co/Qwen/Qwen3-VL-4B-Instruct)
+
+## Citation
+
+```bibtex
+@inproceedings{liang2026robometer,
+title = {Robometer: Scaling General-Purpose Robotic Reward Models via Trajectory Comparisons},
+author={Anthony Liang and Yigit Korkmaz and Jiahui Zhang and Minyoung Hwang and Abrar Anwar and Sidhant Kaushik and Aditya Shah and Alex S. Huang and Luke Zettlemoyer and Dieter Fox and Yu Xiang and Anqi Li and Andreea Bobu and Abhishek Gupta and Stephen Tu and Erdem Biyik and Jesse Zhang},
+year={2026},
+booktitle={Robotics: Science and Systems 2026},
+}
+```
+
+## License
+
+This LeRobot integration follows the **Apache 2.0 License** used by LeRobot. Check the upstream ROBOMETER code and model pages for the licenses of the original implementation and released checkpoints.

docs/source/smolvla.mdx

@@ -97,22 +97,22 @@ Similarly for when recording an episode, it is recommended that you are logged i
 Once you are logged in, you can run inference in your setup by doing:
 
 ```bash
-lerobot-record \
+lerobot-rollout \
+  --strategy.type=base \
   --robot.type=so101_follower \
   --robot.port=/dev/ttyACM0 \ # <- Use your port
   --robot.id=my_blue_follower_arm \ # <- Use your robot id
   --robot.cameras="{ front: {type: opencv, index_or_path: 8, width: 640, height: 480, fps: 30}}" \ # <- Use your cameras
-  --dataset.single_task="Grasp a lego block and put it in the bin." \ # <- Use the same task description you used in your dataset recording
-  --dataset.repo_id=${HF_USER}/eval_DATASET_NAME_test \  # <- This will be the dataset name on HF Hub
-  --dataset.episode_time_s=50 \
-  --dataset.num_episodes=10 \
-  --dataset.streaming_encoding=true \
-  --dataset.encoder_threads=2 \
-  # --dataset.camera_encoder.vcodec=auto \
+  --task="Grasp a lego block and put it in the bin." \ # <- Use the same task description you used in your dataset recording
+  # <- RTC optional, use when running on low power hardware \
+  # --inference.type=rtc \
+  # --inference.rtc.execution_horizon=10 \
+  # --inference.rtc.max_guidance_weight=10.0 \
   # <- Teleop optional if you want to teleoperate in between episodes \
   # --teleop.type=so100_leader \
   # --teleop.port=/dev/ttyACM0 \
   # --teleop.id=my_red_leader_arm \
+  # --display_data=true #optional use if you want to see the camera stream \
   --policy.path=HF_USER/FINETUNE_MODEL_NAME # <- Use your fine-tuned model
 ```
 

docs/source/topreward.mdx

@@ -0,0 +1,177 @@
+# TOPReward
+
+TOPReward is a **zero-shot reward model** that extracts token log-probabilities from an off-the-shelf vision-language model (VLM) as a robotic reward signal. Given a video trajectory and a task instruction, it returns the VLM's log-likelihood that the instruction is true — no fine-tuning required.
+
+**Paper**: [TOPReward: Token Probabilities as Hidden Zero-Shot Rewards for Robotics](https://arxiv.org/abs/2602.19313)
+**Project**: [topreward.github.io](https://topreward.github.io/webpage/)
+**Original code**: [github.com/TOPReward/TOPReward](https://github.com/TOPReward/TOPReward)
+**Default backbone**: [Qwen/Qwen3-VL-8B-Instruct](https://huggingface.co/Qwen/Qwen3-VL-8B-Instruct)
+
+## Overview
+
+TOPReward asks a generic VLM how likely a task instruction is, **conditioned on the video** of a robot trying to complete that task. Concretely, given:
+
+- A trajectory video (a sequence of frames).
+- A task instruction (e.g. _"open the drawer"_).
+
+it builds a chat prompt of the form
+
+```text
+<video>
+"The above video shows a robot manipulation trajectory that completes the
+ following task: <instruction> Decide whether the above statement is True
+ or not. The answer is: True"
+```
+
+forwards it through the VLM, label-masks everything except the very last token, and reads back the log-probability of that token — by default the literal `"True"` that closes the suffix template. The resulting `log P("True" | video + prompt + instruction)` is the reward.
+
+Because the method only depends on a frozen VLM, TOPReward is **zero-shot**: there are no fine-tuned weights to host. The "model" in LeRobot is a small wrapper around `transformers`' `Qwen3VLForConditionalGeneration` plus the label-masking logic. The processor owns the tokeniser and builds the full chat prompt (EO-1/Robometer pattern).
+
+## What the LeRobot integration covers
+
+- Standard `reward_model.type=topreward` configuration through LeRobot.
+- VLM loading via the `transformers` `Qwen3VLForConditionalGeneration` API.
+- Prompt assembly + tokenisation in the processor (matching upstream `QwenClient.compute_instruction_reward`).
+- `compute_reward()` returns one scalar log-prob per sample.
+- LeRobot reward-model save/load — `save_pretrained` writes only `config.json` (the VLM is identified by `vlm_name`).
+- An offline labeling script that writes a `topreward_progress.parquet` (SARM-compatible schema) for RA-BC and overlay.
+
+The current LeRobot port supports the **Qwen3-VL client only**. Other upstream clients (Gemini, OpenAI, Gemma, Molmo) can be added as follow-up extras.
+
+## Installation Requirements
+
+1. Install LeRobot following the [Installation Guide](./installation).
+2. Install the TOPReward optional extra:
+
+```bash
+pip install -e ".[topreward]"
+```
+
+or, with `uv` from a source checkout:
+
+```bash
+uv sync --extra topreward
+```
+
+This pulls in `transformers`. The first time you run TOPReward, Hugging Face will also download the VLM weights from the Hub (~16 GB for Qwen3-VL-8B-Instruct). A GPU is strongly recommended.
+
+## Model Inputs and Outputs
+
+TOPReward expects:
+
+- A trajectory video or sequence of frames.
+- A natural-language task description.
+
+In LeRobot datasets the preprocessor reads:
+
+| Config field              | Default                     | Meaning                                       |
+| ------------------------- | --------------------------- | --------------------------------------------- |
+| `reward_model.image_key`  | `observation.images.top`    | Camera observation used by TOPReward          |
+| `reward_model.task_key`   | `task`                      | Key in complementary data for the task string |
+| `reward_model.max_frames` | `16`                        | Cap on frames per sample                      |
+| `reward_model.fps`        | `2.0`                       | Metadata passed to the Qwen video processor   |
+| `reward_model.vlm_name`   | `Qwen/Qwen3-VL-8B-Instruct` | Hugging Face Hub id of the underlying VLM     |
+
+The model returns:
+
+- `compute_reward(batch)`: one log-probability per sample. Higher = better task-video alignment. When `success_threshold` is finite, returns the binary thresholded value instead.
+
+## Usage
+
+### Load the reward model directly
+
+```python
+from lerobot.rewards.topreward import TOPRewardConfig, TOPRewardModel
+
+cfg = TOPRewardConfig(
+    vlm_name="Qwen/Qwen3-VL-8B-Instruct",
+    device="cuda",
+)
+reward_model = TOPRewardModel(cfg)
+```
+
+### Use the reward factory
+
+```python
+from lerobot.rewards import make_reward_model, make_reward_model_config, make_reward_pre_post_processors
+
+cfg = make_reward_model_config(
+    "topreward",
+    vlm_name="Qwen/Qwen3-VL-8B-Instruct",
+    device="cuda",
+    image_key="observation.images.top",
+)
+reward_model = make_reward_model(cfg)
+preprocessor, postprocessor = make_reward_pre_post_processors(cfg)
+```
+
+The preprocessor tokenises the full prompt (video + prefix + instruction suffix), writes Qwen-VL tensors + `prompt_length` under `observation.topreward.*`. The model reads those tensors, label-masks based on `prompt_length`, and extracts the log-prob reward.
+
+### Offline dataset labeling
+
+Write a `topreward_progress.parquet` for RA-BC training and overlay videos:
+
+```bash
+# Sparse-dense (15 anchors per episode, matches upstream)
+uv run python -m lerobot.rewards.topreward.compute_rabc_weights \
+    --dataset-repo-id lerobot/libero_10_image \
+    --num-samples 15 \
+    --device cuda
+```
+
+Then render the progress overlay for any episode:
+
+```bash
+uv run examples/dataset/create_progress_videos.py \
+    --repo-id lerobot/libero_10_image \
+    --episode 0 \
+    --progress-file topreward_progress.parquet \
+    --gif
+```
+
+## Configuration Notes
+
+### Prompt knobs
+
+The default prompt mirrors the upstream paper:
+
+```text
+prompt_prefix = "The above video shows a robot manipulation trajectory that completes the following task: "
+prompt_suffix_template = "{instruction} Decide whether the above statement is True or not. The answer is: True"
+```
+
+Both are exposed on `TOPRewardConfig` for ablation. The suffix template **must** contain `{instruction}`.
+
+### Chat template
+
+`add_chat_template=True` wraps the full prompt (including instruction) with the tokenizer's chat template before tokenisation. Default is `False`, matching the upstream paper's main experiments.
+
+## Limitations
+
+- The current LeRobot port is **inference-only and zero-shot**; `forward()` is not overridden and `is_trainable` returns `False`.
+- Only the **Qwen3-VL family** is supported; other upstream clients are out of scope.
+- TOPReward inherits the underlying VLM's biases.
+
+## References
+
+- [TOPReward project page](https://topreward.github.io/webpage/)
+- [TOPReward paper](https://arxiv.org/abs/2602.19313)
+- [Original TOPReward code](https://github.com/TOPReward/TOPReward)
+- [Qwen3-VL-8B-Instruct](https://huggingface.co/Qwen/Qwen3-VL-8B-Instruct)
+
+## Citation
+
+```bibtex
+@article{chen2026topreward,
+  title={TOPReward: Token Probabilities as Hidden Zero-Shot Rewards for Robotics},
+  author={Chen, Shirui and Harrison, Cole and Lee, Ying-Chun and Yang, Angela Jin and
+          Ren, Zhongzheng and Ratliff, Lillian J and Duan, Jiafei and Fox, Dieter and
+          Krishna, Ranjay},
+  journal={arXiv preprint arXiv:2602.19313},
+  year={2026}
+}
+```
+
+## License
+
+The original TOPReward codebase is MIT-licensed. The LeRobot port follows the LeRobot Apache 2.0 license; the wrapped Qwen3-VL weights are subject to the original Qwen license.

docs/source/vla_jepa.mdx

@@ -0,0 +1,235 @@
+# VLA-JEPA
+
+This is the LeRobot port of **VLA-JEPA**, a Vision-Language-Action model that combines a Qwen3-VL language backbone with a self-supervised video world model (V-JEPA2) and a flow-matching DiT action head.
+
+---
+
+## Architecture Overview
+
+VLA-JEPA has three main components:
+
+| Component               | Module                            | Role                                                    |
+| ----------------------- | --------------------------------- | ------------------------------------------------------- |
+| **Qwen3-VL backbone**   | `Qwen3VLInterface`                | Fuses images + language instruction into context tokens |
+| **DiT-B action head**   | `VLAJEPAActionHead`               | Flow-matching diffusion over the action chunk           |
+| **V-JEPA2 world model** | `ActionConditionedVideoPredictor` | Self-supervised video prediction loss (training only)   |
+
+### Data flow
+
+**Training:**
+
+1. A video clip of `num_video_frames` frames is encoded by V-JEPA2 into per-frame patch tokens.
+2. The Qwen3-VL backbone processes multi-view images + the task instruction and produces a sequence of context tokens that includes special action tokens (for world model conditioning) and embodied tokens.
+3. The action head receives those context tokens as cross-attention keys/values and predicts a denoised action chunk via flow matching.
+4. The world model predictor uses the action tokens extracted from Qwen to predict future V-JEPA2 frame embeddings; a regression loss on those predictions is added to the action loss.
+
+**Inference:**
+Only Qwen + the action head are used. The world model is not needed at inference time.
+
+### Action head details
+
+Available presets via `action_model_type`:
+
+| Preset  | Hidden dim | Heads | Head dim |
+| ------- | ---------- | ----- | -------- |
+| `DiT-B` | 768        | 12    | 64       |
+| `DiT-L` | 1536       | 32    | 48       |
+
+### World model details
+
+The video predictor is a ViT-style transformer (`ActionConditionedVideoPredictor`) that takes:
+
+- **Frame tokens**: V-JEPA2 patch embeddings projected to `predictor_embed_dim`
+- **Action tokens**: Qwen action token embeddings projected to `predictor_embed_dim`
+
+It uses block-causal attention so each temporal step can attend to all previous steps. The predictor's input `embed_dim` equals `num_views × video_encoder_hidden_size` (e.g. 2 views × 1024 = 2048 for the pretrained checkpoints).
+
+---
+
+## Pretrained Checkpoints
+
+Three checkpoints are available directly inside the LeRobot org here: [`lerobot/VLA-JEPA`](https://huggingface.co/collections/lerobot/vla-jepa), converted from [ginwind/VLA-JEPA](https://huggingface.co/ginwind/VLA-JEPA):
+
+| Checkpoint                    | Dataset           | Cameras                 | World model | Action dim |
+| ----------------------------- | ----------------- | ----------------------- | ----------- | ---------- |
+| `lerobot/VLA-JEPA-LIBERO`     | LIBERO-10         | 2 (agentview + wrist)   | Enabled     | 7          |
+| `lerobot/VLA-JEPA-Pretrain`   | DROID 1.0.1       | 2 (exterior left views) | Enabled     | 7          |
+| `lerobot/VLA-JEPA-SimplerEnv` | OXE Bridge / RT-1 | 1 (view duplicated ×2)  | Enabled     | 7          |
+
+All checkpoints use `Qwen/Qwen3-VL-2B-Instruct` as the language backbone.
+
+---
+
+## Configuration
+
+Key parameters in `VLAJEPAConfig`:
+
+| Parameter                 | Default | Description                                                                                                                                                                     |
+| ------------------------- | ------- | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
+| `chunk_size`              | 7       | Number of actions predicted per inference call                                                                                                                                  |
+| `n_action_steps`          | 7       | Steps executed from the predicted chunk before re-planning                                                                                                                      |
+| `num_video_frames`        | 8       | Video clip length fed to the world model                                                                                                                                        |
+| `enable_world_model`      | `True`  | Whether to load and train the V-JEPA2 predictor                                                                                                                                 |
+| `world_model_loss_weight` | 0.1     | Weight of the JEPA prediction loss relative to the action loss                                                                                                                  |
+| `num_inference_timesteps` | 4       | Euler integration steps for action denoising                                                                                                                                    |
+| `freeze_qwen`             | `False` | Freeze the Qwen3-VL backbone and only train the action head                                                                                                                     |
+| `reinit_modules`          | `None`  | Key prefixes allowed to be randomly re-initialised on load (for cross-embodiment transfer, see [Fine-tuning on a different embodiment](#fine-tuning-on-a-different-embodiment)) |
+| `gripper_dim`             | 6       | Index of the gripper dimension in the action vector (e.g. 6 for a 7-DoF arm with gripper as the last joint)                                                                     |
+| `gripper_threshold`       | 0.5     | Threshold used by `pre_snap_gripper_action` and `binarize_gripper_action` to binarize the gripper dimension                                                                     |
+| `pre_snap_gripper_action` | `True`  | Snap the gripper dim to {0, 1} before unnormalization. Set to `False` for robots without a binary gripper                                                                       |
+| `binarize_gripper_action` | `True`  | Binarize the gripper dim to {-1, 1} after unnormalization. Set to `False` for robots without a binary gripper                                                                   |
+
+---
+
+## Training
+
+Number of training steps may vary based on dataset size and compute budget. The original paper pretrained for 50k on ssv2 + droid jointly, then additional 30k steps for LIBERO, but fewer steps may still yield good performance when fine-tuning from the provided pretrained checkpoints.
+
+### Full training from scratch
+
+```bash
+lerobot-train \
+  policy.type=vla_jepa \
+  policy.repo_id=your_org/your_repo \
+  dataset.repo_id=your_org/your_dataset
+```
+
+### Fine-tuning from a pretrained checkpoint
+
+```bash
+lerobot-train \
+  --policy.path=lerobot/VLA-JEPA-Pretrain \
+  --policy.repo_id=your_org/your_repo \
+  --dataset.repo_id=your_org/your_dataset
+```
+
+If you want to freeze the Qwen backbone and only train the action head, set `policy.freeze_qwen=True`:
+
+```bash
+lerobot-train \
+  --policy.path=lerobot/VLA-JEPA-Pretrain \
+  --policy.repo_id=your_org/your_repo \
+  --policy.freeze_qwen=true \
+  --dataset.repo_id=your_org/your_dataset
+```
+
+### Fine-tuning on a different embodiment
+
+When the target robot has a different action or state dimensionality than the pretrained checkpoint, the input/output projection layers of the action head will have mismatched shapes and cannot be loaded directly. `reinit_modules` lets you list the key prefixes that are allowed to mismatch — those layers are randomly re-initialised while every other weight is reused from the checkpoint. Any shape mismatch outside the listed prefixes raises an error.
+
+The layers that depend on `action_dim` and `state_dim` are:
+
+| Layer                                     | Key prefix                          |
+| ----------------------------------------- | ----------------------------------- |
+| Action encoder (action_dim → inner_dim)   | `model.action_model.action_encoder` |
+| Action decoder (hidden_size → action_dim) | `model.action_model.action_decoder` |
+| State encoder (state_dim → inner_dim)     | `model.action_model.state_encoder`  |
+
+```bash
+lerobot-train \
+  --policy.path=lerobot/VLA-JEPA-Pretrain \
+  --policy.repo_id=your_org/your_repo \
+  --policy.freeze_qwen=true \
+  --policy.reinit_modules='["model.action_model.action_encoder", "model.action_model.action_decoder", "model.action_model.state_encoder"]' \
+  --dataset.repo_id=your_org/your_dataset
+```
+
+If your robot has no proprioceptive state, omit `model.action_model.state_encoder` from the list.
+
+### Reproducing the LIBERO results
+
+**Training on LIBERO:**
+starts the training from the Pretrain checkpoint, trains for 30k steps on the LIBERO dataset.
+Original paper mentions training across 8 GPUs with a batch size of 32, meaning global batch size of 256.
+
+```bash
+lerobot-train \
+  --policy.path=lerobot/VLA-JEPA-Pretrain \
+  --policy.repo_id=your_org/your_repo \
+  --dataset.repo_id=HuggingFaceVLA/libero \
+  --steps=30000
+```
+
+**Evaluating the pretrained LIBERO-10 checkpoint:**
+
+```bash
+lerobot-eval \
+  --policy.path=lerobot/VLA-JEPA-LIBERO \
+  --env.type=libero \
+  --env.task=libero_spatial,libero_object,libero_goal,libero_10 \
+  --eval.n_episodes=10 \
+  --eval.batch_size=5
+```
+
+To evaluate a subset of tasks only:
+
+```bash
+lerobot-eval \
+  --policy.path=lerobot/VLA-JEPA-LIBERO \
+  --env.type=libero \
+  --env.task=libero_10 \
+  --env.task_ids='[0,1,2]' \
+  --eval.n_episodes=10 \
+  --eval.batch_size=5
+```
+
+**Expected results:**
+
+| Suite          | Episodes | Successes | Success Rate |
+| -------------- | -------- | --------- | ------------ |
+| libero_spatial | 100      | 93        | **95.0%**    |
+| libero_object  | 100      | 100       | **100.0%**   |
+| libero_goal    | 100      | 98        | **98.0%**    |
+| libero_10      | 100      | 96        | **93.0%**    |
+| **Overall**    | **400**  | **387**   | **96.5%**    |
+
+---
+
+## Fine-tuning on datasets with a different number of cameras
+
+The pretrained world model predictor was trained with `embed_dim = jepa_tubelet_size × 1024` (default `jepa_tubelet_size=2`).
+
+**Default behaviour — view padding / trimming (no action required)**
+
+When fine-tuning from `VLA-JEPA-Pretrain` the model automatically adjusts the number of views fed to the world model to match `jepa_tubelet_size`:
+
+- **Single-view datasets (e.g. BridgeV2):** the single-view latent is duplicated to produce a two-view world-model input, preserving the JEPA self-supervised signal without any weight mismatch.
+- **>2-view datasets (e.g. DROID with 3 views):** all views are passed to the Qwen backbone (for richer context), but only the first `jepa_tubelet_size` views (one wrist + one third-person, following the configured view order) are used for the world model.
+
+**Option 1 — Disable the world model**
+
+Set `enable_world_model=False` to skip the JEPA loss entirely. Only the Qwen backbone and action head are loaded and trained. This is sufficient for good action performance.
+
+```bash
+lerobot-train \
+  --policy.path=lerobot/VLA-JEPA-Pretrain \
+  --policy.enable_world_model=false \
+  --policy.repo_id=your_org/your_repo \
+  --dataset.repo_id=your_org/single_camera_dataset
+```
+
+**Option 2 — Reinitialize the predictor input projection**
+
+If you want to change `jepa_tubelet_size` to a value other than 2, load the checkpoint with `strict=False` and reinitialize `model.video_predictor.predictor_embed` for the new `embed_dim`. All other predictor block weights (attention, MLP, norm, output projection) are camera-count-agnostic and can be reused from the pretrained checkpoint.
+
+---
+
+## Citation
+
+```bibtex
+@misc{sun2026vlajepaenhancingvisionlanguageactionmodel,
+  title         = {VLA-JEPA: Enhancing Vision-Language-Action Model with Latent World Model},
+  author        = {Jingwen Sun and Wenyao Zhang and Zekun Qi and Shaojie Ren and Zezhi Liu and Hanxin Zhu and Guangzhong Sun and Xin Jin and Zhibo Chen},
+  year          = {2026},
+  eprint        = {2602.10098},
+  archivePrefix = {arXiv},
+  primaryClass  = {cs.RO},
+  url           = {https://arxiv.org/abs/2602.10098},
+}
+```
+
+---
+
+## License
+
+Weights are distributed under the license terms of the original [ginwind/VLA-JEPA](https://huggingface.co/ginwind/VLA-JEPA) repository (**Apache 2.0 License**). The LeRobot integration code follows the **Apache 2.0 License**.

examples/annotations/run_hf_job.py

@@ -0,0 +1,109 @@
+#!/usr/bin/env python
+
+# Copyright 2026 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""Launch ``lerobot-annotate`` on a Hugging Face job (vllm + Qwen3.6-27B VLM).
+
+Spawns one single-GPU ``h200`` job that:
+
+  1. installs ``lerobot`` from ``main`` plus the annotation extras,
+  2. boots one vllm server with Qwen3.6-27B (dense VLM),
+  3. runs the plan / interjections / vqa modules across the dataset
+     in free-form mode (each episode generates its own subtasks +
+     memory),
+  4. uploads the annotated dataset to ``--new_repo_id`` (when set)
+     or back to ``--repo_id``.
+
+Usage:
+
+    HF_TOKEN=hf_... uv run python examples/annotations/run_hf_job.py
+
+Adjust ``CMD`` (dataset, model, hub repo) and ``flavor`` below for your
+run. For larger datasets, scale to ``h200x4`` and raise
+``--vlm.parallel_servers`` / ``--vlm.num_gpus`` to match.
+"""
+
+import os
+
+from huggingface_hub import get_token, run_job
+
+token = os.environ.get("HF_TOKEN") or get_token()
+if not token:
+    raise RuntimeError("No HF token. Run `huggingface-cli login` or `export HF_TOKEN=hf_...`")
+
+CMD = (
+    "apt-get update -qq && apt-get install -y -qq git ffmpeg && "
+    "pip install --no-deps "
+    "'lerobot @ git+https://github.com/huggingface/lerobot.git@main' && "
+    "pip install --upgrade-strategy only-if-needed "
+    "datasets pyarrow av jsonlines draccus gymnasium torchcodec mergedeep pyyaml-include toml typing-inspect "
+    "openai && "
+    "export VLLM_MEMORY_PROFILER_ESTIMATE_CUDAGRAPHS=0 && "
+    "export VLLM_VIDEO_BACKEND=pyav && "
+    "lerobot-annotate "
+    "--repo_id=pepijn223/robocasa_pretrain_human300_v4 "
+    "--new_repo_id=pepijn223/robocasa_pretrain_human300_v4_annotated5 "
+    "--push_to_hub=true "
+    "--vlm.backend=openai "
+    "--vlm.model_id=Qwen/Qwen3.6-27B "
+    "--vlm.parallel_servers=1 "
+    "--vlm.num_gpus=1 "
+    '--vlm.serve_command="vllm serve Qwen/Qwen3.6-27B '
+    "--tensor-parallel-size 1 --max-model-len 32768 "
+    '--gpu-memory-utilization 0.8 --uvicorn-log-level warning --port {port}" '
+    "--vlm.serve_ready_timeout_s=1800 "
+    "--vlm.client_concurrency=128 "
+    "--vlm.max_new_tokens=512 "
+    "--vlm.temperature=0.7 "
+    "--executor.episode_parallelism=16 "
+    "--vlm.chat_template_kwargs='{\"enable_thinking\": false}' "
+    "--vlm.camera_key=observation.images.robot0_agentview_right "
+    # Phase 1 — plan module (subtasks + memory).
+    # Embed decoded frames (not a file:// clip): if clip extraction fails,
+    # the video_url path silently sends no video and the VLM hallucinates.
+    "--plan.use_video_url=false "
+    "--plan.frames_per_second=1.0 "
+    # 32 frames ≈ 8-10k vision tokens, fits the 32768 context. Don't push
+    # toward 128 — that overflows the context (BadRequestError 400).
+    "--plan.max_video_frames=32 "
+    # Window long episodes into 32s chunks (constant 1 fps density) so they
+    # get more subtasks; per-window spans are merged + stitched. 0 disables.
+    "--plan.subtask_window_seconds=32 "
+    # RoboCasa: the dataset task string is authoritative (eval uses it), so
+    # keep it driving subtasks. ``always`` would throw it away and hallucinate.
+    "--plan.derive_task_from_video=off "
+    # No task augmentation: eval conditions on the exact task strings, so
+    # rephrasings are unused at best and harmful when they drift.
+    "--plan.n_task_rephrasings=0 "
+    # Keep subtask decomposition tight for atomic tasks.
+    "--plan.plan_max_steps=10 "
+    # Only subtasks + memory — skip the numbered "plan" rows. true re-enables.
+    "--plan.emit_plan=false "
+    # The describe->segment grounding pass (+1 VLM call/episode) is ON by
+    # default; pass --plan.subtask_describe_first=false to skip it.
+    # Phase 2 — interjections + speech.
+    "--interjections.max_interjections_per_episode=6 "
+    # Phase 4 — general VQA: disabled for this run.
+    "--vqa.enabled=false"
+)
+
+job = run_job(
+    image="vllm/vllm-openai:latest",
+    command=["bash", "-c", CMD],
+    flavor="h200",
+    secrets={"HF_TOKEN": token},
+    timeout="2h",
+)
+print(f"Job URL: {job.url}")
+print(f"Job ID:  {job.id}")

examples/dataset/create_progress_videos.py

@@ -15,10 +15,12 @@
 # limitations under the License.
 
 """
-Create MP4 (or GIF) videos with sarm_progress overlay for specified episodes.
+Create MP4 (or GIF) videos with per-frame progress overlay for specified episodes.
 
 Downloads datasets from HuggingFace, seeks directly into the episode segment
 of the source video, draws a progress line on each frame, and writes the result.
+The progress data is read from a parquet file that lives alongside the dataset
+(configurable via ``--progress-file``).
 
 Usage:
     python examples/dataset/create_progress_videos.py \
@@ -56,22 +58,26 @@ SCORE_FONT_SCALE = 0.8
 TASK_FONT_SCALE = 0.55
 
 
-def download_episode_metadata(repo_id: str, episode: int) -> Path:
-    """Download only the metadata and sarm_progress files for a dataset.
+def download_episode_metadata(
+    repo_id: str, episode: int, progress_file: str = "sarm_progress.parquet"
+) -> Path:
+    """Download only the metadata and per-frame progress file for a dataset.
 
     Args:
         repo_id: HuggingFace dataset repository ID.
         episode: Episode index (used for logging only; all meta is fetched).
+        progress_file: Filename of the per-frame progress parquet inside the
+            dataset repo.
 
     Returns:
         Local cache path for the downloaded snapshot.
     """
-    logging.info("[1/4] Downloading metadata for %s (episode %d) ...", repo_id, episode)
+    logging.info("[1/4] Downloading metadata + %s for %s (episode %d) ...", progress_file, repo_id, episode)
     local_path = Path(
         snapshot_download(
             repo_id=repo_id,
             repo_type="dataset",
-            allow_patterns=["meta/**", "sarm_progress.parquet"],
+            allow_patterns=["meta/**", progress_file],
             ignore_patterns=["*.mp4"],
         )
     )
@@ -215,25 +221,28 @@ def download_video_file(repo_id: str, local_path: Path, video_rel: str) -> Path:
     return video_path
 
 
-def load_progress_data(local_path: Path, episode: int) -> np.ndarray | None:
-    """Load sarm_progress values for an episode.
+def load_progress_data(
+    local_path: Path, episode: int, progress_file: str = "sarm_progress.parquet"
+) -> np.ndarray | None:
+    """Load per-frame progress values for an episode.
 
     Args:
         local_path: Dataset cache root.
         episode: Episode index.
+        progress_file: Filename of the per-frame progress parquet.
 
     Returns:
         Sorted (N, 2) array of (frame_index, progress), or None if unavailable.
     """
-    parquet_path = local_path / "sarm_progress.parquet"
+    parquet_path = local_path / progress_file
     if not parquet_path.exists():
-        logging.warning("sarm_progress.parquet not found")
+        logging.warning("%s not found", progress_file)
         return None
     df = pd.read_parquet(parquet_path)
-    logging.info("   sarm_progress.parquet columns: %s", list(df.columns))
+    logging.info("   %s columns: %s", progress_file, list(df.columns))
     episode_df = df[df["episode_index"] == episode].copy()
     if episode_df.empty:
-        logging.warning("No sarm_progress rows for episode %d", episode)
+        logging.warning("No progress rows for episode %d in %s", episode, progress_file)
         return None
     episode_df = episode_df.sort_values("frame_index")
 
@@ -576,6 +585,7 @@ def process_dataset(
     camera_key: str | None,
     output_dir: Path,
     create_gif: bool = False,
+    progress_file: str = "sarm_progress.parquet",
 ) -> Path | None:
     """Full pipeline: download, extract metadata, composite progress, write output.
 
@@ -585,6 +595,8 @@ def process_dataset(
         camera_key: Camera key to use, or None for auto-selection.
         output_dir: Directory to write output files.
         create_gif: If True, also generate a GIF from the MP4.
+        progress_file: Filename of the per-frame progress parquet inside the
+            dataset repo.
 
     Returns:
         Path to the final output file, or None on failure.
@@ -592,7 +604,7 @@ def process_dataset(
     safe_name = repo_id.replace("/", "_")
     logging.info("Processing: %s  |  episode %d", repo_id, episode)
 
-    local_path = download_episode_metadata(repo_id, episode)
+    local_path = download_episode_metadata(repo_id, episode, progress_file)
     logging.info("   Local cache: %s", local_path)
 
     episode_meta = load_episode_meta(local_path, episode, camera_key)
@@ -600,9 +612,9 @@ def process_dataset(
 
     video_path = download_video_file(repo_id, local_path, episode_meta["video_rel"])
 
-    progress_data = load_progress_data(local_path, episode)
+    progress_data = load_progress_data(local_path, episode, progress_file)
     if progress_data is None:
-        logging.error("Could not load sarm_progress data. Skipping overlay.")
+        logging.error("Could not load progress data from %s. Skipping overlay.", progress_file)
         return None
 
     logging.info("   Progress frames: %d", len(progress_data))
@@ -627,7 +639,7 @@ def process_dataset(
 
 def main() -> None:
     parser = argparse.ArgumentParser(
-        description="Create MP4/GIF videos with sarm_progress overlay for dataset episodes."
+        description="Create MP4/GIF videos with per-frame progress overlay for dataset episodes."
     )
     parser.add_argument(
         "--repo-id",
@@ -658,6 +670,15 @@ def main() -> None:
         action="store_true",
         help="Also generate a GIF from the MP4 output.",
     )
+    parser.add_argument(
+        "--progress-file",
+        type=str,
+        default="sarm_progress.parquet",
+        help=(
+            "Filename of the per-frame progress parquet inside the dataset repo "
+            "(default: 'sarm_progress.parquet')."
+        ),
+    )
     args = parser.parse_args()
 
     logging.basicConfig(level=logging.INFO, format="%(levelname)s: %(message)s")
@@ -670,6 +691,7 @@ def main() -> None:
         camera_key=args.camera_key,
         output_dir=args.output_dir,
         create_gif=args.gif,
+        progress_file=args.progress_file,
     )
 
     if result:

pyproject.toml

@@ -85,6 +85,11 @@ dependencies = [
     "termcolor>=2.4.0,<4.0.0",
     "tqdm>=4.66.0,<5.0.0",
 
+    # Training utilities
+    # EMA of policy parameters (Diffusion Policy / pi05 style). Tiny
+    # pure-python dependency — preferred over a hand-rolled implementation.
+    "ema-pytorch>=0.7.7,<1.0.0",
+
     # Build tools (required by opencv-python-headless on some platforms)
     "cmake>=3.29.0.1,<4.2.0",
     "setuptools>=71.0.0,<81.0.0",
@@ -138,8 +143,11 @@ dataset_viz = ["lerobot[dataset]", "lerobot[viz]"]
 # Common
 av-dep = ["av>=15.0.0,<16.0.0"]
 pygame-dep = ["pygame>=2.5.1,<2.7.0"]
-placo-dep = ["placo>=0.9.6,<0.9.17"]
+# NOTE: 0.9.16 links against liburdfdom_sensor.so.4, which is unavailable on Ubuntu 24.04
+# (noble ships urdfdom 3.x). Cap below 0.9.16 until system urdfdom 4.x is broadly available.
+placo-dep = ["placo>=0.9.6,<0.9.16"]
 transformers-dep = ["transformers>=5.4.0,<5.6.0"]
+sentencepiece-dep = ["sentencepiece>=0.2.0,<0.3.0"] # FAST action tokenizer backend (pi052, pi0_fast)
 grpcio-dep = ["grpcio==1.73.1", "protobuf>=6.31.1,<6.32.0"]
 can-dep = ["python-can>=4.2.0,<5.0.0"]
 peft-dep = ["peft>=0.18.0,<1.0.0"]
@@ -195,7 +203,8 @@ wallx = [
     "torchdiffeq>=0.2.4,<0.3.0",
     "lerobot[qwen-vl-utils-dep]",
 ]
-pi = ["lerobot[transformers-dep]", "lerobot[scipy-dep]"]
+pi = ["lerobot[transformers-dep]", "lerobot[scipy-dep]", "lerobot[sentencepiece-dep]"]
+molmoact2 = ["lerobot[transformers-dep]", "lerobot[peft-dep]", "lerobot[scipy-dep]"]
 smolvla = ["lerobot[transformers-dep]", "num2words>=0.5.14,<0.6.0", "accelerate>=1.7.0,<2.0.0"]
 multi_task_dit = ["lerobot[transformers-dep]", "lerobot[diffusers-dep]"]
 groot = [
@@ -209,14 +218,40 @@ groot = [
     "flash-attn>=2.5.9,<3.0.0 ; sys_platform != 'darwin'"
 ]
 sarm = ["lerobot[transformers-dep]", "pydantic>=2.0.0,<3.0.0", "faker>=33.0.0,<35.0.0", "lerobot[matplotlib-dep]", "lerobot[qwen-vl-utils-dep]"]
+robometer = ["lerobot[transformers-dep]", "lerobot[qwen-vl-utils-dep]", "lerobot[peft-dep]"]
+topreward = ["lerobot[transformers-dep]"]
 xvla = ["lerobot[transformers-dep]"]
 eo1 = ["lerobot[transformers-dep]", "lerobot[qwen-vl-utils-dep]"]
 hilserl = ["lerobot[transformers-dep]", "lerobot[dataset]", "gym-hil>=0.1.13,<0.2.0", "lerobot[grpcio-dep]", "lerobot[placo-dep]"]
+vla_jepa = ["lerobot[transformers-dep]", "lerobot[diffusers-dep]", "lerobot[qwen-vl-utils-dep]"]
 
 # Features
 async = ["lerobot[grpcio-dep]", "lerobot[matplotlib-dep]"]
 peft = ["lerobot[transformers-dep]", "lerobot[peft-dep]"]
 
+# Annotation pipeline (lerobot-annotate). The only backend is ``openai``,
+# which talks to any OpenAI-compatible server (``vllm serve`` /
+# ``transformers serve`` / hosted). Distributed runs use Hugging Face Jobs
+# (see examples/annotations/run_hf_job.py).
+annotations = [
+    "lerobot[dataset]",
+    "lerobot[transformers-dep]",
+    "openai>=1.40,<2.0",
+    # ``vllm`` is intentionally NOT a hard dep: it pins an older torch, and
+    # uv's single unified lock would then cap ``torch`` for every extra
+    # (e.g. forcing 2.8 while ``torchcodec`` in [dataset] needs 2.11 -> ABI
+    # break in CI). The HF Jobs image (``vllm/vllm-openai``) provides vLLM;
+    # install it locally only if you run your own ``vllm serve``.
+]
+
+# Tool implementations under src/lerobot/tools/. Each tool's dependencies
+# are isolated so adding a new tool doesn't bloat the base install.
+# Currently only `say` (Kyutai pocket-tts; CPU-only, ~100M params).
+tools = [
+    "pocket-tts>=1.0.0,<3.0.0",
+    "scipy>=1.11.0,<2.0.0",  # SayTool.output_dir uses scipy.io.wavfile
+]
+
 # Development
 dev = ["pre-commit>=3.7.0,<5.0.0", "debugpy>=1.8.1,<1.9.0", "lerobot[grpcio-dep]", "grpcio-tools==1.73.1", "mypy>=1.19.1", "ruff>=0.14.1", "lerobot[notebook]"]
 notebook = ["jupyter>=1.0.0,<2.0.0", "ipykernel>=6.0.0,<7.0.0"]
@@ -272,10 +307,12 @@ all = [
     "lerobot[multi_task_dit]",
     "lerobot[wallx]",
     "lerobot[pi]",
+    "lerobot[molmoact2]",
     "lerobot[smolvla]",
     # "lerobot[groot]", TODO(Steven): Gr00t requires specific installation instructions for flash-attn
     "lerobot[xvla]",
     "lerobot[hilserl]",
+    "lerobot[vla_jepa]",
     "lerobot[async]",
     "lerobot[dev]",
     "lerobot[test]",
@@ -286,6 +323,8 @@ all = [
     "lerobot[libero]; sys_platform == 'linux'",
     "lerobot[metaworld]",
     "lerobot[sarm]",
+    "lerobot[robometer]",
+    "lerobot[topreward]",
     "lerobot[peft]",
     # "lerobot[unitree_g1]", TODO: Unitree requires specific installation instructions for unitree_sdk2
 ]
@@ -307,7 +346,10 @@ lerobot-find-joint-limits="lerobot.scripts.lerobot_find_joint_limits:main"
 lerobot-imgtransform-viz="lerobot.scripts.lerobot_imgtransform_viz:main"
 lerobot-edit-dataset="lerobot.scripts.lerobot_edit_dataset:main"
 lerobot-setup-can="lerobot.scripts.lerobot_setup_can:main"
+lerobot-annotate="lerobot.scripts.lerobot_annotate:main"
 lerobot-rollout="lerobot.scripts.lerobot_rollout:main"
+# Interactive hierarchical-VLA runtime for PI052 (PaliGemma backbone).
+lerobot-pi052-runtime="lerobot.scripts.lerobot_pi052_runtime:main"
 
 # ---------------- Tool Configurations ----------------
 
@@ -325,7 +367,7 @@ torch = [{ index = "pytorch-cu128", marker = "sys_platform == 'linux'" }]
 torchvision = [{ index = "pytorch-cu128", marker = "sys_platform == 'linux'" }]
 
 [tool.setuptools.package-data]
-lerobot = ["envs/*.json"]
+lerobot = ["envs/*.json", "annotations/steerable_pipeline/prompts/*.txt"]
 
 [tool.setuptools.packages.find]
 where = ["src"]
@@ -401,8 +443,11 @@ default.extend-ignore-identifiers-re = [
     "ein",
     "thw",
     "inpt",
+    "arange",
+    "is_compileable",
     "ROBOTIS",
-    "OT_VALUE"
+    "OT_VALUE",
+    "VanderBilt"
 ]
 
 # TODO: Uncomment when ready to use

src/lerobot/annotations/__init__.py

@@ -0,0 +1,15 @@
+#!/usr/bin/env python
+
+# Copyright 2026 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.

src/lerobot/annotations/steerable_pipeline/__init__.py

@@ -0,0 +1,36 @@
+#!/usr/bin/env python
+
+# Copyright 2026 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""Steerable annotation pipeline producing ``language_persistent`` and
+``language_events`` columns for LeRobot datasets.
+
+The pipeline is decomposed into three independently runnable modules whose
+outputs are staged per-episode before a final parquet rewrite:
+
+- :mod:`.modules.plan_subtasks_memory` (the ``plan`` module) — persistent styles
+- :mod:`.modules.interjections_and_speech` (the ``interjections`` module) — event styles + speech
+- :mod:`.modules.general_vqa` (the ``vqa`` module) — event-style VQA pairs
+"""
+
+from .config import AnnotationPipelineConfig
+from .validator import StagingValidator, ValidationReport
+from .writer import LanguageColumnsWriter
+
+__all__ = [
+    "AnnotationPipelineConfig",
+    "LanguageColumnsWriter",
+    "StagingValidator",
+    "ValidationReport",
+]

src/lerobot/annotations/steerable_pipeline/config.py

@@ -0,0 +1,196 @@
+#!/usr/bin/env python
+
+# Copyright 2026 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from __future__ import annotations
+
+from dataclasses import dataclass, field
+from pathlib import Path
+from typing import Any
+
+
+@dataclass
+class PlanConfig:
+    """``plan`` module: subtasks + plan + memory + task augmentation."""
+
+    enabled: bool = True
+
+    # ``task_aug`` rephrasings at t=0 (renderer rotates ${task} among them); 0 disables.
+    n_task_rephrasings: int = 10
+
+    # Derive the task from video instead of episode_task: off / if_short / always.
+    # Affects prompts only; ``meta/tasks.parquet`` is untouched.
+    derive_task_from_video: str = "if_short"
+    derive_task_min_words: int = 3
+
+    # Frames sampled uniformly, capped at max_video_frames — a hard context cap
+    # (~300 tokens/frame, so 32 fit a 32k VLM; 128 overflow).
+    frames_per_second: float = 1.0
+    max_video_frames: int = 32
+
+    # >0: split long episodes into windows of this length (constant fps density)
+    # instead of subsampling the whole episode; spans merged + stitched. 0 disables.
+    subtask_window_seconds: float = 0.0
+
+    min_subtask_seconds: float = 1.5
+    plan_max_steps: int = 8
+
+    # Narrate-only grounding pass before segmenting — best defense against subtasks
+    # invented from the task text (+1 VLM call/episode).
+    subtask_describe_first: bool = True
+
+    # Emit ``style="plan"`` rows at each boundary; False = subtasks + memory only.
+    emit_plan: bool = True
+
+    # (subtask spans are always stitched to a contiguous full-episode cover; not configurable.)
+
+    # Send a server-side ``video_url`` clip (at use_video_url_fps) instead of embedded frames.
+    use_video_url: bool = False
+    use_video_url_fps: float = 1.0
+
+    # Optional EgoMimic-style 5-axis task augmentation; replaces n_task_rephrasings.
+    task_aug_axes: TaskAugAxesConfig = field(default_factory=lambda: TaskAugAxesConfig())
+
+
+@dataclass
+class TaskAugAxesConfig:
+    """5-axis t=0 task augmentation (EgoMimic-style): synonym / omit_arm /
+    omit_orientation / omit_grasp_method / combined. Replaces n_task_rephrasings
+    when enabled; each variant becomes a ``task_aug`` row. Axes with nothing to
+    omit emit fewer entries. Defaults (3+3+2+2+2) match EgoMimic."""
+
+    enabled: bool = False
+
+    synonym_paraphrase: int = 3
+    omit_arm: int = 3
+    omit_orientation: int = 2
+    omit_grasp_method: int = 2
+    combined_omissions: int = 2
+
+
+@dataclass
+class InterjectionsConfig:
+    """``interjections`` module: interjections + paired speech."""
+
+    enabled: bool = True
+
+    # Each emits a paired (interjection, speech) row + a plan refresh at that ts.
+    max_interjections_per_episode: int = 3
+    interjection_min_t: float = 2.0
+
+    # Frame window centered on the timestamp so the VLM sees motion, not one frame.
+    interjection_window_seconds: float = 2.0
+    interjection_window_frames: int = 4
+
+
+@dataclass
+class VqaConfig:
+    """``vqa`` module: general VQA."""
+
+    enabled: bool = True
+    vqa_emission_hz: float = 1.0
+    K: int = 1
+    """Consecutive frames per emission tick. The VLM grounds on the FIRST frame,
+    so K>1 smears stale labels onto moved frames. Default 1 (no smear)."""
+    question_types: tuple[str, ...] = ("bbox", "keypoint", "count", "attribute", "spatial")
+
+    # True: ground VQA only on --vlm.camera_key (default: every camera).
+    restrict_to_default_camera: bool = False
+
+
+@dataclass
+class VlmConfig:
+    """Shared Qwen-VL client configuration."""
+
+    # Only ``openai`` (OpenAI-compatible vLLM server, auto-spawned when
+    # auto_serve=True); ``stub`` is for tests.
+    backend: str = "openai"
+    model_id: str = "Qwen/Qwen3.6-27B"
+
+    # OpenAI-compatible endpoint; ``EMPTY`` key works for local servers.
+    api_base: str = "http://localhost:8000/v1"
+    api_key: str = "EMPTY"
+
+    # Spawn a server if none answers api_base; False = fail fast on a remote.
+    auto_serve: bool = True
+    serve_port: int = 8000
+    # Override the auto-serve command; ``{port}`` substituted per replica.
+    serve_command: str | None = None
+
+    # Independent servers for round-robin routing (one per GPU). num_gpus=0 = one each.
+    parallel_servers: int = 1
+    num_gpus: int = 0
+    client_concurrency: int = 16
+    serve_ready_timeout_s: float = 600.0
+
+    max_new_tokens: int = 512
+    temperature: float = 0.2
+
+    # Auto-serve context length (None → 32768); other vLLM flags go in serve_command.
+    max_model_len: int | None = None
+
+    # Camera for keyframes; None → first ``observation.images.*`` key.
+    camera_key: str | None = None
+    # Forwarded as extra_body.chat_template_kwargs (e.g. {"enable_thinking": false}).
+    chat_template_kwargs: dict[str, Any] | None = None
+
+
+@dataclass
+class ExecutorConfig:
+    """Executor settings (intra-process episode concurrency; distribution via HF Jobs)."""
+
+    # Episodes processed concurrently per phase; main knob for saturating the servers.
+    episode_parallelism: int = 16
+
+
+@dataclass
+class AnnotationPipelineConfig:
+    """Top-level config for ``lerobot-annotate`` (rewrites data shards in place)."""
+
+    # Hub dataset: download source when ``root`` unset; push target when push_to_hub
+    # is on and ``new_repo_id`` unset.
+    repo_id: str | None = None
+
+    # Separate push target (matches the LeRobot edit tools). Unset → push in place.
+    new_repo_id: str | None = None
+
+    root: Path | None = None
+
+    # Defaults to ``<root>/.annotate_staging/``.
+    staging_dir: Path | None = None
+
+    seed: int = 1729
+
+    plan: PlanConfig = field(default_factory=PlanConfig)
+    interjections: InterjectionsConfig = field(default_factory=InterjectionsConfig)
+    vqa: VqaConfig = field(default_factory=VqaConfig)
+
+    vlm: VlmConfig = field(default_factory=VlmConfig)
+    executor: ExecutorConfig = field(default_factory=ExecutorConfig)
+
+    skip_validation: bool = False
+    only_episodes: tuple[int, ...] | None = None
+
+    # Keyframe decode backend. None → ffmpeg CLI (crash-/thread-safe; torchcodec
+    # SIGSEGVs under concurrent decode). Or ``"torchcodec"`` / ``"pyav"``.
+    video_backend: str | None = None
+
+    # Upload to the Hub (new_repo_id if set, else repo_id; one must be set).
+    push_to_hub: bool = False
+    push_private: bool = False
+    push_commit_message: str | None = None
+
+    def resolved_staging_dir(self, root: Path) -> Path:
+        return self.staging_dir if self.staging_dir is not None else root / ".annotate_staging"

src/lerobot/annotations/steerable_pipeline/executor.py

@@ -0,0 +1,253 @@
+#!/usr/bin/env python
+
+# Copyright 2026 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""In-process executor that runs the annotation phases.
+
+The executor runs **six phases** in dependency order:
+
+    phase 1: ``plan`` module (plan + subtasks + memory)
+    phase 2: ``interjections`` module (interjections + speech)
+    phase 3: ``plan`` plan-update pass — re-runs plan emission at every
+             interjection timestamp produced by phase 2
+    phase 4: ``vqa`` module (VQA)
+    phase 5: validator
+    phase 6: writer
+
+Phase 3 is why the ``plan`` module must be re-entered after the
+``interjections`` module — to refresh ``plan`` rows at interjection
+timestamps.
+
+Distributed execution is provided by Hugging Face Jobs (see
+``examples/annotations/run_hf_job.py``); the runner inside the job
+invokes ``lerobot-annotate`` which uses this in-process executor.
+Episode-level concurrency is controlled by
+``ExecutorConfig.episode_parallelism``.
+"""
+
+from __future__ import annotations
+
+import logging
+import time
+from concurrent.futures import ThreadPoolExecutor, as_completed
+from dataclasses import dataclass
+from pathlib import Path
+from typing import Any
+
+from .config import AnnotationPipelineConfig
+from .reader import EpisodeRecord, iter_episodes
+from .staging import EpisodeStaging
+from .validator import StagingValidator
+from .writer import LanguageColumnsWriter
+
+logger = logging.getLogger(__name__)
+
+
+@dataclass
+class PhaseResult:
+    """Summary of one pipeline phase across all episodes."""
+
+    name: str
+    episodes_processed: int
+    episodes_skipped: int
+
+
+@dataclass
+class PipelineRunSummary:
+    """Aggregated result returned by :meth:`Executor.run`."""
+
+    phases: list[PhaseResult]
+    written_paths: list[Path]
+    validation_report: Any  # ValidationReport, kept Any to avoid import cycle
+
+
+@dataclass
+class Executor:
+    """Run all six phases over a dataset root in-process.
+
+    Episode-level concurrency comes from ``ExecutorConfig.episode_parallelism``
+    (a thread pool); cluster-level concurrency comes from running this
+    executor inside a Hugging Face Job. Tests construct the executor
+    directly with stub modules.
+    """
+
+    config: AnnotationPipelineConfig
+    plan: Any  # PlanSubtasksMemoryModule
+    interjections: Any  # InterjectionsAndSpeechModule
+    vqa: Any  # GeneralVqaModule
+    writer: LanguageColumnsWriter
+    validator: StagingValidator
+
+    def run(self, root: Path) -> PipelineRunSummary:
+        records = list(iter_episodes(root, only_episodes=self.config.only_episodes))
+        n = len(records)
+        if n == 0:
+            raise ValueError(f"No episodes found under {root}/data/")
+
+        print(f"[annotate] {n} episodes total", flush=True)
+
+        staging_dir = self.config.resolved_staging_dir(root)
+        staging_dir.mkdir(parents=True, exist_ok=True)
+
+        phases: list[PhaseResult] = []
+
+        # Phase 1: ``plan`` module (plan + subtasks + memory)
+        phases.append(self._run_module_phase("plan", records, staging_dir, self.plan))
+        # Phase 2: ``interjections`` module (interjections + speech). It
+        # reads the ``plan`` module's subtask rows from the same staging
+        # tree to ground the interjection prompt in the correct local subtask.
+        phases.append(self._run_module_phase("interjections", records, staging_dir, self.interjections))
+        # Phase 3: ``plan`` plan-update pass at interjection timestamps.
+        phases.append(self._run_plan_update_phase(records, staging_dir))
+        # Phase 4: ``vqa`` module (VQA)
+        phases.append(self._run_module_phase("vqa", records, staging_dir, self.vqa))
+
+        print("[annotate] running validator...", flush=True)
+        report = self.validator.validate(records, staging_dir)
+        if not report.ok and not self.config.skip_validation:
+            raise RuntimeError(f"Staging validation failed: {report.summary()}")
+        print(f"[annotate] validator: {report.summary()}", flush=True)
+
+        print(f"[annotate] writing parquet shards into {root}/data/...", flush=True)
+        written = self.writer.write_all(records, staging_dir, root)
+        print(f"[annotate] wrote {len(written)} shard(s); pipeline complete", flush=True)
+
+        # Keep meta/info.json aligned with the parquet schema we just wrote.
+        # Idempotent and additive: existing user metadata is preserved.
+        self._ensure_annotation_metadata_in_info(root)
+
+        return PipelineRunSummary(phases=phases, written_paths=written, validation_report=report)
+
+    @staticmethod
+    def _ensure_annotation_metadata_in_info(root: Path) -> None:
+        """Write language features and canonical tools to ``meta/info.json``.
+
+        ``LanguageColumnsWriter`` adds ``language_persistent`` and
+        ``language_events`` to parquet shards. The metadata must advertise
+        those columns too, otherwise non-streaming ``LeRobotDataset`` loads
+        cast against the old schema and fail on the extra parquet columns.
+        """
+        from lerobot.datasets.io_utils import load_info, write_info  # noqa: PLC0415
+        from lerobot.datasets.language import SAY_TOOL_SCHEMA, language_feature_info  # noqa: PLC0415
+
+        info_path = root / "meta" / "info.json"
+        if not info_path.exists():
+            return
+        try:
+            info = load_info(root)
+        except Exception as exc:  # noqa: BLE001
+            print(f"[annotate] could not read {info_path}: {exc}", flush=True)
+            return
+
+        changed = False
+
+        merged_features = {**info.features, **language_feature_info()}
+        if merged_features != info.features:
+            info.features = merged_features
+            changed = True
+
+        existing = info.tools or []
+        names = {(t.get("function") or {}).get("name") for t in existing if isinstance(t, dict)}
+        if SAY_TOOL_SCHEMA["function"]["name"] not in names:
+            info.tools = [*existing, SAY_TOOL_SCHEMA]
+            changed = True
+
+        if changed:
+            write_info(info, root)
+            print(
+                "[annotate] meta/info.json: "
+                f"language_features={list(language_feature_info())}, "
+                f"tools={[t['function']['name'] for t in (info.tools or [])]}",
+                flush=True,
+            )
+
+    def _run_module_phase(
+        self,
+        name: str,
+        records: list[EpisodeRecord],
+        staging_dir: Path,
+        module: Any,
+    ) -> PhaseResult:
+        if not module.enabled:
+            print(f"[annotate] phase={name} skipped (module disabled)", flush=True)
+            return PhaseResult(name=name, episodes_processed=0, episodes_skipped=len(records))
+        n = len(records)
+        parallelism = max(1, min(self.config.executor.episode_parallelism, n))
+        print(
+            f"[annotate] phase={name} starting on {n} episode(s) (parallelism={parallelism})",
+            flush=True,
+        )
+        t0 = time.time()
+
+        def _do(idx_record: tuple[int, EpisodeRecord]) -> tuple[int, int, float]:
+            i, record = idx_record
+            ep_start = time.time()
+            staging = EpisodeStaging(staging_dir, record.episode_index)
+            module.run_episode(record, staging)
+            return i, record.episode_index, time.time() - ep_start
+
+        processed = 0
+        if parallelism == 1:
+            for i, record in enumerate(records, 1):
+                _, ep_idx, elapsed = _do((i, record))
+                processed += 1
+                print(
+                    f"[annotate]   {name} episode {i}/{n} (idx={ep_idx}) done in {elapsed:.1f}s",
+                    flush=True,
+                )
+        else:
+            with ThreadPoolExecutor(max_workers=parallelism) as pool:
+                futures = [pool.submit(_do, (i, r)) for i, r in enumerate(records, 1)]
+                for fut in as_completed(futures):
+                    i, ep_idx, elapsed = fut.result()
+                    processed += 1
+                    print(
+                        f"[annotate]   {name} episode {processed}/{n} "
+                        f"(idx={ep_idx}, submit_order={i}) done in {elapsed:.1f}s",
+                        flush=True,
+                    )
+        total = time.time() - t0
+        print(f"[annotate] phase={name} complete: {processed}/{n} in {total:.1f}s", flush=True)
+        return PhaseResult(name=name, episodes_processed=processed, episodes_skipped=0)
+
+    def _run_plan_update_phase(  # noqa: PLR0915
+        self, records: list[EpisodeRecord], staging_dir: Path
+    ) -> PhaseResult:
+        """Re-emit ``plan`` rows at each timestamp the ``interjections`` module produced.
+
+        The ``plan`` module owns the prompt; the ``interjections`` module
+        produced the timestamps. This phase therefore calls back into the
+        ``plan`` module with the interjection timestamps so its existing
+        prompt path is reused.
+        """
+        if not self.plan.enabled or not self.interjections.enabled:
+            return PhaseResult(name="plan_update", episodes_processed=0, episodes_skipped=len(records))
+        processed = 0
+        for record in records:
+            staging = EpisodeStaging(staging_dir, record.episode_index)
+            interjection_rows = [
+                row for row in staging.read("interjections") if row.get("style") == "interjection"
+            ]
+            interjection_times = [float(row["timestamp"]) for row in interjection_rows]
+            interjection_texts = [str(row.get("content") or "") for row in interjection_rows]
+            if interjection_times:
+                self.plan.run_plan_updates(record, staging, interjection_times, interjection_texts)
+                processed += 1
+        # Episodes without any interjections are skipped (no plan refresh
+        # needed); count them so the summary's processed+skipped == total.
+        return PhaseResult(
+            name="plan_update",
+            episodes_processed=processed,
+            episodes_skipped=len(records) - processed,
+        )

src/lerobot/annotations/steerable_pipeline/frames.py

@@ -0,0 +1,498 @@
+#!/usr/bin/env python
+
+# Copyright 2026 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""Keyframe extraction for the annotation pipeline.
+
+Modules attach decoded camera frames to their VLM prompts so the model can
+ground subtask decomposition, interjection scenarios, and VQA in actual
+visual content. The pipeline shares one provider across modules and one
+episode at a time, with a small per-episode cache so multiple modules
+querying the same timestamp pay decode cost once.
+"""
+
+from __future__ import annotations
+
+import logging
+import threading
+from dataclasses import dataclass, field
+from pathlib import Path
+from typing import Any, Protocol
+
+import PIL.Image
+import torch
+
+from lerobot.datasets.video_utils import decode_video_frames
+
+from .reader import EpisodeRecord
+
+logger = logging.getLogger(__name__)
+
+
+class FrameProvider(Protocol):
+    """Decodes camera frames at episode-relative timestamps."""
+
+    @property
+    def camera_keys(self) -> list[str]:
+        """All ``observation.images.*`` feature keys this provider can decode."""
+
+    def frames_at(
+        self,
+        record: EpisodeRecord,
+        timestamps: list[float],
+        camera_key: str | None = None,
+    ) -> list[Any]:
+        """Return one decoded frame per timestamp from ``camera_key`` (or default).
+
+        Frames are ``torch.Tensor`` (``C, H, W`` uint8) — the shape
+        :func:`lerobot.datasets.video_utils.decode_video_frames` returns.
+        :func:`to_image_blocks` converts them to PIL only at the VLM-message
+        boundary.
+
+        Empty list if the camera is unavailable. ``camera_key=None`` falls back
+        to the provider's default camera so existing single-camera callers
+        (the ``plan`` and ``interjections`` modules) keep working unchanged.
+        """
+
+    def video_for_episode(
+        self,
+        record: EpisodeRecord,
+        max_frames: int,
+        camera_key: str | None = None,
+    ) -> list[Any]:
+        """Return up to ``max_frames`` decoded frames covering the whole episode.
+
+        Sampling is uniform across the episode duration. Frames are
+        ``torch.Tensor`` (``C, H, W`` uint8); :func:`to_video_block` wraps
+        them into one ``{"type":"video", "video":<list>}`` block for a
+        Qwen-VL-compatible model that pools temporally itself. Empty list if
+        no camera available.
+        """
+
+
+@dataclass
+class _NullProvider:
+    """No-op provider used when the dataset has no video keys or in tests."""
+
+    @property
+    def camera_keys(self) -> list[str]:
+        return []
+
+    def frames_at(
+        self,
+        record: EpisodeRecord,
+        timestamps: list[float],
+        camera_key: str | None = None,
+    ) -> list[Any]:
+        return []
+
+    def video_for_episode(
+        self,
+        record: EpisodeRecord,
+        max_frames: int,
+        camera_key: str | None = None,
+    ) -> list[Any]:
+        return []
+
+
+def null_provider() -> FrameProvider:
+    return _NullProvider()
+
+
+@dataclass
+class VideoFrameProvider:
+    """Decodes frames from the dataset's ``observation.images.*`` streams.
+
+    By default the *first* camera key is used for the ``plan`` module
+    (subtask decomposition) and the ``interjections`` module (interjection
+    scenarios) — those prompts care about *what is happening*, not which
+    angle. The ``vqa`` module instead iterates over every camera in
+    :attr:`camera_keys` so each frame's
+    grounded answer (bbox/keypoint/...) is tagged with the camera it was
+    grounded against.
+
+    ``camera_key`` overrides the default-camera choice but does not restrict
+    :attr:`camera_keys`. Pass ``camera_key`` explicitly to ``frames_at`` /
+    ``video_for_episode`` to read a non-default stream.
+
+    Caches up to ``cache_size`` decoded frames per process to keep
+    co-timestamped ``interjections`` + ``plan`` plan-update calls cheap.
+    """
+
+    root: Path
+    camera_key: str | None = None
+    tolerance_s: float = 1e-2
+    cache_size: int = 256
+    # Keyframe decode backend. ``None`` uses the ffmpeg CLI — the
+    # concurrency- and crash-safe default for the pipeline's threaded
+    # decode. Set to ``"torchcodec"`` or ``"pyav"`` to pin an in-process
+    # decoder when the build is known thread-safe.
+    video_backend: str | None = None
+    _meta: Any = field(default=None, init=False, repr=False)
+    _cache: dict = field(default_factory=dict, init=False, repr=False)
+    _camera_keys: list[str] = field(default_factory=list, init=False, repr=False)
+    # Pipeline runs the three module phases under a ThreadPoolExecutor (see
+    # ``ExecutorConfig.episode_parallelism``); guard the dict cache and the
+    # one-shot warn flag against concurrent updates from worker threads.
+    _lock: threading.Lock = field(default_factory=threading.Lock, init=False, repr=False)
+    _warned_decode_fail: bool = field(default=False, init=False, repr=False)
+
+    def __post_init__(self) -> None:
+        from lerobot.datasets.dataset_metadata import LeRobotDatasetMetadata  # noqa: PLC0415
+
+        self._meta = LeRobotDatasetMetadata(repo_id="local", root=self.root)
+        # Only ``video_keys`` are decodable here: the clip/decode paths read
+        # ``videos/<key>/from_timestamp`` from episode metadata, which exists
+        # only for video-stored cameras. Image-stored cameras (also in
+        # ``camera_keys``) would KeyError, so restrict the list — and the
+        # default — to video keys.
+        keys = list(self._meta.video_keys)
+        # Last-resort fallback: if metadata didn't surface any video keys but
+        # the caller explicitly named a camera (``--vlm.camera_key=...``),
+        # trust them — the key is by definition known to exist on the dataset.
+        if not keys and self.camera_key:
+            keys = [self.camera_key]
+        self._camera_keys = keys
+        if self.camera_key is None:
+            self.camera_key = keys[0] if keys else None
+
+    @property
+    def camera_keys(self) -> list[str]:
+        """All ``observation.images.*`` keys available on this dataset."""
+        return list(self._camera_keys)
+
+    def frames_at(
+        self,
+        record: EpisodeRecord,
+        timestamps: list[float],
+        camera_key: str | None = None,
+    ) -> list[Any]:
+        target = camera_key if camera_key is not None else self.camera_key
+        if not timestamps or target is None:
+            return []
+
+        out: list[Any] = []
+        misses: list[float] = []
+        miss_indices: list[int] = []
+        with self._lock:
+            for i, ts in enumerate(timestamps):
+                key = (record.episode_index, target, round(float(ts), 6))
+                cached = self._cache.get(key)
+                if cached is not None:
+                    out.append(cached)
+                else:
+                    out.append(None)
+                    misses.append(float(ts))
+                    miss_indices.append(i)
+
+        if misses:
+            decoded = self._decode(record.episode_index, misses, target)
+            # ``_decode`` returns exactly one frame per requested timestamp,
+            # or an empty list if decoding failed wholesale. A partial list
+            # would mean a frame/timestamp misalignment, so only pair them up
+            # when the counts match (``strict=True`` then guards regressions).
+            if len(decoded) == len(miss_indices):
+                with self._lock:
+                    for i, frame in zip(miss_indices, decoded, strict=True):
+                        out[i] = frame
+                        key = (record.episode_index, target, round(float(timestamps[i]), 6))
+                        if len(self._cache) >= self.cache_size:
+                            self._cache.pop(next(iter(self._cache)))
+                        self._cache[key] = frame
+        # filter out any None left over from decode failures
+        return [frame for frame in out if frame is not None]
+
+    def video_for_episode(
+        self,
+        record: EpisodeRecord,
+        max_frames: int,
+        camera_key: str | None = None,
+    ) -> list[Any]:
+        """Return up to ``max_frames`` frames uniformly sampled across the episode.
+
+        The whole episode duration is covered; the model picks subtask
+        boundaries from the temporal pooling it does internally. Frames are
+        ``torch.Tensor`` (see :meth:`frames_at`).
+        """
+        target = camera_key if camera_key is not None else self.camera_key
+        if max_frames <= 0 or target is None or not record.frame_timestamps:
+            return []
+        n_frames = min(max_frames, len(record.frame_timestamps))
+        if n_frames == len(record.frame_timestamps):
+            timestamps = list(record.frame_timestamps)
+        else:
+            t0 = record.frame_timestamps[0]
+            t_last = record.frame_timestamps[-1]
+            if t_last <= t0:
+                timestamps = [float(t0)] * n_frames
+            else:
+                step = (t_last - t0) / (n_frames - 1) if n_frames > 1 else 0.0
+                timestamps = [float(t0 + i * step) for i in range(n_frames)]
+        return self.frames_at(record, timestamps, camera_key=target)
+
+    def episode_clip_path(self, record: EpisodeRecord, cache_dir: Path) -> Path | None:
+        """Extract the episode's subclip to ``cache_dir/ep_{idx:06d}.mp4``.
+
+        Returns ``None`` if the dataset has no video tracks. Skips
+        re-extract when the cached clip already exists. Re-encodes to
+        H.264 (libx264) so the resulting mp4 is decodable by every
+        downstream video processor — stream-copy would inherit the
+        source codec (often AV1 in modern LeRobot datasets), which
+        vllm's libav build cannot decode.
+        """
+        import subprocess  # noqa: PLC0415
+
+        if self.camera_key is None:
+            return None
+        cache_dir.mkdir(parents=True, exist_ok=True)
+        out_path = cache_dir / f"ep_{record.episode_index:06d}.mp4"
+        if out_path.exists() and out_path.stat().st_size > 0:
+            return out_path
+        ep = self._meta.episodes[record.episode_index]
+        from_timestamp = float(ep[f"videos/{self.camera_key}/from_timestamp"])
+        to_timestamp = float(ep[f"videos/{self.camera_key}/to_timestamp"])
+        src = self.root / self._meta.get_video_file_path(record.episode_index, self.camera_key)
+        cmd = [
+            "ffmpeg",
+            "-y",
+            "-loglevel",
+            "error",
+            "-ss",
+            f"{from_timestamp:.3f}",
+            "-to",
+            f"{to_timestamp:.3f}",
+            "-i",
+            str(src),
+            "-c:v",
+            "libx264",
+            "-preset",
+            "ultrafast",
+            "-crf",
+            "23",
+            "-pix_fmt",
+            "yuv420p",
+            "-an",
+            str(out_path),
+        ]
+        try:
+            # ffmpeg is invoked by name via PATH lookup (the standard way to
+            # call the CLI); the arg list is fully controlled here, not shell.
+            subprocess.run(cmd, check=True, timeout=300)  # nosec B607
+        except (subprocess.CalledProcessError, subprocess.TimeoutExpired, FileNotFoundError):
+            return None
+        return out_path if out_path.exists() and out_path.stat().st_size > 0 else None
+
+    def _decode(self, episode_index: int, timestamps: list[float], camera_key: str) -> list[Any]:
+        """Decode ``timestamps`` from the episode's video as ``(C, H, W)`` tensors.
+
+        Delegates to :func:`lerobot.datasets.video_utils.decode_video_frames`
+        (torchcodec by default, PyAV fallback) rather than a bespoke decoder.
+        Returns one frame per requested timestamp, or ``[]`` if decoding
+        failed wholesale — callers treat ``[]`` as "no frames available".
+        """
+        ep = self._meta.episodes[episode_index]
+        from_timestamp = ep[f"videos/{camera_key}/from_timestamp"]
+        shifted = [from_timestamp + ts for ts in timestamps]
+        video_path = self.root / self._meta.get_video_file_path(episode_index, camera_key)
+
+        # Default to the ffmpeg CLI. The pipeline decodes under a 16-wide
+        # ThreadPoolExecutor and the in-process decoders are unsafe there:
+        # torchcodec is not thread-safe and SIGSEGVs under concurrent decode
+        # (a crash no try/except can catch), PyAV can likewise segfault on
+        # AV1, and lerobot's ``pyav`` backend routes through the removed
+        # ``torchvision.io.VideoReader``. ``_decode_frames_ffmpeg`` shells
+        # out per frame: each decode is an isolated child process, so it is
+        # both crash-safe and concurrency-safe. ``video_backend`` can pin
+        # ``torchcodec`` / ``pyav`` explicitly for callers that know their
+        # build is safe.
+        chain = [self.video_backend] if self.video_backend else ["ffmpeg"]
+
+        exc: Exception | None = None
+        for backend in chain:
+            try:
+                if backend == "ffmpeg":
+                    return _decode_frames_ffmpeg(video_path, shifted)
+                if backend in ("pyav", "av"):
+                    return _decode_frames_av(video_path, shifted)
+                # Stacked ``(N, C, H, W)`` uint8 tensor; one row per timestamp.
+                decoded = decode_video_frames(
+                    video_path, shifted, self.tolerance_s, backend=backend, return_uint8=True
+                )
+                return list(decoded)
+            except Exception as e:  # noqa: PERF203
+                exc = e
+
+        # Every backend raised. Log loudly the first time so a silent
+        # vqa-module no-op (every prompt skipped because frames_at returned
+        # []) is debuggable from the job log instead of post-hoc parquet
+        # inspection. Subsequent failures stay quiet.
+        with self._lock:
+            already_warned = self._warned_decode_fail
+            if not already_warned:
+                self._warned_decode_fail = True
+        if not already_warned:
+            logger.warning(
+                "VideoFrameProvider._decode failed for episode=%s camera=%s video_path=%s backends=%s: %s",
+                episode_index,
+                camera_key,
+                video_path,
+                chain,
+                exc,
+                exc_info=exc,
+            )
+        return []
+
+
+def make_frame_provider(
+    root: Path, camera_key: str | None = None, video_backend: str | None = None
+) -> FrameProvider:
+    """Build a :class:`VideoFrameProvider` if videos are present, else null."""
+    try:
+        provider = VideoFrameProvider(root=root, camera_key=camera_key, video_backend=video_backend)
+    except Exception:
+        return null_provider()
+    if provider.camera_key is None:
+        return null_provider()
+    return provider
+
+
+def _decode_frames_ffmpeg(video_path: Path, timestamps: list[float]) -> list[Any]:
+    """Decode the frames nearest to ``timestamps`` via the ffmpeg CLI.
+
+    Runs one ``ffmpeg`` process per timestamp, seeking with ``-ss`` and
+    piping a single PNG to stdout. Unlike the in-process decoders this
+    survives a hostile container: a full ffmpeg build decodes AV1 (the codec
+    modern LeRobot datasets use) where torchcodec raises and PyAV can
+    SIGSEGV, and a crash stays isolated to the child process — a non-zero
+    exit is a catchable error, not a segfault of the whole job. Returns one
+    ``(C, H, W)`` uint8 tensor per timestamp.
+    """
+    import io  # noqa: PLC0415
+    import subprocess  # noqa: PLC0415
+
+    import numpy as np  # noqa: PLC0415
+
+    frames: list[Any] = []
+    for ts in timestamps:
+        # ffmpeg invoked by name via PATH lookup; fully-controlled arg list, no shell.
+        proc = subprocess.run(  # nosec B607
+            [
+                "ffmpeg",
+                "-nostdin",
+                "-loglevel",
+                "error",
+                "-ss",
+                f"{max(ts, 0.0):.3f}",
+                "-i",
+                str(video_path),
+                "-frames:v",
+                "1",
+                "-f",
+                "image2pipe",
+                "-vcodec",
+                "png",
+                "pipe:1",
+            ],
+            capture_output=True,
+            check=True,
+            timeout=120,
+        )
+        if not proc.stdout:
+            raise RuntimeError(f"ffmpeg returned no frame for t={ts:.3f}s of {video_path}")
+        img = PIL.Image.open(io.BytesIO(proc.stdout)).convert("RGB")
+        frames.append(torch.from_numpy(np.asarray(img).copy()).permute(2, 0, 1).contiguous())
+    return frames
+
+
+def _decode_frames_av(video_path: Path, timestamps: list[float]) -> list[Any]:
+    """Decode the frames nearest to ``timestamps`` using PyAV directly.
+
+    lerobot's ``decode_video_frames(backend="pyav")`` routes through
+    ``torchvision.io.VideoReader``, removed in torchvision 0.23+. This helper
+    talks to the ``av`` package directly. Note PyAV can SIGSEGV on AV1
+    streams in some builds — prefer ``_decode_frames_ffmpeg`` as the default
+    fallback; this stays available behind ``video_backend="pyav"``. Returns
+    one ``(C, H, W)`` uint8 tensor per timestamp.
+    """
+    import av  # noqa: PLC0415
+
+    first_ts = min(timestamps)
+    last_ts = max(timestamps)
+    loaded_frames: list[torch.Tensor] = []
+    loaded_ts: list[float] = []
+    with av.open(str(video_path)) as container:
+        stream = container.streams.video[0]
+        # Seek to the keyframe at or before the first requested timestamp.
+        offset = max(int(first_ts / stream.time_base), 0) if stream.time_base else 0
+        container.seek(offset, stream=stream, backward=True, any_frame=False)
+        for idx, frame in enumerate(container.decode(stream)):
+            ts = frame.time
+            if ts is None:
+                ts = float(frame.pts * stream.time_base) if frame.pts is not None else float(idx)
+            loaded_ts.append(ts)
+            loaded_frames.append(
+                torch.from_numpy(frame.to_ndarray(format="rgb24")).permute(2, 0, 1).contiguous()
+            )
+            if ts >= last_ts:
+                break
+    if not loaded_frames:
+        raise RuntimeError(f"PyAV decoded no frames from {video_path}")
+    ts_tensor = torch.tensor(loaded_ts)
+    return [loaded_frames[int(torch.argmin((ts_tensor - q).abs()))] for q in timestamps]
+
+
+def _frame_to_pil(frame: Any) -> Any:
+    """Materialise a decoded frame as a ``PIL.Image`` for the VLM message.
+
+    Frames flow through the provider as ``torch.Tensor`` (``C, H, W`` uint8,
+    straight from :func:`decode_video_frames`); PIL is only created here, at
+    the VLM-message boundary, because the chat backends expect PIL images /
+    data URLs. Non-tensor inputs (e.g. test stubs) pass through untouched.
+    """
+    if not isinstance(frame, torch.Tensor):
+        return frame
+    array = frame.detach().cpu()
+    if array.ndim == 3 and array.shape[0] in (1, 3):
+        array = array.permute(1, 2, 0)  # (C, H, W) -> (H, W, C)
+    if array.shape[-1] == 1:
+        array = array.squeeze(-1)
+    return PIL.Image.fromarray(array.to(torch.uint8).numpy())
+
+
+def to_image_blocks(frames: list[Any]) -> list[dict[str, Any]]:
+    """Convert decoded frames to Qwen-VL-compatible image content blocks."""
+    return [{"type": "image", "image": _frame_to_pil(frame)} for frame in frames]
+
+
+def to_video_block(frames: list[Any]) -> list[dict[str, Any]]:
+    """Wrap a list of decoded frames as one Qwen-VL video block.
+
+    Returns ``[]`` when the list is empty, so the caller can splat the result
+    into a content array without a separate emptiness check.
+    """
+    if not frames:
+        return []
+    return [{"type": "video", "video": [_frame_to_pil(frame) for frame in frames]}]
+
+
+def to_video_url_block(url: str | None, fps: float = 2.0) -> list[dict[str, Any]]:
+    """Wrap a video file URL as one ``video_url`` block.
+
+    Used by the ``openai`` backend (transformers serve / vllm serve /
+    ktransformers serve), where the server handles frame sampling.
+    Returns ``[]`` when ``url`` is ``None`` so the caller can splat.
+    """
+    if not url:
+        return []
+    return [{"type": "video_url", "video_url": {"url": url}, "fps": fps}]

src/lerobot/annotations/steerable_pipeline/modules/__init__.py

@@ -0,0 +1,25 @@
+#!/usr/bin/env python
+
+# Copyright 2026 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from .general_vqa import GeneralVqaModule
+from .interjections_and_speech import InterjectionsAndSpeechModule
+from .plan_subtasks_memory import PlanSubtasksMemoryModule
+
+__all__ = [
+    "GeneralVqaModule",
+    "InterjectionsAndSpeechModule",
+    "PlanSubtasksMemoryModule",
+]

src/lerobot/annotations/steerable_pipeline/modules/general_vqa.py

@@ -0,0 +1,248 @@
+#!/usr/bin/env python
+
+# Copyright 2026 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""``vqa`` module: general VQA at a timed cadence.
+
+Every ``1/hz`` seconds an emission tick fires; each tick anchors ``K``
+consecutive frames, and every anchored frame gets its own VQA pair. Each
+pair is grounded on that single anchor frame — there is no per-pair frame
+window. For datasets with multiple cameras, every anchored frame produces
+one ``(vqa, user)`` + ``(vqa, assistant)`` pair *per camera*: each pair is
+generated against that camera's frame and stamped with the matching
+``camera`` field on the emitted rows. The resolver disambiguates via
+``camera=...``; recipes that consume VQA do so through one sub-recipe
+per camera (see ``recipes/pi05_hirobot.yaml``).
+
+Within a single (frame, camera) we still emit at most one ``(vqa, user)``
+and one ``(vqa, assistant)`` row, so the resolver contract stays scalar.
+
+Question types covered (per the plan's ``vqa`` table): bbox, keypoint,
+count, attribute, spatial. The assistant's ``content`` is a JSON string
+whose schema depends on the question type. Malformed JSON triggers one
+retry inside :meth:`VlmClient.generate_json`.
+"""
+
+from __future__ import annotations
+
+import json
+import logging
+import random
+from collections.abc import Sequence
+from dataclasses import dataclass, field
+from typing import Any
+
+from ..config import VqaConfig
+from ..frames import FrameProvider, null_provider, to_image_blocks
+from ..prompts import load as load_prompt
+from ..reader import EpisodeRecord
+from ..staging import EpisodeStaging
+from ..validator import classify_vqa_answer
+from ..vlm_client import VlmClient
+
+
+def _emission_anchor_indices(frame_timestamps: Sequence[float], hz: float, k: int) -> list[int]:
+    """Return the relative frame indices to anchor VQA emissions to.
+
+    For each emission tick (every ``1/hz`` seconds), we anchor ``k``
+    consecutive frames starting at the tick. Ticks fall on the nearest
+    available source frame timestamp.
+    """
+    if hz <= 0 or k <= 0 or not frame_timestamps:
+        return []
+    t0 = frame_timestamps[0]
+    t_last = frame_timestamps[-1]
+    period = 1.0 / hz
+    indices: list[int] = []
+    t = t0
+    while t <= t_last + 1e-9:
+        # find the index of the nearest frame to t
+        nearest_i = min(range(len(frame_timestamps)), key=lambda i: abs(frame_timestamps[i] - t))
+        for offset in range(k):
+            j = nearest_i + offset
+            if j >= len(frame_timestamps):
+                break
+            if not indices or indices[-1] != j:
+                indices.append(j)
+        t += period
+    # dedupe while preserving order
+    seen: set[int] = set()
+    deduped: list[int] = []
+    for i in indices:
+        if i in seen:
+            continue
+        seen.add(i)
+        deduped.append(i)
+    return deduped
+
+
+@dataclass
+class GeneralVqaModule:
+    """Emit grounded VQA pairs at a timed cadence."""
+
+    vlm: VlmClient
+    config: VqaConfig
+    seed: int = 1729
+    frame_provider: FrameProvider = field(default_factory=null_provider)
+    _warned_no_camera: bool = field(default=False, init=False, repr=False)
+
+    @property
+    def enabled(self) -> bool:
+        return self.config.enabled
+
+    def run_episode(self, record: EpisodeRecord, staging: EpisodeStaging) -> None:
+        if not record.frame_timestamps:
+            staging.write("vqa", [])
+            return
+        rng = random.Random(f"{self.seed}:{record.episode_index}:vqa")
+        anchor_idx = _emission_anchor_indices(
+            record.frame_timestamps, self.config.vqa_emission_hz, self.config.K
+        )
+        cameras = self._target_cameras()
+        if not cameras:
+            # No camera available — emit nothing rather than producing
+            # untagged rows that would fail validation. Surface a loud one-
+            # time warning so this is never silently a no-op.
+            if not self._warned_no_camera:
+                logging.getLogger(__name__).warning(
+                    "vqa module found no cameras on the frame provider — "
+                    "every episode will emit zero VQA rows. Check that the "
+                    "dataset declares observation.images.* features in "
+                    "meta/info.json; passing --vlm.camera_key=<key> at the "
+                    "CLI now also seeds the cameras list as a fallback."
+                )
+                self._warned_no_camera = True
+            staging.write("vqa", [])
+            return
+
+        # Build all messages first (one per (frame, camera)), then issue them
+        # as a single batched generate_json call so the client can fan them
+        # out concurrently.
+        per_call: list[tuple[float, str, str, list[dict[str, Any]]]] = []
+        for idx in anchor_idx:
+            ts = float(record.frame_timestamps[idx])
+            qtype = rng.choice(self.config.question_types)
+            for camera in cameras:
+                messages = self._build_messages(record, qtype, ts, camera)
+                # Skip cameras that decoded to zero frames at this ts: no point
+                # asking the VLM to ground a bbox without an image.
+                if not _has_image_block(messages):
+                    continue
+                per_call.append((ts, camera, qtype, messages))
+
+        if not per_call:
+            staging.write("vqa", [])
+            return
+
+        results = self.vlm.generate_json([m for _, _, _, m in per_call])
+
+        rows: list[dict[str, Any]] = []
+        for (ts, camera, _qtype, _messages), result in zip(per_call, results, strict=True):
+            qa = self._postprocess(result)
+            if qa is None:
+                continue
+            question, answer = qa
+            rows.append(
+                {
+                    "role": "user",
+                    "content": question,
+                    "style": "vqa",
+                    "timestamp": ts,
+                    "camera": camera,
+                    "tool_calls": None,
+                }
+            )
+            rows.append(
+                {
+                    "role": "assistant",
+                    "content": json.dumps(answer, sort_keys=True),
+                    "style": "vqa",
+                    "timestamp": ts,
+                    "camera": camera,
+                    "tool_calls": None,
+                }
+            )
+        staging.write("vqa", rows)
+
+    def _target_cameras(self) -> list[str]:
+        """Return the cameras the ``vqa`` module should iterate per anchored frame.
+
+        Defaults to every camera the provider exposes. Datasets with no
+        cameras (or test/null providers) yield an empty list, which makes
+        ``run_episode`` a no-op.
+
+        When ``config.restrict_to_default_camera`` is set, VQA grounds on
+        only the provider's default camera (the single ``--vlm.camera_key``
+        stream), matching the plan / interjection modules so the whole
+        pipeline focuses on one view.
+        """
+        all_cameras = list(getattr(self.frame_provider, "camera_keys", []) or [])
+        if getattr(self.config, "restrict_to_default_camera", False):
+            default = getattr(self.frame_provider, "camera_key", None)
+            if default and default in all_cameras:
+                return [default]
+            # ``restrict_to_default_camera`` is set but the configured default
+            # isn't one the provider exposes. Returning it anyway would make
+            # ``_decode`` raise a KeyError deep in frame extraction, so warn and
+            # fall through to every available camera instead.
+            if default:
+                logging.getLogger(__name__).warning(
+                    "restrict_to_default_camera is set but camera_key=%r is not in the "
+                    "provider's cameras %s; grounding VQA on all available cameras instead.",
+                    default,
+                    all_cameras,
+                )
+        return all_cameras
+
+    def _build_messages(
+        self,
+        record: EpisodeRecord,
+        question_type: str,
+        frame_timestamp: float,
+        camera_key: str,
+    ) -> list[dict[str, Any]]:
+        prompt = load_prompt("vqa").format(
+            episode_task=record.episode_task,
+            question_type=question_type,
+        )
+        images = self.frame_provider.frames_at(record, [frame_timestamp], camera_key=camera_key)
+        content = [*to_image_blocks(images), {"type": "text", "text": prompt}]
+        return [{"role": "user", "content": content}]
+
+    def _postprocess(self, result: Any) -> tuple[str, dict[str, Any]] | None:
+        if not isinstance(result, dict):
+            return None
+        question = result.get("question")
+        answer = result.get("answer")
+        if not isinstance(question, str) or not question.strip():
+            return None
+        if not isinstance(answer, dict):
+            return None
+        # The validator will enforce shape; here we just sanity-check that the
+        # answer matches *some* known shape so we can drop garbage early.
+        if classify_vqa_answer(answer) is None:
+            return None
+        return question.strip(), answer
+
+
+def _has_image_block(messages: list[dict[str, Any]]) -> bool:
+    """Return True if any user content block is a populated image block."""
+    for msg in messages:
+        content = msg.get("content")
+        if not isinstance(content, list):
+            continue
+        for block in content:
+            if isinstance(block, dict) and block.get("type") == "image":
+                return True
+    return False

src/lerobot/annotations/steerable_pipeline/modules/interjections_and_speech.py

@@ -0,0 +1,211 @@
+#!/usr/bin/env python
+
+# Copyright 2026 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""``interjections`` module: interjections + paired speech (EVENT styles + speech atoms).
+
+Two sub-passes:
+
+1. At ``t=0``, emit ONLY a speech tool-call atom (acknowledgement of the
+   canonical task). No interjection row — the canonical task is already the
+   user utterance from ``meta/tasks.parquet``.
+
+2. For mid-episode interruptions, emit a co-timestamped pair:
+       {role:user, style:interjection, content:<text>}
+       speech atom (role:assistant, style:None, tool_calls=[say(...)])
+   Both rows go in ``language_events`` at the same timestamp.
+
+The ``plan`` module's :meth:`run_plan_updates` reuses this module's
+interjection timestamps to refresh the ``plan`` row at the same instant.
+"""
+
+from __future__ import annotations
+
+import random
+from collections.abc import Sequence
+from dataclasses import dataclass, field
+from typing import Any
+
+from ..config import InterjectionsConfig
+from ..frames import FrameProvider, null_provider, to_image_blocks
+from ..prompts import load as load_prompt
+from ..reader import EpisodeRecord, reconstruct_subtask_spans, snap_to_frame
+from ..staging import EpisodeStaging
+from ..vlm_client import VlmClient
+from ..writer import speech_atom
+
+
+@dataclass
+class InterjectionsAndSpeechModule:
+    """Generate task-start speech and mid-episode interjection/speech pairs."""
+
+    vlm: VlmClient
+    config: InterjectionsConfig
+    seed: int = 1729
+    frame_provider: FrameProvider = field(default_factory=null_provider)
+
+    @property
+    def enabled(self) -> bool:
+        return self.config.enabled
+
+    def run_episode(self, record: EpisodeRecord, staging: EpisodeStaging) -> None:
+        rows: list[dict[str, Any]] = []
+        if record.frame_timestamps:
+            t0 = float(record.frame_timestamps[0])
+            initial = self._initial_speech(record)
+            if initial:
+                rows.append(speech_atom(t0, initial))
+        # Pull the ``plan`` module's subtask spans for this episode so the
+        # interjection prompt can ground itself in the actual current
+        # subtask at each chosen timestamp. The ``plan`` module ran first.
+        episode_end_t = float(record.frame_timestamps[-1]) if record.frame_timestamps else None
+        subtask_spans = reconstruct_subtask_spans(staging.read("plan"), episode_end_t=episode_end_t)
+        rows.extend(self._mid_episode_interjections(record, subtask_spans))
+        staging.write("interjections", rows)
+
+    @staticmethod
+    def _subtask_at(spans: Sequence[dict[str, Any]], t: float) -> str | None:
+        current: str | None = None
+        for span in spans:
+            if float(span["start"]) <= t:
+                current = span.get("text")
+            else:
+                break
+        return current
+
+    def _initial_speech(self, record: EpisodeRecord) -> str | None:
+        prompt = load_prompt("interjections_initial_speech").format(
+            episode_task=record.episode_task,
+        )
+        messages = [{"role": "user", "content": [{"type": "text", "text": prompt}]}]
+        result = self.vlm.generate_json([messages])[0]
+        if isinstance(result, dict) and isinstance(result.get("text"), str):
+            text = result["text"].strip()
+            if text:
+                return text
+        return None
+
+    def _mid_episode_interjections(
+        self,
+        record: EpisodeRecord,
+        subtask_spans: Sequence[dict[str, Any]],
+    ) -> list[dict[str, Any]]:
+        """Generate interjections aligned with the actual demo trajectory.
+
+        Teleop data is frozen — the robot already executed every step in
+        the video. A *counterfactual* interjection like "actually skip
+        the wipe" contradicts what then happens in the video, which is
+        what qwen36moe-10/11 surfaced as low-quality interjections.
+
+        Instead, anchor every interjection at a subtask boundary and
+        write it as a natural user request for the *upcoming* subtask.
+        The robot's visible next behavior IS the interjection's effect,
+        so the training signal stays consistent: interjection text →
+        plan refresh → action stream all line up.
+        """
+        if self.config.max_interjections_per_episode <= 0:
+            return []
+        if len(subtask_spans) < 2:
+            # Need at least one transition (subtask 0 → subtask 1).
+            return []
+        # Deterministic per-episode RNG so reruns are stable across SLURM jobs.
+        rng = random.Random(f"{self.seed}:{record.episode_index}:interjection")
+
+        # Boundaries: the start time of every subtask except the first
+        # (which is just t0 and is covered by the initial-task speech atom).
+        boundaries: list[tuple[float, str, str]] = []
+        for i in range(1, len(subtask_spans)):
+            ts = float(subtask_spans[i]["start"])
+            if ts < self.config.interjection_min_t:
+                continue
+            prev_text = (subtask_spans[i - 1].get("text") or "").strip()
+            next_text = (subtask_spans[i].get("text") or "").strip()
+            if not next_text:
+                continue
+            boundaries.append((ts, prev_text, next_text))
+        if not boundaries:
+            return []
+
+        n = min(self.config.max_interjections_per_episode, len(boundaries))
+        chosen = sorted(rng.sample(boundaries, n), key=lambda b: b[0])
+
+        out: list[dict[str, Any]] = []
+        for t, prev_subtask, next_subtask in chosen:
+            t_snap = snap_to_frame(t, record.frame_timestamps)
+            # Window straddles the boundary so the VLM sees the end of the
+            # previous subtask and the start of the next one — same
+            # conditioning the policy will see at training time.
+            window_ts = self._window_timestamps(t_snap, record.frame_timestamps)
+            prompt = load_prompt("interjections_interjection").format(
+                episode_task=record.episode_task,
+                prev_subtask=prev_subtask or "(starting from initial state)",
+                next_subtask=next_subtask,
+                timestamp=t_snap,
+                window_seconds=self.config.interjection_window_seconds,
+            )
+            images = self.frame_provider.frames_at(record, window_ts)
+            content = [*to_image_blocks(images), {"type": "text", "text": prompt}]
+            messages = [{"role": "user", "content": content}]
+            result = self.vlm.generate_json([messages])[0]
+            if not isinstance(result, dict):
+                continue
+            interjection_text = result.get("interjection")
+            speech_text = result.get("speech")
+            if not isinstance(interjection_text, str) or not interjection_text.strip():
+                continue
+            if not isinstance(speech_text, str) or not speech_text.strip():
+                continue
+            out.append(
+                {
+                    "role": "user",
+                    "content": interjection_text.strip(),
+                    "style": "interjection",
+                    "timestamp": t_snap,
+                    "tool_calls": None,
+                }
+            )
+            out.append(speech_atom(t_snap, speech_text.strip()))
+        return out
+
+    def _window_timestamps(self, t_anchor: float, frame_timestamps: Sequence[float]) -> list[float]:
+        """Return a small set of frame timestamps centered on ``t_anchor``.
+
+        The window straddles the subtask boundary the interjection sits
+        on: roughly half the frames cover the end of the previous
+        subtask, half cover the start of the next one. The VLM therefore
+        sees BOTH what just finished AND what's about to start, which is
+        the conditioning we need to write a natural "now please do X"
+        request that matches the visible upcoming behavior.
+        """
+        if not frame_timestamps:
+            return [t_anchor]
+        n = max(1, int(self.config.interjection_window_frames))
+        if n == 1:
+            return [t_anchor]
+        window = float(self.config.interjection_window_seconds)
+        step = window / max(1, n - 1)
+        # Center the window on the anchor so half lands before, half after.
+        start_offset = -window / 2.0
+        targets = [t_anchor + start_offset + step * i for i in range(n)]
+        first_ts = float(frame_timestamps[0])
+        last_ts = float(frame_timestamps[-1])
+        snapped: list[float] = []
+        seen: set[float] = set()
+        for tgt in targets:
+            clamped = min(last_ts, max(first_ts, tgt))
+            t = snap_to_frame(clamped, frame_timestamps)
+            if t not in seen:
+                seen.add(t)
+                snapped.append(t)
+        return snapped or [t_anchor]

src/lerobot/annotations/steerable_pipeline/modules/plan_subtasks_memory.py

@@ -0,0 +1,712 @@
+#!/usr/bin/env python
+
+# Copyright 2026 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""``plan`` module: subtask decomposition + plan + memory (PERSISTENT styles)."""
+
+from __future__ import annotations
+
+import logging
+from collections.abc import Sequence
+from dataclasses import dataclass, field
+from pathlib import Path
+from typing import Any
+
+from ..config import PlanConfig
+from ..frames import (
+    FrameProvider,
+    VideoFrameProvider,
+    null_provider,
+    to_video_block,
+    to_video_url_block,
+)
+from ..prompts import load as load_prompt
+from ..reader import EpisodeRecord, reconstruct_subtask_spans, snap_to_frame
+from ..staging import EpisodeStaging
+from ..vlm_client import VlmClient
+
+logger = logging.getLogger(__name__)
+
+
+@dataclass
+class PlanSubtasksMemoryModule:
+    """Generate subtask spans, plan, and memory rows.
+
+    All output is persistent (lives in ``language_persistent``):
+
+    - ``subtask`` rows: one per span, stamped at the span's *start* timestamp
+      (snapped to an exact frame).
+    - ``plan`` rows: emitted at ``t=0``; refreshed at every interjection
+      timestamp via :meth:`run_plan_updates` (called by the executor after
+      the ``interjections`` module completes).
+    - ``memory`` rows: emitted at each subtask boundary (= subtask start
+      timestamp from the second subtask onward).
+    """
+
+    vlm: VlmClient
+    config: PlanConfig
+    frame_provider: FrameProvider = field(default_factory=null_provider)
+
+    @property
+    def enabled(self) -> bool:
+        return self.config.enabled
+
+    def run_episode(self, record: EpisodeRecord, staging: EpisodeStaging) -> None:
+        rows: list[dict[str, Any]] = []
+        # Task driving every plan-module prompt: canonical episode_task, or a
+        # video-derived one when it's empty/placeholder (see derive_task_*).
+        effective_task = self._resolve_effective_task(record)
+        # task_aug rows at t=0: phrasings the renderer rotates ${task} through.
+        # Either the structured 5-axis taxonomy (task_aug_axes.enabled) or
+        # free-form n_task_rephrasings; the effective task is always emitted
+        # first so the rotation covers the source-of-truth phrasing.
+        t0 = float(record.frame_timestamps[0]) if record.frame_timestamps else 0.0
+        variants: list[str] | None = None
+        if self.config.task_aug_axes.enabled and effective_task:
+            variants = self._generate_task_aug_by_axes(effective_task, self.config.task_aug_axes)
+        elif self.config.n_task_rephrasings > 0 and effective_task:
+            variants = self._generate_task_rephrasings(effective_task, n=self.config.n_task_rephrasings)
+        if variants is not None:
+            rows.extend(self._task_aug_rows([effective_task, *variants], t0))
+
+        subtask_spans = self._generate_subtasks(record, task=effective_task)
+
+        # subtask rows
+        for span in subtask_spans:
+            rows.append(
+                {
+                    "role": "assistant",
+                    "content": span["text"],
+                    "style": "subtask",
+                    "timestamp": snap_to_frame(span["start"], record.frame_timestamps),
+                    "tool_calls": None,
+                }
+            )
+        # Plan rows at every subtask boundary (incl. t=0). The plan is a
+        # numbered list of still-todo subtasks, so re-emitting at each
+        # boundary makes it shrink as work progresses — ${plan} at frame t is
+        # exactly what's left to do.
+        if self.config.emit_plan:
+            for span in subtask_spans:
+                boundary_t = snap_to_frame(span["start"], record.frame_timestamps)
+                plan_text = self._generate_plan(
+                    record, subtask_spans, refresh_t=boundary_t, task=effective_task
+                )
+                if plan_text is not None:
+                    rows.append(
+                        {
+                            "role": "assistant",
+                            "content": plan_text,
+                            "style": "plan",
+                            "timestamp": float(boundary_t),
+                            "tool_calls": None,
+                        }
+                    )
+        # memory rows at every subtask boundary except the very first start
+        prior_memory = ""
+        for i, span in enumerate(subtask_spans[1:], start=1):
+            completed = subtask_spans[i - 1]["text"]
+            remaining = [s["text"] for s in subtask_spans[i:]]
+            mem_text = self._generate_memory(record, prior_memory, completed, remaining, task=effective_task)
+            if mem_text:
+                ts = snap_to_frame(span["start"], record.frame_timestamps)
+                rows.append(
+                    {
+                        "role": "assistant",
+                        "content": mem_text,
+                        "style": "memory",
+                        "timestamp": ts,
+                        "tool_calls": None,
+                    }
+                )
+                prior_memory = mem_text
+        staging.write("plan", rows)
+
+    # ------------------------------------------------------------------
+    # Task derivation + rephrasings
+    # ------------------------------------------------------------------
+
+    _PLACEHOLDER_TASKS: frozenset[str] = frozenset(
+        {
+            "debug",
+            "test",
+            "tbd",
+            "todo",
+            "n/a",
+            "na",
+            "untitled",
+            "unnamed",
+            "default",
+            "placeholder",
+        }
+    )
+
+    def _resolve_effective_task(self, record: EpisodeRecord) -> str:
+        """Decide which task string drives the ``plan`` module for this episode.
+
+        Returns the user-supplied ``record.episode_task`` unless
+        ``derive_task_from_video`` says otherwise (see config docstring).
+        Falls back gracefully to the canonical task if video derivation
+        fails.
+        """
+        canonical = (record.episode_task or "").strip()
+        mode = (self.config.derive_task_from_video or "off").strip().lower()
+        if mode == "always":
+            derived = self._derive_task_from_video(record)
+            return derived or canonical
+        if mode == "if_short" and self._task_seems_bad(canonical):
+            derived = self._derive_task_from_video(record)
+            if derived:
+                return derived
+        return canonical
+
+    def _task_seems_bad(self, task: str) -> bool:
+        if not task:
+            return True
+        if len(task.split()) < int(self.config.derive_task_min_words):
+            return True
+        return task.lower() in self._PLACEHOLDER_TASKS
+
+    @staticmethod
+    def _task_aug_rows(phrasings: Sequence[str], t0: float) -> list[dict[str, Any]]:
+        """Build deduplicated ``task_aug`` rows (role=user) at ``t0``."""
+        seen: set[str] = set()
+        rows: list[dict[str, Any]] = []
+        for phrasing in phrasings:
+            key = phrasing.strip()
+            if not key or key in seen:
+                continue
+            seen.add(key)
+            rows.append(
+                {"role": "user", "content": key, "style": "task_aug", "timestamp": t0, "tool_calls": None}
+            )
+        return rows
+
+    # ------------------------------------------------------------------
+    # VLM call helpers — every plan-module prompt follows the same shape:
+    # build messages → single VLM call → pull a named field.
+    # ------------------------------------------------------------------
+
+    def _vlm_field(self, messages: list[dict[str, Any]], field: str) -> Any:
+        """Run a single VLM call and return ``result[field]`` or ``None``.
+
+        Centralizes the ``vlm.generate_json([m])[0]`` + ``isinstance(dict)``
+        dance every prompt-call site needs.
+        """
+        result = self.vlm.generate_json([messages])[0]
+        if isinstance(result, dict):
+            return result.get(field)
+        return None
+
+    @staticmethod
+    def _text_message(text: str) -> list[dict[str, Any]]:
+        """One-shot text-only user message wrapped for ``generate_json``."""
+        return [{"role": "user", "content": [{"type": "text", "text": text}]}]
+
+    def _video_message(
+        self,
+        record: EpisodeRecord,
+        prompt: str,
+        window: tuple[float, float] | None = None,
+    ) -> list[dict[str, Any]]:
+        """User message combining the (optionally windowed) video block with ``prompt``."""
+        content = [*self._episode_video_block(record, window=window), {"type": "text", "text": prompt}]
+        return [{"role": "user", "content": content}]
+
+    def _derive_task_from_video(self, record: EpisodeRecord) -> str | None:
+        """Ask the VLM "what is this video about" with no task hint at all."""
+        text = self._vlm_field(self._video_message(record, load_prompt("plan_video_task")), "task")
+        return text.strip() if isinstance(text, str) and text.strip() else None
+
+    def _generate_task_rephrasings(self, base_task: str, *, n: int) -> list[str]:
+        """Generate ``n`` text-only paraphrases of ``base_task``."""
+        if n <= 0 or not base_task:
+            return []
+        prompt = load_prompt("plan_task_rephrasings").format(base_task=base_task, n=n)
+        raw = self._vlm_field(self._text_message(prompt), "rephrasings")
+        if not isinstance(raw, list):
+            return []
+        out = [item.strip().strip('"').strip("'") for item in raw if isinstance(item, str)]
+        return [s for s in out if s][:n]
+
+    # ------------------------------------------------------------------
+    # Structured 5-axis task augmentation (EgoMimic-style taxonomy)
+    # ------------------------------------------------------------------
+
+    def _generate_task_aug_by_axes(self, base_task: str, axes_cfg: Any) -> list[str]:
+        """One VLM call → variants along the 5-axis taxonomy.
+
+        Variants from all axes are flattened into a single list (the
+        downstream pipeline doesn't need to know about the per-axis
+        bucketing — every variant becomes a ``task_aug`` row). Order
+        is preserved for reproducibility: synonym_paraphrase first,
+        then omit_arm, then omit_orientation, then omit_grasp_method,
+        then combined_omissions.
+        """
+        if not base_task:
+            return []
+        prompt = load_prompt("plan_task_aug_axes").format(
+            base_task=base_task,
+            n_synonym=axes_cfg.synonym_paraphrase,
+            n_omit_arm=axes_cfg.omit_arm,
+            n_omit_orientation=axes_cfg.omit_orientation,
+            n_omit_grasp_method=axes_cfg.omit_grasp_method,
+            n_combined=axes_cfg.combined_omissions,
+        )
+        result = self.vlm.generate_json([self._text_message(prompt)])[0]
+        if not isinstance(result, dict):
+            return []
+        ordered_axes = (
+            "synonym_paraphrase",
+            "omit_arm",
+            "omit_orientation",
+            "omit_grasp_method",
+            "combined_omissions",
+        )
+        flat: list[str] = []
+        seen: set[str] = set()
+        for axis in ordered_axes:
+            entries = result.get(axis)
+            if not isinstance(entries, list):
+                continue
+            for item in entries:
+                if not isinstance(item, str):
+                    continue
+                key = item.strip().strip('"').strip("'")
+                if not key or key in seen:
+                    continue
+                seen.add(key)
+                flat.append(key)
+        return flat
+
+    def _episode_video_block(
+        self, record: EpisodeRecord, window: tuple[float, float] | None = None
+    ) -> list[dict[str, Any]]:
+        """Video block for the segmentation / describe prompts.
+
+        Always returns a block that actually carries the video. When
+        ``use_video_url`` is set we try the server-side ``video_url``
+        path first, but if clip extraction fails we FALL BACK to
+        decoding + embedding frames rather than returning an empty
+        block — an empty block would leave the VLM with no visual
+        grounding at all and it would hallucinate subtasks purely from
+        the task text.
+
+        When ``window=(w0, w1)`` is given (windowed subtask generation,
+        ``subtask_window_seconds > 0``), embed frames sampled at the FIXED
+        ``frames_per_second`` rate within ``[w0, w1]`` — constant temporal
+        density regardless of episode length, so long episodes are split
+        into windows rather than subsampled to a sparse 32-frame whole-
+        episode view. The ``video_url`` path is skipped for windows (it is
+        a whole-episode clip). ``max_video_frames`` still caps each window
+        as a context-budget safety net.
+        """
+        if not record.frame_timestamps:
+            return []
+        if window is not None:
+            w0, w1 = float(window[0]), float(window[1])
+            dur = max(0.0, w1 - w0)
+            n = max(1, int(round(dur * self.config.frames_per_second)) + 1)
+            n = min(n, self.config.max_video_frames)
+            if n <= 1 or dur <= 0.0:
+                timestamps = [0.5 * (w0 + w1)]
+            else:
+                step = dur / (n - 1)
+                timestamps = [w0 + i * step for i in range(n)]
+            return to_video_block(self.frame_provider.frames_at(record, timestamps))
+        if self.config.use_video_url and isinstance(self.frame_provider, VideoFrameProvider):
+            cache_dir = Path(self.frame_provider.root) / ".annotate_staging" / ".video_clips"
+            clip = self.frame_provider.episode_clip_path(record, cache_dir)
+            if clip is not None:
+                return to_video_url_block(f"file://{clip}", fps=self.config.use_video_url_fps)
+            logger.warning(
+                "episode %d: video_url clip extraction failed — falling back to "
+                "embedded frames so the VLM still sees the demonstration",
+                record.episode_index,
+            )
+        episode_duration = record.frame_timestamps[-1] - record.frame_timestamps[0]
+        target_count = max(1, int(round(episode_duration * self.config.frames_per_second)))
+        target_count = min(target_count, self.config.max_video_frames)
+        video_frames = self.frame_provider.video_for_episode(record, target_count)
+        return to_video_block(video_frames)
+
+    def run_plan_updates(
+        self,
+        record: EpisodeRecord,
+        staging: EpisodeStaging,
+        interjection_times: Sequence[float],
+        interjection_texts: Sequence[str] | None = None,
+    ) -> None:
+        """Append additional ``plan`` rows at every interjection timestamp.
+
+        Plans refresh ONLY on user interjections (event-driven). The
+        interjection text is forwarded into the prompt so the refreshed plan
+        reflects the user's correction.
+        """
+        if not self.config.emit_plan:
+            return
+        existing = staging.read("plan")
+        # Pass the last frame timestamp so the final span is closed (else its
+        # end == start, zero duration, and a refresh inside it is missed).
+        episode_end_t = float(record.frame_timestamps[-1]) if record.frame_timestamps else None
+        spans = reconstruct_subtask_spans(existing, episode_end_t=episode_end_t)
+        already_planned: set[float] = {float(r["timestamp"]) for r in existing if r.get("style") == "plan"}
+        new_rows = list(existing)
+
+        texts: list[str | None] = (
+            [None] * len(interjection_times)
+            if interjection_texts is None
+            else [str(t) if t else None for t in interjection_texts]
+        )
+        for raw_t, inter_text in zip(interjection_times, texts, strict=True):
+            t = snap_to_frame(raw_t, record.frame_timestamps)
+            if t in already_planned:
+                continue
+            already_planned.add(t)
+            plan_text = self._generate_plan(record, spans, refresh_t=t, interjection=inter_text)
+            if plan_text is not None:
+                new_rows.append(
+                    {
+                        "role": "assistant",
+                        "content": plan_text,
+                        "style": "plan",
+                        "timestamp": t,
+                        "tool_calls": None,
+                    }
+                )
+        staging.write("plan", new_rows)
+
+    def _generate_subtasks(self, record: EpisodeRecord, *, task: str | None = None) -> list[dict[str, Any]]:
+        """Generate subtask spans, optionally via a multi-call quality chain.
+
+        Single call (default): watch video → emit subtask JSON.
+
+        Multi-call (opt-in, higher quality, more VLM calls):
+          1. ``subtask_describe_first`` — a grounding pass that narrates
+             ONLY what is visible (no JSON commitment to subtasks yet);
+             its description is injected into the segmentation prompt so
+             the model segments its own grounded observations instead of
+             pattern-matching the task text.
+          2. segmentation — emit subtask JSON (as before).
+        """
+        if record.row_count == 0 or not record.frame_timestamps:
+            return []
+        episode_duration = record.frame_timestamps[-1] - record.frame_timestamps[0]
+        effective_task = task if task is not None else record.episode_task
+
+        # ---- Windowed path (constant temporal density) ---------------
+        # If subtask_window_seconds > 0 and the episode exceeds one window,
+        # process fixed-length windows so the VLM always sees
+        # frames_per_second density; results are merged + stitched.
+        window_s = float(getattr(self.config, "subtask_window_seconds", 0.0) or 0.0)
+        if window_s > 0.0 and episode_duration > window_s:
+            return self._generate_subtasks_windowed(record, effective_task, window_s)
+
+        # ---- Pass 1 (optional): grounding description ----------------
+        observation_block = ""
+        if getattr(self.config, "subtask_describe_first", False):
+            description = self._describe_episode(record, effective_task)
+            if description:
+                observation_block = (
+                    "You watched this video and described, chronologically, "
+                    "ONLY what the robot actually does:\n"
+                    f'"""{description}"""\n\n'
+                    "Segment THAT grounded description (cross-checked against "
+                    "the video) into atomic subtasks. Do not introduce any "
+                    "action that is not in your description above.\n\n"
+                )
+
+        # ---- Pass 2: segmentation ------------------------------------
+        prompt = load_prompt("plan_subtasks").format(
+            episode_task=effective_task,
+            min_subtask_seconds=self.config.min_subtask_seconds,
+            max_steps=self.config.plan_max_steps,
+            episode_duration=f"{episode_duration:.3f}",
+            observation_block=observation_block,
+        )
+        spans = self._vlm_field(self._video_message(record, prompt), "subtasks")
+        cleaned = self._clean_spans(spans, record)
+        if not cleaned:
+            return []
+
+        # ---- Full-episode coverage stitch ----------------------------
+        # The VLM can start after t0 or leave gaps, so frames fall through
+        # with no active subtask. Always stitch into a contiguous
+        # [t0, t_last] cover.
+        cleaned = self._stitch_full_coverage(cleaned, record)
+
+        return cleaned
+
+    def _generate_subtasks_windowed(
+        self, record: EpisodeRecord, task: str, window_s: float
+    ) -> list[dict[str, Any]]:
+        """Subtask generation in fixed-length windows at constant fps.
+
+        Splits ``[t0, t_last]`` into consecutive windows of ``window_s``
+        seconds, runs the describe -> segment chain on each window's own
+        frames (sampled at ``frames_per_second``), offsets
+        each window's spans back to absolute episode time, then merges +
+        stitches into a contiguous whole-episode cover.
+        """
+        t0 = float(record.frame_timestamps[0])
+        t_last = float(record.frame_timestamps[-1])
+        all_spans: list[dict[str, Any]] = []
+        w0 = t0
+        n_windows = 0
+        while w0 < t_last - 1e-6:
+            w1 = min(w0 + window_s, t_last)
+            all_spans.extend(self._subtasks_for_window(record, task, w0, w1))
+            n_windows += 1
+            w0 = w1
+        logger.info(
+            "episode %d: windowed subtask gen over %d window(s) of %.1fs -> %d raw spans",
+            record.episode_index,
+            n_windows,
+            window_s,
+            len(all_spans),
+        )
+        # Merge across windows: clamp to the absolute episode, sort, and
+        # frame-snap to distinct starts (handles any boundary collisions).
+        cleaned = self._clean_spans(all_spans, record)
+        if not cleaned:
+            return []
+        return self._stitch_full_coverage(cleaned, record)
+
+    def _subtasks_for_window(
+        self, record: EpisodeRecord, task: str, w0: float, w1: float
+    ) -> list[dict[str, Any]]:
+        """Run describe -> segment on one ``[w0, w1]`` window.
+
+        The model works in window-RELATIVE time ``[0, L]`` (it perceives
+        the window as a clip starting at 0); spans are offset back to
+        absolute ``[w0, w1]`` before returning.
+        """
+        window = (w0, w1)
+        win_len = max(0.0, w1 - w0)
+
+        observation_block = ""
+        if getattr(self.config, "subtask_describe_first", False):
+            description = self._describe_episode(record, task, window=window)
+            if description:
+                observation_block = (
+                    "You watched this video clip and described, chronologically, "
+                    "ONLY what the robot actually does:\n"
+                    f'"""{description}"""\n\n'
+                    "Segment THAT grounded description (cross-checked against "
+                    "the clip) into atomic subtasks. Do not introduce any "
+                    "action that is not in your description above.\n\n"
+                )
+
+        prompt = load_prompt("plan_subtasks").format(
+            episode_task=task,
+            min_subtask_seconds=self.config.min_subtask_seconds,
+            max_steps=self.config.plan_max_steps,
+            episode_duration=f"{win_len:.3f}",
+            observation_block=observation_block,
+        )
+        spans = self._vlm_field(self._video_message(record, prompt, window=window), "subtasks")
+        # Window-relative clamp; no frame-snap dedupe yet (done on the
+        # merged absolute set).
+        cleaned = self._clean_spans(spans, record, bounds=(0.0, win_len), dedupe=False)
+        if not cleaned:
+            return []
+
+        # Offset window-relative spans back to absolute episode time.
+        for s in cleaned:
+            s["start"] = w0 + float(s["start"])
+            s["end"] = w0 + float(s["end"])
+        return cleaned
+
+    def _stitch_full_coverage(
+        self, spans: list[dict[str, Any]], record: EpisodeRecord
+    ) -> list[dict[str, Any]]:
+        """Make subtask spans tile the full episode with no gaps.
+
+        * The first subtask starts at the episode's first frame ``t0``
+          (any idle / approach before the first labelled action is folded
+          into it), so every early frame has an active subtask.
+        * Each subtask's ``end`` is snapped to the next subtask's
+          ``start`` (gaps between spans are closed), and the final
+          subtask's ``end`` extends to the last frame ``t_last``.
+
+        Starts are otherwise left as the (already frame-snapped, distinct)
+        values the VLM produced — only the FIRST start is pulled
+        back to ``t0``, which can't collide with a later span because it
+        was already the earliest. Purely deterministic; runs after the
+        VLM passes.
+        """
+        if not spans or not record.frame_timestamps:
+            return spans
+        t0 = float(record.frame_timestamps[0])
+        t_last = float(record.frame_timestamps[-1])
+        spans = sorted(spans, key=lambda s: float(s["start"]))
+        spans[0]["start"] = t0
+        for i in range(len(spans) - 1):
+            spans[i]["end"] = float(spans[i + 1]["start"])
+        spans[-1]["end"] = t_last
+        for s in spans:
+            if float(s["end"]) < float(s["start"]):
+                s["end"] = float(s["start"])
+        return spans
+
+    def _clean_spans(
+        self,
+        spans: Any,
+        record: EpisodeRecord,
+        bounds: tuple[float, float] | None = None,
+        dedupe: bool = True,
+    ) -> list[dict[str, Any]]:
+        """Clamp / sort / (optionally) dedupe raw VLM subtask spans into valid rows.
+
+        ``bounds`` overrides the clamp range — pass the window's
+        ``(w_lo, w_hi)`` when cleaning window-relative spans, or leave
+        ``None`` to clamp to the whole episode ``[t0, t_last]``.
+        ``dedupe`` runs the frame-snap distinct-start step; skip it for
+        window-relative spans (frame snapping is done once on the merged,
+        absolute-time set).
+        """
+        if not spans:
+            return []
+        if bounds is not None:
+            lo, hi = float(bounds[0]), float(bounds[1])
+        else:
+            lo = record.frame_timestamps[0]
+            hi = record.frame_timestamps[-1]
+        cleaned: list[dict[str, Any]] = []
+        for span in spans:
+            try:
+                start = float(span["start"])
+                end = float(span["end"])
+                text = str(span["text"]).strip()
+            except (KeyError, ValueError, TypeError):
+                continue
+            start = max(lo, min(start, hi))
+            end = max(lo, min(end, hi))
+            if end < start:
+                start, end = end, start
+            if not text:
+                continue
+            cleaned.append({"text": text, "start": start, "end": end})
+        cleaned.sort(key=lambda s: s["start"])
+        if dedupe:
+            return self._dedupe_starts_to_distinct_frames(cleaned, record)
+        return cleaned
+
+    def _describe_episode(
+        self, record: EpisodeRecord, task: str, window: tuple[float, float] | None = None
+    ) -> str:
+        """Grounding pass: free-form chronological description of the (windowed) video."""
+        prompt = load_prompt("plan_subtask_describe").format(episode_task=task)
+        text = self._vlm_field(self._video_message(record, prompt, window=window), "description")
+        return text.strip() if isinstance(text, str) and text.strip() else ""
+
+    @staticmethod
+    def _dedupe_starts_to_distinct_frames(
+        spans: list[dict[str, Any]], record: EpisodeRecord
+    ) -> list[dict[str, Any]]:
+        """Bump same-frame subtask starts onto distinct frames.
+
+        Two consecutive VLM spans whose ``start`` rounds to the same
+        source frame (after :func:`snap_to_frame`) would otherwise emit
+        two ``style=subtask`` rows at the identical persistent
+        timestamp. The training-time renderer's ``active_at(t,
+        style=subtask)`` resolver can't disambiguate that and raises
+        ``Ambiguous resolver for style='subtask'``.
+
+        Walk the (sorted-by-start) spans, snap each to its frame, and
+        if the snapped frame is already taken push the span onto the
+        next unused frame so both subtasks survive on distinct
+        timestamps. If the episode ends before a free frame is found,
+        the trailing span is dropped with a warning — better than
+        poisoning the render.
+        """
+        if not spans:
+            return spans
+        frames = record.frame_timestamps
+        if not frames:
+            return spans
+        used: set[float] = set()
+        out: list[dict[str, Any]] = []
+        for span in spans:
+            ts = snap_to_frame(span["start"], frames)
+            if ts in used:
+                next_ts = next((f for f in frames if f > ts and f not in used), None)
+                if next_ts is None:
+                    logger.warning(
+                        "episode %d: subtask %r snapped to occupied frame "
+                        "%.3f and no free later frame exists — dropping",
+                        record.episode_index,
+                        span.get("text"),
+                        ts,
+                    )
+                    continue
+                ts = next_ts
+            used.add(ts)
+            new_span = {**span, "start": ts}
+            if float(new_span.get("end", ts)) < ts:
+                new_span["end"] = ts
+            out.append(new_span)
+        return out
+
+    def _generate_plan(
+        self,
+        record: EpisodeRecord,  # noqa: ARG002  (kept for signature stability)
+        subtask_spans: Sequence[dict[str, Any]],
+        *,
+        refresh_t: float | None = None,
+        interjection: str | None = None,  # noqa: ARG002
+        task: str | None = None,  # noqa: ARG002
+    ) -> str | None:
+        """Deterministic plan = numbered list of *still-todo* subtasks.
+
+        No VLM call: a plain numbered list keeps the plan aligned with the
+        upcoming subtasks (the old VLM "compact hierarchical plan" prompt
+        cost a round-trip per episode/refresh and could diverge).
+
+            1. <subtask 1>
+            2. <subtask 2>
+
+        On a refresh at ``refresh_t`` (from ``run_plan_updates`` on
+        interjections, and ``run_episode`` at each boundary), only subtasks
+        starting at or after ``refresh_t`` are included — so it always
+        describes what's left.
+        """
+        if not subtask_spans:
+            return None
+        remaining = [
+            s for s in subtask_spans if refresh_t is None or float(s.get("start", 0.0)) >= float(refresh_t)
+        ]
+        if not remaining:
+            # Past the last subtask boundary on a late refresh — nothing
+            # left to plan; emit None so the caller skips the row.
+            return None
+        return "\n".join(f"{i}. {span.get('text', '').strip()}" for i, span in enumerate(remaining, start=1))
+
+    def _generate_memory(
+        self,
+        record: EpisodeRecord,
+        prior_memory: str,
+        completed: str,
+        remaining: Sequence[str],
+        *,
+        task: str | None = None,
+    ) -> str:
+        prompt = load_prompt("plan_memory").format(
+            episode_task=(task if task is not None else record.episode_task),
+            prior_memory=prior_memory or "(none)",
+            completed_subtask=completed,
+            remaining_subtasks=", ".join(remaining) if remaining else "(none)",
+        )
+        memory = self._vlm_field(self._text_message(prompt), "memory")
+        return memory.strip() if isinstance(memory, str) else ""

src/lerobot/annotations/steerable_pipeline/prompts/__init__.py

@@ -0,0 +1,33 @@
+#!/usr/bin/env python
+
+# Copyright 2026 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""Prompt templates loaded as plain text.
+
+One file per use site. Templates use ``str.format(**vars)`` substitution; we
+intentionally avoid jinja2 here so the templates remain inspectable in
+plain editors and roundtrip cleanly through ``ruff format``.
+"""
+
+from __future__ import annotations
+
+from pathlib import Path
+
+_DIR = Path(__file__).parent
+
+
+def load(name: str) -> str:
+    """Read prompt template ``name.txt`` from the ``prompts/`` directory."""
+    path = _DIR / f"{name}.txt"
+    return path.read_text(encoding="utf-8")

src/lerobot/annotations/steerable_pipeline/prompts/interjections_initial_speech.txt

@@ -0,0 +1,12 @@
+The user just asked the robot: "{episode_task}".
+
+Generate a short verbal acknowledgement the robot would speak back before
+beginning the task. Style: compact, confident, friendly.
+
+Examples (Hi Robot, Shi 2025): "Sure, I won't put cheese on it.",
+"OK, starting with the sponge.", "Got it.".
+
+Prefer very short replies: "Got it.", "On it.", "OK."
+
+Output strictly valid JSON:
+  {{ "text": "<the spoken acknowledgement>" }}

src/lerobot/annotations/steerable_pipeline/prompts/interjections_interjection.txt

@@ -0,0 +1,46 @@
+You are generating training data for a Hi Robot-style hierarchical
+robot policy. The robot in this demonstration has ALREADY executed
+every step shown in the video — we cannot retroactively change the
+action stream. To keep training data consistent with the video, the
+"interjection" must align with what the robot is *about to do next* in
+the demonstration, framed as a natural mid-task user request.
+
+The episode's overall task: "{episode_task}".
+
+The images above show roughly {window_seconds:.1f} seconds straddling a
+subtask boundary in the demonstration:
+
+- Subtask the robot just finished: "{prev_subtask}"
+- Subtask the robot is about to start: "{next_subtask}"
+- Time into episode: {timestamp:.2f}s
+
+Write ONE compact interjection the user would naturally say at this
+moment to prompt / confirm / encourage the robot to do "{next_subtask}".
+Keep it like a mid-task coaching cue, not a full instruction paragraph.
+Also write the robot's compact verbal acknowledgement.
+
+Hard rules:
+
+- The interjection MUST be consistent with the next subtask. The user
+  cannot ask for something different from what the robot then does in
+  the video. If you're tempted to say "actually skip X" or "do Y
+  instead", DO NOT — those would contradict the demonstration.
+- The interjection must reference an object, location, or action that
+  is plausible given the visible scene and the next subtask text.
+- One short phrase or sentence each. Conversational, not robotic.
+- Prefer direct cues: "{next_subtask}, please."; "Now {next_subtask}."
+- Keep robot speech very short: "OK.", "On it.", "Doing that."
+
+Style examples (vary the phrasing — don't reuse these verbatim):
+  - "Now go ahead and {next_subtask}."
+  - "Great, can you {next_subtask} next?"
+  - "{next_subtask}, please."
+  - "Before you continue, please {next_subtask}."
+  - "Looking good — {next_subtask} now."
+  - "Okay, {next_subtask}."
+
+Output strictly valid JSON:
+  {{
+    "interjection": "<short cue from the user, asking for the next subtask>",
+    "speech":       "<short robot acknowledgement>"
+  }}

src/lerobot/annotations/steerable_pipeline/prompts/plan_memory.txt

@@ -0,0 +1,36 @@
+You are updating the robot's compressed semantic memory at the boundary of
+a completed subtask.
+
+Reference (verbatim from MEM, Torne 2026):
+"Remove or compress information in the language memory whenever
+appropriate. Keep ONLY the minimal set of relevant information for future
+task execution. Specific object attributes (colors, precise quantities of
+each item) get discarded when their details won't affect subsequent
+actions. Functional outcomes (where items went, how many) are preserved."
+
+Episode task: "{episode_task}"
+Previous memory: {prior_memory}
+Just-completed subtask: "{completed_subtask}"
+Remaining subtasks (for relevance judgement only): {remaining_subtasks}
+
+Write the memory as a short FIRST-PERSON, PAST-TENSE narrative of what the
+robot has accomplished so far — the running story it would tell itself.
+
+Authoring rules:
+- First person, past tense. Every sentence starts with "I": "I picked
+  up...", "I opened...", "I moved to...".
+- One or two short sentences. Extend the previous memory with the
+  just-completed subtask; do not rewrite it from scratch.
+- Keep WHAT happened (functional outcomes — where items went, how many),
+  drop HOW (grasp details, motions).
+- Compress completed steps and drop object attributes (colors, exact
+  counts) once they no longer affect the remaining subtasks.
+
+Example (MEM, Torne 2026):
+  Before: "I prepared the pot and got the potatoes, milk, and butter. I
+           moved to the drawer."
+  After:  "I prepared the pot and got the ingredients. I opened the
+           drawer with the masher."
+
+Output strictly valid JSON:
+  {{ "memory": "<one or two short first-person past-tense sentences>" }}

src/lerobot/annotations/steerable_pipeline/prompts/plan_subtask_describe.txt

@@ -0,0 +1,27 @@
+You are watching a teleoperated robot demonstration from a single
+camera. The user asked the robot to: "{episode_task}"
+
+This is an OBSERVATION pass. Watch the entire clip and describe, in
+chronological order, ONLY what the robot physically does — the concrete
+motions, approaches, contacts, grasps, releases, and relocations you can
+actually SEE in the frames.
+
+Hard rules:
+- Describe only motion visible in the video. Do NOT use the task
+  instruction to guess steps that aren't shown. The instruction is the
+  goal; the video is ground truth.
+- Do NOT segment into named subtasks yet and do NOT output JSON beyond
+  the single field below. Just narrate what happens.
+- Give an approximate timestamp (in seconds) for each distinct event,
+  e.g. "0.0-1.4s: the base drives forward toward the stove".
+- Do NOT invent objects, grasps, destinations, or steps. If the robot
+  only does one thing (e.g. it just navigates and the clip ends), say
+  exactly that and nothing more.
+- Be concrete and literal. "the gripper closes on the mug" — not "the
+  robot prepares to make coffee".
+
+Output strictly valid JSON:
+
+  {{
+    "description": "<chronological, timestamped description of ONLY what is visible>"
+  }}

src/lerobot/annotations/steerable_pipeline/prompts/plan_subtasks.txt

@@ -0,0 +1,112 @@
+You are labeling a teleoperated robot demonstration.
+
+The user originally asked: "{episode_task}"
+
+You are shown the entire demonstration as a single video. Watch the
+whole clip, then segment it into a list of consecutive atomic subtasks
+the robot performs.
+
+{observation_block}GROUNDING — read this first, it overrides everything below:
+- Label ONLY what the robot actually does in the video. Every subtask
+  you emit must correspond to motion you can SEE in specific frames.
+- Do NOT invent, anticipate, or pad. If the robot only does one thing
+  (e.g. it just navigates to a location and the clip ends), emit
+  EXACTLY ONE subtask. Many demonstrations are a single atomic skill.
+- ``max_steps`` below is a hard CEILING, not a target. Emitting fewer
+  subtasks than the ceiling is not just allowed, it is expected for
+  short / atomic demonstrations. One correct subtask is far better
+  than several invented ones.
+- If the video does not clearly show the action implied by the task,
+  describe what you actually see — do NOT fabricate the task's steps
+  from the instruction text. The instruction tells you the goal; the
+  VIDEO is the ground truth for what happened.
+
+Authoring rules — Hi Robot atom granularity, pi0.7-style short prompts:
+
+- Each subtask = one COMPOSITE atomic skill the low-level policy can
+  execute end-to-end. A "skill" bundles its own approach motion with
+  its terminal action — do NOT split the approach off as its own
+  subtask. The whole-arm policy already learns to reach as part of
+  every manipulation primitive.
+- Write each subtask as an IMPERATIVE COMMAND, starting with one of
+  these verbs (extend only when none fits):
+    pick up <obj>           — approach + grasp + lift in one subtask
+    put <obj> on/in <loc>   — transport + release in one subtask
+    place <obj> on/in <loc> — synonym of "put"; pick one and stay consistent
+    push <obj>              — contact + linear shove
+    pull <obj>              — contact + linear retract
+    turn <knob/dial/handle> — rotary actuation
+    press <button>          — single-press contact
+    open <drawer/door/lid>  — full open motion
+    close <drawer/door/lid> — full close motion
+    pour <src> into <dst>   — tilt + flow
+    insert <obj> into <slot>— alignment + push-fit
+    go to <loc>             — ONLY when no grasp / actuation follows
+                             (e.g. a pure relocation between phases).
+                             If the next subtask grasps something at
+                             that location, drop "go to ..." and just
+                             write "pick up ..." instead.
+- Forbidden ultra-fine splits — the VLM is NOT allowed to emit these
+  as standalone subtasks; fold them into the parent composite:
+    "move to X"   → fold into "pick up X" (or whatever follows)
+    "reach for X" → fold into "pick up X"
+    "grasp X"     → fold into "pick up X"
+    "lift X"      → fold into "pick up X" (or "put X on Y" if it's
+                    the transport phase of a place)
+    "release X"   → fold into "put X on Y" (or "place X in Y")
+- Keep it SHORT — a verb phrase, not a sentence. Drop articles
+  ("the", "a") and adverbs ("carefully", "slowly"). Add a "how"
+  detail (which hand, which grasp point) ONLY when it is needed to
+  disambiguate. Every subtask must begin with one of the verbs
+  above (no leading nouns, no "then", no "first").
+- NEVER use third person. Never write "the robot", "the arm", "the
+  gripper moves", "it picks up" — the robot is implied. Command it,
+  do not describe it.
+- Use the exact object nouns from the task above. If the task says
+  "cube", every subtask says "cube" — never switch to "block". If it
+  says "box", never switch to "bin"/"container". Keep vocabulary
+  consistent across the whole episode.
+- Good: "pick up blue cube", "put blue cube in box", "open drawer",
+  "turn red knob", "press start button", "go to sink".
+- Bad: "move to blue cube" (approach as its own subtask — forbidden,
+  must be folded into "pick up blue cube"); "the robot arm moves
+  towards the blue cube" (third person, too long); "carefully pick
+  up the cube" (adverb, article); "release the yellow block"
+  ("block" when the task said "cube", and "release" must be folded
+  into a "put"/"place" subtask).
+- Subtasks are non-overlapping and cover the full episode in order.
+  Choose the cut points yourself based on what you see in the video
+  (gripper open/close events, contact, regrasps, transitions).
+- Each subtask spans at least {min_subtask_seconds} seconds. If a
+  candidate span would be shorter, merge it into its neighbour
+  rather than emitting it.
+- Do not exceed {max_steps} subtasks total. Fewer, larger composites
+  are preferred over many micro-steps.
+- Every subtask's [start_time, end_time] must lie within
+  [0.0, {episode_duration}] seconds.
+
+SPECIAL CASES — verb disambiguation (each rule is narrowly visual and
+fires ONLY on the spatial situation it names; it must not change how you
+label any other situation):
+- STACK vs PUT: if an object is placed ON TOP OF another specific object
+  (not on a flat table / shelf / counter), use "stack ... on ...", not
+  "put". "stack blue book on green book", NOT "put blue book on table".
+- INSERT vs PUT: if an object goes INTO a fitted slot / hole / socket /
+  receptacle (push-fit), use "insert ... into ...", not "put".
+- RETRIEVE/PICK-UP vs PUT (direction): watch the gripper. If it CLOSES
+  on the object and the object moves WITH the hand, it is "pick up" /
+  "retrieve" (object leaves its location). If the gripper OPENS and the
+  object stays where the hand left it, it is "put" / "place" (object
+  arrives at a location). Decide by which way the object moves, not by
+  where the hand ends up.
+- POUR vs PUT: only use "pour" when the source is tilted and contents
+  flow out; moving a full container without tilting is "put"/"place".
+
+Output strictly valid JSON of shape:
+
+  {{
+    "subtasks": [
+      {{"text": "<short imperative verb phrase>", "start": <float>, "end": <float>}},
+      ...
+    ]
+  }}

src/lerobot/annotations/steerable_pipeline/prompts/plan_task_aug_axes.txt

@@ -0,0 +1,67 @@
+You are generating structured augmentations of a robot task instruction
+for training a language-conditioned policy. Unlike free-form rephrasing,
+your variants follow a NAMED 5-axis taxonomy — each axis omits or varies
+a specific element of the task while preserving its meaning.
+
+Original task: "{base_task}"
+
+Produce variants along five named axes. Each axis has a target count.
+The whole batch should expose the policy to maximum linguistic diversity
+WITHOUT changing what the robot is supposed to do.
+
+Axes and target counts:
+
+  synonym_paraphrase ({n_synonym}):
+    Different wording / verbs / sentence structure. ALL information
+    from the original task is preserved — same object, same arm
+    specification if present, same orientation if present, same grasp
+    if present.
+
+  omit_arm ({n_omit_arm}):
+    Drop the left/right/both arm specification from the task. Skip
+    entirely (emit 0 entries) if the original task does NOT mention an
+    arm. Do not invent an arm specification just to omit it.
+
+  omit_orientation ({n_omit_orientation}):
+    Drop orientation cues (upright, sideways, facing the user,
+    long-edge-first, etc.). Skip entirely if no orientation cue is
+    present in the original task.
+
+  omit_grasp_method ({n_omit_grasp_method}):
+    Drop the grip / grasp method specification (pinch, wrap, hold by
+    the rim, etc.). Skip entirely if no grasp method is mentioned.
+
+  combined_omissions ({n_combined}):
+    Combine TWO of the above omissions simultaneously (e.g. drop both
+    arm and orientation). Skip entirely if fewer than two of (arm,
+    orientation, grasp_method) appear in the original task.
+
+Hard rules:
+- Each variant MUST preserve the core action, the target object, AND
+  the goal / destination. Do not change which object is involved, where
+  it goes, or the high-level action. "Navigate to the stove" may become
+  "go to the stove" or "head over to the stove" — it must NEVER become
+  "wander around the kitchen", "explore the room", or anything that
+  drops or generalises the stove destination. If you cannot vary the
+  wording without changing the goal, emit fewer variants.
+- Only the FIVE listed elements (wording, arm, orientation, grasp
+  method, or a combination) may be varied or omitted. The verb's
+  meaning, the object, and the destination are fixed.
+- Each variant is plain prose, no markdown, no quotes, no list numbers.
+- Each variant must be DISTINCT from every other variant in the entire
+  output, both within and across axes. Near-duplicates are not allowed.
+- If an axis cannot reach its target count because the original task
+  lacks the omittable element, emit fewer entries — do NOT pad the
+  axis with paraphrases that belong to a different axis.
+- Variants should not all start with verbs — vary sentence structure
+  (some imperative, some polite request, some question).
+
+Output strictly valid JSON of shape:
+
+  {{
+    "synonym_paraphrase": ["<v1>", "<v2>", ...],
+    "omit_arm": ["<v1>", "<v2>", ...],
+    "omit_orientation": ["<v1>", ...],
+    "omit_grasp_method": ["<v1>", ...],
+    "combined_omissions": ["<v1>", ...]
+  }}

src/lerobot/annotations/steerable_pipeline/prompts/plan_task_rephrasings.txt

@@ -0,0 +1,32 @@
+You are generating training data for a Hi Robot-style policy. We need
+{n} alternative phrasings of the same robot task so the policy sees
+diverse user prompts during training instead of the same canonical
+string repeated every frame.
+
+Original task:
+"{base_task}"
+
+Generate exactly {n} alternative phrasings of the same task. Vary:
+
+- formality (casual / polite / curt)
+- verbosity (mostly short imperative; occasional polite request)
+- word choice (synonyms, different verbs)
+- sentence structure (imperative / question / suggestion)
+
+Hard rules:
+- Each phrasing MUST preserve the exact meaning of the original task.
+  Do not change which object is involved, the destination, or the
+  action. Do not add extra steps. Do not invent new objects.
+- Each phrasing must be a short phrase or sentence, plain prose, no
+  markdown, no quotes, no list numbers.
+- Phrasings must be distinct — no near-duplicates.
+- Output exactly {n} entries.
+
+Output strictly valid JSON:
+  {{
+    "rephrasings": [
+      "<phrasing 1>",
+      "<phrasing 2>",
+      ...
+    ]
+  }}

src/lerobot/annotations/steerable_pipeline/prompts/plan_video_task.txt

@@ -0,0 +1,17 @@
+The video above shows a robot manipulation episode in full. Look at
+the entire video and describe in ONE concise sentence what the robot
+is doing.
+
+Rules:
+- One sentence, in natural English, like a user instruction.
+- Capture the goal of the demonstration, not low-level motions.
+  Example: "place the yellow cube into the red bin" — not "move the
+  end-effector down 5cm and close the gripper".
+- 4 to 15 words. Plain prose, no markdown, no bullets, no quotes.
+- Do not invent objects or actions that aren't visible.
+- Do not output anything other than the JSON object below.
+
+Output strictly valid JSON:
+  {{
+    "task": "<single concise sentence describing what the robot does in this video>"
+  }}

src/lerobot/annotations/steerable_pipeline/prompts/vqa.txt

@@ -0,0 +1,32 @@
+You are generating a frame-grounded visual question/answer pair for
+chain-of-thought training. Reference: ECoT (Zawalski 2024) and Steerable
+Policies — both train policies on grounded features such as bounding box
+pixel coordinates, keypoints, counts, attributes, and spatial relations.
+
+The frame shows a robot working on: "{episode_task}".
+
+Question types and the EXACT answer JSON shape required for each:
+
+  bbox       => {{"detections": [{{"label": "<obj>", "bbox_format": "xyxy",
+                                    "bbox": [x1, y1, x2, y2]}}, ...]}}
+                bbox is in pixel coordinates (x_min, y_min, x_max, y_max).
+                ECoT example: "a white cup [124, 25, 176, 113]".
+
+  keypoint   => {{"label": "<point>", "point_format": "xy",
+                  "point": [x, y]}}
+
+  count      => {{"label": "<obj>", "count": <int>,
+                  "note": "<optional short note>"}}
+
+  attribute  => {{"label": "<obj>", "attribute": "<color|shape|state|...>",
+                  "value": "<observed value>"}}
+
+  spatial    => {{"subject": "<obj>", "relation": "<left_of|right_of|on|in|"
+                  "above|below|near>", "object": "<obj>"}}
+
+Generate a question of type "{question_type}". Output strictly valid JSON:
+
+  {{
+    "question": "<short, frame-grounded question>",
+    "answer":   <object whose shape matches the schema above>
+  }}

src/lerobot/annotations/steerable_pipeline/reader.py

@@ -0,0 +1,216 @@
+#!/usr/bin/env python
+
+# Copyright 2026 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""Datatrove-shaped reader.
+
+The reader walks ``data/chunk-*/file-*.parquet`` and yields one record per
+episode containing:
+
+- ``episode_index``: int
+- ``frame_timestamps``: tuple[float, ...]
+- ``frame_indices``: tuple[int, ...]
+- ``episode_task``: str (canonical task from ``meta/tasks.parquet``)
+- ``data_path``: pathlib.Path of the source parquet shard
+- ``frames_df``: pandas.DataFrame slice for the episode (only loaded on demand)
+
+This shape lets each module operate per-episode without loading all parquet
+rows into memory at once.
+"""
+
+from __future__ import annotations
+
+from collections.abc import Iterator, Sequence
+from dataclasses import dataclass, field
+from pathlib import Path
+from typing import Any
+
+import pyarrow.parquet as pq
+
+from lerobot.datasets.io_utils import load_tasks
+from lerobot.datasets.utils import DEFAULT_TASKS_PATH
+
+
+@dataclass
+class EpisodeRecord:
+    """Per-episode record yielded by the reader."""
+
+    episode_index: int
+    episode_task: str
+    frame_timestamps: tuple[float, ...]
+    frame_indices: tuple[int, ...]
+    data_path: Path
+    row_offset: int  # row offset within the parquet file where this episode starts
+    row_count: int  # number of rows for this episode
+
+    # Memoized parquet slice — populated on first ``frames_df()`` call so
+    # repeat queries from different modules don't re-read the whole shard.
+    _frames_df_cache: Any = field(default=None, init=False, repr=False, compare=False)
+
+    def frames_df(self):  # type: ignore[no-untyped-def]
+        """Lazy-load the pandas slice for this episode (memoized)."""
+        if self._frames_df_cache is None:
+            import pandas as pd  # noqa: PLC0415  - deferred for optional dataset extra
+
+            table = pq.read_table(self.data_path)
+            df: pd.DataFrame = table.to_pandas()
+            self._frames_df_cache = df.iloc[self.row_offset : self.row_offset + self.row_count].reset_index(
+                drop=True
+            )
+        return self._frames_df_cache
+
+
+def reconstruct_subtask_spans(
+    rows: Sequence[dict[str, Any]],
+    *,
+    episode_end_t: float | None = None,
+) -> list[dict[str, Any]]:
+    """Turn ``style="subtask"`` rows into ``{text, start, end}`` spans.
+
+    Each span's ``end`` is the next span's ``start``. The final span's
+    ``end`` defaults to its own ``start`` (zero-duration) — pass
+    ``episode_end_t`` to extend it to the episode's last frame instead,
+    which is what downstream consumers (memory, interjection boundary
+    selection) expect.
+
+    Used by the ``plan`` module (plan-update pass) and the
+    ``interjections`` module (interjection anchoring), which both need the
+    same span shape.
+    """
+    sorted_rows = sorted(
+        (r for r in rows if r.get("style") == "subtask"),
+        key=lambda r: float(r["timestamp"]),
+    )
+    spans: list[dict[str, Any]] = []
+    for r in sorted_rows:
+        t = float(r["timestamp"])
+        if spans:
+            spans[-1]["end"] = t
+        spans.append({"text": r.get("content") or "", "start": t, "end": t})
+    if spans and episode_end_t is not None and float(episode_end_t) > spans[-1]["start"]:
+        spans[-1]["end"] = float(episode_end_t)
+    return spans
+
+
+def snap_to_frame(t: float, frame_timestamps: Sequence[float]) -> float:
+    """Snap an arbitrary float to the nearest exact source frame timestamp.
+
+    Modules use this when emitting event-style rows so the row's
+    timestamp matches a real parquet frame: event rows must land on an
+    exact frame, otherwise the per-frame event lookup the writer does
+    would never match them.
+    """
+    if not frame_timestamps:
+        return float(t)
+    nearest = min(frame_timestamps, key=lambda f: abs(f - t))
+    return float(nearest)
+
+
+def _load_tasks_lookup(root: Path) -> dict[int, str]:
+    """Map ``task_index -> task`` from ``meta/tasks.parquet``.
+
+    Returns an empty dict when the file is absent — the task description is
+    derived later from the video if needed. Reuses the library-level
+    :func:`lerobot.datasets.io_utils.load_tasks`, which returns the tasks
+    frame indexed by task string with a ``task_index`` column.
+    """
+    if not (root / DEFAULT_TASKS_PATH).exists():
+        return {}
+    tasks = load_tasks(root)
+    return {int(idx): str(task) for task, idx in zip(tasks.index, tasks["task_index"], strict=True)}
+
+
+def iter_episodes(root: Path, *, only_episodes: tuple[int, ...] | None = None) -> Iterator[EpisodeRecord]:
+    """Yield :class:`EpisodeRecord` for every episode under ``root/data/``.
+
+    Episodes are yielded in ascending ``episode_index`` order. The reader does
+    not assume a specific chunk/file layout: it scans every ``*.parquet``
+    under ``data/`` and groups by ``episode_index``.
+    """
+    tasks = _load_tasks_lookup(root)
+    data_dir = root / "data"
+    parquet_files = sorted(data_dir.rglob("*.parquet"))
+
+    only_set = set(only_episodes) if only_episodes is not None else None
+
+    for path in parquet_files:
+        yield from _iter_one_path(path, tasks, only_set)
+
+
+def _iter_one_path(path: Path, tasks: dict[int, str], only_set: set[int] | None) -> Iterator[EpisodeRecord]:
+    table = pq.read_table(path)
+    names = table.column_names
+    if "episode_index" not in names:
+        return
+    episode_col = table.column("episode_index").to_pylist()
+    timestamp_col = (
+        table.column("timestamp").to_pylist() if "timestamp" in names else [0.0] * len(episode_col)
+    )
+    frame_col = (
+        table.column("frame_index").to_pylist() if "frame_index" in names else list(range(len(episode_col)))
+    )
+    task_col = table.column("task_index").to_pylist() if "task_index" in names else None
+
+    def _build(
+        ep: int,
+        start: int,
+        end: int,
+        task_idx: int | None,
+        ts_buf: list[float],
+        fi_buf: list[int],
+    ) -> EpisodeRecord | None:
+        if only_set is not None and ep not in only_set:
+            return None
+        task = tasks.get(task_idx, "") if task_idx is not None else ""
+        return EpisodeRecord(
+            episode_index=ep,
+            episode_task=task,
+            frame_timestamps=tuple(ts_buf),
+            frame_indices=tuple(fi_buf),
+            data_path=path,
+            row_offset=start,
+            row_count=end - start,
+        )
+
+    cur_ep: int | None = None
+    start_offset = 0
+    ts_buf: list[float] = []
+    fi_buf: list[int] = []
+    cur_task_idx: int | None = None
+
+    for i, ep in enumerate(episode_col):
+        if cur_ep is None:
+            cur_ep = ep
+            start_offset = i
+            ts_buf = [timestamp_col[i]]
+            fi_buf = [frame_col[i]]
+            cur_task_idx = task_col[i] if task_col is not None else None
+            continue
+        if ep != cur_ep:
+            rec = _build(cur_ep, start_offset, i, cur_task_idx, ts_buf, fi_buf)
+            if rec is not None:
+                yield rec
+            cur_ep = ep
+            start_offset = i
+            ts_buf = [timestamp_col[i]]
+            fi_buf = [frame_col[i]]
+            cur_task_idx = task_col[i] if task_col is not None else None
+        else:
+            ts_buf.append(timestamp_col[i])
+            fi_buf.append(frame_col[i])
+
+    if cur_ep is not None:
+        rec = _build(cur_ep, start_offset, len(episode_col), cur_task_idx, ts_buf, fi_buf)
+        if rec is not None:
+            yield rec

src/lerobot/annotations/steerable_pipeline/staging.py

@@ -0,0 +1,92 @@
+#!/usr/bin/env python
+
+# Copyright 2026 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""Per-episode staging.
+
+Each module writes its raw output as a JSONL file under
+``<staging_dir>/episode_{ep:06d}/<module>.jsonl``. The writer reads back this
+staging tree and partitions rows into the two language columns.
+
+JSONL is preferred over parquet here because the staging artifact is meant to
+be human-inspectable, easy to diff between prompt iterations, and trivially
+appended to. The final dataset format is parquet; staging is just an
+intermediate.
+"""
+
+from __future__ import annotations
+
+import json
+from collections.abc import Iterable
+from dataclasses import dataclass
+from pathlib import Path
+from typing import Any
+
+ModuleName = str
+
+_MODULES: tuple[ModuleName, ...] = (
+    "plan",
+    "interjections",
+    "vqa",
+)
+
+
+@dataclass
+class EpisodeStaging:
+    """Filesystem layout for a single episode's staged module outputs."""
+
+    root: Path
+    episode_index: int
+
+    @property
+    def episode_dir(self) -> Path:
+        return self.root / f"episode_{self.episode_index:06d}"
+
+    def path_for(self, module: ModuleName) -> Path:
+        if module not in _MODULES:
+            raise ValueError(f"Unknown module {module!r}; expected one of {_MODULES}")
+        return self.episode_dir / f"{module}.jsonl"
+
+    def write(self, module: ModuleName, rows: Iterable[dict[str, Any]]) -> Path:
+        path = self.path_for(module)
+        path.parent.mkdir(parents=True, exist_ok=True)
+        # Atomic replace: a crash mid-write would otherwise leave a
+        # half-written JSONL file that ``read()`` would then fail to
+        # parse. Write to a sibling .tmp and rename so the target path
+        # only ever points at a complete file.
+        tmp_path = path.with_suffix(path.suffix + ".tmp")
+        with tmp_path.open("w", encoding="utf-8") as f:
+            for row in rows:
+                f.write(json.dumps(row, ensure_ascii=False, sort_keys=True))
+                f.write("\n")
+        tmp_path.replace(path)
+        return path
+
+    def read(self, module: ModuleName) -> list[dict[str, Any]]:
+        path = self.path_for(module)
+        if not path.exists():
+            return []
+        out: list[dict[str, Any]] = []
+        with path.open(encoding="utf-8") as f:
+            for line in f:
+                line = line.strip()
+                if line:
+                    out.append(json.loads(line))
+        return out
+
+    def read_all(self) -> dict[ModuleName, list[dict[str, Any]]]:
+        return {m: self.read(m) for m in _MODULES}
+
+    def has(self, module: ModuleName) -> bool:
+        return self.path_for(module).exists()

src/lerobot/annotations/steerable_pipeline/validator.py

@@ -0,0 +1,332 @@
+#!/usr/bin/env python
+
+# Copyright 2026 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""Pre-write validation against staged outputs.
+
+Runs after all three modules have written their per-episode artifacts but
+*before* the writer rewrites parquet shards. The validator never touches
+parquet; it only inspects the staging tree and the source frame timestamps
+exposed by :class:`EpisodeRecord`.
+
+Checks (per the plan's "Intermediate staging and validation" section):
+
+- exact timestamp alignment against source frame timestamps
+- no orphan speech / interjection pairs
+- plan / memory emission consistency (events have a paired persistent row)
+- VQA assistant ``content`` is valid JSON (one of bbox / keypoint / count /
+  attribute / spatial)
+- every row maps to its correct column under :func:`column_for_style`
+"""
+
+from __future__ import annotations
+
+import json
+import logging
+from collections.abc import Iterable, Sequence
+from dataclasses import dataclass, field
+from pathlib import Path
+from typing import Any
+
+from lerobot.datasets.language import (
+    LANGUAGE_EVENTS,
+    LANGUAGE_PERSISTENT,
+    column_for_style,
+    is_view_dependent_style,
+    validate_camera_field,
+)
+
+from .reader import EpisodeRecord
+from .staging import EpisodeStaging
+
+logger = logging.getLogger(__name__)
+
+
+@dataclass
+class ValidationReport:
+    """Outcome of one validation pass across all episodes."""
+
+    errors: list[str] = field(default_factory=list)
+    warnings: list[str] = field(default_factory=list)
+    episodes_checked: int = 0
+
+    @property
+    def ok(self) -> bool:
+        return not self.errors
+
+    def add_error(self, message: str) -> None:
+        self.errors.append(message)
+
+    def add_warning(self, message: str) -> None:
+        self.warnings.append(message)
+
+    def summary(self) -> str:
+        return f"checked={self.episodes_checked} errors={len(self.errors)} warnings={len(self.warnings)}"
+
+
+VQA_ANSWER_SHAPES: dict[str, set[str]] = {
+    "bbox": {"detections"},
+    "keypoint": {"label", "point_format", "point"},
+    "count": {"label", "count"},
+    "attribute": {"label", "attribute", "value"},
+    "spatial": {"subject", "relation", "object"},
+}
+
+
+def classify_vqa_answer(payload: Any) -> str | None:
+    """Best-effort classification of a VQA answer payload to a question type."""
+    if not isinstance(payload, dict):
+        return None
+    keys = set(payload.keys())
+    for kind, required in VQA_ANSWER_SHAPES.items():
+        if required.issubset(keys):
+            return kind
+    return None
+
+
+@dataclass
+class StagingValidator:
+    """Walks the staging tree and produces a :class:`ValidationReport`."""
+
+    timestamp_atol: float = 0.0  # exact-match by default
+    dataset_camera_keys: tuple[str, ...] | None = None
+    """Known ``observation.images.*`` keys on the dataset. When set, the
+    validator additionally enforces that every view-dependent row's
+    ``camera`` field references one of these keys. Pass ``None`` (default)
+    to skip that cross-check (e.g. in unit tests with no real dataset)."""
+
+    def validate(
+        self,
+        records: Sequence[EpisodeRecord],
+        staging_dir: Path,
+    ) -> ValidationReport:
+        report = ValidationReport()
+        for record in records:
+            self._validate_episode(record, staging_dir, report)
+            report.episodes_checked += 1
+        return report
+
+    def _validate_episode(
+        self,
+        record: EpisodeRecord,
+        staging_dir: Path,
+        report: ValidationReport,
+    ) -> None:
+        staging = EpisodeStaging(staging_dir, record.episode_index)
+        staged = staging.read_all()
+        all_rows: list[dict[str, Any]] = []
+        for module_name, rows in staged.items():
+            for row in rows:
+                row = {**row, "_module": module_name}
+                all_rows.append(row)
+
+        frame_ts = set(record.frame_timestamps)
+
+        events: list[dict[str, Any]] = []
+        persistent: list[dict[str, Any]] = []
+        for row in all_rows:
+            self._check_column_routing(row, report, record.episode_index)
+            self._check_camera_field(row, report, record.episode_index, self.dataset_camera_keys)
+            # ``_check_column_routing`` already recorded any unknown-style error;
+            # don't let the same ``column_for_style`` lookup raise here uncaught.
+            try:
+                column = column_for_style(row.get("style"))
+            except ValueError:
+                continue
+            if column == LANGUAGE_PERSISTENT:
+                persistent.append(row)
+            else:
+                events.append(row)
+
+        for row in events:
+            self._check_event_timestamp_alignment(row, frame_ts, report, record.episode_index)
+
+        self._check_speech_interjection_pairs(events, report, record.episode_index)
+        self._check_plan_memory_consistency(persistent, events, report, record.episode_index)
+        self._check_vqa_json(events, report, record.episode_index)
+        self._check_vqa_uniqueness_per_frame_camera(events, report, record.episode_index)
+
+    def _check_camera_field(
+        self,
+        row: dict[str, Any],
+        report: ValidationReport,
+        episode_index: int,
+        dataset_camera_keys: Sequence[str] | None,
+    ) -> None:
+        """Enforce the camera invariant + that the key matches the dataset's cameras."""
+        style = row.get("style")
+        camera = row.get("camera")
+        try:
+            validate_camera_field(style, camera)
+        except ValueError as exc:
+            report.add_error(f"ep={episode_index} module={row.get('_module')}: {exc}")
+            return
+        if is_view_dependent_style(style) and dataset_camera_keys and camera not in dataset_camera_keys:
+            report.add_error(
+                f"ep={episode_index} module={row.get('_module')}: camera {camera!r} on style "
+                f"{style!r} is not one of the dataset's video keys {sorted(dataset_camera_keys)!r}"
+            )
+
+    def _check_vqa_uniqueness_per_frame_camera(
+        self,
+        events: Iterable[dict[str, Any]],
+        report: ValidationReport,
+        episode_index: int,
+    ) -> None:
+        """Ensure at most one (vqa, user) and one (vqa, assistant) per (t, camera)."""
+        counts: dict[tuple[float, str, str], int] = {}
+        for row in events:
+            if row.get("style") != "vqa":
+                continue
+            ts = row.get("timestamp")
+            camera = row.get("camera")
+            role = row.get("role")
+            if ts is None or camera is None or role is None:
+                continue  # other validators flag these
+            key = (float(ts), str(camera), str(role))
+            counts[key] = counts.get(key, 0) + 1
+        for (ts, camera, role), n in counts.items():
+            if n > 1:
+                report.add_error(
+                    f"ep={episode_index}: {n} duplicate vqa rows at t={ts} "
+                    f"camera={camera!r} role={role!r}; expected at most one per (t, camera, role)"
+                )
+
+    def _check_column_routing(
+        self,
+        row: dict[str, Any],
+        report: ValidationReport,
+        episode_index: int,
+    ) -> None:
+        style = row.get("style")
+        module = row.get("_module")
+        try:
+            target_col = column_for_style(style)
+        except ValueError:
+            report.add_error(f"ep={episode_index} module={module}: unknown style {style!r}")
+            return
+        if module == "plan" and target_col != LANGUAGE_PERSISTENT:
+            report.add_error(
+                f"ep={episode_index} module=plan emitted style {style!r} that routes to {target_col} (must be persistent)"
+            )
+        if module in {"interjections", "vqa"} and target_col != LANGUAGE_EVENTS:
+            report.add_error(
+                f"ep={episode_index} module={module} emitted style {style!r} that routes to {target_col} (must be events)"
+            )
+
+    def _check_event_timestamp_alignment(
+        self,
+        row: dict[str, Any],
+        frame_ts: set[float],
+        report: ValidationReport,
+        episode_index: int,
+    ) -> None:
+        ts = row.get("timestamp")
+        if ts is None:
+            report.add_error(f"ep={episode_index}: event row missing timestamp: {row!r}")
+            return
+        if self.timestamp_atol == 0.0:
+            if float(ts) not in frame_ts:
+                report.add_error(
+                    f"ep={episode_index}: event row timestamp {ts!r} does not match any source frame timestamp"
+                )
+        else:
+            if not any(abs(float(ts) - f) <= self.timestamp_atol for f in frame_ts):
+                report.add_error(
+                    f"ep={episode_index}: event row timestamp {ts!r} not within {self.timestamp_atol}s of any frame"
+                )
+
+    def _check_speech_interjection_pairs(
+        self,
+        events: Iterable[dict[str, Any]],
+        report: ValidationReport,
+        episode_index: int,
+    ) -> None:
+        speech_ts: dict[float, int] = {}
+        interjection_ts: dict[float, int] = {}
+        for row in events:
+            ts = row.get("timestamp")
+            if ts is None:
+                continue
+            ts_f = float(ts)
+            if row.get("style") is None and row.get("role") == "assistant":
+                speech_ts[ts_f] = speech_ts.get(ts_f, 0) + 1
+            if row.get("style") == "interjection":
+                interjection_ts[ts_f] = interjection_ts.get(ts_f, 0) + 1
+
+        for ts in interjection_ts:
+            if ts not in speech_ts:
+                report.add_error(f"ep={episode_index}: interjection at t={ts} has no paired speech atom")
+
+    def _check_plan_memory_consistency(
+        self,
+        persistent: Sequence[dict[str, Any]],
+        events: Sequence[dict[str, Any]],
+        report: ValidationReport,
+        episode_index: int,
+    ) -> None:
+        plan_ts = sorted({float(r["timestamp"]) for r in persistent if r.get("style") == "plan"})
+        memory_ts = sorted({float(r["timestamp"]) for r in persistent if r.get("style") == "memory"})
+        subtask_ts = sorted({float(r["timestamp"]) for r in persistent if r.get("style") == "subtask"})
+        interjection_ts = sorted(
+            {
+                float(r["timestamp"])
+                for r in events
+                if r.get("style") == "interjection" and r.get("timestamp") is not None
+            }
+        )
+
+        if persistent and not plan_ts:
+            report.add_warning(f"ep={episode_index}: persistent rows present but no plan emitted")
+        # every interjection should have a same-timestamp plan refresh
+        for ts in interjection_ts:
+            if ts not in set(plan_ts):
+                report.add_error(
+                    f"ep={episode_index}: interjection at t={ts} has no co-timestamped plan update"
+                )
+        # memory should be emitted at subtask boundaries (subset relation)
+        if memory_ts and subtask_ts:
+            mem_set = set(memory_ts)
+            sub_set = set(subtask_ts)
+            stray = sorted(mem_set - sub_set)
+            if stray:
+                report.add_warning(f"ep={episode_index}: memory rows at {stray} not at any subtask boundary")
+
+    def _check_vqa_json(
+        self,
+        events: Iterable[dict[str, Any]],
+        report: ValidationReport,
+        episode_index: int,
+    ) -> None:
+        for row in events:
+            if row.get("style") != "vqa" or row.get("role") != "assistant":
+                continue
+            content = row.get("content")
+            if content is None:
+                report.add_error(
+                    f"ep={episode_index}: VQA assistant row at t={row.get('timestamp')} has null content"
+                )
+                continue
+            try:
+                payload = json.loads(content)
+            except (TypeError, ValueError) as exc:
+                report.add_error(
+                    f"ep={episode_index}: VQA assistant content not valid JSON at t={row.get('timestamp')}: {exc}"
+                )
+                continue
+            shape = classify_vqa_answer(payload)
+            if shape is None:
+                report.add_error(
+                    f"ep={episode_index}: VQA assistant payload at t={row.get('timestamp')} does not match any known shape: keys={list(payload) if isinstance(payload, dict) else type(payload).__name__}"
+                )

src/lerobot/annotations/steerable_pipeline/vlm_client.py

@@ -0,0 +1,599 @@
+#!/usr/bin/env python
+
+# Copyright 2026 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""Shared Qwen-VL client.
+
+The pipeline uses a single shared VLM across modules. vLLM is preferred when
+available (high throughput, JSON-guided decoding); transformers is the
+fallback. A ``stub`` backend is used for unit tests so fixtures never call
+into a real model.
+
+The client speaks one method, :meth:`VlmClient.generate_json`, which:
+
+- accepts a list of OpenAI/HF-style multimodal messages,
+- requests JSON output from the server,
+- batches requests transparently,
+- and reprompts once on a JSON parse failure with an inline correction
+  message before raising.
+"""
+
+from __future__ import annotations
+
+import atexit
+import base64
+import io
+import json
+import os
+import shlex
+import signal
+import subprocess
+import sys
+import threading
+import time
+import urllib.request
+from collections.abc import Callable, Sequence
+from concurrent.futures import ThreadPoolExecutor
+from dataclasses import dataclass
+from typing import Any, Protocol
+
+from .config import VlmConfig
+
+
+class VlmClient(Protocol):
+    """Protocol every backend must implement."""
+
+    def generate_json(
+        self,
+        messages_batch: Sequence[Sequence[dict[str, Any]]],
+        *,
+        max_new_tokens: int | None = None,
+        temperature: float | None = None,
+    ) -> list[Any]:
+        """Generate one JSON-decoded response per messages list."""
+
+
+@dataclass
+class StubVlmClient:
+    """Deterministic stub used in unit tests.
+
+    A test passes a callable that maps the *last user message text* (or, if
+    that is empty, the full message list) to a JSON-serializable response.
+    """
+
+    responder: Callable[[Sequence[dict[str, Any]]], Any]
+
+    def generate_json(
+        self,
+        messages_batch: Sequence[Sequence[dict[str, Any]]],
+        *,
+        max_new_tokens: int | None = None,
+        temperature: float | None = None,
+    ) -> list[Any]:
+        return [self.responder(list(messages)) for messages in messages_batch]
+
+
+def _strip_to_json(text: str) -> Any:
+    text = text.strip()
+    # Strip <think>...</think> blocks (Qwen3 Thinking style)
+    while "<think>" in text and "</think>" in text:
+        start = text.find("<think>")
+        end = text.find("</think>", start) + len("</think>")
+        text = (text[:start] + text[end:]).strip()
+    # Strip ```json ... ``` fences from chat-tuned backbones
+    if text.startswith("```"):
+        first = text.find("\n")
+        last = text.rfind("```")
+        if first != -1 and last != -1 and last > first:
+            text = text[first + 1 : last].strip()
+    try:
+        return json.loads(text)
+    except (ValueError, json.JSONDecodeError):
+        pass
+    # Fall back to extracting the first balanced {...} block.
+    obj_text = _extract_first_json_object(text)
+    if obj_text is None:
+        raise json.JSONDecodeError("No JSON object found", text, 0)
+    return json.loads(obj_text)
+
+
+def _extract_first_json_object(text: str) -> str | None:
+    """Return the first balanced ``{...}`` substring, ignoring braces in
+    string literals. Returns ``None`` if no balanced block is found."""
+    start = text.find("{")
+    if start < 0:
+        return None
+    depth = 0
+    in_string = False
+    escape = False
+    for i in range(start, len(text)):
+        ch = text[i]
+        if escape:
+            escape = False
+            continue
+        if ch == "\\":
+            escape = True
+            continue
+        # Note: ``escape`` is always False here — the ``if escape`` branch
+        # above already handled and reset it.
+        if ch == '"':
+            in_string = not in_string
+            continue
+        if in_string:
+            continue
+        if ch == "{":
+            depth += 1
+        elif ch == "}":
+            depth -= 1
+            if depth == 0:
+                return text[start : i + 1]
+    return None
+
+
+@dataclass
+class _GenericTextClient:
+    """Wraps any text-generation callable in JSON-mode + one-retry semantics."""
+
+    generate_text: Callable[[Sequence[Sequence[dict[str, Any]]], int, float], list[str]]
+    config: VlmConfig
+
+    def generate_json(
+        self,
+        messages_batch: Sequence[Sequence[dict[str, Any]]],
+        *,
+        max_new_tokens: int | None = None,
+        temperature: float | None = None,
+    ) -> list[Any]:
+        max_tok = max_new_tokens if max_new_tokens is not None else self.config.max_new_tokens
+        temp = temperature if temperature is not None else self.config.temperature
+        raw = self.generate_text(messages_batch, max_tok, temp)
+        out: list[Any] = []
+        for messages, text in zip(messages_batch, raw, strict=True):
+            try:
+                out.append(_strip_to_json(text))
+                continue
+            except (ValueError, json.JSONDecodeError):
+                pass
+            retry = list(messages) + [
+                {"role": "assistant", "content": text},
+                {
+                    "role": "user",
+                    "content": (
+                        "Your previous reply was not valid JSON. "
+                        "Reply with strictly valid JSON, no prose, no fences."
+                    ),
+                },
+            ]
+            retry_text = self.generate_text([retry], max_tok, temp)[0]
+            try:
+                out.append(_strip_to_json(retry_text))
+            except (ValueError, json.JSONDecodeError):
+                # After retry: log preview and return None instead of crashing
+                # the whole pipeline. Modules treat None as "skip".
+                preview = retry_text.strip().replace("\n", " ")[:200]
+                print(
+                    f"[vlm] WARNING: failed to parse JSON after retry; preview: {preview!r}",
+                    flush=True,
+                )
+                out.append(None)
+        return out
+
+
+def make_vlm_client(config: VlmConfig) -> VlmClient:
+    """Build the shared VLM client.
+
+    Only the ``openai`` backend is supported for now. The shipped workflow
+    is Hugging Face Jobs (``examples/annotations/run_hf_job.py``): it boots
+    a vLLM server inside the ``vllm/vllm-openai`` image and the pipeline
+    talks to it over the OpenAI-compatible API (``--vlm.backend=openai``,
+    optionally auto-spawning the server via ``auto_serve`` /
+    ``serve_command``). The former in-process ``vllm`` / ``transformers``
+    backends were removed to keep the support surface to the HF Jobs path.
+
+    For ``stub``, construct :class:`StubVlmClient` directly with a responder
+    callable; it is rejected here to make accidental misuse obvious.
+    """
+    if config.backend == "openai":
+        return _make_openai_client(config)
+    if config.backend == "stub":
+        raise ValueError(
+            "Use StubVlmClient(...) directly for the stub backend; make_vlm_client builds real clients."
+        )
+    if config.backend in {"vllm", "transformers"}:
+        raise ValueError(
+            f"backend={config.backend!r} (in-process local model) is not supported for now — "
+            "only backend='openai' (the Hugging Face Jobs flow) is. Run the pipeline via "
+            "examples/annotations/run_hf_job.py, which serves the model with vLLM in the "
+            "vllm/vllm-openai image and talks to it over the OpenAI-compatible API."
+        )
+    raise ValueError(f"Unknown VLM backend: {config.backend!r}")
+
+
+def _make_openai_client(config: VlmConfig) -> VlmClient:
+    """Backend that talks to any OpenAI-compatible server.
+
+    Compatible with ``vllm serve``, ``transformers serve``,
+    ``ktransformers serve``, and hosted endpoints. By default the server
+    is expected to be already running. Set ``auto_serve=True`` to have
+    this client spawn one (default: ``transformers serve``), wait until
+    it's ready, and tear it down on process exit.
+
+    Image blocks ``{"type":"image", "image":<PIL.Image>}`` are
+    auto-converted to ``image_url`` data-URLs. Video blocks
+    ``{"type":"video", "video":[<PIL>...]}`` are forwarded as
+    multi-frame ``video_url`` items where supported.
+    """
+    try:
+        from openai import OpenAI  # type: ignore[import-not-found]
+    except ImportError as exc:
+        raise ImportError(
+            "openai package is required for backend='openai'. Install with `pip install openai`."
+        ) from exc
+
+    api_base = config.api_base
+    api_key = config.api_key
+    auto_serve = config.auto_serve
+    api_bases: list[str] = [api_base]
+
+    print(
+        f"[lerobot-annotate] backend=openai model={config.model_id} "
+        f"api_base={api_base} auto_serve={auto_serve}",
+        flush=True,
+    )
+    if auto_serve:
+        if config.parallel_servers > 1:
+            print(
+                f"[lerobot-annotate] spawning {config.parallel_servers} parallel servers",
+                flush=True,
+            )
+            api_bases = _spawn_parallel_inference_servers(config)
+        elif _server_is_up(api_base):
+            print(f"[lerobot-annotate] reusing server already up at {api_base}", flush=True)
+        else:
+            print("[lerobot-annotate] no server reachable; spawning one", flush=True)
+            api_base = _spawn_inference_server(config)
+            api_bases = [api_base]
+            print(f"[lerobot-annotate] server ready at {api_base}", flush=True)
+
+    clients = [OpenAI(base_url=base, api_key=api_key) for base in api_bases]
+    # round-robin counter for parallel mode
+    rr_counter = {"i": 0}
+
+    # ``mm_processor_kwargs`` is a vllm-specific extra; transformers serve
+    # rejects it with HTTP 422. Send it only when explicitly opted in via
+    # an env var (e.g. ``LEROBOT_OPENAI_SEND_MM_KWARGS=1`` for vllm).
+    send_mm_kwargs = os.environ.get("LEROBOT_OPENAI_SEND_MM_KWARGS", "").lower() in {"1", "true", "yes"}
+
+    rr_lock = threading.Lock()
+
+    def _one_call(messages: Sequence[dict[str, Any]], max_tok: int, temp: float) -> str:
+        api_messages, mm_kwargs = _to_openai_messages(messages)
+        kwargs: dict[str, Any] = {
+            "model": config.model_id,
+            "messages": api_messages,
+            "max_tokens": max_tok,
+            "temperature": temp,
+        }
+        extra_body: dict[str, Any] = {}
+        if send_mm_kwargs and mm_kwargs:
+            extra_body["mm_processor_kwargs"] = {**mm_kwargs, "do_sample_frames": True}
+        if config.chat_template_kwargs:
+            extra_body["chat_template_kwargs"] = config.chat_template_kwargs
+        if extra_body:
+            kwargs["extra_body"] = extra_body
+        with rr_lock:
+            chosen = clients[rr_counter["i"] % len(clients)]
+            rr_counter["i"] += 1
+        response = chosen.chat.completions.create(**kwargs)
+        return response.choices[0].message.content or ""
+
+    def _gen(batch: Sequence[Sequence[dict[str, Any]]], max_tok: int, temp: float) -> list[str]:
+        if len(batch) <= 1 or config.client_concurrency <= 1:
+            return [_one_call(messages, max_tok, temp) for messages in batch]
+        # Parallel fan-out — vllm batches these on the server side.
+        max_workers = min(config.client_concurrency, len(batch))
+        with ThreadPoolExecutor(max_workers=max_workers) as pool:
+            futures = [pool.submit(_one_call, messages, max_tok, temp) for messages in batch]
+            return [f.result() for f in futures]
+
+    return _GenericTextClient(_gen, config)
+
+
+def _spawn_parallel_inference_servers(config: VlmConfig) -> list[str]:
+    """Spawn ``config.parallel_servers`` independent vllm replicas.
+
+    Each replica:
+    - is pinned to a single GPU via ``CUDA_VISIBLE_DEVICES``
+    - listens on ``serve_port + i``
+    - is shut down via the same atexit hook as the single-server path
+
+    Returns the list of ``api_base`` URLs the client should round-robin
+    across.
+    """
+    n = config.parallel_servers
+    api_bases: list[str] = []
+    procs: list[subprocess.Popen] = []
+    ready_events: list[threading.Event] = []
+    # Multiple readiness signals — uvicorn's own banner is suppressed at
+    # ``--uvicorn-log-level warning``, so we also accept vllm's own
+    # "Starting vLLM API server" line and the route-listing line. The
+    # HTTP probe below is the ultimate fallback.
+    ready_markers = (
+        "Uvicorn running",
+        "Application startup complete",
+        "Starting vLLM API server",
+        "Available routes are",
+    )
+    # Single lock for all server-stream threads so multibyte chars from
+    # different servers don't interleave and tear UTF-8 sequences.
+    print_lock = threading.Lock()
+
+    base_cmd = config.serve_command or (
+        f"vllm serve {shlex.quote(config.model_id)} "
+        f"--tensor-parallel-size 1 "
+        f"--max-model-len {config.max_model_len or 32768} "
+        f"--uvicorn-log-level warning"
+    )
+
+    num_gpus = config.num_gpus if config.num_gpus > 0 else n
+    for i in range(n):
+        port = config.serve_port + i
+        gpu = i % num_gpus
+        env = os.environ.copy()
+        env["CUDA_VISIBLE_DEVICES"] = str(gpu)
+        cmd = base_cmd.replace("{port}", str(port)) if "{port}" in base_cmd else f"{base_cmd} --port {port}"
+        api_base = f"http://localhost:{port}/v1"
+        api_bases.append(api_base)
+        print(f"[server-{i}] launching on GPU {gpu} port {port}: {cmd}", flush=True)
+        proc = subprocess.Popen(
+            shlex.split(cmd),
+            stdout=subprocess.PIPE,
+            stderr=subprocess.STDOUT,
+            text=True,
+            bufsize=1,
+            env=env,
+        )
+        procs.append(proc)
+        ready = threading.Event()
+        ready_events.append(ready)
+
+        def _stream(idx: int, p: subprocess.Popen, ev: threading.Event) -> None:
+            # Read whole lines and emit each line atomically under the
+            # shared print_lock so output from N servers stays readable.
+            assert p.stdout is not None
+            for line in iter(p.stdout.readline, ""):
+                with print_lock:
+                    sys.stdout.write(f"[server-{idx}] {line}")
+                    if not line.endswith(("\n", "\r")):
+                        sys.stdout.write("\n")
+                    sys.stdout.flush()
+                if any(m in line for m in ready_markers):
+                    ev.set()
+
+        threading.Thread(target=_stream, args=(i, proc, ready), daemon=True).start()
+
+        def _probe(idx: int, base: str, ev: threading.Event, p: subprocess.Popen) -> None:
+            while not ev.is_set() and p.poll() is None:
+                if _server_is_up(base):
+                    print(f"[server-{idx}] ready (http probe)", flush=True)
+                    ev.set()
+                    return
+                time.sleep(2)
+
+        threading.Thread(target=_probe, args=(i, api_base, ready, proc), daemon=True).start()
+
+    def _shutdown() -> None:
+        for i, p in enumerate(procs):
+            if p.poll() is None:
+                print(f"[server-{i}] stopping pid={p.pid}", flush=True)
+                p.send_signal(signal.SIGINT)
+        for p in procs:
+            try:
+                p.wait(timeout=15)
+            except subprocess.TimeoutExpired:
+                p.kill()
+                p.wait(timeout=5)
+
+    atexit.register(_shutdown)
+
+    deadline = time.monotonic() + config.serve_ready_timeout_s
+    while any(not ev.is_set() for ev in ready_events) and time.monotonic() < deadline:
+        for i, p in enumerate(procs):
+            if p.poll() is not None:
+                raise RuntimeError(
+                    f"[server-{i}] inference server exited unexpectedly with rc={p.returncode}"
+                )
+        time.sleep(2)
+    if any(not ev.is_set() for ev in ready_events):
+        raise RuntimeError(f"[server] not all replicas became ready within {config.serve_ready_timeout_s}s")
+    print(f"[lerobot-annotate] all {n} servers ready: {api_bases}", flush=True)
+    return api_bases
+
+
+def _server_is_up(api_base: str) -> bool:
+    """Return True if ``api_base/models`` answers 200 within 2 seconds."""
+    url = api_base.rstrip("/") + "/models"
+    # ``api_base`` is the user-configured local-server URL we just spawned
+    # or the user passed in via ``--vlm.api_base``; the bandit B310 warning
+    # is for arbitrary user-controlled URLs with file:/ schemes which
+    # cannot reach this code path.
+    try:
+        with urllib.request.urlopen(url, timeout=2) as resp:  # noqa: S310  # nosec B310
+            return resp.status == 200
+    except Exception:  # noqa: BLE001
+        return False
+
+
+def _spawn_inference_server(config: VlmConfig) -> str:
+    """Spawn ``transformers serve`` (or ``serve_command``), wait until it
+    accepts ``/v1/models``, and register a shutdown hook.
+
+    Streams the server's stdout/stderr to the parent terminal in
+    real-time on a background thread so users can see model-load
+    progress and errors as they happen.
+
+    Returns the full ``api_base`` URL the OpenAI client should use.
+    """
+    cmd = config.serve_command
+    if not cmd:
+        cmd = (
+            f"transformers serve {shlex.quote(config.model_id)} "
+            f"--port {config.serve_port} --continuous-batching"
+        )
+    api_base = f"http://localhost:{config.serve_port}/v1"
+    print(f"[server] launching: {cmd}", flush=True)
+    proc = subprocess.Popen(
+        shlex.split(cmd),
+        stdout=subprocess.PIPE,
+        stderr=subprocess.STDOUT,
+        text=True,
+        bufsize=1,
+    )
+
+    # Watch the server output for the uvicorn readiness banner. This is
+    # more reliable than polling /v1/models because transformers serve
+    # rescans its cache on every model-list request, which can exceed
+    # the urllib timeout and trigger an infinite probe loop.
+    ready_event = threading.Event()
+    # See _spawn_parallel_inference_servers for why we accept these.
+    ready_markers = (
+        "Uvicorn running",
+        "Application startup complete",
+        "Starting vLLM API server",
+        "Available routes are",
+    )
+
+    def _probe() -> None:
+        while not ready_event.is_set() and proc.poll() is None:
+            if _server_is_up(api_base):
+                print("[server] ready (http probe)", flush=True)
+                ready_event.set()
+                return
+            time.sleep(2)
+
+    threading.Thread(target=_probe, daemon=True).start()
+
+    def _stream_output() -> None:
+        # Read raw chunks instead of iterating lines so tqdm progress
+        # bars (which overwrite using \r) flush in real time.
+        assert proc.stdout is not None
+        buf = ""
+        prefix_started = False
+        while True:
+            ch = proc.stdout.read(1)
+            if ch == "":
+                # process exited; flush any tail
+                if buf:
+                    sys.stdout.write(buf)
+                    sys.stdout.flush()
+                return
+            if not prefix_started:
+                sys.stdout.write("[server] ")
+                prefix_started = True
+            sys.stdout.write(ch)
+            sys.stdout.flush()
+            buf += ch
+            if ch in ("\n", "\r"):
+                if any(marker in buf for marker in ready_markers):
+                    ready_event.set()
+                buf = ""
+                prefix_started = False
+
+    threading.Thread(target=_stream_output, daemon=True).start()
+
+    def _shutdown() -> None:
+        if proc.poll() is None:
+            print(f"[server] stopping pid={proc.pid}", flush=True)
+            proc.send_signal(signal.SIGINT)
+            try:
+                proc.wait(timeout=15)
+            except subprocess.TimeoutExpired:
+                proc.kill()
+                proc.wait(timeout=5)
+
+    atexit.register(_shutdown)
+
+    deadline = time.monotonic() + config.serve_ready_timeout_s
+    while time.monotonic() < deadline:
+        if proc.poll() is not None:
+            raise RuntimeError(
+                f"[server] inference server exited unexpectedly with rc={proc.returncode}. "
+                f"See [server] log lines above for the cause."
+            )
+        if ready_event.wait(timeout=2):
+            return api_base
+    proc.terminate()
+    raise RuntimeError(f"[server] did not become ready within {config.serve_ready_timeout_s}s")
+
+
+def _to_openai_messages(
+    messages: Sequence[dict[str, Any]],
+) -> tuple[list[dict[str, Any]], dict[str, Any]]:
+    """Convert internal messages to OpenAI chat format.
+
+    Returns ``(api_messages, mm_kwargs)``. Multimodal-processor kwargs
+    (``fps`` from ``video_url`` blocks) are extracted out so the caller
+    can pass them via ``extra_body.mm_processor_kwargs`` rather than
+    inside the content blocks (which transformers serve rejects).
+
+    File-URL video blocks are inlined as base64 data URLs.
+    """
+    out_messages: list[dict[str, Any]] = []
+    mm_kwargs: dict[str, Any] = {}
+    for message in messages:
+        content = message.get("content")
+        if not isinstance(content, list):
+            out_messages.append({"role": message["role"], "content": content})
+            continue
+        out_blocks: list[dict[str, Any]] = []
+        for block in content:
+            block_type = block.get("type") if isinstance(block, dict) else None
+            if block_type == "text":
+                out_blocks.append({"type": "text", "text": block.get("text", "")})
+            elif block_type == "image":
+                out_blocks.append(
+                    {"type": "image_url", "image_url": {"url": _pil_to_data_url(block["image"])}}
+                )
+            elif block_type == "video":
+                frames = block.get("video", [])
+                for img in frames:
+                    out_blocks.append({"type": "image_url", "image_url": {"url": _pil_to_data_url(img)}})
+            elif block_type == "video_url":
+                video_url = dict(block["video_url"])
+                url = video_url.get("url", "")
+                if url.startswith("file://"):
+                    video_url["url"] = _file_to_data_url(url[len("file://") :])
+                out_blocks.append({"type": "video_url", "video_url": video_url})
+                fps = block.get("fps")
+                if fps is not None:
+                    mm_kwargs["fps"] = fps
+            else:
+                out_blocks.append(block)
+        out_messages.append({"role": message["role"], "content": out_blocks})
+    return out_messages, mm_kwargs
+
+
+def _file_to_data_url(path: str) -> str:
+    """Read a local video file and return a base64 ``data:video/mp4`` URL."""
+    with open(path, "rb") as f:
+        b64 = base64.b64encode(f.read()).decode("ascii")
+    return f"data:video/mp4;base64,{b64}"
+
+
+def _pil_to_data_url(image: Any) -> str:
+    """Encode a PIL.Image as a base64 data URL."""
+    buf = io.BytesIO()
+    image.save(buf, format="PNG")
+    b64 = base64.b64encode(buf.getvalue()).decode("ascii")
+    return f"data:image/png;base64,{b64}"

src/lerobot/annotations/steerable_pipeline/writer.py

@@ -0,0 +1,341 @@
+#!/usr/bin/env python
+
+# Copyright 2026 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""Final parquet rewrite.
+
+For every episode the writer:
+
+1. reads the staged module outputs,
+2. partitions them into a persistent slice (PERSISTENT_STYLES) and an event
+   slice (EVENT_ONLY_STYLES + style=None tool-call atoms),
+3. sorts each slice deterministically,
+4. broadcasts the persistent slice across every frame in the episode,
+5. for each frame, materializes the sublist of event rows whose timestamp
+   exactly equals that frame's timestamp,
+6. drops the legacy ``subtask_index`` column,
+7. writes the parquet shard back in place.
+
+The writer does NOT add a dataset-level ``tools`` column. Tool *calls* are
+emitted per-row via the existing ``tool_calls`` field on the v3.1 row
+struct for every speech atom. The tool *schema* (the description
+of the ``say`` function and its parameters) is a fixed code constant —
+``SAY_TOOL_SCHEMA`` below — and downstream chat-template consumers import
+it directly rather than reading a redundant per-row column.
+
+Invariants enforced here (and re-checked by the validator):
+
+- per-episode persistent slice is byte-identical across every frame;
+- ``language_events`` rows on a frame all have ``timestamp == frame_ts``
+  (timestamps come straight from the source parquet — never recomputed);
+- every row passes ``column_for_style(style)``.
+"""
+
+from __future__ import annotations
+
+import logging
+from collections import defaultdict
+from collections.abc import Sequence
+from dataclasses import dataclass
+from pathlib import Path
+from typing import Any
+
+import pyarrow as pa
+import pyarrow.parquet as pq
+
+from lerobot.datasets.language import (
+    EVENT_ONLY_STYLES,
+    LANGUAGE_EVENTS,
+    LANGUAGE_PERSISTENT,
+    PERSISTENT_STYLES,
+    column_for_style,
+    validate_camera_field,
+)
+
+from .reader import EpisodeRecord
+from .staging import EpisodeStaging
+
+logger = logging.getLogger(__name__)
+
+
+# Tool schema constants live in lerobot.datasets.language — single
+# source of truth. Re-exported here so existing imports
+# (``from lerobot.annotations.steerable_pipeline.writer import SAY_TOOL_SCHEMA``)
+# keep working.
+from lerobot.datasets.language import DEFAULT_TOOLS, SAY_TOOL_SCHEMA  # noqa: F401, E402
+
+
+def _row_persistent_sort_key(row: dict[str, Any]) -> tuple:
+    return (float(row["timestamp"]), row.get("style") or "", row.get("role") or "")
+
+
+def _row_event_sort_key(row: dict[str, Any]) -> tuple:
+    # events are bucketed per-frame, but within a frame we still want determinism
+    return (
+        row.get("style") or "",
+        row.get("role") or "",
+        row.get("camera") or "",
+    )
+
+
+def _normalize_row(row: dict[str, Any], style: str | None, *, with_timestamp: bool) -> dict[str, Any]:
+    """Coerce a staged row into the language-column struct shape.
+
+    Key order matches ``PERSISTENT_ROW_FIELDS`` / ``EVENT_ROW_FIELDS`` — the
+    writer infers the parquet struct schema from insertion order, so
+    ``timestamp`` (persistent rows only) sits between ``style`` and ``camera``.
+    """
+    camera = row.get("camera")
+    validate_camera_field(style, camera)
+    out: dict[str, Any] = {
+        "role": str(row["role"]),
+        "content": None if row.get("content") is None else str(row["content"]),
+        "style": style,
+    }
+    if with_timestamp:
+        out["timestamp"] = float(row["timestamp"])
+    out["camera"] = None if camera is None else str(camera)
+    out["tool_calls"] = _normalize_tool_calls(row.get("tool_calls"))
+    return out
+
+
+def _normalize_persistent_row(row: dict[str, Any]) -> dict[str, Any]:
+    """Coerce a staged row into the persistent column's struct shape."""
+    style = row.get("style")
+    if style not in PERSISTENT_STYLES:
+        raise ValueError(
+            f"persistent slice contains row with non-persistent style {style!r}; "
+            "row would be misrouted under column_for_style()"
+        )
+    if "timestamp" not in row:
+        raise ValueError(f"persistent row missing timestamp: {row!r}")
+    if "role" not in row:
+        # Friendly error from the writer instead of a raw KeyError below;
+        # the validator doesn't check ``role`` yet.
+        raise ValueError(f"persistent row missing role: {row!r}")
+    return _normalize_row(row, style, with_timestamp=True)
+
+
+def _normalize_event_row(row: dict[str, Any]) -> dict[str, Any]:
+    """Coerce a staged row into the event column's struct shape (no timestamp)."""
+    style = row.get("style")
+    if style is not None and style not in EVENT_ONLY_STYLES:
+        raise ValueError(
+            f"event slice contains row with style {style!r}; expected None or one of {EVENT_ONLY_STYLES}"
+        )
+    if column_for_style(style) != LANGUAGE_EVENTS:
+        raise ValueError(f"event row with style {style!r} would not route to language_events")
+    if "role" not in row:
+        raise ValueError(f"event row missing role: {row!r}")
+    return _normalize_row(row, style, with_timestamp=False)
+
+
+def _normalize_tool_calls(value: Any) -> list[Any] | None:
+    if value is None:
+        return None
+    if not isinstance(value, list):
+        raise ValueError(f"tool_calls must be a list or None, got {type(value).__name__}")
+    return list(value)
+
+
+def _validate_atom_invariants(row: dict[str, Any]) -> None:
+    """At-least-one of content/tool_calls; style=None implies tool_calls."""
+    has_content = row.get("content") is not None
+    has_tools = row.get("tool_calls") is not None
+    if not (has_content or has_tools):
+        raise ValueError(f"row has neither content nor tool_calls: {row!r}")
+    if row.get("style") is None and not has_tools:
+        raise ValueError(f"style=None requires tool_calls: {row!r}")
+
+
+def _validate_speech_atom(row: dict[str, Any]) -> None:
+    """Speech atoms: role=assistant, style=None, content=None, say tool call."""
+    if row.get("style") is not None:
+        return  # not a speech atom
+    if row.get("role") != "assistant":
+        raise ValueError(f"speech atom must have role=assistant: {row!r}")
+    if row.get("content") is not None:
+        raise ValueError(f"speech atom must have content=null: {row!r}")
+    tool_calls = row.get("tool_calls")
+    if not tool_calls or not isinstance(tool_calls, list):
+        raise ValueError(f"speech atom must have non-empty tool_calls list: {row!r}")
+    first = tool_calls[0]
+    if not isinstance(first, dict):
+        raise ValueError(f"speech atom tool_calls[0] must be a dict: {row!r}")
+    if first.get("type") != "function":
+        raise ValueError(f"speech atom tool_calls[0].type must be 'function': {row!r}")
+    fn = first.get("function") or {}
+    if fn.get("name") != "say":
+        raise ValueError(f"speech atom tool_calls[0].function.name must be 'say': {row!r}")
+    args = fn.get("arguments") or {}
+    if not isinstance(args, dict) or "text" not in args or not isinstance(args["text"], str):
+        raise ValueError(f"speech atom must carry 'text' string in arguments: {row!r}")
+
+
+@dataclass
+class LanguageColumnsWriter:
+    """Rewrite ``data/chunk-*/file-*.parquet`` with the two language columns."""
+
+    drop_existing_subtask_index: bool = True
+
+    def write_all(
+        self,
+        records: Sequence[EpisodeRecord],
+        staging_dir: Path,
+        root: Path,
+    ) -> list[Path]:
+        episodes_by_path: dict[Path, list[EpisodeRecord]] = defaultdict(list)
+        for record in records:
+            episodes_by_path[record.data_path].append(record)
+
+        written: list[Path] = []
+        for path, eps in episodes_by_path.items():
+            self._rewrite_one(path, eps, staging_dir, root)
+            written.append(path)
+        return written
+
+    def _rewrite_one(
+        self,
+        path: Path,
+        episodes: Sequence[EpisodeRecord],
+        staging_dir: Path,
+        root: Path,
+    ) -> None:
+        table = pq.read_table(path)
+        n_rows = table.num_rows
+
+        # Ensure we cover every episode in the file. Episodes that don't have
+        # staging artifacts are passed through with empty annotation lists —
+        # this keeps the writer idempotent and safe for partial reruns.
+        staged_per_ep: dict[int, dict[str, list[dict[str, Any]]]] = {}
+        for record in episodes:
+            staging = EpisodeStaging(staging_dir, record.episode_index)
+            staged_per_ep[record.episode_index] = staging.read_all()
+
+        persistent_by_ep: dict[int, list[dict[str, Any]]] = {}
+        events_by_ep_ts: dict[int, dict[float, list[dict[str, Any]]]] = {}
+
+        for ep_index, ep_staged in staged_per_ep.items():
+            persistent_rows: list[dict[str, Any]] = []
+            event_rows: list[dict[str, Any]] = []  # carry timestamp until bucketed
+            for _module_name, rows in ep_staged.items():
+                for row in rows:
+                    style = row.get("style")
+                    if column_for_style(style) == LANGUAGE_PERSISTENT:
+                        persistent_rows.append(row)
+                    else:
+                        event_rows.append(row)
+
+            persistent_rows.sort(key=_row_persistent_sort_key)
+            normalized_persistent = []
+            for r in persistent_rows:
+                _validate_atom_invariants(r)
+                _validate_speech_atom(r)
+                normalized_persistent.append(_normalize_persistent_row(r))
+            persistent_by_ep[ep_index] = normalized_persistent
+
+            buckets: dict[float, list[dict[str, Any]]] = defaultdict(list)
+            for r in event_rows:
+                _validate_atom_invariants(r)
+                _validate_speech_atom(r)
+                ts = float(r["timestamp"])
+                buckets[ts].append(_normalize_event_row(r))
+            for ts in list(buckets.keys()):
+                buckets[ts].sort(key=_row_event_sort_key)
+            events_by_ep_ts[ep_index] = buckets
+
+        episode_col = (
+            table.column("episode_index").to_pylist() if "episode_index" in table.column_names else None
+        )
+        ts_col = table.column("timestamp").to_pylist() if "timestamp" in table.column_names else None
+        if episode_col is None or ts_col is None:
+            raise ValueError(f"{path} is missing 'episode_index' or 'timestamp' — required by the writer.")
+
+        per_row_persistent: list[list[dict[str, Any]]] = []
+        per_row_events: list[list[dict[str, Any]]] = []
+        for i in range(n_rows):
+            ep = episode_col[i]
+            ts = float(ts_col[i])
+            per_row_persistent.append(persistent_by_ep.get(ep, []))
+            buckets = events_by_ep_ts.get(ep, {})
+            per_row_events.append(buckets.get(ts, []))
+
+        new_table = self._materialize_table(
+            table, per_row_persistent, per_row_events, drop_old=self.drop_existing_subtask_index
+        )
+        # Atomic replace: write to a sibling tmp path and rename so a crash
+        # mid-write can't leave a half-written shard that ``pq.read_table``
+        # would then fail to open. ``Path.replace`` is atomic on POSIX +
+        # Windows when source and target sit on the same filesystem.
+        tmp_path = path.with_suffix(path.suffix + ".tmp")
+        pq.write_table(new_table, tmp_path)
+        tmp_path.replace(path)
+
+    def _materialize_table(
+        self,
+        table: pa.Table,
+        persistent: list[list[dict[str, Any]]],
+        events: list[list[dict[str, Any]]],
+        *,
+        drop_old: bool,
+    ) -> pa.Table:
+        cols = []
+        names = []
+        for name in table.column_names:
+            if drop_old and name == "subtask_index":
+                continue
+            if name in (LANGUAGE_PERSISTENT, LANGUAGE_EVENTS):
+                continue  # we'll re-add canonical versions
+            # Strip any legacy ``tools`` column previously emitted by older
+            # writers — the schema no longer uses it (constant lives in
+            # SAY_TOOL_SCHEMA / DEFAULT_TOOLS).
+            if name == "tools":
+                continue
+            cols.append(table.column(name))
+            names.append(name)
+
+        # We let pyarrow infer struct/list schema rather than passing the
+        # canonical type from `lerobot.datasets.language` directly: that type
+        # uses `pa.json_()` for the `tool_calls` element type, which
+        # `pa.array(..., type=...)` cannot materialize from Python lists on
+        # current pyarrow versions. The inferred schema round-trips through
+        # parquet and `LeRobotDataset` correctly — `tests/datasets/test_language.py`
+        # exercises the same flow.
+        persistent_arr = pa.array(persistent)
+        events_arr = pa.array(events)
+
+        cols.extend([persistent_arr, events_arr])
+        names.extend([LANGUAGE_PERSISTENT, LANGUAGE_EVENTS])
+
+        return pa.Table.from_arrays(cols, names=names)
+
+
+def speech_atom(timestamp: float, text: str) -> dict[str, Any]:
+    """Build a canonical speech tool-call atom for the events column."""
+    return {
+        "role": "assistant",
+        "content": None,
+        "style": None,
+        "timestamp": float(timestamp),
+        "camera": None,
+        "tool_calls": [
+            {
+                "type": "function",
+                "function": {
+                    "name": "say",
+                    "arguments": {"text": text},
+                },
+            }
+        ],
+    }

src/lerobot/common/control_utils.py

@@ -18,6 +18,7 @@ from __future__ import annotations
 # Utilities
 ########################################################################################
 import logging
+import time
 import traceback
 from contextlib import nullcontext
 from copy import copy
@@ -243,3 +244,72 @@ def sanity_check_dataset_robot_compatibility(
         raise ValueError(
             "Dataset metadata compatibility check failed with mismatches:\n" + "\n".join(mismatches)
         )
+
+
+########################################################################################
+# Teleoperator smooth handover helpers
+# NOTE(Maxime): These functions use minimal type hints to maintain compatibility with utils
+# being a root module.
+########################################################################################
+
+
+def teleop_supports_feedback(teleop) -> bool:
+    """Return True when the teleop can receive position feedback (is actuated).
+
+    Actuated teleops (e.g. SO-101, OpenArmMini) have non-empty ``feedback_features``
+    and expose ``enable_torque`` / ``disable_torque`` motor-control methods.
+
+    TODO(Maxime): See if it is possible to unify this interface across teleops instead of duck-typing.
+    """
+    return (
+        bool(teleop.feedback_features)
+        and hasattr(teleop, "disable_torque")
+        and hasattr(teleop, "enable_torque")
+    )
+
+
+def teleop_smooth_move_to(teleop, target_pos: dict, duration_s: float = 2.0, fps: int = 30) -> None:
+    """Smoothly move an actuated teleop to ``target_pos`` via linear interpolation.
+
+    Requires the teleoperator to support feedback (i.e. have non-empty
+    ``feedback_features`` and implement ``disable_torque`` / ``enable_torque``).
+
+    ``target_pos`` is expected to be in the teleop's action/feedback key space.
+    For homogeneous setups (e.g. SO-101 leader + SO-101 follower) this matches
+    the robot action key space directly.
+
+    TODO(Maxime): This blocks up to ``duration_s`` seconds; during this time the
+    follower robot does not receive new actions, which could be an issue on LeKiwi.
+    """
+    teleop.enable_torque()
+    current = teleop.get_action()
+    steps = max(int(duration_s * fps), 1)
+
+    for step in range(steps + 1):
+        t = step / steps
+        interp = {
+            k: current[k] * (1 - t) + target_pos[k] * t if k in target_pos else current[k] for k in current
+        }
+        teleop.send_feedback(interp)
+        time.sleep(1 / fps)
+
+
+def follower_smooth_move_to(
+    robot, current: dict, target: dict, duration_s: float = 1.0, fps: int = 30
+) -> None:
+    """Smoothly move the follower robot from ``current`` to ``target`` action.
+
+    Used when the teleop is non-actuated: instead of driving the leader arm to
+    the follower, the follower is brought to the teleop's current pose so the
+    robot meets the operator's hand rather than jumping to it on the first frame.
+
+    Both ``current`` and ``target`` must be in the robot action key space
+    (i.e. the output of ``robot_action_processor``).
+    """
+    steps = max(int(duration_s * fps), 1)
+
+    for step in range(steps + 1):
+        t = step / steps
+        interp = {k: current[k] * (1 - t) + target[k] * t if k in target else current[k] for k in current}
+        robot.send_action(interp)
+        time.sleep(1 / fps)

src/lerobot/common/wandb_utils.py

@@ -205,3 +205,149 @@ class WandBLogger:
 
         wandb_video = self._wandb.Video(video_path, fps=self.env_fps, format="mp4")
         self._wandb.log({f"{mode}/video": wandb_video}, step=step)
+
+    def log_training_examples(
+        self,
+        batch: dict,
+        step: int,
+        *,
+        camera_keys: list[str],
+        n_samples: int = 4,
+        policy=None,
+        predict_actions: bool = False,
+        mode: str = "train",
+    ) -> None:
+        """Push a ``wandb.Table`` of training-example rows for the current batch.
+
+        Each row is one batch element with:
+          * one ``wandb.Image`` column per camera in ``camera_keys`` (CHW or
+            HWC, uint8 or float in [0,1] — auto-detected),
+          * any text fields present in the batch (``task`` / ``subtask`` /
+            ``memory`` / ``instruction``),
+          * ground-truth action first/last frame (the action chunk's
+            endpoints — gives a quick sense of trajectory direction),
+          * if ``predict_actions=True`` and ``policy`` is supplied, the model's
+            ``predict_action_chunk`` first/last frame alongside.
+
+        This is opt-in via ``--wandb.log_examples_freq=N`` on the CLI; the
+        training loop calls it once every N steps. Cheap to keep on: with
+        N=4 samples and 3 cameras you upload 12 small PNGs per dump and (if
+        enabled) run one extra inference forward pass.
+        """
+        import logging  # noqa: PLC0415
+        import numpy as np  # noqa: PLC0415
+        import torch  # noqa: PLC0415
+
+        if mode not in {"train", "eval"}:
+            raise ValueError(mode)
+
+        # Batch size — first tensor-like value wins.
+        bsz = next(
+            (int(v.shape[0]) for v in batch.values() if hasattr(v, "shape") and v.ndim > 0),
+            None,
+        )
+        if not bsz:
+            return
+        n = min(int(n_samples), bsz)
+
+        # Optional predicted-action forward pass on the first n samples.
+        pred_actions: np.ndarray | None = None
+        if predict_actions and policy is not None:
+            was_training = policy.training
+            try:
+                policy.eval()
+                sub_batch = {}
+                for k, v in batch.items():
+                    if isinstance(v, torch.Tensor):
+                        sub_batch[k] = v[:n]
+                    elif isinstance(v, (list, tuple)):
+                        sub_batch[k] = list(v[:n])
+                    else:
+                        sub_batch[k] = v
+                with torch.no_grad():
+                    pred = policy.predict_action_chunk(sub_batch)
+                pred_actions = pred.detach().cpu().float().numpy()
+            except Exception as exc:  # noqa: BLE001
+                logging.warning(
+                    "log_training_examples: predict_action_chunk failed (%s) — "
+                    "skipping predicted-action columns",
+                    exc,
+                )
+                pred_actions = None
+            finally:
+                if was_training:
+                    policy.train()
+
+        present_cameras = [c for c in camera_keys if c in batch]
+        text_keys = [k for k in ("task", "subtask", "memory", "instruction") if k in batch]
+
+        columns = ["sample"]
+        columns.extend(c.removeprefix("observation.images.") or c for c in present_cameras)
+        columns.extend(text_keys)
+        columns.append("gt_action_first")
+        columns.append("gt_action_last")
+        if pred_actions is not None:
+            columns.append("pred_action_first")
+            columns.append("pred_action_last")
+
+        table = self._wandb.Table(columns=columns)
+
+        def _to_uint8_hwc(t: torch.Tensor) -> np.ndarray:
+            # Strip an outer time dim if present: (T, C, H, W) -> first frame.
+            if t.ndim == 4:
+                t = t[0]
+            # CHW -> HWC.
+            if t.ndim == 3 and t.shape[0] in (1, 3, 4) and t.shape[-1] not in (1, 3, 4):
+                t = t.permute(1, 2, 0)
+            arr = t.detach().cpu().float().numpy()
+            if arr.size and float(arr.max()) <= 1.5:
+                arr = arr * 255.0
+            return np.clip(arr, 0, 255).astype(np.uint8)
+
+        def _action_endpoints(a: torch.Tensor) -> tuple[str, str]:
+            arr = a.detach().cpu().float().numpy()
+            if arr.ndim == 2:  # (T, D)
+                return (
+                    str(np.round(arr[0], 3).tolist()),
+                    str(np.round(arr[-1], 3).tolist()),
+                )
+            if arr.ndim == 1:
+                rounded = np.round(arr, 3).tolist()
+                return (str(rounded), str(rounded))
+            return (str(arr.tolist()), str(arr.tolist()))
+
+        for i in range(n):
+            row: list = [i]
+            for cam in present_cameras:
+                try:
+                    row.append(self._wandb.Image(_to_uint8_hwc(batch[cam][i])))
+                except Exception as exc:  # noqa: BLE001
+                    logging.warning(
+                        "log_training_examples: camera %s sample %d failed (%s)",
+                        cam,
+                        i,
+                        exc,
+                    )
+                    row.append(None)
+            for tk in text_keys:
+                v = batch[tk]
+                if isinstance(v, (list, tuple)):
+                    row.append(str(v[i]) if i < len(v) else "")
+                else:
+                    row.append(str(v))
+            action = batch.get("action")
+            if isinstance(action, torch.Tensor) and action.ndim >= 1:
+                first, last = _action_endpoints(action[i])
+                row.append(first)
+                row.append(last)
+            else:
+                row.append("")
+                row.append("")
+            if pred_actions is not None:
+                p = torch.from_numpy(pred_actions[i])
+                pfirst, plast = _action_endpoints(p)
+                row.append(pfirst)
+                row.append(plast)
+            table.add_data(*row)
+
+        self._wandb.log({f"{mode}/examples": table}, step=step)

src/lerobot/configs/dataset.py

@@ -41,8 +41,8 @@ class DatasetRecordConfig:
     video: bool = True
     # Upload dataset to Hugging Face hub.
     push_to_hub: bool = True
-    # Upload on private repository on the Hugging Face hub.
-    private: bool = False
+    # If True, upload as private; if None, defer to the org default on the Hub (only affects orgs).
+    private: bool | None = None
     # Add tags to your dataset on the hub.
     tags: list[str] | None = None
     # Number of subprocesses handling the saving of frames as PNG. Set to 0 to use threads only;

src/lerobot/configs/default.py

@@ -62,6 +62,72 @@ class WandBConfig:
     run_id: str | None = None
     mode: str | None = None  # Allowed values: 'online', 'offline' 'disabled'. Defaults to 'online'
     add_tags: bool = True  # If True, save configuration as tags in the WandB run.
+    # Periodic training-example dump (independent of ``log_freq``). When > 0,
+    # every ``log_examples_freq`` steps the trainer pushes a ``wandb.Table``
+    # with one row per sampled batch element containing each camera view
+    # (rendered as ``wandb.Image``), any text fields present in the batch
+    # (``task`` / ``subtask`` / ``memory`` / ``instruction``), and the
+    # ground-truth action chunk's first + last frames. Defaults to 5000 — set
+    # to 0 to disable. Only fires when ``enable=True``, so runs without wandb
+    # are unaffected.
+    log_examples_freq: int = 5000
+    # Number of batch elements to include in each example dump.
+    log_examples_n: int = 4
+    # If True (default), also run ``policy.predict_action_chunk`` on the logged
+    # samples (in eval mode, no_grad) and add predicted vs ground-truth action
+    # columns to the table. Costs one extra forward pass per dump — negligible
+    # at the 5k-step default cadence. Set to ``False`` if your policy doesn't
+    # implement ``predict_action_chunk`` or you want to skip the extra forward.
+    log_examples_predict_actions: bool = True
+
+
+@dataclass
+class EMAConfig:
+    """Exponential Moving Average of trainable policy parameters.
+
+    Diffusion / flow-matching policies (Diffusion Policy, π0/π0.5,
+    pi052) benefit substantially from averaging late-training
+    parameter oscillations — see Chi et al. 2023 §V.D. The official
+    JAX openpi trainer ships EMA with ``ema_decay=0.99`` (default) and
+    ``0.999`` for its pi05_libero config; the openpi PyTorch port
+    explicitly lists EMA as unsupported, and LeRobot main inherited
+    that gap. Enabling this flag plugs ema-pytorch
+    (https://github.com/lucidrains/ema-pytorch) into the LeRobot
+    training loop with a shadow ``nn.Module`` clone of the policy.
+
+    Cost: 1× model params in fp32 shadow (~13 GB for pi052's 3.3B
+    params) + one elementwise update per training step (~1% step time).
+
+    Off by default (opt-in): EMA is only beneficial for flow-matching /
+    diffusion policies (pi0/pi05/pi052), and the fp32 shadow copy is pure
+    overhead for other policies (e.g. VLA-JEPA). Set ``--ema.enable=true``
+    to turn it on (the pi05/pi052 training recipes do this). openpi (JAX)
+    ships EMA on for every config; enable it explicitly to match that.
+    """
+
+    enable: bool = False
+    # Target EMA decay β in θ_ema ← β·θ_ema + (1-β)·θ_live (passed to
+    # ema-pytorch as ``beta``).
+    #   0.999  — last ~1000 steps; pi05_libero default in openpi
+    #   0.99   — last ~100 steps; openpi top-level default
+    #   0.75   — very fast EMA (Diffusion Policy original setting)
+    #   0.9999 — very slow EMA (long classification runs)
+    decay: float = 0.99
+    # Skip the first N calls to ``ema.update()``; during this window
+    # the shadow is just a hard copy of the live weights (no averaging).
+    # Lets early-training rapid changes settle before averaging begins.
+    # Maps to ema-pytorch's ``update_after_step`` (NOT a smooth decay
+    # ramp like older lerobot EMA implementations).
+    warmup_steps: int = 0
+    # When True, the periodic eval block uses the EMA shadow model
+    # directly (``ema.ema_model``) instead of the live policy. Standard
+    # practice for diffusion-style policies — eval scores are usually
+    # 1–3% higher than the live policy at the same step.
+    use_for_eval: bool = True
+    # When True, the periodic wandb training-example dump uses the EMA
+    # shadow for the optional predicted-action columns (so what you see
+    # in W&B matches eval behavior).
+    use_for_wandb_examples: bool = True
 
 
 @dataclass

src/lerobot/configs/parser.py

@@ -255,8 +255,7 @@ def extract_path_fields_from_config(config_path: str, path_fields: list[str]) ->
             remaining = config_data[field]
             if remaining:
                 _config_yaml_overrides[field] = _flatten_to_cli_args(remaining)
-            else:
-                del config_data[field]
+            del config_data[field]
             modified = True
 
     if not modified:
@@ -311,7 +310,13 @@ def wrap(config_path: Path | None = None) -> Callable[[F], F]:
                     cli_args = filter_arg("config_path", cli_args)
                     cfg = argtype.from_pretrained(config_path_cli, cli_args=cli_args)
                 else:
-                    cfg = draccus.parse(config_class=argtype, config_path=config_path, args=cli_args)
+                    if config_path_cli:
+                        cli_args = filter_arg("config_path", cli_args)
+                    cfg = draccus.parse(
+                        config_class=argtype,
+                        config_path=config_path_cli or config_path,
+                        args=cli_args,
+                    )
             response = fn(cfg, *args, **kwargs)
             return response
 

src/lerobot/configs/recipe.py

@@ -147,7 +147,16 @@ class TrainingRecipe:
         return cls.from_dict(data)
 
     def _validate_message_recipe(self) -> None:
-        """Ensure every templated binding is known and at least one turn is a target."""
+        """Ensure every templated binding is known and the recipe supervises something.
+
+        A recipe is valid if it has at least one of:
+
+        * a ``target: true`` assistant turn (drives text-CE supervision), or
+        * a ``stream: low_level`` turn (drives flow / action supervision via
+          ``predict_actions=True``, even when no assistant turn is targeted —
+          e.g. π0.5-style ``low_level_execution`` where the action expert
+          conditions on a user-only ``${subtask}`` prompt).
+        """
         assert self.messages is not None
         known_bindings = set(DEFAULT_BINDINGS) | set(self.bindings or {}) | {"task"}
 
@@ -156,8 +165,14 @@ class TrainingRecipe:
             if missing:
                 raise ValueError(f"MessageTurn references unknown binding(s): {sorted(missing)}")
 
-        if not any(turn.target for turn in self.messages):
-            raise ValueError("Message recipes must contain at least one target turn.")
+        has_target = any(turn.target for turn in self.messages)
+        has_low_level = any(turn.stream == "low_level" for turn in self.messages)
+        if not (has_target or has_low_level):
+            raise ValueError(
+                "Message recipes must contain at least one supervised turn — "
+                "either ``target: true`` (text CE) or ``stream: low_level`` "
+                "(flow/action loss)."
+            )
 
     def _validate_blend_recipe(self) -> None:
         """Ensure each blend component is a non-empty, weighted message recipe."""

src/lerobot/configs/recipes/subtask_mem.yaml

@@ -0,0 +1,68 @@
+# subtask_mem_vqa_speech — Hi-Robot blend + memory + spoken responses.
+#
+# Superset of subtasks_vqa.yaml. Keeps the core subtask + action + VQA
+# training, and adds two text-supervised tasks:
+#
+#   high_level_subtask         — predict the subtask from the task.
+#   low_level_execution        — flow loss with [images, subtask, state].
+#   memory_update              — compress progress into a memory note.
+#   user_interjection_response — reply to a user interjection with a
+#                                spoken `say` tool call (no plan, no
+#                                subtask text — just the spoken reply).
+#   ask_vqa_{top,wrist}        — camera-grounded VQA.
+#
+# Plan is intentionally left out — memory is the only persistent
+# high-level state here, keeping the prompt short.
+#
+# Requires the dataset to carry `memory`, `interjection` and `say`-tool
+# annotations (the annotation pipeline's memory + interjection modules)
+# in addition to `subtask` and `vqa`. Sub-recipes whose `if_present`
+# bindings are missing simply don't render for that sample, so a
+# dataset without interjections still trains the rest of the blend.
+#
+# Tool-call note: the `say` tool call on the interjection-response turn
+# is flattened to a `<say>...</say>` text marker by the tokenizer step
+# (`_flatten_say_tool_calls`) so the LM head learns to emit exactly the
+# marker the runtime parses back (`_split_plan_and_say`).
+
+blend:
+
+  high_level_subtask:
+    weight: 0.30
+    messages:
+      - {role: user, content: "${task}", stream: high_level}
+      - {role: assistant, content: "${subtask}", stream: high_level, target: true, if_present: subtask}
+
+  low_level_execution:
+    weight: 0.55
+    messages:
+      # The action expert is conditioned on the SUBTASK — at inference
+      # `HighLevelSubtaskFwd` generates it via the LM head and feeds it
+      # here. `stream: low_level` flips `predict_actions=True` so the
+      # flow loss fires; no text-CE target (subtask prediction is owned
+      # by `high_level_subtask`).
+      - {role: user, content: "${subtask}", stream: low_level, if_present: subtask}
+
+  memory_update:
+    # At inference, `MemoryUpdateFwd` is triggered only on
+    # `subtask_change` events (sparse). Training densely with
+    # `active_at` — i.e. on every frame inside a subtask interval,
+    # not just the boundary frame — supervises the same
+    # (prior_memory, completed_subtask) → current_memory mapping
+    # against varied observations within the interval. The model
+    # learns a stateless transformation; the *when* to emit lives in
+    # the inference trigger, not the model. Annotations only exist
+    # for ~1% of frames as boundary events, so `emitted_at` would
+    # waste 99% of the blend draws (and silently leak them into a
+    # task-conditioned fallback); `active_at` lifts the renderable
+    # rate to ~87% on this dataset.
+    weight: 0.15
+    bindings:
+      prior_memory: "nth_prev(style=memory, offset=1)"
+      current_memory: "active_at(t, style=memory)"
+      completed_subtask: "nth_prev(style=subtask, offset=1)"
+    messages:
+      - {role: user, content: "${task}", stream: high_level}
+      - {role: assistant, content: "Previous memory: ${prior_memory}", stream: high_level, if_present: prior_memory}
+      - {role: user, content: "Completed subtask: ${completed_subtask}", stream: high_level, if_present: completed_subtask}
+      - {role: assistant, content: "${current_memory}", stream: high_level, target: true, if_present: current_memory}

src/lerobot/configs/recipes/subtask_mem_vqa_robocasa.yaml

@@ -0,0 +1,99 @@
+# subtask_mem_vqa_robocasa — Hi-Robot blend tuned for RoboCasa cameras.
+#
+# Same supervision as ``subtask_mem.yaml`` (subtask + memory) plus
+# camera-grounded VQA across the three RoboCasa camera keys produced
+# by ``slurm_build_robocasa_composite_seen.py``:
+#
+#   observation.images.robot0_agentview_left   (left scene view)
+#   observation.images.robot0_agentview_right  (right scene view)
+#   observation.images.robot0_eye_in_hand      (wrist)
+#
+# The annotation pipeline (``examples/annotations/run_hf_job.py``) emits
+# VQA per camera, so each anchor frame produces three (user, assistant)
+# rows tagged with their source camera. Each VQA sub-recipe consumes
+# the rows for one camera via ``camera=...`` resolver bindings.
+#
+# Spatial VQA targets (bbox / point) are rewritten from JSON to
+# PaliGemma ``<locDDDD>`` tokens by ``_messages_vqa_to_loc`` —
+# ``register_paligemma_loc_tokens`` already collapses them to single
+# detection-vocab ids so the LM head learns the pretrained pointing /
+# detection prior, not a 7-piece BPE salad.
+#
+# Interjections / spoken responses are intentionally absent — the
+# annotation job runs with ``--interjections.enabled=false``.
+
+blend:
+
+  high_level_subtask:
+    weight: 0.25
+    messages:
+      - {role: user, content: "${task}", stream: high_level}
+      - {role: assistant, content: "${subtask}", stream: high_level, target: true, if_present: subtask}
+
+  low_level_execution:
+    weight: 0.45
+    messages:
+      # Action expert is conditioned on the SUBTASK; at inference the
+      # high-level loop generates it via the LM head and feeds it here.
+      # ``stream: low_level`` flips ``predict_actions=True`` so the flow
+      # loss fires; subtask CE is owned by ``high_level_subtask``.
+      - {role: user, content: "${subtask}", stream: low_level, if_present: subtask}
+
+  memory_update:
+    # Trained densely with ``active_at`` — every frame inside a subtask
+    # interval — so the (prior_memory, completed_subtask) → current_memory
+    # mapping is supervised against varied observations. The *when* to
+    # emit lives in the inference trigger (subtask_change), not the
+    # model. See ``subtask_mem.yaml`` for the long version of this note.
+    weight: 0.15
+    bindings:
+      prior_memory: "nth_prev(style=memory, offset=1)"
+      current_memory: "active_at(t, style=memory)"
+      completed_subtask: "nth_prev(style=subtask, offset=1)"
+    messages:
+      - {role: user, content: "${task}", stream: high_level}
+      - {role: assistant, content: "Previous memory: ${prior_memory}", stream: high_level, if_present: prior_memory}
+      - {role: user, content: "Completed subtask: ${completed_subtask}", stream: high_level, if_present: completed_subtask}
+      - {role: assistant, content: "${current_memory}", stream: high_level, target: true, if_present: current_memory}
+
+  ask_vqa_agentview_left:
+    weight: 0.05
+    bindings:
+      vqa_query: "emitted_at(t, style=vqa, role=user, camera=observation.images.robot0_agentview_left)"
+      vqa: "emitted_at(t, style=vqa, role=assistant, camera=observation.images.robot0_agentview_left)"
+    messages:
+      - role: user
+        stream: high_level
+        if_present: vqa_query
+        content:
+          - {type: image, feature: observation.images.robot0_agentview_left}
+          - {type: text, text: "${vqa_query}"}
+      - {role: assistant, content: "${vqa}", stream: high_level, target: true, if_present: vqa}
+
+  ask_vqa_agentview_right:
+    weight: 0.05
+    bindings:
+      vqa_query: "emitted_at(t, style=vqa, role=user, camera=observation.images.robot0_agentview_right)"
+      vqa: "emitted_at(t, style=vqa, role=assistant, camera=observation.images.robot0_agentview_right)"
+    messages:
+      - role: user
+        stream: high_level
+        if_present: vqa_query
+        content:
+          - {type: image, feature: observation.images.robot0_agentview_right}
+          - {type: text, text: "${vqa_query}"}
+      - {role: assistant, content: "${vqa}", stream: high_level, target: true, if_present: vqa}
+
+  ask_vqa_wrist:
+    weight: 0.05
+    bindings:
+      vqa_query: "emitted_at(t, style=vqa, role=user, camera=observation.images.robot0_eye_in_hand)"
+      vqa: "emitted_at(t, style=vqa, role=assistant, camera=observation.images.robot0_eye_in_hand)"
+    messages:
+      - role: user
+        stream: high_level
+        if_present: vqa_query
+        content:
+          - {type: image, feature: observation.images.robot0_eye_in_hand}
+          - {type: text, text: "${vqa_query}"}
+      - {role: assistant, content: "${vqa}", stream: high_level, target: true, if_present: vqa}

src/lerobot/configs/recipes/subtask_mem_vqa_speech.yaml

@@ -0,0 +1,114 @@
+# subtask_mem_vqa_speech — Hi-Robot blend + memory + spoken responses.
+#
+# Superset of subtasks_vqa.yaml. Keeps the core subtask + action + VQA
+# training, and adds two text-supervised tasks:
+#
+#   high_level_subtask         — predict the subtask from the task.
+#   low_level_execution        — flow loss with [images, subtask, state].
+#   memory_update              — compress progress into a memory note.
+#   user_interjection_response — reply to a user interjection with a
+#                                spoken `say` tool call (no plan, no
+#                                subtask text — just the spoken reply).
+#   ask_vqa_{top,wrist}        — camera-grounded VQA.
+#
+# Plan is intentionally left out — memory is the only persistent
+# high-level state here, keeping the prompt short.
+#
+# Requires the dataset to carry `memory`, `interjection` and `say`-tool
+# annotations (the annotation pipeline's memory + interjection modules)
+# in addition to `subtask` and `vqa`. Sub-recipes whose `if_present`
+# bindings are missing simply don't render for that sample, so a
+# dataset without interjections still trains the rest of the blend.
+#
+# Tool-call note: the `say` tool call on the interjection-response turn
+# is flattened to a `<say>...</say>` text marker by the tokenizer step
+# (`_flatten_say_tool_calls`) so the LM head learns to emit exactly the
+# marker the runtime parses back (`_split_plan_and_say`).
+
+blend:
+
+  high_level_subtask:
+    weight: 0.25
+    messages:
+      - {role: user, content: "${task}", stream: high_level}
+      - {role: assistant, content: "${subtask}", stream: high_level, target: true, if_present: subtask}
+
+  low_level_execution:
+    weight: 0.40
+    messages:
+      # The action expert is conditioned on the SUBTASK — at inference
+      # `HighLevelSubtaskFwd` generates it via the LM head and feeds it
+      # here. `stream: low_level` flips `predict_actions=True` so the
+      # flow loss fires; no text-CE target (subtask prediction is owned
+      # by `high_level_subtask`).
+      - {role: user, content: "${subtask}", stream: low_level, if_present: subtask}
+
+  memory_update:
+    # At inference, `MemoryUpdateFwd` is triggered only on
+    # `subtask_change` events (sparse). Training densely with
+    # `active_at` — i.e. on every frame inside a subtask interval,
+    # not just the boundary frame — supervises the same
+    # (prior_memory, completed_subtask) → current_memory mapping
+    # against varied observations within the interval. The model
+    # learns a stateless transformation; the *when* to emit lives in
+    # the inference trigger, not the model. Annotations only exist
+    # for ~1% of frames as boundary events, so `emitted_at` would
+    # waste 99% of the blend draws (and silently leak them into the
+    # task-conditioned fallback); `active_at` lifts the renderable
+    # rate to ~87% on Hi-Robot-style datasets.
+    weight: 0.10
+    bindings:
+      prior_memory: "nth_prev(style=memory, offset=1)"
+      current_memory: "active_at(t, style=memory)"
+      completed_subtask: "nth_prev(style=subtask, offset=1)"
+    messages:
+      - {role: user, content: "${task}", stream: high_level}
+      - {role: assistant, content: "Previous memory: ${prior_memory}", stream: high_level, if_present: prior_memory}
+      - {role: user, content: "Completed subtask: ${completed_subtask}", stream: high_level, if_present: completed_subtask}
+      - {role: assistant, content: "${current_memory}", stream: high_level, target: true, if_present: current_memory}
+
+  user_interjection_response:
+    weight: 0.10
+    bindings:
+      interjection: "emitted_at(t, style=interjection)"
+      speech: "emitted_at(t, role=assistant, tool_name=say)"
+    messages:
+      - {role: user, content: "${task}", stream: high_level}
+      - {role: user, content: "${interjection}", stream: high_level, if_present: interjection}
+      # Spoken reply only: the assistant turn carries no text content,
+      # just a `say` tool call (`tool_calls_from: speech`). The chat
+      # tokenizer flattens it to a `<say>...</say>` marker, so the
+      # supervised target trains the model to respond to an
+      # interjection with a spoken acknowledgement.
+      - {role: assistant, stream: high_level, target: true, if_present: speech, tool_calls_from: speech}
+
+  # VQA is view-dependent — each camera gets its own sub-recipe so the
+  # resolver disambiguates via `camera=...`. Camera keys match
+  # subtasks_vqa.yaml (`front` + `wrist`); adjust to your dataset.
+  ask_vqa_top:
+    weight: 0.075
+    bindings:
+      vqa_query: "emitted_at(t, style=vqa, role=user, camera=observation.images.front)"
+      vqa: "emitted_at(t, style=vqa, role=assistant, camera=observation.images.front)"
+    messages:
+      - role: user
+        stream: high_level
+        if_present: vqa_query
+        content:
+          - {type: image, feature: observation.images.front}
+          - {type: text, text: "${vqa_query}"}
+      - {role: assistant, content: "${vqa}", stream: high_level, target: true, if_present: vqa}
+
+  ask_vqa_wrist:
+    weight: 0.075
+    bindings:
+      vqa_query: "emitted_at(t, style=vqa, role=user, camera=observation.images.wrist)"
+      vqa: "emitted_at(t, style=vqa, role=assistant, camera=observation.images.wrist)"
+    messages:
+      - role: user
+        stream: high_level
+        if_present: vqa_query
+        content:
+          - {type: image, feature: observation.images.wrist}
+          - {type: text, text: "${vqa_query}"}
+      - {role: assistant, content: "${vqa}", stream: high_level, target: true, if_present: vqa}

src/lerobot/configs/recipes/subtasks_vqa.yaml

@@ -0,0 +1,61 @@
+# subtasks_vqa — Hi-Robot blend for PI052 (PaliGemma backbone).
+#
+#   Trains two things only: subtasks and VQA. Plan and memory are
+#   intentionally left out — keeps the prompt short and the training
+#   surface small. The fuller blend with memory + spoken replies is
+#   ``subtask_mem_vqa_speech.yaml``.
+#
+#     high_level_subtask  — predict the subtask from the task.
+#     low_level_execution — flow loss with [images, subtask, state].
+#     ask_vqa_{top,wrist} — camera-grounded VQA.
+#
+# PI052's text tokenizer renders these messages as plain
+# ``Role: content`` text (PaliGemma is not chat-pretrained).
+
+blend:
+
+  high_level_subtask:
+    weight: 0.40
+    messages:
+      - {role: user, content: "${task}", stream: high_level}
+      - {role: assistant, content: "${subtask}", stream: high_level, target: true, if_present: subtask}
+
+  low_level_execution:
+    weight: 0.40
+    messages:
+      # The action expert is conditioned on the SUBTASK — at inference
+      # the high-level loop (``HighLevelSubtaskFwd``) generates the
+      # subtask via the LM head and feeds it here. The action expert's
+      # prefix is [images, subtask, state]. ``stream: low_level`` flips
+      # ``predict_actions=True`` so the flow loss fires; no text-CE
+      # target here (subtask prediction is owned by
+      # ``high_level_subtask``).
+      - {role: user, content: "${subtask}", stream: low_level, if_present: subtask}
+
+  ask_vqa_top:
+    weight: 0.10
+    bindings:
+      vqa_query: "emitted_at(t, style=vqa, role=user, camera=observation.images.front)"
+      vqa: "emitted_at(t, style=vqa, role=assistant, camera=observation.images.front)"
+    messages:
+      - role: user
+        stream: high_level
+        if_present: vqa_query
+        content:
+          - {type: image, feature: observation.images.front}
+          - {type: text, text: "${vqa_query}"}
+      - {role: assistant, content: "${vqa}", stream: high_level, target: true, if_present: vqa}
+
+  ask_vqa_wrist:
+    weight: 0.10
+    bindings:
+      vqa_query: "emitted_at(t, style=vqa, role=user, camera=observation.images.wrist)"
+      vqa: "emitted_at(t, style=vqa, role=assistant, camera=observation.images.wrist)"
+    messages:
+      - role: user
+        stream: high_level
+        if_present: vqa_query
+        content:
+          - {type: image, feature: observation.images.wrist}
+          - {type: text, text: "${vqa_query}"}
+      - {role: assistant, content: "${vqa}", stream: high_level, target: true, if_present: vqa}

src/lerobot/configs/train.py

@@ -30,7 +30,7 @@ from lerobot.utils.hub import HubMixin
 from lerobot.utils.sample_weighting import SampleWeightingConfig
 
 from . import parser
-from .default import DatasetConfig, EvalConfig, PeftConfig, WandBConfig
+from .default import DatasetConfig, EMAConfig, EvalConfig, PeftConfig, WandBConfig
 from .policies import PreTrainedConfig
 from .rewards import RewardModelConfig
 
@@ -111,9 +111,20 @@ class TrainPipelineConfig(HubMixin):
     scheduler: LRSchedulerConfig | None = None
     eval: EvalConfig = field(default_factory=EvalConfig)
     wandb: WandBConfig = field(default_factory=WandBConfig)
+    ema: EMAConfig = field(default_factory=EMAConfig)
     peft: PeftConfig | None = None
 
-    # Sample weighting configuration (e.g., for RA-BC training)
+    # VQA oversampling. When set (a fraction in (0, 1)), the training
+    # dataloader uses a WeightedEpisodeAwareSampler that draws frames
+    # carrying a `vqa` language annotation often enough that they make
+    # up roughly this fraction of the training stream. VQA annotations
+    # are typically sparse, so without this they are underrepresented.
+    # `None` (default) keeps uniform episode-aware sampling.
+    vqa_target_fraction: float | None = None
+
+    # Sample weighting configuration (e.g., for RA-BC training). Old
+    # inline ``use_rabc`` / ``rabc_*`` params are migrated to this
+    # field by ``_migrate_legacy_rabc_keys`` above.
     sample_weighting: SampleWeightingConfig | None = None
 
     # Rename map for the observation to override the image and state keys
@@ -177,6 +188,12 @@ class TrainPipelineConfig(HubMixin):
             )
 
         active_cfg = self.trainable_config
+        if self.rename_map and active_cfg.pretrained_path is None:
+            raise ValueError(
+                "`rename_map` requires a pretrained policy checkpoint. "
+                "Fresh initialization derives feature names from the current dataset, so no rename is applied."
+            )
+
         if not self.job_name:
             if self.env is None:
                 self.job_name = f"{active_cfg.type}"

src/lerobot/datasets/__init__.py

@@ -35,7 +35,6 @@ from .dataset_tools import (
     remove_feature,
     split_dataset,
 )
-from .factory import make_dataset, resolve_delta_timestamps
 from .image_writer import safe_stop_image_writer
 from .io_utils import load_episodes, write_stats
 from .language import (
@@ -50,11 +49,24 @@ from .lerobot_dataset import LeRobotDataset
 from .multi_dataset import MultiLeRobotDataset
 from .pipeline_features import aggregate_pipeline_dataset_features, create_initial_features
 from .pyav_utils import check_video_encoder_parameters_pyav, detect_available_encoders_pyav
-from .sampler import EpisodeAwareSampler
+from .sampler import EpisodeAwareSampler, WeightedEpisodeAwareSampler
 from .streaming_dataset import StreamingLeRobotDataset
 from .utils import DEFAULT_EPISODES_PATH, create_lerobot_dataset_card
 from .video_utils import VideoEncodingManager
 
+
+def make_dataset(*args, **kwargs):
+    from .factory import make_dataset as _make_dataset
+
+    return _make_dataset(*args, **kwargs)
+
+
+def resolve_delta_timestamps(*args, **kwargs):
+    from .factory import resolve_delta_timestamps as _resolve_delta_timestamps
+
+    return _resolve_delta_timestamps(*args, **kwargs)
+
+
 # NOTE: Low-level I/O functions (cast_stats_to_numpy, get_parquet_file_size_in_mb, etc.)
 # and legacy migration constants are intentionally NOT re-exported here.
 # Import directly: ``from lerobot.datasets.io_utils import ...``
@@ -65,6 +77,7 @@ __all__ = [
     "DEFAULT_QUANTILES",
     "EVENT_ONLY_STYLES",
     "EpisodeAwareSampler",
+    "WeightedEpisodeAwareSampler",
     "LANGUAGE_EVENTS",
     "LANGUAGE_PERSISTENT",
     "LeRobotDataset",

src/lerobot/datasets/dataset_reader.py

@@ -126,10 +126,53 @@ class DatasetReader:
     def _load_hf_dataset(self) -> datasets.Dataset:
         """hf_dataset contains all the observations, states, actions, rewards, etc."""
         features = get_hf_features_from_features(self._meta.features)
+        # Datasets annotated with the PR1 language columns may have been
+        # written without registering those columns in ``meta/info.json``
+        # (e.g. they predate ``CODEBASE_VERSION="v3.1"`` and were
+        # back-filled by ``lerobot-annotate``). Probe a single parquet
+        # shard and graft the column features on so the strict
+        # ``Dataset.from_parquet`` cast doesn't fail with
+        # ``column names don't match``.
+        features = self._extend_features_with_language_columns(features)
         hf_dataset = load_nested_dataset(self.root / "data", features=features, episodes=self.episodes)
         hf_dataset.set_transform(hf_transform_to_torch)
         return hf_dataset
 
+    def _extend_features_with_language_columns(
+        self, features: datasets.Features
+    ) -> datasets.Features:
+        """Add ``language_persistent`` / ``language_events`` to ``features``
+        when the underlying parquet shards declare them but the metadata
+        doesn't. No-op when neither column is present or both are
+        already registered.
+        """
+        # Find any one parquet to peek at; bail if there are none yet
+        # (the dataset will fail later for an unrelated reason and we
+        # want that error to surface as-is).
+        try:
+            sample = next((self.root / "data").glob("*/*.parquet"))
+        except StopIteration:
+            return features
+
+        from pyarrow import parquet as _pq  # noqa: PLC0415
+
+        schema_names = set(_pq.read_schema(sample).names)
+        from .language import (  # noqa: PLC0415
+            LANGUAGE_EVENTS,
+            LANGUAGE_PERSISTENT,
+            language_events_column_feature,
+            language_persistent_column_feature,
+        )
+
+        extra: dict[str, object] = {}
+        if LANGUAGE_PERSISTENT in schema_names and LANGUAGE_PERSISTENT not in features:
+            extra[LANGUAGE_PERSISTENT] = language_persistent_column_feature()
+        if LANGUAGE_EVENTS in schema_names and LANGUAGE_EVENTS not in features:
+            extra[LANGUAGE_EVENTS] = language_events_column_feature()
+        if not extra:
+            return features
+        return datasets.Features({**features, **extra})
+
     def _check_cached_episodes_sufficient(self) -> bool:
         """Check if the cached dataset contains all requested episodes and their video files."""
         if self.hf_dataset is None or len(self.hf_dataset) == 0:

src/lerobot/datasets/language_render.py

@@ -170,6 +170,29 @@ def render_sample(
     """
     persistent_rows = _normalize_rows(persistent or [])
     event_rows = _normalize_rows(events or [])
+
+    # VQA-priority routing. A ``vqa`` annotation is sparse and
+    # view-dependent; the plain weighted blend would (a) waste a draw
+    # whenever it picks an ``ask_vqa*`` sub-recipe for a frame that has
+    # no VQA, and (b) silently drop a VQA-annotated frame whenever it
+    # picks a non-VQA sub-recipe. So: if the blend has ``ask_vqa*``
+    # sub-recipes and *this* frame carries one of their VQA bindings,
+    # render VQA here regardless of the weighted draw. That makes VQA's
+    # recipe-side training share equal the VQA-annotation density (the
+    # maximum reachable without a dataset-level oversampling sampler).
+    if recipe.blend is not None:
+        vqa_rendered = _render_vqa_if_present(
+            recipe,
+            persistent=persistent_rows,
+            events=event_rows,
+            t=t,
+            sample_idx=sample_idx,
+            task=task,
+            dataset_ctx=dataset_ctx,
+        )
+        if vqa_rendered is not None:
+            return vqa_rendered
+
     selected_recipe = _select_recipe(recipe, sample_idx)
     bindings = _resolve_bindings(
         selected_recipe,
@@ -183,6 +206,59 @@ def render_sample(
     return _render_message_recipe(selected_recipe, bindings)
 
 
+def _render_vqa_if_present(
+    recipe: TrainingRecipe,
+    *,
+    persistent: Sequence[LanguageRow],
+    events: Sequence[LanguageRow],
+    t: float,
+    sample_idx: int,
+    task: str | None,
+    dataset_ctx: Any | None,
+) -> RenderedMessages | None:
+    """Render an ``ask_vqa*`` sub-recipe iff this frame carries a VQA
+    annotation; otherwise return ``None`` so the caller falls back to the
+    normal weighted blend.
+
+    When several VQA sub-recipes resolve (e.g. a frame annotated for more
+    than one camera), one is chosen deterministically by relative weight.
+    """
+    assert recipe.blend is not None
+    renderable: list[tuple[float, RenderedMessages]] = []
+    for name, component in recipe.blend.items():
+        if not name.startswith("ask_vqa"):
+            continue
+        bindings = _resolve_bindings(
+            component,
+            persistent=persistent,
+            events=events,
+            t=t,
+            sample_idx=sample_idx,
+            task=task,
+            dataset_ctx=dataset_ctx,
+        )
+        rendered = _render_message_recipe(component, bindings)
+        if rendered is not None:
+            renderable.append((float(component.weight or 0.0), rendered))
+
+    if not renderable:
+        return None
+    if len(renderable) == 1:
+        return renderable[0][1]
+
+    # Multiple cameras have a VQA for this frame — deterministic pick by
+    # relative weight (fall back to a uniform draw if all weights are 0).
+    total = sum(w for w, _ in renderable) or float(len(renderable))
+    digest = hashlib.blake2b(f"vqa:{sample_idx}".encode(), digest_size=8).digest()
+    draw = int.from_bytes(digest, "big") / 2**64 * total
+    cumulative = 0.0
+    for w, rendered in renderable:
+        cumulative += w or (total / len(renderable))
+        if draw < cumulative:
+            return rendered
+    return renderable[-1][1]
+
+
 def _select_recipe(recipe: TrainingRecipe, sample_idx: int) -> TrainingRecipe:
     """Pick a deterministic blend component for ``sample_idx`` (or return ``recipe``)."""
     if recipe.blend is None:
@@ -346,7 +422,15 @@ def _render_message_recipe(
         if turn.target:
             target_indices.append(message_idx)
 
-    if not target_indices:
+    # A render is meaningful if it supervises *something*: either a
+    # text-CE target turn, or a ``low_level`` stream turn (flow / action
+    # supervision — e.g. the flow-only ``low_level_execution`` recipe,
+    # ``user(${subtask})`` with ``stream: low_level`` and no target).
+    # Without this, a flow-only recipe renders to ``None`` every time
+    # the blend draws it → ``predict_actions`` is never True → the
+    # action expert never receives a flow loss.
+    has_low_level = any(stream == "low_level" for stream in streams)
+    if not target_indices and not has_low_level:
         return None
 
     rendered = {
@@ -403,8 +487,10 @@ def _validate_rendered(rendered: RenderedMessages) -> None:
 
     if len(streams) != len(messages):
         raise ValueError("message_streams must be aligned with messages.")
-    if not target_indices:
-        raise ValueError("Rendered samples must contain at least one target message.")
+    # Valid iff it supervises something: a text-CE target turn OR a
+    # ``low_level`` stream turn (flow / action supervision).
+    if not target_indices and not any(s == "low_level" for s in streams):
+        raise ValueError("Rendered samples must contain a target message or a low_level-stream message.")
     for idx in target_indices:
         if idx < 0 or idx >= len(messages):
             raise ValueError(f"Target message index {idx} is out of bounds.")

src/lerobot/datasets/lerobot_dataset.py

@@ -524,7 +524,7 @@ class LeRobotDataset(torch.utils.data.Dataset):
         license: str | None = "apache-2.0",
         tag_version: bool = True,
         push_videos: bool = True,
-        private: bool = False,
+        private: bool | None = None,
         allow_patterns: list[str] | str | None = None,
         upload_large_folder: bool = False,
         **card_kwargs,
@@ -543,7 +543,8 @@ class LeRobotDataset(torch.utils.data.Dataset):
             tag_version: If ``True``, create a Git tag for the current codebase
                 version.
             push_videos: If ``False``, skip uploading the ``videos/`` directory.
-            private: If ``True``, create a private repository.
+            private: If ``True``, create a private repository. If ``None``
+                (default), defer to the org default on the Hub (only affects orgs).
             allow_patterns: Glob pattern(s) restricting which files to upload.
             upload_large_folder: If ``True``, use ``upload_large_folder`` instead
                 of ``upload_folder`` for very large datasets.

src/lerobot/datasets/sampler.py

@@ -84,3 +84,66 @@ class EpisodeAwareSampler:
 
     def __len__(self) -> int:
         return len(self.indices)
+
+
+class WeightedEpisodeAwareSampler(EpisodeAwareSampler):
+    """``EpisodeAwareSampler`` that draws frames *with replacement* in
+    proportion to per-frame weights.
+
+    Used to oversample frames carrying a sparse annotation (e.g. a VQA
+    question) so the policy sees them more often than their natural
+    dataset density. One epoch still yields ``len(self.indices)``
+    samples — the weights only change the *composition* of the stream,
+    not its length. Each epoch re-draws, so the oversampled subset
+    varies run to run.
+    """
+
+    def __init__(
+        self,
+        dataset_from_indices: list[int],
+        dataset_to_indices: list[int],
+        frame_weights,
+        *,
+        episode_indices_to_use: list | None = None,
+        drop_n_first_frames: int = 0,
+        drop_n_last_frames: int = 0,
+    ):
+        """
+        Args:
+            dataset_from_indices: Episode start indices (see ``EpisodeAwareSampler``).
+            dataset_to_indices: Episode end indices.
+            frame_weights: 1-D sequence/tensor of non-negative weights, one per
+                dataset frame (length == total dataset frames). Higher weight ⇒
+                that frame is sampled more often.
+            episode_indices_to_use / drop_n_first_frames / drop_n_last_frames:
+                Same meaning as ``EpisodeAwareSampler`` — the episode-boundary
+                frame filtering is applied first, then weighting is restricted
+                to the surviving frames.
+        """
+        super().__init__(
+            dataset_from_indices,
+            dataset_to_indices,
+            episode_indices_to_use=episode_indices_to_use,
+            drop_n_first_frames=drop_n_first_frames,
+            drop_n_last_frames=drop_n_last_frames,
+            shuffle=False,
+        )
+        weights = torch.as_tensor(frame_weights, dtype=torch.double).flatten()
+        idx = torch.tensor(self.indices, dtype=torch.long)
+        if weights.numel() <= int(idx.max()):
+            raise ValueError(
+                f"frame_weights has {weights.numel()} entries but the sampler "
+                f"references frame index {int(idx.max())}."
+            )
+        selected = weights[idx]
+        if not torch.isfinite(selected).all() or bool((selected < 0).any()):
+            raise ValueError("frame_weights must be finite and non-negative.")
+        if float(selected.sum()) <= 0.0:
+            # All surviving frames have zero weight — fall back to uniform.
+            selected = torch.ones_like(selected)
+        self._weights = selected
+
+    def __iter__(self) -> Iterator[int]:
+        picks = torch.multinomial(self._weights, num_samples=len(self.indices), replacement=True)
+        for i in picks.tolist():
+            yield self.indices[i]

src/lerobot/datasets/utils.py

@@ -366,17 +366,24 @@ def get_safe_version(repo_id: str, version: str | packaging.version.Version) ->
     hub_versions = get_repo_versions(repo_id)
 
     if not hub_versions:
-        raise RevisionNotFoundError(
-            f"""Your dataset must be tagged with a codebase version.
-            Assuming _version_ is the codebase_version value in the info.json, you can run this:
-            ```python
-            from huggingface_hub import HfApi
-
-            hub_api = HfApi()
-            hub_api.create_tag("{repo_id}", tag="_version_", repo_type="dataset")
-            ```
-            """
+        msg = (
+            f"Repo {repo_id!r} has no codebase-version tags. The dataset "
+            f"either doesn't exist on the Hub yet, or it was uploaded "
+            f"without a ``v3.x``-style tag. To tag an existing dataset run:\n"
+            f"  from huggingface_hub import HfApi\n"
+            f"  HfApi().create_tag({repo_id!r}, tag='v3.0', repo_type='dataset', exist_ok=True)"
         )
+        # ``RevisionNotFoundError`` extends ``HfHubHTTPError`` whose
+        # ``__init__`` indexes ``response.headers`` unconditionally on
+        # current ``huggingface_hub`` versions. Constructing it without
+        # a real ``Response`` object crashes with either
+        # ``TypeError: missing 1 required keyword-only argument`` (old
+        # builds) or ``AttributeError: 'NoneType' object has no attribute
+        # 'headers'`` (new builds). Skip that path entirely — this isn't
+        # really an HTTP error, it's a configuration issue — and raise a
+        # plain ``RuntimeError`` so the message actually reaches the
+        # caller.
+        raise RuntimeError(msg)
 
     if target_version in hub_versions:
         return f"v{target_version}"

src/lerobot/envs/robocasa.py

@@ -33,8 +33,8 @@ logger = logging.getLogger(__name__)
 
 # Dimensions for the flat action/state vectors used by the LeRobot wrapper.
 # These correspond to the PandaOmron robot in RoboCasa365.
-OBS_STATE_DIM = 16  # base_pos(3) + base_quat(4) + ee_pos_rel(3) + ee_quat_rel(4) + gripper_qpos(2)
-ACTION_DIM = 12  # base_motion(4) + control_mode(1) + ee_pos(3) + ee_rot(3) + gripper(1)
+OBS_STATE_DIM = 16  # ee_pos_rel(3) + ee_quat_rel(4) + base_pos(3) + base_quat(4) + gripper_qpos(2)
+ACTION_DIM = 12  # ee_pos(3) + ee_rot(3) + gripper(1) + base_motion(4) + control_mode(1)
 ACTION_LOW = -1.0
 ACTION_HIGH = 1.0
 
@@ -101,14 +101,15 @@ def _resolve_tasks(task: str) -> tuple[list[str], str | None]:
 def convert_action(flat_action: np.ndarray) -> dict[str, Any]:
     """Split a flat (12,) action vector into a RoboCasa action dict.
 
-    Layout: base_motion(4) + control_mode(1) + ee_pos(3) + ee_rot(3) + gripper(1)
+    Layout (openpi / robocasa.utils.env_utils.convert_action order):
+        ee_pos(3) + ee_rot(3) + gripper(1) + base_motion(4) + control_mode(1)
     """
     return {
-        "action.base_motion": flat_action[0:4],
-        "action.control_mode": flat_action[4:5],
-        "action.end_effector_position": flat_action[5:8],
-        "action.end_effector_rotation": flat_action[8:11],
-        "action.gripper_close": flat_action[11:12],
+        "action.end_effector_position": flat_action[0:3],
+        "action.end_effector_rotation": flat_action[3:6],
+        "action.gripper_close": flat_action[6:7],
+        "action.base_motion": flat_action[7:11],
+        "action.control_mode": flat_action[11:12],
     }
 
 
@@ -230,12 +231,14 @@ class RoboCasaEnv(gym.Env):
             return {"pixels": images}
 
         # `state.*` keys come from PandaOmronKeyConverter inside the wrapper.
+        # openpi state order: ee first, then base, then gripper (matches the
+        # openpi robocasa pipeline / examples/robocasa/main.py state layout).
         agent_pos = np.concatenate(
             [
-                raw_obs.get("state.base_position", np.zeros(3)),
-                raw_obs.get("state.base_rotation", np.zeros(4)),
                 raw_obs.get("state.end_effector_position_relative", np.zeros(3)),
                 raw_obs.get("state.end_effector_rotation_relative", np.zeros(4)),
+                raw_obs.get("state.base_position", np.zeros(3)),
+                raw_obs.get("state.base_rotation", np.zeros(4)),
                 raw_obs.get("state.gripper_qpos", np.zeros(2)),
             ],
             axis=-1,

src/lerobot/model/kinematics.py

@@ -18,12 +18,25 @@ from typing import TYPE_CHECKING
 
 import numpy as np
 
-from lerobot.utils.import_utils import _placo_available, require_package
+from lerobot.utils.import_utils import require_package
 
-if TYPE_CHECKING or _placo_available:
+_placo_runtime_error: ImportError | None = None
+
+if TYPE_CHECKING:
     import placo  # type: ignore[import-not-found]
 else:
-    placo = None
+    try:
+        import placo  # type: ignore[import-not-found]
+    except ImportError as _placo_import_err:
+        placo = None
+        _placo_runtime_error = _placo_import_err
+
+
+def _raise_if_placo_unusable() -> None:
+    if placo is None and _placo_runtime_error is not None:
+        raise ImportError(
+            f"placo is installed but failed to import: {_placo_runtime_error!s}"
+        ) from _placo_runtime_error
 
 
 class RobotKinematics:
@@ -44,6 +57,7 @@ class RobotKinematics:
             joint_names (list[str] | None): List of joint names to use for the kinematics solver
         """
         require_package("placo", extra="placo-dep")
+        _raise_if_placo_unusable()
 
         self.robot = placo.RobotWrapper(urdf_path)
         self.solver = placo.KinematicsSolver(self.robot)

src/lerobot/motors/robstride/robstride.py

@@ -43,6 +43,7 @@ from .tables import (
     CAN_CMD_SET_ZERO,
     DEFAULT_BAUDRATE,
     DEFAULT_TIMEOUT_MS,
+    HANDSHAKE_TIMEOUT_S,
     MODEL_RESOLUTION,
     MOTOR_LIMIT_PARAMS,
     NORMALIZED_DATA,
@@ -215,14 +216,16 @@ class RobstrideMotorsBus(MotorsBusBase):
             self._is_connected = False
             raise ConnectionError(f"Failed to connect to CAN bus: {e}") from e
 
-    def _query_status_via_clear_fault(self, motor: NameOrID) -> tuple[bool, can.Message | None]:
+    def _query_status_via_clear_fault(
+        self, motor: NameOrID, timeout: float = RUNNING_TIMEOUT
+    ) -> tuple[bool, can.Message | None]:
         motor_name = self._get_motor_name(motor)
         motor_id = self._get_motor_id(motor_name)
         recv_id = self._get_motor_recv_id(motor_name)
         data = [0xFF] * 7 + [CAN_CMD_CLEAR_FAULT]
         msg = can.Message(arbitration_id=motor_id, data=data, is_extended_id=False)
         self._bus().send(msg)
-        return self._recv_status_via_clear_fault(expected_recv_id=recv_id)
+        return self._recv_status_via_clear_fault(expected_recv_id=recv_id, timeout=timeout)
 
     def _recv_status_via_clear_fault(
         self, expected_recv_id: int | None = None, timeout: float = RUNNING_TIMEOUT
@@ -280,7 +283,7 @@ class RobstrideMotorsBus(MotorsBusBase):
         faulted_motors = []
 
         for motor_name in self.motors:
-            has_fault, msg = self._query_status_via_clear_fault(motor_name)
+            has_fault, msg = self._query_status_via_clear_fault(motor_name, timeout=HANDSHAKE_TIMEOUT_S)
             if msg is None:
                 missing_motors.append(motor_name)
             elif has_fault:
@@ -505,6 +508,87 @@ class RobstrideMotorsBus(MotorsBusBase):
 
         return responses
 
+    def _recv_all_messages_until_quiet(
+        self,
+        *,
+        timeout: float = RUNNING_TIMEOUT,
+        max_messages: int = 4096,
+    ) -> list[can.Message]:
+        """
+        Receive frames until the bus goes quiet.
+
+        Args:
+            timeout: Poll timeout used for each recv() call. Collection stops
+                when one recv() times out (quiet gap).
+            max_messages: Safety cap to prevent unbounded loops.
+        """
+        out: list[can.Message] = []
+        max_messages = max(1, max_messages)
+        timeout = max(0.0, timeout)
+
+        try:
+            while len(out) < max_messages:
+                msg = self._bus().recv(timeout=timeout)
+                if msg is None:
+                    break
+                out.append(msg)
+        except (can.CanError, OSError) as e:
+            logger.debug(f"Error draining CAN RX queue on {self.port}: {e}")
+
+        return out
+
+    def _process_feedback_messages(self, messages: list[can.Message]) -> set[int]:
+        """
+        Decode all received feedback frames and update cached motor states.
+
+        Returns:
+            Set of payload recv_ids that were successfully mapped to motors.
+        """
+        processed_recv_ids: set[int] = set()
+        for msg in messages:
+            if len(msg.data) < 1:
+                logger.debug(
+                    f"Dropping short CAN frame on {self.port} "
+                    f"(arb=0x{int(msg.arbitration_id):02X}, data={bytes(msg.data).hex()})"
+                )
+                continue
+
+            recv_id = int(msg.data[0])
+            motor_name = self._recv_id_to_motor.get(recv_id)
+            if motor_name is None:
+                logger.debug(
+                    f"Unmapped CAN frame on {self.port} "
+                    f"(arb=0x{int(msg.arbitration_id):02X}, recv_id=0x{recv_id:02X}, data={bytes(msg.data).hex()})"
+                )
+                continue
+
+            self._process_response(motor_name, msg)
+            processed_recv_ids.add(recv_id)
+
+        return processed_recv_ids
+
+    def flush_rx_queue(self, poll_timeout_s: float = 0.0005, max_messages: int = 4096) -> int:
+        """
+        Drain pending RX frames from the CAN interface.
+
+        This is used by higher-level controllers to drop stale feedback before issuing
+        a fresh read cycle, so subsequent state reads are based on most recent replies.
+        It should also be called once when a controller instance is created/connected,
+        to clear residual frames left on the interface from previous sessions.
+        """
+        drained = 0
+        poll_timeout_s = max(0.0, poll_timeout_s)
+        max_messages = max(1, max_messages)
+        try:
+            while drained < max_messages:
+                msg = self._bus().recv(timeout=poll_timeout_s)
+                if msg is None:
+                    break
+                drained += 1
+        except (can.CanError, OSError) as e:
+            logger.debug(f"Failed to flush CAN RX queue on {self.port}: {e}")
+        return drained
+
     def _speed_control(
         self,
         motor: NameOrID,
@@ -644,11 +728,14 @@ class RobstrideMotorsBus(MotorsBusBase):
             msg = can.Message(arbitration_id=motor_id, data=data, is_extended_id=False)
             self._bus().send(msg)
             recv_id_to_motor[self._get_motor_recv_id(motor)] = motor_name
+        # Read every feedback frame until RX goes quiet, then decode all of them.
+        # This avoids dropping useful frames when responses from different motors interleave.
+        messages = self._recv_all_messages_until_quiet()
+        processed_recv_ids = self._process_feedback_messages(messages)
 
-        responses = self._recv_all_responses(list(recv_id_to_motor.keys()), timeout=RUNNING_TIMEOUT)
         for recv_id, motor_name in recv_id_to_motor.items():
-            if msg := responses.get(recv_id):
-                self._process_response(motor_name, msg)
+            if recv_id not in processed_recv_ids:
+                logger.warning(f"Packet drop: {motor_name} (ID: 0x{recv_id:02X}). Using last known state.")
 
     def _float_to_uint(self, x: float, x_min: float, x_max: float, bits: int) -> int:
         """Convert float to unsigned integer for CAN transmission."""
@@ -711,7 +798,10 @@ class RobstrideMotorsBus(MotorsBusBase):
         try:
             self._decode_motor_state(msg.data)
         except Exception as e:
-            logger.warning(f"Failed to decode response from {motor}: {e}")
+            logger.warning(
+                f"Failed to decode response from {motor} "
+                f"(arb=0x{int(msg.arbitration_id):02X}, data={bytes(msg.data).hex()}): {e}"
+            )
 
     def _get_cached_value(self, motor: str, data_name: str) -> Value:
         """Retrieve a specific value from the state cache."""
@@ -848,20 +938,12 @@ class RobstrideMotorsBus(MotorsBusBase):
             self._bus().send(msg)
             updated_motors.append(motor)
 
-        expected_recv_ids = [self._get_motor_recv_id(motor) for motor in updated_motors]
-        responses = self._recv_all_responses(expected_recv_ids, timeout=RUNNING_TIMEOUT)
-
-        for response in responses.values():
-            payload_motor_name = self._recv_id_to_motor.get(response.data[0])
-            if payload_motor_name is not None:
-                self._process_response(payload_motor_name, response)
-            else:
-                # Fallback: still attempt to decode based on payload byte0 mapping.
-                self._decode_motor_state(response.data)
+        messages = self._recv_all_messages_until_quiet()
+        processed_recv_ids = self._process_feedback_messages(messages)
 
         for motor in updated_motors:
             recv_id = self._get_motor_recv_id(motor)
-            if recv_id not in responses:
+            if recv_id not in processed_recv_ids:
                 logger.warning(f"Packet drop: {motor} (ID: 0x{recv_id:02X}). Using last known state.")
 
     def read_calibration(self) -> dict[str, MotorCalibration]:

src/lerobot/motors/robstride/tables.py

@@ -114,7 +114,8 @@ CAN_CMD_SAVE_PARAM = 0xAA
 CAN_PARAM_ID = 0x7FF
 
 
-RUNNING_TIMEOUT = 0.001
+RUNNING_TIMEOUT = 0.003
+HANDSHAKE_TIMEOUT_S = 0.05
 PARAM_TIMEOUT = 0.01
 
 STATE_CACHE_TTL_S = 0.02

src/lerobot/optim/optimizers.py

@@ -104,6 +104,8 @@ class AdamWConfig(OptimizerConfig):
     eps: float = 1e-8
     weight_decay: float = 1e-2
     grad_clip_norm: float = 10.0
+    foreach: bool | None = None
+    fused: bool | None = None
 
     def build(self, params: OptimizerParams) -> torch.optim.Optimizer:
         kwargs = asdict(self)

src/lerobot/policies/__init__.py

@@ -20,10 +20,12 @@ from .eo1.configuration_eo1 import EO1Config as EO1Config
 from .factory import get_policy_class, make_policy, make_policy_config, make_pre_post_processors
 from .gaussian_actor.configuration_gaussian_actor import GaussianActorConfig as GaussianActorConfig
 from .groot.configuration_groot import GrootConfig as GrootConfig
+from .molmoact2.configuration_molmoact2 import MolmoAct2Config as MolmoAct2Config
 from .multi_task_dit.configuration_multi_task_dit import MultiTaskDiTConfig as MultiTaskDiTConfig
 from .pi0.configuration_pi0 import PI0Config as PI0Config
 from .pi0_fast.configuration_pi0_fast import PI0FastConfig as PI0FastConfig
 from .pi05.configuration_pi05 import PI05Config as PI05Config
+from .pi052.configuration_pi052 import PI052Config as PI052Config
 from .pretrained import PreTrainedPolicy as PreTrainedPolicy
 from .smolvla.configuration_smolvla import SmolVLAConfig as SmolVLAConfig
 from .tdmpc.configuration_tdmpc import TDMPCConfig as TDMPCConfig
@@ -43,10 +45,12 @@ __all__ = [
     "EO1Config",
     "GaussianActorConfig",
     "GrootConfig",
+    "MolmoAct2Config",
     "MultiTaskDiTConfig",
     "PI0Config",
     "PI0FastConfig",
     "PI05Config",
+    "PI052Config",
     "SmolVLAConfig",
     "TDMPCConfig",
     "VQBeTConfig",

src/lerobot/policies/factory.py

@@ -49,6 +49,7 @@ from .diffusion.configuration_diffusion import DiffusionConfig
 from .eo1.configuration_eo1 import EO1Config
 from .gaussian_actor.configuration_gaussian_actor import GaussianActorConfig
 from .groot.configuration_groot import GrootConfig
+from .molmoact2.configuration_molmoact2 import MolmoAct2Config
 from .multi_task_dit.configuration_multi_task_dit import MultiTaskDiTConfig
 from .pi0.configuration_pi0 import PI0Config
 from .pi05.configuration_pi05 import PI05Config
@@ -56,11 +57,85 @@ from .pretrained import PreTrainedPolicy
 from .smolvla.configuration_smolvla import SmolVLAConfig
 from .tdmpc.configuration_tdmpc import TDMPCConfig
 from .utils import validate_visual_features_consistency
+from .vla_jepa.configuration_vla_jepa import VLAJEPAConfig
 from .vqbet.configuration_vqbet import VQBeTConfig
 from .wall_x.configuration_wall_x import WallXConfig
 from .xvla.configuration_xvla import XVLAConfig
 
 
+def _restore_pi052_pretrained_state(
+    preprocessor: PolicyProcessorPipeline,
+    postprocessor: PolicyProcessorPipeline,
+    pretrained_path: str,
+) -> None:
+    """Transplant saved stateful blobs from a pi052 checkpoint into fresh pipelines.
+
+    pi052's preprocessor includes steps whose constructor args don't
+    JSON-roundtrip (``RenderMessagesStep.recipe`` is a Python object,
+    ``ActionTokenizerProcessorStep.action_tokenizer_name`` is a
+    fitted-tokenizer path that may not exist at eval time). We rebuild
+    those pipelines fresh from ``config.recipe_path`` and then walk
+    over the saved ``policy_{pre,post}processor.json`` files to find
+    each step's ``state_file`` reference and load the bytes back into
+    the corresponding fresh step. Today that's only the
+    NormalizerProcessorStep / UnnormalizerProcessorStep (the action /
+    state quantile stats), but the loop is generic so any future
+    stateful step picks up its blob automatically.
+
+    Pairing is by ``registry_name`` AND position so a benign reorder
+    on the saved side surfaces a warning rather than silently feeding
+    the wrong tensors into the wrong step.
+    """
+    import json  # noqa: PLC0415
+    import logging  # noqa: PLC0415
+    from pathlib import Path  # noqa: PLC0415
+
+    from safetensors.torch import load_file  # noqa: PLC0415
+
+    base = Path(pretrained_path)
+    if not base.exists():
+        return
+
+    log = logging.getLogger(__name__)
+
+    for pipeline, config_filename in [
+        (preprocessor, f"{POLICY_PREPROCESSOR_DEFAULT_NAME}.json"),
+        (postprocessor, f"{POLICY_POSTPROCESSOR_DEFAULT_NAME}.json"),
+    ]:
+        config_path = base / config_filename
+        if not config_path.exists():
+            continue
+        saved = json.loads(config_path.read_text())
+
+        for idx, (saved_step, fresh_step) in enumerate(
+            zip(saved.get("steps", []), pipeline.steps, strict=False)
+        ):
+            state_file = saved_step.get("state_file")
+            if not state_file:
+                continue
+            saved_name = saved_step.get("registry_name")
+            fresh_name = getattr(type(fresh_step), "_registry_name", None)
+            if saved_name and fresh_name and saved_name != fresh_name:
+                log.warning(
+                    "PI052 state restore: %s step %d registry name mismatch "
+                    "(saved=%s, fresh=%s); skipping %s",
+                    config_filename, idx, saved_name, fresh_name, state_file,
+                )
+                continue
+            state_path = base / state_file
+            if not state_path.exists():
+                log.warning(
+                    "PI052 state restore: %s missing at %s; %s left at fresh init",
+                    state_file, base, fresh_name,
+                )
+                continue
+            fresh_step.load_state_dict(load_file(str(state_path)))
+            log.info(
+                "PI052 state restore: loaded %s into %s (step %d)",
+                state_file, fresh_name, idx,
+            )
+
+
 def _reconnect_relative_absolute_steps(
     preprocessor: PolicyProcessorPipeline, postprocessor: PolicyProcessorPipeline
 ) -> None:
@@ -88,7 +163,8 @@ def get_policy_class(name: str) -> type[PreTrainedPolicy]:
 
     Args:
         name: The name of the policy. Supported names are "tdmpc", "diffusion", "act",
-            "multi_task_dit", "vqbet", "pi0", "pi05", "gaussian_actor", "smolvla", "wall_x".
+            "multi_task_dit", "vqbet", "pi0", "pi05", "gaussian_actor", "smolvla", "wall_x",
+            "molmoact2".
     Returns:
         The policy class corresponding to the given name.
 
@@ -127,6 +203,10 @@ def get_policy_class(name: str) -> type[PreTrainedPolicy]:
         from .pi05.modeling_pi05 import PI05Policy
 
         return PI05Policy
+    elif name == "pi052":
+        from .pi052.modeling_pi052 import PI052Policy
+
+        return PI052Policy
     elif name == "gaussian_actor":
         from .gaussian_actor.modeling_gaussian_actor import GaussianActorPolicy
 
@@ -151,6 +231,14 @@ def get_policy_class(name: str) -> type[PreTrainedPolicy]:
         from .eo1.modeling_eo1 import EO1Policy
 
         return EO1Policy
+    elif name == "molmoact2":
+        from .molmoact2.modeling_molmoact2 import MolmoAct2Policy
+
+        return MolmoAct2Policy
+    elif name == "vla_jepa":
+        from .vla_jepa.modeling_vla_jepa import VLAJEPAPolicy
+
+        return VLAJEPAPolicy
     else:
         try:
             return _get_policy_cls_from_policy_name(name=name)
@@ -167,8 +255,8 @@ def make_policy_config(policy_type: str, **kwargs) -> PreTrainedConfig:
 
     Args:
         policy_type: The type of the policy. Supported types include "tdmpc",
-                     "multi_task_dit", "diffusion", "act", "vqbet", "pi0", "pi05", "gaussian_actor",
-                     "smolvla", "wall_x".
+                     "multi_task_dit", "diffusion", "act", "vqbet", "pi0", "pi05",
+                     "pi052", "gaussian_actor", "smolvla", "wall_x", "molmoact2".
         **kwargs: Keyword arguments to be passed to the configuration class constructor.
 
     Returns:
@@ -191,6 +279,10 @@ def make_policy_config(policy_type: str, **kwargs) -> PreTrainedConfig:
         return PI0Config(**kwargs)
     elif policy_type == "pi05":
         return PI05Config(**kwargs)
+    elif policy_type == "pi052":
+        from .pi052.configuration_pi052 import PI052Config
+
+        return PI052Config(**kwargs)
     elif policy_type == "gaussian_actor":
         return GaussianActorConfig(**kwargs)
     elif policy_type == "smolvla":
@@ -203,6 +295,10 @@ def make_policy_config(policy_type: str, **kwargs) -> PreTrainedConfig:
         return WallXConfig(**kwargs)
     elif policy_type == "eo1":
         return EO1Config(**kwargs)
+    elif policy_type == "molmoact2":
+        return MolmoAct2Config(**kwargs)
+    elif policy_type == "vla_jepa":
+        return VLAJEPAConfig(**kwargs)
     else:
         try:
             config_cls = PreTrainedConfig.get_choice_class(policy_type)
@@ -231,6 +327,13 @@ class ProcessorConfigKwargs(TypedDict, total=False):
     preprocessor_overrides: dict[str, Any] | None
     postprocessor_overrides: dict[str, Any] | None
     dataset_stats: dict[str, dict[str, torch.Tensor]] | None
+    # Optional: HF Hub repo id of the dataset the policy is being
+    # trained on. Used by policies that auto-fit pieces of their
+    # preprocessing (e.g. pi052's FAST action tokenizer per
+    # Pertsch et al. 2025 [64], π0.5 §III.C). When omitted, those
+    # policies fall back to their universal pre-fitted tokenizers.
+    dataset_repo_id: str | None
+    dataset_meta: Any | None
 
 
 def make_pre_post_processors(
@@ -263,6 +366,29 @@ def make_pre_post_processors(
         NotImplementedError: If a processor factory is not implemented for the given
             policy configuration type.
     """
+    if pretrained_path and getattr(policy_cfg, "type", None) == "pi052":
+        # pi052 pipelines don't roundtrip through the saved
+        # ``policy_preprocessor.json``: ``RenderMessagesStep`` holds a
+        # Python ``TrainingRecipe`` (not JSON-serializable; saved as
+        # ``{}``) and ``ActionTokenizerProcessorStep`` saves a host-only
+        # FAST tokenizer path. Generic ``from_pretrained`` then dies
+        # with ``RenderMessagesStep.__init__() missing 1 required
+        # positional argument: 'recipe'`` (job 22164494).
+        #
+        # Mirror ``lerobot_pi052_runtime``'s bootstrap: build pipelines
+        # fresh from ``config.recipe_path`` and transplant the saved
+        # stateful blobs (normalizer stats) from the checkpoint dir.
+        from .pi052.processor_pi052 import make_pi052_pre_post_processors
+
+        preprocessor, postprocessor = make_pi052_pre_post_processors(
+            config=policy_cfg,
+            dataset_stats=kwargs.get("dataset_stats"),
+            dataset_repo_id=kwargs.get("dataset_repo_id"),
+        )
+        _restore_pi052_pretrained_state(preprocessor, postprocessor, pretrained_path)
+        _reconnect_relative_absolute_steps(preprocessor, postprocessor)
+        return preprocessor, postprocessor
+
     if pretrained_path:
         # TODO(Steven): Temporary patch, implement correctly the processors for Gr00t
         if isinstance(policy_cfg, GrootConfig):
@@ -357,6 +483,22 @@ def make_pre_post_processors(
             dataset_stats=kwargs.get("dataset_stats"),
         )
 
+    elif policy_cfg.type == "pi052":
+        # NOTE: PI052Config subclasses PI05Config, so this branch MUST
+        # come before the PI05Config isinstance check below (otherwise
+        # pi052 would silently pick up π0.5's processor).
+        from .pi052.processor_pi052 import make_pi052_pre_post_processors
+
+        processors = make_pi052_pre_post_processors(
+            config=policy_cfg,
+            dataset_stats=kwargs.get("dataset_stats"),
+            # ``dataset_repo_id`` flows in via kwargs when FAST CE is
+            # enabled — the train loop sets it from ``--dataset.repo_id``.
+            # When ``None``, ``make_pi052_pre_post_processors`` skips
+            # the auto-fit and uses the universal tokenizer.
+            dataset_repo_id=kwargs.get("dataset_repo_id"),
+        )
+
     elif isinstance(policy_cfg, PI05Config):
         from .pi05.processor_pi05 import make_pi05_pre_post_processors
 
@@ -406,6 +548,7 @@ def make_pre_post_processors(
             config=policy_cfg,
             dataset_stats=kwargs.get("dataset_stats"),
         )
+
     elif isinstance(policy_cfg, EO1Config):
         from .eo1.processor_eo1 import make_eo1_pre_post_processors
 
@@ -414,6 +557,23 @@ def make_pre_post_processors(
             dataset_stats=kwargs.get("dataset_stats"),
         )
 
+    elif isinstance(policy_cfg, MolmoAct2Config):
+        from .molmoact2.processor_molmoact2 import make_molmoact2_pre_post_processors
+
+        processors = make_molmoact2_pre_post_processors(
+            config=policy_cfg,
+            dataset_stats=kwargs.get("dataset_stats"),
+            dataset_meta=kwargs.get("dataset_meta"),
+        )
+
+    elif isinstance(policy_cfg, VLAJEPAConfig):
+        from .vla_jepa.processor_vla_jepa import make_vla_jepa_pre_post_processors
+
+        processors = make_vla_jepa_pre_post_processors(
+            config=policy_cfg,
+            dataset_stats=kwargs.get("dataset_stats"),
+        )
+
     else:
         try:
             processors = _make_processors_from_policy_config(
@@ -499,6 +659,10 @@ def make_policy(
         action_names = ds_meta.features.get(ACTION, {}).get("names")
         if action_names is not None:
             cfg.action_feature_names = list(action_names)
+    if ds_meta is not None:
+        set_dataset_feature_metadata = getattr(cfg, "set_dataset_feature_metadata", None)
+        if callable(set_dataset_feature_metadata):
+            set_dataset_feature_metadata(ds_meta.features)
 
     kwargs["config"] = cfg
 

src/lerobot/policies/groot/eagle2_hg_model/modeling_eagle2_5_vl.py

@@ -60,6 +60,7 @@ class Eagle25VLPreTrainedModel(PreTrainedModel):
         "SiglipEncoderLayer",
     ]
     _skip_keys_device_placement = "past_key_values"
+    _supports_flash_attn = True
     _supports_flash_attn_2 = True
     _supports_cache_class = True
     _supports_static_cache = True

src/lerobot/policies/groot/eagle2_hg_model/processing_eagle2_5_vl.py

@@ -124,7 +124,6 @@ class Eagle25VLProcessor(ProcessorMixin):
         "videos_kwargs",
         "text_kwargs",
     ]
-    image_processor_class = "AutoImageProcessor"
     tokenizer_class = "AutoTokenizer"
 
     def __init__(

src/lerobot/policies/groot/groot_n1.py

@@ -14,7 +14,7 @@
 # limitations under the License.
 
 from pathlib import Path
-from typing import TYPE_CHECKING
+from typing import TYPE_CHECKING, Any
 
 import numpy as np
 import torch
@@ -26,9 +26,14 @@ from lerobot.utils.import_utils import _transformers_available
 
 # Conditional import for type checking and lazy loading
 if TYPE_CHECKING or _transformers_available:
+    from huggingface_hub.dataclasses import strict
     from transformers import AutoConfig, AutoModel, PretrainedConfig, PreTrainedModel
     from transformers.feature_extraction_utils import BatchFeature
 else:
+
+    def strict(cls):
+        return cls
+
     AutoConfig = None
     AutoModel = None
     PretrainedConfig = object
@@ -176,16 +181,16 @@ N_COLOR_CHANNELS = 3
 class GR00TN15Config(PretrainedConfig):
     model_type = "gr00t_n1_5"
 
-    backbone_cfg: dict
-    action_head_cfg: dict
-    action_horizon: int
-    action_dim: int
+    backbone_cfg: dict[str, Any] | None = None
+    action_head_cfg: dict[str, Any] | None = None
+    action_horizon: int = 0
+    action_dim: int = 0
     compute_dtype: str = "float32"
 
-    def __init__(self, **kwargs):
-        super().__init__(**kwargs)
-        for key, value in kwargs.items():
-            setattr(self, key, value)
+    def __post_init__(self, **kwargs):
+        self.backbone_cfg = {} if self.backbone_cfg is None else self.backbone_cfg
+        self.action_head_cfg = {} if self.action_head_cfg is None else self.action_head_cfg
+        super().__post_init__(**kwargs)
 
 
 # real model

src/lerobot/policies/groot/processor_groot.py

@@ -206,7 +206,11 @@ def _build_eagle_processor(tokenizer_assets_repo: str = DEFAULT_TOKENIZER_ASSETS
             "Vendor files are copied during model creation. Create the policy/model first, "
             "or call ensure_eagle_cache_ready() before building processors."
         )
-    proc = AutoProcessor.from_pretrained(str(cache_dir), trust_remote_code=True, use_fast=True)
+    proc = AutoProcessor.from_pretrained(
+        str(cache_dir),
+        trust_remote_code=True,
+        fix_mistral_regex=False,
+    )
     proc.tokenizer.padding_side = "left"
     return proc
 

src/lerobot/policies/molmoact2/README.md

@@ -0,0 +1 @@
+../../../../docs/source/policy_molmoact2_README.md

src/lerobot/policies/molmoact2/__init__.py

@@ -0,0 +1,21 @@
+#!/usr/bin/env python
+
+# Copyright 2026 The Allen Institute for Artificial Intelligence and The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from .configuration_molmoact2 import MolmoAct2Config
+from .modeling_molmoact2 import MolmoAct2Policy
+from .processor_molmoact2 import make_molmoact2_pre_post_processors
+
+__all__ = ["MolmoAct2Config", "MolmoAct2Policy", "make_molmoact2_pre_post_processors"]

src/lerobot/policies/molmoact2/configuration_molmoact2.py

@@ -0,0 +1,519 @@
+#!/usr/bin/env python
+
+# Copyright 2026 The Allen Institute for Artificial Intelligence and The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from __future__ import annotations
+
+import json
+import math
+import os
+from contextlib import suppress
+from dataclasses import dataclass, field
+from pathlib import Path
+from typing import Any
+
+from huggingface_hub import snapshot_download
+
+from lerobot.configs import FeatureType, NormalizationMode, PolicyFeature, PreTrainedConfig
+from lerobot.optim import (
+    AdamWConfig,
+    CosineDecayWithWarmupSchedulerConfig,
+    LRSchedulerConfig,
+    OptimizerConfig,
+)
+from lerobot.utils.constants import ACTION, OBS_STATE
+
+from ..rtc.configuration_rtc import RTCConfig
+
+MOLMOACT2_DEFAULT_NUM_IMAGES = 2
+MOLMOACT2_IMAGE_TOKENS_PER_IMAGE = 196
+MOLMOACT2_FIXED_PROMPT_TOKEN_BUDGET = 80
+MOLMOACT2_TASK_TOKEN_BUDGET = 32
+MOLMOACT2_SEQUENCE_LENGTH_MARGIN = 32
+MOLMOACT2_SEQUENCE_LENGTH_MULTIPLE = 64
+MOLMOACT2_DISCRETE_ACTION_WRAPPER_TOKENS = 4
+MOLMOACT2_MIN_DISCRETE_ACTION_TOKENS_PER_STEP = 6
+MOLMOACT2_DISCRETE_ACTION_TOKENS_PER_DIM = 0.95
+
+
+def _hf_token() -> str | None:
+    return os.environ.get("HF_TOKEN") or os.environ.get("HF_ACCESS_TOKEN")
+
+
+def _resolve_checkpoint_location(
+    checkpoint_path: str,
+    *,
+    revision: str | None = None,
+    force_download: bool = False,
+) -> str:
+    checkpoint_path = str(checkpoint_path or "").strip()
+    if not checkpoint_path:
+        raise ValueError("MolmoAct2 policy requires `checkpoint_path`.")
+    local_path = Path(checkpoint_path).expanduser()
+    if local_path.exists():
+        return str(local_path)
+    return snapshot_download(
+        repo_id=checkpoint_path,
+        repo_type="model",
+        revision=revision,
+        force_download=force_download,
+        ignore_patterns=["*.py", "*.pyc", "__pycache__/*"],
+        token=_hf_token(),
+    )
+
+
+def _load_hf_norm_metadata_for_tag(
+    checkpoint_path: str,
+    *,
+    revision: str | None,
+    force_download: bool,
+    norm_tag: str | None,
+) -> dict[str, Any]:
+    norm_tag = str(norm_tag or "").strip()
+    if not norm_tag:
+        return {}
+    checkpoint_location = Path(
+        _resolve_checkpoint_location(
+            checkpoint_path,
+            revision=revision,
+            force_download=force_download,
+        )
+    )
+    norm_stats_filename = "norm_stats.json"
+    config_path = checkpoint_location / "config.json"
+    if config_path.exists():
+        with suppress(OSError, json.JSONDecodeError):
+            norm_stats_filename = str(
+                json.loads(config_path.read_text()).get("norm_stats_filename") or norm_stats_filename
+            )
+    stats_path = checkpoint_location / norm_stats_filename
+    if not stats_path.exists():
+        raise FileNotFoundError(
+            f"MolmoAct2 HF checkpoint is missing {norm_stats_filename!r}; cannot resolve norm_tag={norm_tag!r}."
+        )
+    payload = json.loads(stats_path.read_text())
+    metadata_by_tag = payload.get("metadata_by_tag")
+    if not isinstance(metadata_by_tag, dict):
+        raise ValueError(f"MolmoAct2 norm stats file {stats_path} has no metadata_by_tag mapping.")
+    metadata = metadata_by_tag.get(norm_tag)
+    if not isinstance(metadata, dict):
+        available = sorted(str(tag) for tag in metadata_by_tag)
+        raise ValueError(f"Unknown MolmoAct2 norm_tag={norm_tag!r}. Available tags: {available}.")
+    return metadata
+
+
+@LRSchedulerConfig.register_subclass("molmoact2_cosine_decay_with_warmup")
+@dataclass
+class MolmoAct2CosineDecayWithWarmupSchedulerConfig(CosineDecayWithWarmupSchedulerConfig):
+    """MolmoAct2-local cosine scheduler with optional decay-step auto-match.
+
+    LeRobot's generic cosine scheduler keeps an explicit integer decay length.
+    For MolmoAct2, leaving num_decay_steps unset means "decay across this run's
+    training steps"; build() is the first point where num_training_steps is known.
+    """
+
+    num_decay_steps: int | None
+
+    def build(self, optimizer, num_training_steps: int):
+        return CosineDecayWithWarmupSchedulerConfig(
+            peak_lr=self.peak_lr,
+            decay_lr=self.decay_lr,
+            num_warmup_steps=self.num_warmup_steps,
+            num_decay_steps=num_training_steps if self.num_decay_steps is None else self.num_decay_steps,
+        ).build(optimizer, num_training_steps=num_training_steps)
+
+
+def _round_up(value: int, multiple: int) -> int:
+    return int(math.ceil(value / multiple) * multiple)
+
+
+def infer_molmoact2_max_sequence_length(
+    *,
+    num_images: int,
+    state_dim: int,
+    action_dim: int,
+    action_horizon: int,
+    include_discrete_action: bool,
+) -> int:
+    """Infer the padded text/image sequence cap from MolmoAct2's fixed token layout."""
+    if num_images < 1:
+        num_images = MOLMOACT2_DEFAULT_NUM_IMAGES
+    if state_dim < 0:
+        state_dim = 0
+    if action_dim < 1:
+        action_dim = 1
+    if action_horizon < 1:
+        action_horizon = 1
+
+    image_tokens = num_images * MOLMOACT2_IMAGE_TOKENS_PER_IMAGE
+    prompt_tokens = (
+        MOLMOACT2_FIXED_PROMPT_TOKEN_BUDGET
+        + MOLMOACT2_TASK_TOKEN_BUDGET
+        + state_dim
+        + MOLMOACT2_SEQUENCE_LENGTH_MARGIN
+    )
+    action_tokens = 0
+    if include_discrete_action:
+        action_tokens_per_step = max(
+            MOLMOACT2_MIN_DISCRETE_ACTION_TOKENS_PER_STEP,
+            math.ceil(action_dim * MOLMOACT2_DISCRETE_ACTION_TOKENS_PER_DIM),
+        )
+        action_tokens = MOLMOACT2_DISCRETE_ACTION_WRAPPER_TOKENS + action_horizon * action_tokens_per_step
+
+    return _round_up(
+        image_tokens + prompt_tokens + action_tokens,
+        MOLMOACT2_SEQUENCE_LENGTH_MULTIPLE,
+    )
+
+
+@PreTrainedConfig.register_subclass("molmoact2")
+@dataclass
+class MolmoAct2Config(PreTrainedConfig):
+    """MolmoAct2 policy backed by the converted HF checkpoint implementation."""
+
+    checkpoint_path: str = "allenai/MolmoAct2"
+    checkpoint_revision: str | None = None
+    checkpoint_force_download: bool = False
+
+    n_obs_steps: int = 1
+    chunk_size: int = 30
+    n_action_steps: int = 30
+
+    action_mode: str = "both"
+    inference_action_mode: str | None = None
+    discrete_action_tokenizer: str = "allenai/MolmoAct2-FAST-Tokenizer"
+    discrete_generation_max_steps: int | None = None
+    norm_tag: str | None = None
+
+    setup_type: str = ""
+    control_mode: str = ""
+    image_keys: list[str] = field(default_factory=list)
+    normalize_language: bool = True
+    add_setup_tokens: bool = True
+    add_control_tokens: bool = True
+    normalize_gripper: bool = False
+    num_state_tokens: int = 256
+    # Leave unset for the default MolmoAct2 sequence budget inferred from the fixed
+    # image/prompt/state/action token layout. Override only for unusual long prompts.
+    max_sequence_length: int | None = None
+
+    # Fixed by released MolmoAct2 checkpoints. We validate this at model load.
+    expected_max_action_dim: int = 32
+
+    # Flow-matching training knobs copied from the original MolmoAct2 training path.
+    num_flow_timesteps: int = 8
+    flow_matching_cutoff: float = 1.0
+    flow_matching_time_offset: float = 0.001
+    flow_matching_time_scale: float = 0.999
+    flow_matching_beta_alpha: float = 1.0
+    flow_matching_beta_beta: float = 1.5
+    num_inference_steps: int | None = None
+    mask_action_dim_padding: bool = True
+    enable_inference_cuda_graph: bool = True
+    # MolmoAct2-local eval option. When enabled, stochastic continuous action
+    # generation uses a rollout-local generator derived from eval_seed.
+    per_episode_seed: bool = False
+    eval_seed: int | None = None
+    rtc_config: RTCConfig | None = None
+
+    # Default is full finetuning with gradients from the action expert flowing into the VLM.
+    enable_lora_vlm: bool = False
+    lora_rank: int = 64
+    lora_alpha: int = 16
+    lora_dropout: float = 0.05
+    lora_bias: str = "none"
+    enable_lora_action_expert: bool = False
+    enable_knowledge_insulation: bool = False
+    freeze_embedding: bool = True
+    train_action_expert_only: bool = False
+    gradient_checkpointing: bool = False
+
+    model_dtype: str = "bfloat16"
+    softmax_auxiliary_loss: bool = True
+    softmax_auxiliary_loss_scale: float = 1e-4
+    discrete_loss_token_weighting: str = "root_subsegments_root_tokens"
+
+    optimizer_lr: float = 1e-5
+    optimizer_vit_lr: float = 5e-6
+    optimizer_connector_lr: float = 5e-6
+    optimizer_action_expert_lr: float = 5e-5
+    optimizer_betas: tuple[float, float] = (0.9, 0.95)
+    optimizer_eps: float = 1e-6
+    optimizer_weight_decay: float = 0.0
+    optimizer_grad_clip_norm: float = 1.0
+
+    scheduler_warmup_steps: int = 200
+    scheduler_decay_steps: int | None = None
+    scheduler_decay_lr: float = 1e-6
+
+    normalization_mapping: dict[str, NormalizationMode] = field(
+        default_factory=lambda: {
+            "VISUAL": NormalizationMode.IDENTITY,
+            "STATE": NormalizationMode.QUANTILES,
+            "ACTION": NormalizationMode.QUANTILES,
+        }
+    )
+
+    input_features: dict[str, PolicyFeature] = field(default_factory=dict)
+    output_features: dict[str, PolicyFeature] = field(default_factory=dict)
+    dataset_feature_names: dict[str, Any] = field(default_factory=dict)
+
+    def __post_init__(self) -> None:
+        super().__post_init__()
+        if self.action_mode not in {"continuous", "discrete", "both"}:
+            raise ValueError(
+                f"Unsupported action_mode={self.action_mode!r}. "
+                "Expected one of {'continuous', 'discrete', 'both'}."
+            )
+        if self.inference_action_mode not in {None, "continuous", "discrete"}:
+            raise ValueError(
+                f"Unsupported inference_action_mode={self.inference_action_mode!r}. "
+                "Expected one of {None, 'continuous', 'discrete'}."
+            )
+        if self.inference_action_mode == "continuous" and self.action_mode == "discrete":
+            raise ValueError("MolmoAct2 action_mode='discrete' cannot run continuous inference.")
+        if self.inference_action_mode == "discrete" and self.action_mode == "continuous":
+            raise ValueError("MolmoAct2 action_mode='continuous' cannot run discrete inference.")
+        if self.train_action_expert_only and self.action_mode != "continuous":
+            raise ValueError("MolmoAct2 train_action_expert_only requires action_mode='continuous'.")
+        if self.train_action_expert_only and self.enable_lora_vlm:
+            raise ValueError("MolmoAct2 train_action_expert_only is incompatible with enable_lora_vlm.")
+        if self.enable_lora_action_expert and not self.enable_lora_vlm:
+            raise ValueError("MolmoAct2 enable_lora_action_expert requires enable_lora_vlm.")
+        if self.chunk_size < 1:
+            raise ValueError(f"chunk_size must be >= 1, got {self.chunk_size}.")
+        if self.n_action_steps < 1:
+            raise ValueError(f"n_action_steps must be >= 1, got {self.n_action_steps}.")
+        if self.n_action_steps > self.chunk_size:
+            raise ValueError(
+                f"n_action_steps ({self.n_action_steps}) cannot exceed chunk_size ({self.chunk_size})."
+            )
+        if self.expected_max_action_dim != 32:
+            raise ValueError("MolmoAct2 released checkpoints use expected_max_action_dim=32.")
+        if self.model_dtype not in {"float32", "bfloat16", "float16"}:
+            raise ValueError(
+                f"Unsupported model_dtype={self.model_dtype!r}. Expected 'float32', 'bfloat16', or 'float16'."
+            )
+        if self.lora_rank < 1:
+            raise ValueError(f"lora_rank must be >= 1, got {self.lora_rank}.")
+        if self.lora_alpha < 1:
+            raise ValueError(f"lora_alpha must be >= 1, got {self.lora_alpha}.")
+        if not 0 <= self.lora_dropout <= 1:
+            raise ValueError(f"lora_dropout must be in [0, 1], got {self.lora_dropout}.")
+        if self.lora_bias not in {"none", "all", "lora_only"}:
+            raise ValueError(
+                f"Unsupported lora_bias={self.lora_bias!r}. Expected one of 'none', 'all', or 'lora_only'."
+            )
+        if self.discrete_loss_token_weighting not in {
+            "none",
+            "token",
+            "root_tokens",
+            "root_subsegments",
+            "root_subsegments_root_tokens",
+        }:
+            raise ValueError(
+                f"Unsupported discrete_loss_token_weighting={self.discrete_loss_token_weighting!r}."
+            )
+        if self.discrete_generation_max_steps is not None and self.discrete_generation_max_steps < 1:
+            raise ValueError(
+                f"discrete_generation_max_steps must be >= 1 or None, got {self.discrete_generation_max_steps}."
+            )
+        if self.max_sequence_length is not None and self.max_sequence_length < 1:
+            raise ValueError(f"max_sequence_length must be >= 1 or None, got {self.max_sequence_length}.")
+
+    def inferred_max_sequence_length(
+        self,
+        *,
+        num_images: int | None = None,
+        state_dim: int | None = None,
+        action_dim: int | None = None,
+        action_horizon: int | None = None,
+        include_discrete_action: bool | None = None,
+    ) -> int:
+        if self.max_sequence_length is not None:
+            return int(self.max_sequence_length)
+
+        if num_images is None:
+            num_images = len(self.image_keys) or len(self.image_features) or MOLMOACT2_DEFAULT_NUM_IMAGES
+        if state_dim is None:
+            state_feature = self.robot_state_feature
+            state_dim = int(state_feature.shape[0]) if state_feature is not None else 0
+        if action_dim is None:
+            action_feature = self.action_feature
+            action_dim = (
+                int(action_feature.shape[0]) if action_feature is not None else self.expected_max_action_dim
+            )
+        if action_horizon is None:
+            action_horizon = self.chunk_size
+        if include_discrete_action is None:
+            include_discrete_action = self.action_mode in {"discrete", "both"}
+
+        return infer_molmoact2_max_sequence_length(
+            num_images=int(num_images),
+            state_dim=int(state_dim),
+            action_dim=int(action_dim),
+            action_horizon=int(action_horizon),
+            include_discrete_action=bool(include_discrete_action),
+        )
+
+    @property
+    def observation_delta_indices(self) -> None:
+        return None
+
+    @property
+    def action_delta_indices(self) -> list[int]:
+        return list(range(self.chunk_size))
+
+    @property
+    def reward_delta_indices(self) -> None:
+        return None
+
+    def get_optimizer_preset(self) -> OptimizerConfig:
+        return AdamWConfig(
+            lr=self.optimizer_lr,
+            betas=self.optimizer_betas,
+            eps=self.optimizer_eps,
+            weight_decay=self.optimizer_weight_decay,
+            grad_clip_norm=self.optimizer_grad_clip_norm,
+        )
+
+    def get_scheduler_preset(self) -> LRSchedulerConfig | None:
+        return MolmoAct2CosineDecayWithWarmupSchedulerConfig(
+            peak_lr=self.optimizer_lr,
+            decay_lr=self.scheduler_decay_lr,
+            num_warmup_steps=self.scheduler_warmup_steps,
+            num_decay_steps=self.scheduler_decay_steps,
+        )
+
+    def set_dataset_feature_metadata(self, features: dict[str, Any]) -> None:
+        self.dataset_feature_names = {}
+        for key in (ACTION, OBS_STATE):
+            feature = features.get(key) if isinstance(features, dict) else None
+            if isinstance(feature, dict) and feature.get("names") is not None:
+                self.dataset_feature_names[key] = feature["names"]
+
+    def validate_features(self) -> None:
+        """Validate and set up MolmoAct2 input and output features."""
+        image_features = [key for key, feat in self.input_features.items() if feat.type == FeatureType.VISUAL]
+        if not image_features:
+            raise ValueError(
+                "MolmoAct2 policy requires at least one visual input feature. "
+                "No features of type FeatureType.VISUAL found in input_features."
+            )
+
+        if OBS_STATE not in self.input_features:
+            state_feature = PolicyFeature(
+                type=FeatureType.STATE,
+                shape=(0,),
+            )
+            self.input_features[OBS_STATE] = state_feature
+
+        if ACTION not in self.output_features:
+            action_feature = PolicyFeature(
+                type=FeatureType.ACTION,
+                shape=(self.expected_max_action_dim,),
+            )
+            self.output_features[ACTION] = action_feature
+
+    def apply_norm_tag_metadata(self) -> None:
+        if not str(self.norm_tag or "").strip():
+            return
+        metadata = _load_hf_norm_metadata_for_tag(
+            self.checkpoint_path,
+            revision=self.checkpoint_revision,
+            force_download=bool(self.checkpoint_force_download),
+            norm_tag=self.norm_tag,
+        )
+        if metadata.get("action_horizon") is not None:
+            self.chunk_size = int(metadata["action_horizon"])
+        if metadata.get("n_action_steps") is not None:
+            self.n_action_steps = int(metadata["n_action_steps"])
+        if not self.setup_type and metadata.get("setup_type") is not None:
+            self.setup_type = str(metadata["setup_type"])
+        if not self.control_mode and metadata.get("control_mode") is not None:
+            self.control_mode = str(metadata["control_mode"])
+
+    def saved_policy_action_mode(self) -> str | None:
+        pretrained_path = getattr(self, "pretrained_path", None)
+        if pretrained_path is None:
+            return None
+        config_path = Path(pretrained_path) / "config.json"
+        if not config_path.exists():
+            return None
+        try:
+            mode = json.loads(config_path.read_text()).get("action_mode")
+        except (OSError, json.JSONDecodeError):
+            return None
+        if mode in {"continuous", "discrete", "both"}:
+            return str(mode)
+        return None
+
+    def training_action_mode(self, saved_policy_action_mode: str | None = None) -> str:
+        return saved_policy_action_mode or self.action_mode
+
+    def validate_inference_action_mode(self, saved_policy_action_mode: str | None = None) -> None:
+        requested_mode = self.inference_action_mode
+        if requested_mode is None:
+            return
+        training_mode = self.training_action_mode(saved_policy_action_mode)
+        if requested_mode == "continuous" and training_mode == "discrete":
+            raise ValueError(
+                "MolmoAct2 checkpoint was trained with action_mode='discrete' and cannot run "
+                "continuous inference."
+            )
+        if requested_mode == "discrete" and training_mode == "continuous":
+            raise ValueError(
+                "MolmoAct2 checkpoint was trained with action_mode='continuous' and cannot run "
+                "discrete inference. Train with action_mode='both' or action_mode='discrete' first."
+            )
+
+    def validate_checkpoint_action_mode(
+        self,
+        checkpoint_action_mode: str,
+        *,
+        has_action_expert: bool,
+    ) -> None:
+        if self.action_mode == "both" and checkpoint_action_mode != "both":
+            raise ValueError(
+                f"action_mode='both' requires checkpoint action_mode='both', got {checkpoint_action_mode!r}."
+            )
+        if self.action_mode == "discrete" and checkpoint_action_mode not in {"discrete", "both"}:
+            raise ValueError(
+                f"action_mode='discrete' requires checkpoint action_mode in {{'discrete', 'both'}}, "
+                f"got {checkpoint_action_mode!r}."
+            )
+        if self.action_mode in {"continuous", "both"} and not has_action_expert:
+            raise ValueError("Continuous MolmoAct2 training requires an action expert checkpoint.")
+
+    def resolve_inference_action_mode(
+        self,
+        requested_mode: str | None,
+        saved_policy_action_mode: str | None = None,
+    ) -> str:
+        training_mode = self.training_action_mode(saved_policy_action_mode)
+        if requested_mode is None:
+            requested_mode = self.inference_action_mode
+        if requested_mode is None:
+            raise ValueError(
+                "MolmoAct2 inference requires `inference_action_mode` to be set explicitly "
+                "to either 'continuous' or 'discrete'."
+            )
+        if requested_mode not in {"continuous", "discrete"}:
+            raise ValueError("MolmoAct2 inference_action_mode must be either 'continuous' or 'discrete'.")
+        if requested_mode == "continuous" and training_mode == "discrete":
+            raise ValueError("MolmoAct2 action_mode='discrete' checkpoint cannot run continuous inference.")
+        if requested_mode == "discrete" and training_mode == "continuous":
+            raise ValueError("MolmoAct2 action_mode='continuous' checkpoint cannot run discrete inference.")
+        return requested_mode

src/lerobot/policies/molmoact2/hf_model/__init__.py

@@ -0,0 +1,17 @@
+#!/usr/bin/env python
+
+# Copyright 2026 The Allen Institute for Artificial Intelligence and The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+# ruff: noqa

src/lerobot/policies/molmoact2/hf_model/action_tokenizer.py

@@ -0,0 +1,237 @@
+#!/usr/bin/env python
+
+# Copyright 2026 The Allen Institute for Artificial Intelligence and The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+# ruff: noqa
+
+import logging
+import os
+from pathlib import Path
+from typing import ClassVar
+
+import numpy as np
+from tokenizers import ByteLevelBPETokenizer
+from tokenizers.trainers import BpeTrainer
+from huggingface_hub import snapshot_download
+from transformers import PreTrainedTokenizerFast
+from transformers.processing_utils import ProcessorMixin
+
+
+def _hf_token() -> str | None:
+    return os.environ.get("HF_TOKEN") or os.environ.get("HF_ACCESS_TOKEN")
+
+
+def _resolve_tokenizer_location(
+    tokenizer_path: str,
+    *,
+    revision: str | None = None,
+    force_download: bool = False,
+) -> str:
+    local_path = Path(str(tokenizer_path)).expanduser()
+    if local_path.exists():
+        return str(local_path)
+    return snapshot_download(
+        repo_id=str(tokenizer_path),
+        repo_type="model",
+        revision=revision,
+        force_download=force_download,
+        ignore_patterns=["*.py", "*.pyc", "__pycache__/*"],
+        token=_hf_token(),
+    )
+
+
+class UniversalActionProcessor(ProcessorMixin):
+    attributes: ClassVar[list[str]] = ["tokenizer"]
+    tokenizer_class: str = "AutoTokenizer"
+
+    def __init__(
+        self,
+        tokenizer: PreTrainedTokenizerFast,
+        scale: float = 10,
+        vocab_size: int = 1024,
+        min_token: int = 0,
+        *,
+        action_dim: int | None = None,
+        time_horizon: int | None = None,
+    ):
+        self.scale = scale
+        self.vocab_size = vocab_size
+        self.min_token = min_token
+
+        # Action horizon and dimension needed during decoding. These can be specified
+        # in three ways (in order of priority):
+        # 1. passed in as kwargs to decode()
+        # 2. in the constructor
+        # 3. cached from the last time decode() was called
+        self.time_horizon = time_horizon
+        self.action_dim = action_dim
+        self.called_time_horizon = time_horizon
+        self.called_action_dim = action_dim
+
+        super().__init__(tokenizer)
+        self.bpe_tokenizer = self.tokenizer
+
+    def __call__(self, action_chunk: np.array) -> np.array:
+        from scipy.fft import dct
+
+        assert action_chunk.ndim <= 3, "Only 3 dimensions supported: [batch, timesteps, action_dim]"
+        if action_chunk.ndim == 2:
+            action_chunk = action_chunk[None, ...]
+
+        # Cache the time horizon and action dimension for decoding
+        self.called_time_horizon = action_chunk.shape[-2]
+        self.called_action_dim = action_chunk.shape[-1]
+
+        dct_coeff = dct(action_chunk, axis=1, norm="ortho")
+        dct_coeff = np.around(dct_coeff * self.scale)
+        tokens = []
+        for elem in dct_coeff:
+            token_str = "".join(map(chr, np.maximum(elem.flatten() - self.min_token, 0).astype(int)))
+            tokens.append(self.bpe_tokenizer(token_str)["input_ids"])
+        return tokens
+
+    def decode(
+        self,
+        tokens: list[list[int]],
+        *,
+        time_horizon: int | None = None,
+        action_dim: int | None = None,
+    ) -> np.array:
+        from scipy.fft import idct
+
+        self.time_horizon = time_horizon or self.time_horizon or self.called_time_horizon
+        self.action_dim = action_dim or self.action_dim or self.called_action_dim
+
+        # Cache the time horizon and action dimension for the next call
+        self.called_time_horizon = self.time_horizon
+        self.called_action_dim = self.action_dim
+
+        assert self.time_horizon is not None and self.action_dim is not None, (
+            "Tokenizer not initialized, call encode() once or pass in time_horizon and action_dim."
+        )
+
+        decoded_actions = []
+        for token in tokens:
+            try:
+                decoded_tokens = self.bpe_tokenizer.decode(token)
+                decoded_dct_coeff = np.array(list(map(ord, decoded_tokens))) + self.min_token
+                decoded_dct_coeff = decoded_dct_coeff.reshape(-1, self.action_dim)
+                assert decoded_dct_coeff.shape == (
+                    self.time_horizon,
+                    self.action_dim,
+                ), (
+                    f"Decoded DCT coefficients have shape {decoded_dct_coeff.shape}, expected ({self.time_horizon}, {self.action_dim})"
+                )
+            except Exception as e:
+                print(f"Error decoding tokens: {e}")
+                print(f"Tokens: {token}")
+                decoded_dct_coeff = np.zeros((self.time_horizon, self.action_dim))
+            decoded_actions.append(idct(decoded_dct_coeff / self.scale, axis=0, norm="ortho"))
+        return np.stack(decoded_actions)
+
+    @classmethod
+    def fit(
+        cls,
+        action_data: list[np.array],
+        scale: float = 10,
+        vocab_size: int = 1024,
+        *,
+        time_horizon: int | None = None,
+        action_dim: int | None = None,
+    ) -> "UniversalActionProcessor":
+        from scipy.fft import dct
+
+        # Run DCT over all inputs
+        dct_tokens = [dct(a, axis=0, norm="ortho").flatten() for a in action_data]
+
+        # Quantize and find min token
+        max_token = int(np.around(np.concatenate(dct_tokens) * scale).max())
+        min_token = int(np.around(np.concatenate(dct_tokens) * scale).min())
+        min_vocab_size = max_token - min_token
+
+        assert min_vocab_size <= vocab_size, (
+            f"Vocab size {vocab_size} is too small for the range of tokens {min_vocab_size}"
+        )
+        if min_vocab_size + 100 > vocab_size:
+            logging.warning(
+                f"Initial alphabet size {min_vocab_size} is almost as large as the vocab"
+                f"size {vocab_size}, consider increasing vocab size"
+            )
+
+        # Make token iterator for BPE training
+        def _token_iter():
+            for tokens in dct_tokens:
+                rounded_tokens = np.around(tokens * scale) - min_token
+                rounded_tokens = rounded_tokens.astype(int)
+                string = "".join(map(chr, rounded_tokens))
+                yield string
+
+        # Train BPE tokenizer
+        bpe = ByteLevelBPETokenizer()
+
+        # Set up the entire range of possible tokens as the initial alphabet
+        alphabet = [chr(i) for i in range(max_token - min_token + 1)]
+        trainer = BpeTrainer(
+            vocab_size=vocab_size,
+            min_frequency=2,
+            show_progress=True,
+            special_tokens=[],
+            initial_alphabet=alphabet,
+            max_token_length=10000,
+        )
+
+        # Train the inner tokenizer (don't use ByteLevelBPETokenizer.train_from_iterator()
+        # because it doesn't support custom alphabets)
+        bpe._tokenizer.train_from_iterator(_token_iter(), trainer=trainer)
+
+        return cls(
+            PreTrainedTokenizerFast(tokenizer_object=bpe, clean_up_tokenization_spaces=False),
+            scale=scale,
+            vocab_size=vocab_size,
+            min_token=min_token,
+            time_horizon=time_horizon,
+            action_dim=action_dim,
+        )
+
+    @classmethod
+    def from_pretrained_local(
+        cls,
+        pretrained_model_name_or_path: str,
+        *,
+        revision: str | None = None,
+        force_download: bool = False,
+    ) -> "UniversalActionProcessor":
+        location = Path(
+            _resolve_tokenizer_location(
+                pretrained_model_name_or_path,
+                revision=revision,
+                force_download=force_download,
+            )
+        )
+        processor_config = {}
+        processor_config_path = location / "processor_config.json"
+        if processor_config_path.exists():
+            import json
+
+            processor_config = json.loads(processor_config_path.read_text())
+        tokenizer = PreTrainedTokenizerFast.from_pretrained(str(location))
+        return cls(
+            tokenizer,
+            scale=processor_config.get("scale", 10),
+            vocab_size=processor_config.get("vocab_size", 1024),
+            min_token=processor_config.get("min_token", 0),
+            action_dim=processor_config.get("action_dim"),
+            time_horizon=processor_config.get("time_horizon"),
+        )

src/lerobot/policies/molmoact2/hf_model/configuration_molmoact2.py

@@ -0,0 +1,553 @@
+#!/usr/bin/env python
+
+# Copyright 2026 The Allen Institute for Artificial Intelligence and The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+# ruff: noqa
+
+"""
+MolmoAct2 configuration
+"""
+
+from typing import Optional, Any
+
+from transformers import PretrainedConfig
+from transformers.modeling_rope_utils import rope_config_validation
+from transformers.utils import logging
+
+logger = logging.get_logger(__name__)
+
+
+class MolmoAct2VitConfig(PretrainedConfig):
+    r"""
+    This is the configuration class to store the configuration of a [`MolmoAct2VisionTransformer`].
+    It is used to instantiate a `MolmoAct2VisionTransformer` according to the specified arguments,
+    defining the model architecture.
+
+    Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
+    documentation from [`PretrainedConfig`] for more information.
+
+    Example:
+    ```python
+    >>> from transformers import MolmoAct2VitConfig, MolmoAct2VisionTransformer
+
+    >>> # Initializing a MolmoAct2VitConfig
+    >>> configuration = MolmoAct2VitConfig()
+
+    >>> # Initializing a MolmoAct2VisionTransformer (with random weights)
+    >>> model = MolmoAct2VisionTransformer(configuration)
+
+    >>> # Accessing the model configuration
+    >>> configuration = model.config
+    ```"""
+
+    model_type = "molmoact2"
+    base_config_key = "vit_config"
+
+    def __init__(
+        self,
+        hidden_size: int = 1152,
+        intermediate_size: int = 4304,
+        num_hidden_layers: int = 27,
+        num_attention_heads: int = 16,
+        num_key_value_heads: int = 16,
+        head_dim: int = 72,
+        hidden_act: str = "gelu_pytorch_tanh",
+        layer_norm_eps: float = 1e-6,
+        image_default_input_size: tuple[int, int] = (378, 378),
+        image_patch_size: int = 14,
+        image_num_pos: int = 577,
+        attention_dropout: float = 0.0,
+        residual_dropout: float = 0.0,
+        initializer_range: float = 0.02,
+        float32_attention: bool = True,
+        attn_implementation: str = "eager",
+        **kwargs,
+    ):
+        self.attn_implementation = attn_implementation
+        super().__init__(attn_implementation=attn_implementation, **kwargs)
+        self.hidden_size = hidden_size
+        self.intermediate_size = intermediate_size
+        self.num_hidden_layers = num_hidden_layers
+        self.num_attention_heads = num_attention_heads
+        self.num_key_value_heads = num_key_value_heads
+        self.head_dim = head_dim
+        self.hidden_act = hidden_act
+        self.layer_norm_eps = layer_norm_eps
+        self.image_default_input_size = image_default_input_size
+        self.image_patch_size = image_patch_size
+        self.image_num_pos = image_num_pos
+        self.attention_dropout = attention_dropout
+        self.residual_dropout = residual_dropout
+        self.initializer_range = initializer_range
+        self.float32_attention = float32_attention
+
+    @property
+    def image_num_patch(self):
+        h, w = self.image_default_input_size
+        return h // self.image_patch_size, w // self.image_patch_size
+
+
+class MolmoAct2AdapterConfig(PretrainedConfig):
+    r"""
+    This is the configuration class to store the configuration of MolmoAct2Adapter. With MolmoAct2VitConfig,
+    It is used to instantiate an MolmoAct2VisionBackbone according to the specified arguments,
+    defining the model architecture.
+
+    Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
+    documentation from [`PretrainedConfig`] for more information.
+
+    Example:
+
+    ```python
+    >>> from transformers import MolmoAct2VitConfig, MolmoAct2AdapterConfig, MolmoAct2VisionBackbone
+
+    >>> # Initializing a MolmoAct2VitConfig and a MolmoAct2AdapterConfig
+    >>> vit_config = MolmoAct2VitConfig()
+    >>> adapter_config = MolmoPoolingConfig()
+
+    >>> # Initializing a MolmoAct2VisionBackbone (with random weights)
+    >>> model = MolmoAct2VisionBackbone(vit_config, adapter_config)
+
+    >>> # Accessing the model configuration
+    >>> vit_configuration = model.vit_config
+    >>> adapter_configuration = model.adapter_config
+    ```"""
+
+    model_type = "molmoact2"
+    base_config_key = "adapter_config"
+
+    def __init__(
+        self,
+        vit_layers: tuple = (-3, -9),
+        pooling_attention_mask: bool = False,
+        hidden_size: int = 1152,
+        num_attention_heads: int = 16,
+        num_key_value_heads: int = 16,
+        head_dim: int = 72,
+        float32_attention: bool = True,
+        attention_dropout: float = 0.0,
+        residual_dropout: float = 0.0,
+        hidden_act: str = "silu",
+        intermediate_size: int = 18944,
+        text_hidden_size: int = 3584,
+        image_feature_dropout: float = 0.0,
+        initializer_range: float = 0.02,
+        attn_implementation: str = "eager",
+        **kwargs,
+    ):
+        self.attn_implementation = attn_implementation
+        super().__init__(attn_implementation=attn_implementation, **kwargs)
+        self.vit_layers = vit_layers
+        self.pooling_attention_mask = pooling_attention_mask
+        self.hidden_size = hidden_size
+        self.num_attention_heads = num_attention_heads
+        self.num_key_value_heads = num_key_value_heads
+        self.head_dim = head_dim
+        self.float32_attention = float32_attention
+        self.attention_dropout = attention_dropout
+        self.residual_dropout = residual_dropout
+        self.hidden_act = hidden_act
+        self.intermediate_size = intermediate_size
+        self.text_hidden_size = text_hidden_size
+        self.image_feature_dropout = image_feature_dropout
+        self.initializer_range = initializer_range
+
+
+class MolmoAct2TextConfig(PretrainedConfig):
+    r"""
+    This is the configuration class to store the configuration of a [`MolmoAct2TextModel`]. It is used to instantiate a
+    `MolmoAct2TextModel` according to the specified arguments, defining the model architecture.
+
+    Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
+    documentation from [`PretrainedConfig`] for more information.
+
+    Example:
+    ```python
+    >>> from transformers import MolmoAct2TextConfig, MolmoAct2TextModel
+
+    >>> # Initializing a MolmoAct2TextConfig
+    >>> configuration = MolmoAct2TextConfig()
+
+    >>> # Initializing a MolmoAct2TextModel (with random weights)
+    >>> model = MolmoAct2TextModel(configuration)
+
+    >>> # Accessing the model configuration
+    >>> configuration = model.config
+    ```"""
+
+    model_type = "molmoact2_text"
+    base_config_key = "text_config"
+    keys_to_ignore_at_inference = ["past_key_values"]
+    base_model_tp_plan = {
+        "blocks.*.self_attn.att_proj": "colwise",
+        "blocks.*.self_attn.attn_out": "rowwise",
+        "blocks.*.mlp.ff_proj": "colwise",
+        "blocks.*.mlp.ff_out": "rowwise",
+    }
+    base_model_pp_plan = {
+        "wte": (["input_ids"], ["inputs_embeds"]),
+        "blocks": (["hidden_states", "attention_mask"], ["hidden_states"]),
+        "ln_f": (["hidden_states"], ["hidden_states"]),
+    }
+
+    def __init__(
+        self,
+        hidden_size: int = 3584,
+        num_attention_heads: int = 28,
+        num_key_value_heads: int | None = 4,
+        head_dim: int = 128,
+        vocab_size: int = 152064,
+        additional_vocab_size: int = 128,
+        qkv_bias: bool = True,
+        num_hidden_layers: int = 48,
+        intermediate_size: int = 18944,
+        hidden_act: str = "silu",
+        embedding_dropout: float = 0.0,
+        attention_dropout: float = 0.0,
+        residual_dropout: float = 0.0,
+        max_position_embeddings: int = 4096,
+        rope_theta: float = 1000000.0,
+        rope_scaling: dict[str, Any] = None,
+        rope_scaling_layers: list[int] | None = None,
+        use_qk_norm: bool = False,
+        qk_norm_type: str = "olmo",
+        layer_norm_eps: int = 1e-6,
+        norm_after: bool = False,
+        initializer_range: float = 0.02,
+        use_cache=True,
+        tie_word_embeddings=False,
+        attn_implementation: str = "eager",
+        **kwargs,
+    ):
+        self.attn_implementation = attn_implementation
+        super().__init__(
+            tie_word_embeddings=tie_word_embeddings, attn_implementation=attn_implementation, **kwargs
+        )
+        self.hidden_size = hidden_size
+        self.num_attention_heads = num_attention_heads
+        if num_key_value_heads is None:
+            num_key_value_heads = num_attention_heads
+        self.num_key_value_heads = num_key_value_heads
+        self.head_dim = head_dim
+        self.vocab_size = vocab_size
+        self.additional_vocab_size = additional_vocab_size
+        self.qkv_bias = qkv_bias
+        self.num_hidden_layers = num_hidden_layers
+        self.intermediate_size = intermediate_size
+        self.hidden_act = hidden_act
+        self.embedding_dropout = embedding_dropout
+        self.attention_dropout = attention_dropout
+        self.residual_dropout = residual_dropout
+        self.max_position_embeddings = max_position_embeddings
+        self.rope_theta = rope_theta
+        self.rope_scaling = rope_scaling
+        self.rope_scaling_layers = rope_scaling_layers
+        self.use_qk_norm = use_qk_norm
+        self.qk_norm_type = qk_norm_type
+        self.layer_norm_eps = layer_norm_eps
+        self.norm_after = norm_after
+        self.initializer_range = initializer_range
+        self.use_cache = use_cache
+
+        # Validate the correctness of rotary position embeddings parameters
+        rope_config_validation(self)
+
+
+class MolmoAct2ActionExpertConfig(PretrainedConfig):
+    r"""Configuration for the MolmoAct2 modern action expert."""
+
+    model_type = "molmoact2_action_expert"
+    base_config_key = "action_expert_config"
+
+    def __init__(
+        self,
+        max_action_horizon: int = 32,
+        max_action_dim: int = 32,
+        hidden_size: int = 1024,
+        num_layers: int = 32,
+        num_heads: int = 16,
+        mlp_ratio: float = 8.0 / 3.0,
+        ffn_multiple_of: int = 256,
+        timestep_embed_dim: int = 256,
+        dropout: float = 0.0,
+        attn_dropout: float = 0.0,
+        context_layer_norm: bool = True,
+        qk_norm: bool = True,
+        qk_norm_eps: float = 1e-6,
+        rope: bool = True,
+        causal_attn: bool = False,
+        **kwargs,
+    ):
+        super().__init__(**kwargs)
+        self.max_action_horizon = max_action_horizon
+        self.max_action_dim = max_action_dim
+        self.hidden_size = hidden_size
+        self.num_layers = num_layers
+        self.num_heads = num_heads
+        self.mlp_ratio = mlp_ratio
+        self.ffn_multiple_of = ffn_multiple_of
+        self.timestep_embed_dim = timestep_embed_dim
+        self.dropout = dropout
+        self.attn_dropout = attn_dropout
+        self.context_layer_norm = context_layer_norm
+        self.qk_norm = qk_norm
+        self.qk_norm_eps = qk_norm_eps
+        self.rope = rope
+        self.causal_attn = causal_attn
+
+    def to_dict(self):
+        output = super().to_dict()
+        # These are derived from the parent MolmoAct2Config for HF exports. Keeping
+        # them out of the public nested config avoids duplicated sources of truth.
+        output.pop("max_action_horizon", None)
+        output.pop("max_action_dim", None)
+        return output
+
+
+class MolmoAct2Config(PretrainedConfig):
+    r"""
+    This is the configuration class to store the configuration of a [`MolmoAct2ForConditionalGeneration`].
+    It is used to instantiate an MolmoAct2 model according to the specified arguments, defining the model architecture.
+
+    Example:
+
+    ```python
+    >>> from transformers import MolmoAct2Config, MolmoAct2VitConfig, MolmoAct2AdapterConfig, MolmoAct2TextConfig
+
+    >>> # Initializing a MolmoAct2VitConfig
+    >>> vit_config = MolmoAct2VitConfig()
+
+    >>> # Initializing a MolmoAct2AdapterConfig
+    >>> adapter_config = MolmoAct2AdapterConfig()
+
+    >>> # Initializing a MolmoAct2TextConfig
+    >>> text_config = MolmoAct2TextConfig()
+
+    >>> # Initializing a MolmoAct2Config
+    >>> configuration = MolmoAct2Config(
+    >>>     vit_config=vit_config,
+    >>>     adapter_config=adapter_config,
+    >>>     text_config=text_config,
+    >>>     image_start_token_id=151936,
+    >>>     image_end_token_id=151937,
+    >>>     image_patch_id=151938,
+    >>>     image_col_id=151939,
+    >>>     low_res_image_start_token_id=151940,
+    >>>     image_low_res_id=151942,
+    >>>     frame_start_token_id=151943,
+    >>>     frame_end_token_id=151944,
+    >>> )
+
+    >>> # Initializing a model
+    >>> model = MolmoAct2ForConditionalGeneration(configuration)
+
+    >>> # Accessing the model configuration
+    >>> configuration = model.config
+    ```"""
+
+    model_type = "molmoact2"
+    sub_configs = {
+        "text_config": MolmoAct2TextConfig,
+        "vit_config": MolmoAct2VitConfig,
+        "adapter_config": MolmoAct2AdapterConfig,
+        "action_expert_config": MolmoAct2ActionExpertConfig,
+    }
+
+    def __init__(
+        self,
+        vit_config: MolmoAct2VitConfig = None,
+        adapter_config: MolmoAct2AdapterConfig = None,
+        text_config: MolmoAct2TextConfig = None,
+        action_expert_config: MolmoAct2ActionExpertConfig = None,
+        image_start_token_id: int = None,
+        low_res_image_start_token_id: int = None,
+        image_end_token_id: int = None,
+        image_low_res_id: int = None,
+        image_patch_id: int = None,
+        image_col_id: int = None,
+        frame_start_token_id: int = None,
+        frame_end_token_id: int = None,
+        use_frame_special_tokens: bool = True,
+        initializer_range: float = 0.02,
+        add_action_expert: bool = True,
+        max_action_dim: int = 32,
+        max_action_horizon: int = 30,
+        n_obs_steps: int = 30,
+        action_mode: str = "both",
+        state_format: str = "discrete",
+        flow_matching_num_steps: int = 10,
+        flow_matching_cutoff: float = 1.0,
+        flow_matching_time_offset: float = 0.001,
+        flow_matching_time_scale: float = 0.999,
+        flow_matching_beta_alpha: float = 1.0,
+        flow_matching_beta_beta: float = 1.5,
+        mask_action_dim_padding: bool = True,
+        enable_depth_reasoning: bool = False,
+        depth_mode: int = 2,
+        num_depth_codes: int = 100,
+        action_expert_depth_gate: bool = False,
+        action_expert_depth_gate_per_layer: bool = False,
+        action_expert_depth_gate_init_bias: float = -4.0,
+        action_output_token_id: int = None,
+        action_start_token_id: int = None,
+        action_end_token_id: int = None,
+        action_token_start_id: int = None,
+        num_action_tokens: int = 0,
+        depth_output_token_id: int = None,
+        depth_start_token_id: int = None,
+        depth_end_token_id: int = None,
+        depth_token_start_id: int = None,
+        num_depth_tokens: int = 0,
+        state_start_token_id: int = None,
+        state_end_token_id: int = None,
+        state_token_start_id: int = None,
+        num_state_tokens: int = 0,
+        add_setup_tokens: bool = True,
+        add_control_tokens: bool = True,
+        norm_stats_filename: str = "norm_stats.json",
+        **kwargs,
+    ):
+        super().__init__(**kwargs)
+        if vit_config is None:
+            self.vit_config = MolmoAct2VitConfig()
+        elif isinstance(vit_config, dict):
+            self.vit_config = MolmoAct2VitConfig(**vit_config)
+        else:
+            self.vit_config = vit_config
+        if adapter_config is None:
+            self.adapter_config = MolmoAct2AdapterConfig()
+        elif isinstance(adapter_config, dict):
+            self.adapter_config = MolmoAct2AdapterConfig(**adapter_config)
+        else:
+            self.adapter_config = adapter_config
+        if text_config is None:
+            self.text_config = MolmoAct2TextConfig()
+        elif isinstance(text_config, dict):
+            self.text_config = MolmoAct2TextConfig(**text_config)
+        else:
+            self.text_config = text_config
+        self.add_action_expert = bool(add_action_expert)
+        if not self.add_action_expert:
+            self.action_expert_config = None
+        elif action_expert_config is None:
+            self.action_expert_config = MolmoAct2ActionExpertConfig(
+                max_action_horizon=max_action_horizon,
+                max_action_dim=max_action_dim,
+                num_layers=self.text_config.num_hidden_layers,
+            )
+        elif isinstance(action_expert_config, dict):
+            self.action_expert_config = MolmoAct2ActionExpertConfig(**action_expert_config)
+        else:
+            self.action_expert_config = action_expert_config
+        if self.add_action_expert:
+            self.action_expert_config.max_action_dim = int(max_action_dim)
+            self.action_expert_config.max_action_horizon = int(max_action_horizon)
+            self._validate_release_action_config(
+                state_format=state_format,
+            )
+        self.image_start_token_id = image_start_token_id
+        self.low_res_image_start_token_id = low_res_image_start_token_id
+        self.image_end_token_id = image_end_token_id
+        self.image_low_res_id = image_low_res_id
+        self.image_high_res_id = image_patch_id
+        self.image_patch_id = image_patch_id
+        self.image_col_id = image_col_id
+        self.frame_start_token_id = frame_start_token_id
+        self.frame_end_token_id = frame_end_token_id
+        self.use_frame_special_tokens = use_frame_special_tokens
+        self.initializer_range = initializer_range
+        self.max_action_dim = max_action_dim
+        self.max_action_horizon = max_action_horizon
+        self.n_obs_steps = n_obs_steps
+        self.action_mode = action_mode
+        self.state_format = state_format
+        self.flow_matching_num_steps = flow_matching_num_steps
+        self.flow_matching_cutoff = flow_matching_cutoff
+        self.flow_matching_time_offset = flow_matching_time_offset
+        self.flow_matching_time_scale = flow_matching_time_scale
+        self.flow_matching_beta_alpha = flow_matching_beta_alpha
+        self.flow_matching_beta_beta = flow_matching_beta_beta
+        self.mask_action_dim_padding = mask_action_dim_padding
+        self.enable_depth_reasoning = enable_depth_reasoning
+        self.depth_mode = depth_mode
+        self.num_depth_codes = num_depth_codes
+        self.action_expert_depth_gate = action_expert_depth_gate
+        self.action_expert_depth_gate_per_layer = action_expert_depth_gate_per_layer
+        self.action_expert_depth_gate_init_bias = action_expert_depth_gate_init_bias
+        self.action_output_token_id = action_output_token_id
+        self.action_start_token_id = action_start_token_id
+        self.action_end_token_id = action_end_token_id
+        self.action_token_start_id = action_token_start_id
+        self.num_action_tokens = num_action_tokens
+        self.depth_output_token_id = depth_output_token_id
+        self.depth_start_token_id = depth_start_token_id
+        self.depth_end_token_id = depth_end_token_id
+        self.depth_token_start_id = depth_token_start_id
+        self.num_depth_tokens = num_depth_tokens
+        self.state_start_token_id = state_start_token_id
+        self.state_end_token_id = state_end_token_id
+        self.state_token_start_id = state_token_start_id
+        self.num_state_tokens = num_state_tokens
+        self.add_setup_tokens = add_setup_tokens
+        self.add_control_tokens = add_control_tokens
+        self.norm_stats_filename = norm_stats_filename
+
+    @staticmethod
+    def _validate_release_action_config(
+        *,
+        state_format: str,
+    ) -> None:
+        if state_format != "discrete":
+            raise ValueError("MolmoAct2 HF export supports only state_format='discrete'.")
+
+    @property
+    def image_num_patch(self):
+        assert self.vit_config is not None
+        return self.vit_config.image_num_patch
+
+    @property
+    def num_attention_heads(self):
+        return self.text_config.num_attention_heads
+
+    @property
+    def num_key_value_heads(self):
+        return self.text_config.num_key_value_heads
+
+    @property
+    def head_dim(self):
+        return self.text_config.head_dim
+
+    @property
+    def num_hidden_layers(self):
+        return self.text_config.num_hidden_layers
+
+    @property
+    def hidden_size(self):
+        return self.text_config.hidden_size
+
+    @property
+    def vocab_size(self):
+        return self.text_config.vocab_size
+
+    @property
+    def max_position_embeddings(self):
+        return self.text_config.max_position_embeddings
+
+
+MolmoAct2VitConfig.register_for_auto_class()
+MolmoAct2AdapterConfig.register_for_auto_class()
+MolmoAct2TextConfig.register_for_auto_class()
+MolmoAct2ActionExpertConfig.register_for_auto_class()
+MolmoAct2Config.register_for_auto_class()

src/lerobot/policies/molmoact2/hf_model/image_processing_molmoact2.py

@@ -0,0 +1,564 @@
+#!/usr/bin/env python
+
+# Copyright 2026 The Allen Institute for Artificial Intelligence and The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+# ruff: noqa
+
+"""Image processor class for MolmoAct2"""
+
+from typing import Optional, Union
+import numpy as np
+import einops
+import torch
+import torchvision.transforms
+
+from transformers.image_utils import (
+    IMAGENET_STANDARD_MEAN,
+    IMAGENET_STANDARD_STD,
+    ImageInput,
+    PILImageResampling,
+    make_flat_list_of_images,
+    valid_images,
+    to_numpy_array,
+)
+from transformers.image_transforms import convert_to_rgb
+from transformers.processing_utils import ImagesKwargs
+from transformers.image_processing_utils import BaseImageProcessor, get_size_dict
+from transformers.utils import logging
+from transformers.feature_extraction_utils import BatchFeature
+from transformers.utils import TensorType, logging
+
+
+logger = logging.get_logger(__name__)
+
+
+def normalize_image(
+    image: np.ndarray,
+    image_mean: list[float],
+    image_std: list[float],
+) -> np.ndarray:
+    if np.allclose(image_mean, [0.5, 0.5, 0.5]) and np.allclose(image_std, [0.5, 0.5, 0.5]):
+        return image * np.asarray(2.0, dtype=np.float32) - np.asarray(1.0, dtype=np.float32)
+    image -= np.array(image_mean, dtype=np.float32)[None, None, :]
+    image /= np.array(image_std, dtype=np.float32)[None, None, :]
+    return image
+
+
+def resize_image(
+    image: np.ndarray,
+    desired_output_size: list[int],
+    resample: PILImageResampling,
+) -> np.ndarray:
+    image = torch.permute(torch.from_numpy(image), [2, 0, 1])
+    dtype = image.dtype
+    if torch.is_floating_point(image):
+        in_min = 0.0
+        in_max = 1.0
+        resized = torchvision.transforms.Resize(
+            desired_output_size,
+            resample,
+            antialias=False,
+        )(image)
+        resized = torch.clip(resized, 0.0, 1.0).to(dtype)
+    else:
+        assert image.dtype == torch.uint8, "SigLIP expects float images or uint8 images, but got {}".format(
+            image.dtype
+        )
+        in_min = 0.0
+        in_max = 255.0
+        resized = torchvision.transforms.Resize(
+            desired_output_size,
+            resample,
+            antialias=False,
+        )(image)
+        resized = torch.clip(resized, 0, 255).to(dtype)
+
+    resized = resized.to(torch.float32)
+    resized = (resized - in_min) / (in_max - in_min)
+
+    resized = torch.permute(resized, [1, 2, 0]).numpy()
+
+    return resized
+
+
+def select_tiling(h, w, patch_size, max_num_crops):
+    """Divide in image of size [w, h] in up to max_num_patches of size patch_size"""
+    original_size = np.stack([h, w])  # [1, 2]
+    original_res = h * w
+    tilings = []
+    for i in range(1, max_num_crops + 1):
+        for j in range(1, max_num_crops + 1):
+            if i * j <= max_num_crops:
+                tilings.append((i, j))
+    # sort so argmin and argmax favour smaller tilings in the event of a tie
+    tilings.sort(key=lambda x: (x[0] * x[1], x[0]))
+    candidate_tilings = np.array(tilings, dtype=np.int32)  # [n_resolutions, 2]
+    candidate_resolutions = candidate_tilings * patch_size  # [n_resolutions, 2]
+
+    # How much we would need to scale the image to fit exactly in each tiling
+    original_size = np.stack([h, w], dtype=np.float32)  # [1, 2]
+
+    # The original size can be zero in rare cases if the image is smaller than the margin
+    # In those cases letting the scale become infinite means the tiling is based on the
+    # other side, or falls back to the smallest tiling
+    with np.errstate(divide="ignore"):
+        required_scale_d = (candidate_resolutions.astype(np.float32) / original_size,)
+    required_scale = np.min(required_scale_d, axis=-1, keepdims=True)  # [n_resolutions, 1]
+    if np.all(required_scale < 1):
+        # We are forced to downscale, so try to minimize the amount of downscaling
+        ix = np.argmax(required_scale)
+    else:
+        # Pick the resolution that required the least upscaling so that it most closely fits the image
+        required_scale = np.where(required_scale < 1.0, 10e9, required_scale)
+        ix = np.argmin(required_scale)
+    return candidate_tilings[ix]
+
+
+def build_resized_image(
+    image: np.ndarray,
+    base_image_input_size: list[int],
+    resample: PILImageResampling,
+    image_mean: list[float],
+    image_std: list[float],
+    image_patch_size: int,
+) -> tuple[np.ndarray, np.ndarray]:
+    resized = resize_image(
+        image,
+        base_image_input_size,
+        resample,
+    )
+    resized = normalize_image(resized, image_mean, image_std)
+    if len(resized.shape) == 3:
+        resized = np.expand_dims(resized, 0)
+    crop_patch_w = base_image_input_size[1] // image_patch_size
+    crop_patch_h = base_image_input_size[0] // image_patch_size
+    resize_idx = np.arange(crop_patch_w * crop_patch_h).reshape([crop_patch_h, crop_patch_w])
+    return resized, resize_idx
+
+
+def build_overlapping_crops(
+    image: np.ndarray,
+    max_crops: int,
+    overlap_margins: list[int],
+    base_image_input_size: list[int],
+    resample: PILImageResampling,
+    image_mean: list[float],
+    image_std: list[float],
+    image_patch_size: int,
+) -> tuple[np.ndarray, np.ndarray]:
+    """Decompose an image into a set of overlapping crops
+
+    :return crop_arr: [n_crops, h, w, 3] The crops
+    :return patch_idx: [overlap_patch_h, overlap_patch_w] For each patch in the resized image
+                        the crops were extracted from, what patch in `crop_arr` it corresponds to
+    """
+    original_image_h, original_image_w = image.shape[:2]
+    crop_size = base_image_input_size[0]
+    assert base_image_input_size[0] == base_image_input_size[1]
+
+    left_margin, right_margin = overlap_margins
+    total_margin_pixels = image_patch_size * (right_margin + left_margin)  # pixels removed per dim
+    crop_patches = base_image_input_size[0] // image_patch_size  # patches per crop dim
+    crop_window_patches = crop_patches - (right_margin + left_margin)  # usable patches
+    crop_window_size = crop_window_patches * image_patch_size
+    crop_patch_w = base_image_input_size[1] // image_patch_size
+    crop_patch_h = base_image_input_size[0] // image_patch_size
+    original_image_h, original_image_w = image.shape[:2]
+    crop_size = base_image_input_size[0]
+
+    # Decide how to tile the image, to account for the overlap margins we compute the tiling
+    # as if we had an image without the margins and were using a crop size without the margins
+    tiling = select_tiling(
+        original_image_h - total_margin_pixels,
+        original_image_w - total_margin_pixels,
+        crop_window_size,
+        max_crops,
+    )
+
+    src = resize_image(
+        image,
+        [
+            tiling[0] * crop_window_size + total_margin_pixels,
+            tiling[1] * crop_window_size + total_margin_pixels,
+        ],
+        resample,
+    )
+    src = normalize_image(src, image_mean, image_std)
+
+    # Now we have to split the image into crops, and track what patches came from
+    # where in `patch_idx_arr`
+    n_crops = tiling[0] * tiling[1]
+    crop_arr = np.zeros([n_crops, crop_size, crop_size, 3], dtype=src.dtype)
+    patch_idx_arr = np.zeros([n_crops, crop_patch_h, crop_patch_w], dtype=np.int32)
+    on_crop = 0
+    for i in range(tiling[0]):
+        # Slide over `src` by `crop_window_size` steps, but extract crops of size `crops_size`
+        # which results in overlapping crop windows
+        y0 = i * crop_window_size
+        for j in range(tiling[1]):
+            x0 = j * crop_window_size
+            crop_arr[on_crop] = src[y0 : y0 + crop_size, x0 : x0 + crop_size]
+            patch_idx = np.arange(crop_patch_w * crop_patch_h).reshape(crop_patch_h, crop_patch_w)
+            patch_idx += on_crop * crop_patch_h * crop_patch_w
+
+            # Mask out idx that are in the overlap region
+            if i != 0:
+                patch_idx[:left_margin, :] = -1
+            if j != 0:
+                patch_idx[:, :left_margin] = -1
+            if i != tiling[0] - 1:
+                patch_idx[-right_margin:, :] = -1
+            if j != tiling[1] - 1:
+                patch_idx[:, -right_margin:] = -1
+            patch_idx_arr[on_crop] = patch_idx
+            on_crop += 1
+
+    # `patch_idx_arr` is ordered crop-by-crop, here we transpose `patch_idx_arr`
+    # so it is ordered left-to-right order
+    patch_idx_arr = np.reshape(patch_idx_arr, [tiling[0], tiling[1], crop_patch_h, crop_patch_w])
+    patch_idx_arr = np.transpose(patch_idx_arr, [0, 2, 1, 3])
+    patch_idx_arr = np.reshape(patch_idx_arr, [-1])
+
+    # Now get the parts not in the overlap region, so it should map each patch in `src`
+    # to the correct patch it should come from in `crop_arr`
+    patch_idx_arr = patch_idx_arr[patch_idx_arr >= 0].reshape(
+        src.shape[0] // image_patch_size,
+        src.shape[1] // image_patch_size,
+    )
+    return crop_arr, patch_idx_arr
+
+
+def batch_pixels_to_patches(array: np.ndarray, patch_size: int) -> np.ndarray:
+    """Reshape images of [n_images, h, w, 3] -> [n_images, n_patches, pixels_per_patch]"""
+    if len(array.shape) == 3:
+        n_crops, h, w = array.shape
+        h_patches = h // patch_size
+        w_patches = w // patch_size
+        array = np.reshape(array, [n_crops, h_patches, patch_size, w_patches, patch_size])
+        array = np.transpose(array, [0, 1, 3, 2, 4])
+        array = np.reshape(array, [n_crops, h_patches * w_patches, patch_size * patch_size])
+        return array
+    else:
+        n_crops, h, w, c = array.shape
+        h_patches = h // patch_size
+        w_patches = w // patch_size
+        array = np.reshape(array, [n_crops, h_patches, patch_size, w_patches, patch_size, c])
+        array = np.transpose(array, [0, 1, 3, 2, 4, 5])
+        array = np.reshape(array, [n_crops, h_patches * w_patches, patch_size * patch_size * c])
+        return array
+
+
+def arange_for_pooling(
+    idx_arr: np.ndarray,
+    pool_h: int,
+    pool_w: int,
+) -> np.ndarray:
+    h_pad = pool_h * ((idx_arr.shape[0] + pool_h - 1) // pool_h) - idx_arr.shape[0]
+    w_pad = pool_w * ((idx_arr.shape[1] + pool_w - 1) // pool_w) - idx_arr.shape[1]
+    idx_arr = np.pad(
+        idx_arr,
+        [[h_pad // 2, (h_pad + 1) // 2], [w_pad // 2, (w_pad + 1) // 2]],
+        mode="constant",
+        constant_values=-1,
+    )
+    return einops.rearrange(idx_arr, "(h dh) (w dw) -> h w (dh dw)", dh=pool_h, dw=pool_w)
+
+
+def image_to_patches_and_grids(
+    image: np.ndarray,
+    max_crops: int,
+    overlap_margins: list[int],
+    base_image_input_size: list[int],
+    resample: PILImageResampling,
+    image_mean: list[float],
+    image_std: list[float],
+    image_patch_size: int,
+    image_pooling_w: int,
+    image_pooling_h: int,
+    crop_mode: str = "overlap-and-resize-c2",
+) -> tuple[np.ndarray, np.ndarray, np.ndarray]:
+    """
+    :return image_grids, the shape of each (low-res, high-res) image after pooling
+    :return crops, the image crops to processes with the ViT
+    :return pooled_patch_idx, for each patch_id tokens in `image_tokens`, the indices of the
+                                patches in `crops` to pool for that token, masked with -1
+    """
+    if isinstance(base_image_input_size, int):
+        base_image_input_size = (base_image_input_size, base_image_input_size)
+
+    base_image_input_d = image_patch_size
+    pooling_w = image_pooling_w
+    pooling_h = image_pooling_h
+    crop_patch_w = base_image_input_size[1] // base_image_input_d
+    crop_patch_h = base_image_input_size[0] // base_image_input_d
+
+    if crop_mode == "resize":
+        resized, resize_idx = build_resized_image(
+            image,
+            base_image_input_size,
+            resample,
+            image_mean,
+            image_std,
+            image_patch_size,
+        )
+        resize_idx = arange_for_pooling(resize_idx, pooling_h, pooling_w)
+        resized_h, resized_w = resize_idx.shape[:2]
+        resize_idx = resize_idx.reshape([-1, pooling_h * pooling_w])
+        image_grid = [np.array([resized_h, resized_w, 0, 0])]
+        return (
+            np.stack(image_grid, 0),
+            batch_pixels_to_patches(resized, image_patch_size),
+            resize_idx,
+        )
+
+    if crop_mode not in {"overlap-and-resize-c2", "overlap-and-resize"}:
+        raise ValueError(f"Unsupported MolmoAct2 image crop_mode {crop_mode!r}.")
+
+    crop_arr, patch_idx_arr = build_overlapping_crops(
+        image,
+        max_crops,
+        overlap_margins,
+        base_image_input_size,
+        resample,
+        image_mean,
+        image_std,
+        image_patch_size,
+    )
+    pooling_idx = arange_for_pooling(patch_idx_arr, pooling_h, pooling_w)
+    h, w = pooling_idx.shape[:2]
+    pooling_idx = pooling_idx.reshape([-1, pooling_h * pooling_w])
+
+    # Finally do the same for the global image
+    resized, resize_idx = build_resized_image(
+        image,
+        base_image_input_size,
+        resample,
+        image_mean,
+        image_std,
+        image_patch_size,
+    )
+    crop_arr = np.concatenate([resized, crop_arr], 0)
+
+    resize_idx = arange_for_pooling(resize_idx, pooling_h, pooling_w)
+    resized_h, resized_w = resize_idx.shape[:2]
+    resize_idx = resize_idx.reshape([-1, pooling_h * pooling_w])
+
+    # Global image goes first, so the order of patches in previous crops gets increased
+    pooling_idx = np.where(pooling_idx >= 0, pooling_idx + crop_patch_h * crop_patch_w, -1)
+    pooling_idx = np.concatenate([resize_idx, pooling_idx])
+    image_grid = [np.array([resized_h, resized_w, h, w])]
+
+    return (np.stack(image_grid, 0), batch_pixels_to_patches(crop_arr, image_patch_size), pooling_idx)
+
+
+class MolmoAct2ImagesKwargs(ImagesKwargs, total=False):
+    max_crops: int | None
+    overlap_margins: list[int] | None
+    crop_mode: str | None
+    patch_size: int | None
+    pooling_size: list[int] | None
+
+
+class MolmoAct2ImageProcessor(BaseImageProcessor):
+    r"""
+    Constructs a MolmoAct2 image processor that preprocesses images for the model.
+
+    Args:
+        size (`dict[str, int]` *optional*, defaults to `{"height": 378, "width": 378}`):
+            Size of the image after resizing.
+        resample (`PILImageResampling`, *optional*, defaults to `Resampling.BILINEAR`):
+            Resampling filter to use when resizing the image.
+        image_mean (`float` or `list[float]`, *optional*, defaults to `[0.5, 0.5, 0.5]`):
+            Mean to use if normalizing the image. This is a float or list of floats for each channel in the image.
+        image_std (`float` or `list[float]`, *optional*, defaults to `[0.5, 0.5, 0.5]`):
+            Standard deviation to use if normalizing the image. This is a float or list of floats for each channel in the image.
+        do_convert_rgb (`bool`, *optional*, defaults to `True`):
+            Whether to convert the image to RGB.
+        max_crops (`int`, *optional*, defaults to `8`):
+            Maximum number of crops to use per image.
+        overlap_margins (`list[int]`, *optional*, defaults to `[4, 4]`):
+            Overlap margins to use.
+        patch_size (`int`, *optional*, defaults to 14):
+            The spatial patch size of the vision encoder.
+        pooling_size (`list[int]`, *optional*, defaults to `[2, 2]`):
+            The pooling size of the vision adapter.
+    """
+
+    model_input_names = ["pixel_values", "image_token_pooling", "image_grids", "image_num_crops"]
+
+    def __init__(
+        self,
+        size: dict[str, int] | None = None,
+        resample: PILImageResampling = PILImageResampling.BILINEAR,
+        image_mean: float | list[float] | None = None,
+        image_std: float | list[float] | None = None,
+        do_convert_rgb: bool = True,
+        max_crops: int = 8,
+        overlap_margins: list[int] = [4, 4],
+        crop_mode: str = "overlap-and-resize-c2",
+        patch_size: int = 14,
+        pooling_size: list[int] = [2, 2],
+        **kwargs,
+    ) -> None:
+        super().__init__(**kwargs)
+        size = size if size is not None else {"height": 378, "width": 378}
+        size = get_size_dict(size, default_to_square=True)
+        self.size = size
+
+        self.resample = resample
+        self.image_mean = image_mean if image_mean is not None else IMAGENET_STANDARD_MEAN
+        self.image_std = image_std if image_std is not None else IMAGENET_STANDARD_STD
+        self.do_convert_rgb = do_convert_rgb
+
+        self.max_crops = max_crops
+        self.overlap_margins = overlap_margins
+        self.crop_mode = crop_mode
+        self.patch_size = patch_size
+        self.pooling_size = pooling_size
+
+    def preprocess(
+        self,
+        images: ImageInput,
+        size: dict[str, int] | None = None,
+        resample: PILImageResampling | None = None,
+        image_mean: float | list[float] | None = None,
+        image_std: float | list[float] | None = None,
+        do_convert_rgb: bool | None = None,
+        max_crops: int | None = None,
+        overlap_margins: list[int] | None = None,
+        crop_mode: str | None = None,
+        patch_size: int | None = None,
+        pooling_size: list[int] | None = None,
+        return_tensors: str | TensorType | None = None,
+        **kwargs,
+    ) -> BatchFeature:
+        """
+        Args:
+            images (`ImageInput`):
+                Image to preprocess.
+            size (`dict[str, int]`, *optional*, defaults to `self.size`):
+                Size of the image after resizing.
+            resample (`PILImageResampling`, *optional*, defaults to `self.resample`):
+                Resampling filter to use when resizing the image. This can be one of the enum `PILImageResampling`. Only
+                has an effect if `do_resize` is set to `True`.
+            image_mean (`float` or `list[float]`, *optional*, defaults to `self.image_mean`):
+                Image mean to use for normalization. Only has an effect if `do_normalize` is set to `True`.
+            image_std (`float` or `list[float]`, *optional*, defaults to `self.image_std`):
+                Image standard deviation to use for normalization. Only has an effect if `do_normalize` is set to
+                `True`.
+            do_convert_rgb (`bool`, *optional*, defaults to `self.do_convert_rgb`):
+                Whether to convert the image to RGB.
+            max_crops (`int`, *optional*, defaults to `self.max_crops`):
+                Maximum number of crops to use per image.
+            overlap_margins (`list[int]`, *optional*, defaults to `self.overlap_margins`):
+                Overlap margins to use.
+            patch_size (`int`, *optional*, defaults to `self.patch_size`):
+                The spatial patch size of the vision encoder.
+            pooling_size (`list[int]`, *optional*, defaults to `self.pooling_size`):
+                The pooling size of the vision adapter.
+            return_tensors (`str` or `TensorType`, *optional*):
+                The type of tensors to return. Can be one of:
+                - Unset: Return a list of `np.ndarray`.
+                - `TensorType.TENSORFLOW` or `'tf'`: Return a batch of type `tf.Tensor`.
+                - `TensorType.PYTORCH` or `'pt'`: Return a batch of type `torch.Tensor`.
+                - `TensorType.NUMPY` or `'np'`: Return a batch of type `np.ndarray`.
+                - `TensorType.JAX` or `'jax'`: Return a batch of type `jax.numpy.ndarray`.
+
+        Returns:
+            A `BatchFeature` containing the following keys:
+                - `pixel_values`: The preprocessed images.
+                - `image_token_pooling`: The indices of the patches in `crops` to pool for each token in `image_tokens`.
+                - `image_grids`: The image grids.
+                - `image_num_crops`: The number of crops for each image.
+        """
+        if size is not None:
+            if "height" not in size or "width" not in size:
+                raise ValueError("size must contain 'height' and 'width' keys.")
+        else:
+            size = {**self.size}
+
+        base_image_input_size = [size["height"], size["width"]]
+
+        resample = resample or self.resample
+        image_mean = image_mean or self.image_mean
+        image_std = image_std or self.image_std
+        do_convert_rgb = do_convert_rgb or self.do_convert_rgb
+
+        max_crops = max_crops or self.max_crops
+        overlap_margins = overlap_margins or self.overlap_margins
+        crop_mode = crop_mode or self.crop_mode
+        patch_size = patch_size or self.patch_size
+        pooling_size = pooling_size or self.pooling_size
+
+        image_pooling_h, image_pooling_w = pooling_size
+
+        if images is not None:
+            images = self.fetch_images(images)
+            images = make_flat_list_of_images(images)
+
+        if images is not None and not valid_images(images):
+            raise ValueError(
+                "Invalid image type. Must be of type PIL.Image.Image, numpy.ndarray, "
+                "torch.Tensor, tf.Tensor or jax.ndarray."
+            )
+
+        if do_convert_rgb:
+            images = [convert_to_rgb(image) for image in images]
+
+        # All transformations expect numpy arrays.
+        images = [to_numpy_array(image) for image in images]
+
+        data = {}
+        if images is not None:
+            batch_grids = []
+            batch_crops = []
+            batch_pooled_patches_idx = []
+            batch_num_crops = []
+
+            for image in images:
+                image_grid, crops, pooled_idx = image_to_patches_and_grids(
+                    image,
+                    max_crops,
+                    overlap_margins,
+                    base_image_input_size,
+                    resample,
+                    image_mean,
+                    image_std,
+                    patch_size,
+                    image_pooling_w,
+                    image_pooling_h,
+                    crop_mode,
+                )
+                batch_grids.append(image_grid)
+                batch_crops.append(crops)
+                batch_pooled_patches_idx.append(pooled_idx)
+                batch_num_crops.append(crops.shape[0])
+
+            pixel_values = np.concatenate(batch_crops, 0)
+            image_token_pooling = np.concatenate(batch_pooled_patches_idx, 0)
+            image_grids = np.concatenate(batch_grids, 0)
+            image_num_crops = np.array(batch_num_crops)
+
+            data.update(
+                pixel_values=pixel_values,
+                image_token_pooling=image_token_pooling,
+                image_grids=image_grids,
+                image_num_crops=image_num_crops,
+            )
+
+        return BatchFeature(data, tensor_type=return_tensors)
+
+
+MolmoAct2ImageProcessor.register_for_auto_class()

src/lerobot/policies/molmoact2/hf_model/inference.py

@@ -0,0 +1,748 @@
+#!/usr/bin/env python
+
+# Copyright 2026 The Allen Institute for Artificial Intelligence and The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+# ruff: noqa
+
+"""Inference utilities for MolmoAct2"""
+
+from dataclasses import dataclass
+from typing import Any, Optional, Tuple
+from collections.abc import Iterable, Sequence
+
+import torch
+from torch.nn import functional as F
+from transformers.cache_utils import Cache
+from transformers.configuration_utils import PretrainedConfig
+
+
+@dataclass
+class _ActionFlowInputs:
+    trajectory: torch.Tensor
+    context: Any
+    modulations: Sequence[Any]
+    action_dim_is_pad: torch.Tensor | None
+
+
+@dataclass
+class _ActionFlowCudaGraph:
+    key: tuple[Any, ...]
+    graph: torch.cuda.CUDAGraph
+    static_inputs: _ActionFlowInputs
+    output: torch.Tensor
+
+
+@dataclass
+class _DepthDecodeCudaGraphLayerStage:
+    residual: torch.Tensor
+    query: torch.Tensor
+    key: torch.Tensor
+    value: torch.Tensor
+
+
+@dataclass
+class _DepthDecodeCudaGraphPostStage:
+    graph: torch.cuda.CUDAGraph
+    attn_context: torch.Tensor
+
+
+@dataclass
+class _DepthDecodeCudaGraph:
+    cache_key: tuple[Any, ...]
+    pre_graph: torch.cuda.CUDAGraph
+    token_ids: torch.Tensor
+    cos: torch.Tensor
+    sin: torch.Tensor
+    positions: torch.Tensor
+    stages: Sequence[_DepthDecodeCudaGraphLayerStage]
+    post_graphs: Sequence[_DepthDecodeCudaGraphPostStage]
+    output: torch.Tensor
+
+
+@dataclass
+class _DepthDecodeCudaGraphSpec:
+    eligible: bool
+    cache_key_prefix: tuple[Any, ...]
+    num_hidden_layers: int
+    head_dim: int
+    num_attention_heads: int
+
+
+def _cache_seq_len_int(past_key_values: Cache | None) -> int:
+    if past_key_values is None:
+        return 0
+    seq_len = past_key_values.get_seq_length()
+    if torch.is_tensor(seq_len):
+        return int(seq_len.item())
+    return int(seq_len)
+
+
+def _cache_max_len_int(past_key_values: Cache | None) -> int:
+    if past_key_values is None:
+        return -1
+    max_len = past_key_values.get_max_cache_shape()
+    if torch.is_tensor(max_len):
+        return int(max_len.item())
+    return int(max_len)
+
+
+def _iter_cache_key_values(
+    past_key_values: Cache,
+) -> Iterable[tuple[torch.Tensor | None, torch.Tensor | None]]:
+    layers = getattr(past_key_values, "layers", None)
+    if layers is not None:
+        for layer in layers:
+            yield getattr(layer, "keys", None), getattr(layer, "values", None)
+        return
+    for layer in past_key_values:
+        yield layer[0], layer[1]
+
+
+class _DepthDecodeStaticLayerCache:
+    is_compileable = False
+    is_sliding = False
+
+    def __init__(self, max_cache_len: int) -> None:
+        self.max_cache_len = int(max_cache_len)
+        self.cumulative_length = 0
+        self.keys: torch.Tensor | None = None
+        self.values: torch.Tensor | None = None
+
+    def _allocate(self, key_states: torch.Tensor, value_states: torch.Tensor) -> None:
+        bsz, n_heads = key_states.shape[:2]
+        self.keys = torch.empty(
+            (bsz, n_heads, self.max_cache_len, key_states.shape[-1]),
+            dtype=key_states.dtype,
+            device=key_states.device,
+        )
+        self.values = torch.empty(
+            (bsz, n_heads, self.max_cache_len, value_states.shape[-1]),
+            dtype=value_states.dtype,
+            device=value_states.device,
+        )
+
+    def update(
+        self,
+        key_states: torch.Tensor,
+        value_states: torch.Tensor,
+        *args,
+        **kwargs,
+    ) -> tuple[torch.Tensor, torch.Tensor]:
+        if self.keys is None:
+            self._allocate(key_states, value_states)
+        start = self.cumulative_length
+        end = start + key_states.shape[-2]
+        if end > self.max_cache_len:
+            raise RuntimeError(f"KV cache length {end} exceeds max_cache_len={self.max_cache_len}.")
+        self.keys[:, :, start:end, :].copy_(key_states)
+        self.values[:, :, start:end, :].copy_(value_states)
+        self.cumulative_length = end
+        return self.keys[:, :, :end, :], self.values[:, :, :end, :]
+
+    def get_seq_length(self) -> int:
+        return self.cumulative_length
+
+    def get_max_cache_shape(self) -> int:
+        return -1
+
+    def reset(self) -> None:
+        self.cumulative_length = 0
+
+
+class _DepthDecodeStaticCache(Cache):
+    def __init__(self, config: PretrainedConfig, max_cache_len: int) -> None:
+        text_config = config.get_text_config(decoder=True)
+        super().__init__(
+            layers=[
+                _DepthDecodeStaticLayerCache(max_cache_len=max_cache_len)
+                for _ in range(text_config.num_hidden_layers)
+            ]
+        )
+
+    def get_seq_length(self, layer_idx: int = 0) -> int:
+        return self.layers[layer_idx].get_seq_length()
+
+    def get_max_cache_shape(self, layer_idx: int = 0) -> int:
+        return self.layers[layer_idx].get_max_cache_shape()
+
+    def reset(self) -> None:
+        for layer in self.layers:
+            layer.reset()
+
+
+class ActionCudaGraphManager:
+    def __init__(self, model: Any) -> None:
+        self.model = model
+        self.enabled = True
+        self.action_flow_graph: _ActionFlowCudaGraph | None = None
+
+    def set_enabled(self, enabled: bool) -> None:
+        self.enabled = bool(enabled)
+
+    def can_use_action_flow(self, inputs: _ActionFlowInputs) -> bool:
+        action_model = self.model
+        if not self.enabled:
+            return False
+        if action_model.training or action_model._require_action_expert().training:
+            return False
+        if inputs.trajectory.device.type != "cuda":
+            return False
+
+        def all_on_cuda():
+            yield inputs.trajectory
+            for k, v in inputs.context.kv_contexts:
+                yield k
+                yield v
+            for t in (
+                inputs.context.cross_mask,
+                inputs.context.self_mask,
+                inputs.context.valid_action,
+                inputs.action_dim_is_pad,
+            ):
+                if t is not None:
+                    yield t
+            if inputs.context.rope_cache is not None:
+                yield from inputs.context.rope_cache
+            for step in inputs.modulations:
+                yield step.conditioning
+                for block_modulation in step.block_modulations:
+                    yield from block_modulation
+                yield from step.final_modulation
+
+        return all(t.device.type == "cuda" for t in all_on_cuda())
+
+    def run_action_flow(
+        self,
+        inputs: _ActionFlowInputs,
+        steps: int,
+        run_loop,
+    ) -> torch.Tensor:
+        key = _cuda_graph_key(inputs, steps)
+        cache = self.action_flow_graph
+        if cache is None or cache.key != key:
+            static_inputs = _clone_static_inputs(inputs)
+            graph, output = _capture_cuda_graph(
+                lambda: run_loop(static_inputs, steps),
+                inputs.trajectory.device,
+                after_warmup=lambda: static_inputs.trajectory.copy_(inputs.trajectory),
+            )
+            cache = _ActionFlowCudaGraph(
+                key=key,
+                graph=graph,
+                static_inputs=static_inputs,
+                output=output,
+            )
+            self.action_flow_graph = cache
+        else:
+            _copy_inputs_(cache.static_inputs, inputs)
+
+        cache.graph.replay()
+        return cache.output.clone()
+
+
+class DepthDecodeCudaGraphManager:
+    def __init__(self, model: Any) -> None:
+        self.model = model
+        self.backbone = model.model
+        self.enabled = True
+        self.graph: _DepthDecodeCudaGraph | None = None
+        self.graph_spec: _DepthDecodeCudaGraphSpec | None = None
+
+    def set_enabled(self, enabled: bool) -> None:
+        self.enabled = bool(enabled)
+
+    def make_static_cache(self, max_cache_len: int) -> _DepthDecodeStaticCache:
+        return _DepthDecodeStaticCache(
+            config=self.model.config.text_config,
+            max_cache_len=max_cache_len,
+        )
+
+    def _depth_decode_spec(self) -> _DepthDecodeCudaGraphSpec:
+        static = self.graph_spec
+        if static is None:
+            cfg = self.backbone.transformer.config
+            rotary_emb = getattr(self.backbone.transformer, "rotary_emb", None)
+            static = _DepthDecodeCudaGraphSpec(
+                eligible=(
+                    not cfg.norm_after
+                    and cfg.rope_scaling_layers is None
+                    and getattr(rotary_emb, "rope_type", None) == "default"
+                    and cfg._attn_implementation == "sdpa"
+                ),
+                cache_key_prefix=(
+                    cfg.hidden_size,
+                    cfg.num_attention_heads,
+                    cfg.num_key_value_heads,
+                    cfg.head_dim,
+                    cfg.num_hidden_layers,
+                    cfg.use_qk_norm,
+                    cfg.qk_norm_type,
+                    cfg._attn_implementation,
+                ),
+                num_hidden_layers=cfg.num_hidden_layers,
+                head_dim=cfg.head_dim,
+                num_attention_heads=cfg.num_attention_heads,
+            )
+            self.graph_spec = static
+        return static
+
+    def can_use(
+        self,
+        next_input_ids: torch.Tensor,
+        *,
+        past_key_values: Cache,
+        attention_bias: torch.Tensor,
+    ) -> bool:
+        if not self.enabled or self.model.training or self.backbone.transformer.training:
+            return False
+        if next_input_ids.device.type != "cuda":
+            return False
+        if next_input_ids.ndim != 2 or next_input_ids.shape[0] != 1 or next_input_ids.shape[1] != 1:
+            return False
+        if not isinstance(past_key_values, _DepthDecodeStaticCache):
+            return False
+        if not torch.is_tensor(attention_bias) or attention_bias.device != next_input_ids.device:
+            return False
+        return self._depth_decode_spec().eligible
+
+    def _depth_decode_key(
+        self,
+        next_input_ids: torch.Tensor,
+        attention_bias: torch.Tensor,
+    ) -> tuple[Any, ...]:
+        device = next_input_ids.device
+        return (
+            self._depth_decode_spec().cache_key_prefix,
+            device.type,
+            device.index,
+            self.model.lm_head.weight.dtype,
+            attention_bias.shape[-1],
+        )
+
+    def _select_depth_decode_rope(self, cos: torch.Tensor, sin: torch.Tensor, *, past_length: int) -> None:
+        emb = self.backbone.transformer.rotary_emb
+        cos.copy_(emb._pos_cos_cache[0, :, past_length : past_length + 1, :])
+        sin.copy_(emb._pos_sin_cache[0, :, past_length : past_length + 1, :])
+
+    def _depth_decode_pre_layer(
+        self,
+        layer_idx: int,
+        hidden_states: torch.Tensor,
+        cos: torch.Tensor,
+        sin: torch.Tensor,
+    ) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor]:
+        block = self.backbone.transformer.blocks[layer_idx]
+        attention = block.self_attn
+        residual = hidden_states
+        hidden_states = block.attn_norm(hidden_states)
+
+        input_shape = hidden_states.shape[:-1]
+        hidden_shape = (*input_shape, -1, attention.head_dim)
+        qkv = attention.att_proj(hidden_states)
+        query_states, key_states, value_states = qkv.split(attention.fused_dims, dim=-1)
+        value_states = value_states.view(hidden_shape)
+
+        apply_qk_norm = attention.q_norm is not None and attention.k_norm is not None
+        norm_after_view = apply_qk_norm and attention.qk_norm_type == "qwen3"
+
+        if apply_qk_norm and not norm_after_view:
+            query_states = attention.q_norm(query_states)
+            key_states = attention.k_norm(key_states)
+
+        query_states = query_states.view(hidden_shape)
+        key_states = key_states.view(hidden_shape)
+
+        if norm_after_view:
+            query_states = attention.q_norm(query_states)
+            key_states = attention.k_norm(key_states)
+
+        query_states = query_states.transpose(1, 2)
+        key_states = key_states.transpose(1, 2)
+        value_states = value_states.transpose(1, 2)
+        query_states, key_states = _apply_rotary_pos_emb(query_states, key_states, cos, sin)
+        return residual, query_states, key_states, value_states
+
+    def _depth_decode_pre0(
+        self,
+        token_ids: torch.Tensor,
+        cos: torch.Tensor,
+        sin: torch.Tensor,
+    ) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor]:
+        inputs_embeds = self.model._embed_base_tokens(token_ids)
+        return self._depth_decode_pre_layer(0, inputs_embeds, cos, sin)
+
+    def _depth_decode_post_layer(
+        self,
+        layer_idx: int,
+        residual: torch.Tensor,
+        attn_context: torch.Tensor,
+    ) -> torch.Tensor:
+        block = self.backbone.transformer.blocks[layer_idx]
+        attention = block.self_attn
+        input_shape = residual.shape[:-1]
+        attn_output = attn_context.reshape(*input_shape, -1).contiguous()
+        attn_output = attention.attn_out(attn_output)
+        hidden_states = residual + block.dropout(attn_output)
+
+        residual = hidden_states
+        hidden_states = block.ff_norm(hidden_states)
+        hidden_states = block.mlp(hidden_states)
+        hidden_states = residual + block.dropout(hidden_states)
+        return hidden_states
+
+    def _depth_decode_post_and_pre_next(
+        self,
+        layer_idx: int,
+        residual: torch.Tensor,
+        attn_context: torch.Tensor,
+        cos: torch.Tensor,
+        sin: torch.Tensor,
+    ) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor]:
+        hidden_states = self._depth_decode_post_layer(layer_idx, residual, attn_context)
+        return self._depth_decode_pre_layer(layer_idx + 1, hidden_states, cos, sin)
+
+    def _depth_decode_last_post(
+        self,
+        layer_idx: int,
+        residual: torch.Tensor,
+        attn_context: torch.Tensor,
+    ) -> torch.Tensor:
+        hidden_states = self._depth_decode_post_layer(layer_idx, residual, attn_context)
+        return self.backbone.transformer.ln_f(hidden_states)
+
+    def _build_depth_decode_graph(
+        self,
+        next_input_ids: torch.Tensor,
+        *,
+        past_length: int,
+        attention_bias: torch.Tensor,
+    ) -> _DepthDecodeCudaGraph:
+        text_config = self.backbone.transformer.config
+        device = next_input_ids.device
+        dtype = self.model.lm_head.weight.dtype
+        static = self._depth_decode_spec()
+        num_layers = static.num_hidden_layers
+        head_dim = static.head_dim
+        max_cache_len = int(attention_bias.shape[-1])
+        max_rope_len = max(int(text_config.max_position_embeddings or 0), max_cache_len)
+        self.backbone.transformer.prepare_rope_cache(device=device, max_seq_len=max_rope_len)
+
+        token_ids = torch.empty((1, 1), device=device, dtype=torch.long)
+        cos = torch.empty((1, 1, head_dim), device=device, dtype=dtype)
+        sin = torch.empty_like(cos)
+        positions = torch.arange(max_cache_len, device=device, dtype=torch.long)
+        context_shape = (1, 1, static.num_attention_heads, head_dim)
+
+        token_ids.copy_(next_input_ids)
+        self._select_depth_decode_rope(cos, sin, past_length=past_length)
+
+        pre_graph, pre_output = _capture_cuda_graph(
+            lambda: self._depth_decode_pre0(token_ids, cos, sin),
+            device,
+        )
+        stages = [_DepthDecodeCudaGraphLayerStage(*pre_output)]
+        post_graphs = []
+        for layer_idx in range(num_layers - 1):
+            stage = stages[-1]
+            attn_context = torch.empty(context_shape, device=device, dtype=dtype)
+            graph, output = _capture_cuda_graph(
+                lambda layer_idx=layer_idx, stage=stage, attn_context=attn_context: (
+                    self._depth_decode_post_and_pre_next(
+                        layer_idx,
+                        stage.residual,
+                        attn_context,
+                        cos,
+                        sin,
+                    )
+                ),
+                device,
+            )
+            post_graphs.append(_DepthDecodeCudaGraphPostStage(graph=graph, attn_context=attn_context))
+            stages.append(_DepthDecodeCudaGraphLayerStage(*output))
+
+        last_stage = stages[-1]
+        last_attn_context = torch.empty(context_shape, device=device, dtype=dtype)
+        last_graph, last_output = _capture_cuda_graph(
+            lambda: self._depth_decode_last_post(
+                num_layers - 1,
+                last_stage.residual,
+                last_attn_context,
+            ),
+            device,
+        )
+        post_graphs.append(_DepthDecodeCudaGraphPostStage(graph=last_graph, attn_context=last_attn_context))
+        return _DepthDecodeCudaGraph(
+            cache_key=self._depth_decode_key(next_input_ids, attention_bias),
+            pre_graph=pre_graph,
+            token_ids=token_ids,
+            cos=cos,
+            sin=sin,
+            positions=positions,
+            stages=tuple(stages),
+            post_graphs=tuple(post_graphs),
+            output=last_output,
+        )
+
+    def _get_depth_decode_graph(
+        self,
+        next_input_ids: torch.Tensor,
+        *,
+        past_length: int,
+        attention_bias: torch.Tensor,
+    ) -> _DepthDecodeCudaGraph:
+        key = self._depth_decode_key(next_input_ids, attention_bias)
+        decode_graph = self.graph
+        if decode_graph is None or decode_graph.cache_key != key:
+            decode_graph = self._build_depth_decode_graph(
+                next_input_ids,
+                past_length=past_length,
+                attention_bias=attention_bias,
+            )
+            self.graph = decode_graph
+        else:
+            decode_graph.token_ids.copy_(next_input_ids)
+            self._select_depth_decode_rope(decode_graph.cos, decode_graph.sin, past_length=past_length)
+        return decode_graph
+
+    def _run_depth_decode_attention_core(
+        self,
+        layer_idx: int,
+        stage: _DepthDecodeCudaGraphLayerStage,
+        *,
+        past_key_values: Cache,
+        attention_bias: torch.Tensor,
+        cache_position: torch.Tensor,
+        cos: torch.Tensor,
+        sin: torch.Tensor,
+    ) -> torch.Tensor:
+        attention = self.backbone.transformer.blocks[layer_idx].self_attn
+        cache_kwargs = {"sin": sin, "cos": cos, "cache_position": cache_position}
+        key_states, value_states = past_key_values.update(
+            stage.key,
+            stage.value,
+            layer_idx,
+            cache_kwargs,
+        )
+        key_states = _repeat_kv(key_states, attention.num_key_value_groups)
+        value_states = _repeat_kv(value_states, attention.num_key_value_groups)
+        attn_output = F.scaled_dot_product_attention(
+            stage.query,
+            key_states,
+            value_states,
+            attn_mask=attention_bias,
+            dropout_p=0.0,
+            is_causal=False,
+        )
+        return attn_output.transpose(1, 2)
+
+    def run(
+        self,
+        next_input_ids: torch.Tensor,
+        *,
+        past_key_values: Cache,
+        attention_bias: torch.Tensor,
+        past_length: int,
+    ) -> tuple[torch.Tensor, Cache]:
+        end = past_length + 1
+        decode_graph = self._get_depth_decode_graph(
+            next_input_ids,
+            past_length=past_length,
+            attention_bias=attention_bias,
+        )
+        cache_position = decode_graph.positions[past_length:end]
+        attention_bias_q = attention_bias[:, :, past_length:end, :end]
+
+        decode_graph.pre_graph.replay()
+
+        for layer_idx, post_graph in enumerate(decode_graph.post_graphs):
+            attn_context = self._run_depth_decode_attention_core(
+                layer_idx,
+                decode_graph.stages[layer_idx],
+                past_key_values=past_key_values,
+                attention_bias=attention_bias_q,
+                cache_position=cache_position,
+                cos=decode_graph.cos,
+                sin=decode_graph.sin,
+            )
+            post_graph.attn_context.copy_(attn_context)
+            post_graph.graph.replay()
+
+        return decode_graph.output, past_key_values
+
+
+def _cuda_graph_tensor_signature(
+    tensor: torch.Tensor | None,
+) -> tuple[Any, ...] | None:
+    if tensor is None:
+        return None
+    return (
+        tuple(tensor.shape),
+        tuple(tensor.stride()),
+        str(tensor.dtype),
+        str(tensor.device),
+    )
+
+
+def _cuda_graph_context_signature(context: Any) -> tuple[Any, ...]:
+    sig = _cuda_graph_tensor_signature
+    return (
+        tuple((sig(k), sig(v)) for k, v in context.kv_contexts),
+        sig(context.cross_mask),
+        sig(context.self_mask),
+        sig(context.valid_action),
+        None if context.rope_cache is None else tuple(sig(t) for t in context.rope_cache),
+    )
+
+
+def _cuda_graph_modulation_signature(modulations: Sequence[Any]) -> tuple[Any, ...]:
+    sig = _cuda_graph_tensor_signature
+    return tuple(
+        (
+            sig(step.conditioning),
+            tuple(tuple(sig(t) for t in block_modulation) for block_modulation in step.block_modulations),
+            tuple(sig(t) for t in step.final_modulation),
+        )
+        for step in modulations
+    )
+
+
+def _cuda_graph_key(inputs: _ActionFlowInputs, steps: int) -> tuple[Any, ...]:
+    sig = _cuda_graph_tensor_signature
+    return (
+        sig(inputs.trajectory),
+        _cuda_graph_context_signature(inputs.context),
+        _cuda_graph_modulation_signature(inputs.modulations),
+        sig(inputs.action_dim_is_pad),
+        int(steps),
+    )
+
+
+def _clone_static_tensor(tensor: torch.Tensor | None) -> torch.Tensor | None:
+    if tensor is None:
+        return None
+    static = torch.empty_strided(
+        tuple(tensor.shape),
+        tuple(tensor.stride()),
+        device=tensor.device,
+        dtype=tensor.dtype,
+    )
+    static.copy_(tensor)
+    return static
+
+
+def _clone_static_context(context: Any) -> Any:
+    rope_cache = None
+    if context.rope_cache is not None:
+        rope_cache = tuple(_clone_static_tensor(t) for t in context.rope_cache)
+    return context.__class__(
+        kv_contexts=tuple((_clone_static_tensor(k), _clone_static_tensor(v)) for k, v in context.kv_contexts),
+        cross_mask=_clone_static_tensor(context.cross_mask),
+        self_mask=_clone_static_tensor(context.self_mask),
+        valid_action=_clone_static_tensor(context.valid_action),
+        rope_cache=rope_cache,
+    )
+
+
+def _clone_static_modulations(modulations: Sequence[Any]) -> Sequence[Any]:
+    return tuple(
+        step.__class__(
+            conditioning=_clone_static_tensor(step.conditioning),
+            block_modulations=tuple(
+                tuple(_clone_static_tensor(t) for t in block_modulation)
+                for block_modulation in step.block_modulations
+            ),
+            final_modulation=tuple(_clone_static_tensor(t) for t in step.final_modulation),
+        )
+        for step in modulations
+    )
+
+
+def _clone_static_inputs(inputs: _ActionFlowInputs) -> _ActionFlowInputs:
+    return _ActionFlowInputs(
+        trajectory=_clone_static_tensor(inputs.trajectory),
+        context=_clone_static_context(inputs.context),
+        modulations=_clone_static_modulations(inputs.modulations),
+        action_dim_is_pad=_clone_static_tensor(inputs.action_dim_is_pad),
+    )
+
+
+def _copy_context_(dst: Any, src: Any) -> None:
+    for (dst_k, dst_v), (src_k, src_v) in zip(dst.kv_contexts, src.kv_contexts):
+        dst_k.copy_(src_k)
+        dst_v.copy_(src_v)
+    if src.cross_mask is not None:
+        dst.cross_mask.copy_(src.cross_mask)
+    if src.self_mask is not None:
+        dst.self_mask.copy_(src.self_mask)
+    if src.valid_action is not None:
+        dst.valid_action.copy_(src.valid_action)
+    if src.rope_cache is not None:
+        for dst_tensor, src_tensor in zip(dst.rope_cache, src.rope_cache):
+            dst_tensor.copy_(src_tensor)
+
+
+def _copy_inputs_(dst: _ActionFlowInputs, src: _ActionFlowInputs) -> None:
+    dst.trajectory.copy_(src.trajectory)
+    _copy_context_(dst.context, src.context)
+    if src.action_dim_is_pad is not None:
+        dst.action_dim_is_pad.copy_(src.action_dim_is_pad)
+
+
+def _rotate_half(x: torch.Tensor) -> torch.Tensor:
+    x1 = x[..., : x.shape[-1] // 2]
+    x2 = x[..., x.shape[-1] // 2 :]
+    return torch.cat((-x2, x1), dim=-1)
+
+
+def _apply_rotary_pos_emb(
+    q: torch.Tensor,
+    k: torch.Tensor,
+    cos: torch.Tensor,
+    sin: torch.Tensor,
+    unsqueeze_dim: int = 1,
+) -> tuple[torch.Tensor, torch.Tensor]:
+    cos = cos.unsqueeze(unsqueeze_dim)
+    sin = sin.unsqueeze(unsqueeze_dim)
+    q_embed = (q * cos) + (_rotate_half(q) * sin)
+    k_embed = (k * cos) + (_rotate_half(k) * sin)
+    return q_embed, k_embed
+
+
+def _repeat_kv(hidden_states: torch.Tensor, n_rep: int) -> torch.Tensor:
+    batch, num_key_value_heads, slen, head_dim = hidden_states.shape
+    if n_rep == 1:
+        return hidden_states
+    hidden_states = hidden_states[:, :, None, :, :].expand(batch, num_key_value_heads, n_rep, slen, head_dim)
+    return hidden_states.reshape(batch, num_key_value_heads * n_rep, slen, head_dim)
+
+
+def _capture_cuda_graph(
+    fn,
+    device: torch.device,
+    *,
+    after_warmup=None,
+) -> tuple[torch.cuda.CUDAGraph, Any]:
+    warmup_stream = torch.cuda.Stream(device=device)
+    warmup_stream.wait_stream(torch.cuda.current_stream(device))
+    with torch.cuda.stream(warmup_stream):
+        fn()
+    torch.cuda.current_stream(device).wait_stream(warmup_stream)
+    if after_warmup is not None:
+        after_warmup()
+
+    graph = torch.cuda.CUDAGraph()
+    with torch.cuda.graph(graph):
+        output = fn()
+    return graph, output

src/lerobot/policies/molmoact2/hf_model/processing_molmoact2.py

@@ -0,0 +1,431 @@
+#!/usr/bin/env python
+
+# Copyright 2026 The Allen Institute for Artificial Intelligence and The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+# ruff: noqa
+
+"""
+Processor class for MolmoAct2.
+"""
+
+from typing import Optional, Union
+import dataclasses
+
+import numpy as np
+
+from transformers.image_utils import ImageInput
+from transformers.video_utils import VideoInput
+from transformers.processing_utils import (
+    Unpack,
+    ProcessingKwargs,
+    ProcessorMixin,
+)
+from transformers.feature_extraction_utils import BatchFeature
+from transformers.tokenization_utils_base import TextInput, PreTokenizedInput
+from transformers.utils import logging
+
+from transformers import AutoTokenizer
+from .image_processing_molmoact2 import MolmoAct2ImagesKwargs, MolmoAct2ImageProcessor
+from .video_processing_molmoact2 import MolmoAct2VideoProcessorKwargs, MolmoAct2VideoProcessor
+
+
+logger = logging.get_logger(__name__)
+
+
+# Special tokens, these should be present in any tokenizer we use since the preprocessor uses them
+IMAGE_PATCH_TOKEN = f"<im_patch>"  # Where to insert high-res tokens
+IMAGE_LOW_RES_TOKEN = f"<im_low>"  # Where to insert low-res tokens
+IM_START_TOKEN = f"<im_start>"
+LOW_RES_IMAGE_START_TOKEN = f"<low_res_im_start>"
+FRAME_START_TOKEN = f"<frame_start>"
+IM_END_TOKEN = f"<im_end>"
+FRAME_END_TOKEN = f"<frame_end>"
+IM_COL_TOKEN = f"<im_col>"
+IMAGE_PROMPT = "<|image|>"
+VIDEO_PROMPT = "<|video|>"
+
+IMAGE_TOKENS = [
+    IMAGE_PATCH_TOKEN,
+    IM_COL_TOKEN,
+    IM_START_TOKEN,
+    LOW_RES_IMAGE_START_TOKEN,
+    FRAME_START_TOKEN,
+    IM_END_TOKEN,
+    FRAME_END_TOKEN,
+    IMAGE_LOW_RES_TOKEN,
+]
+
+
+class MolmoAct2ProcessorKwargs(ProcessingKwargs, total=False):
+    """MolmoAct2 processor kwargs"""
+
+    images_kwargs: MolmoAct2ImagesKwargs
+    videos_kwargs: MolmoAct2VideoProcessorKwargs
+    _defaults = {
+        "text_kwargs": {
+            "padding": False,
+            "return_mm_token_type_ids": True,
+        },
+        "videos_kwargs": {"return_metadata": True},
+    }
+
+
+class MolmoAct2Processor(ProcessorMixin):
+    attributes = ["image_processor", "video_processor", "tokenizer"]
+    optional_attributes = [
+        "chat_template",
+        "time_mode",
+        "image_use_col_tokens",
+        "use_single_crop_col_tokens",
+        "use_single_crop_start_token",
+        "video_use_col_tokens",
+        "use_frame_special_tokens",
+    ]
+    image_processor_class = "AutoImageProcessor"
+    video_processor_class = "AutoVideoProcessor"
+    tokenizer_class = "AutoTokenizer"
+
+    def __init__(
+        self,
+        image_processor: MolmoAct2ImageProcessor = None,
+        video_processor: MolmoAct2VideoProcessor = None,
+        tokenizer: AutoTokenizer = None,
+        chat_template: str | None = None,
+        image_use_col_tokens: bool | None = True,
+        use_single_crop_col_tokens: bool | None = None,
+        use_single_crop_start_token: bool | None = True,
+        video_use_col_tokens: bool | None = False,
+        use_frame_special_tokens: bool | None = True,
+        **kwargs,
+    ) -> None:
+        super().__init__(
+            image_processor,
+            video_processor,
+            tokenizer,
+            chat_template=chat_template,
+        )
+        self.image_use_col_tokens = image_use_col_tokens
+        self.use_single_crop_col_tokens = use_single_crop_col_tokens
+        self.use_single_crop_start_token = use_single_crop_start_token
+        self.video_use_col_tokens = video_use_col_tokens
+        self.use_frame_special_tokens = use_frame_special_tokens
+
+        self.image_placeholder_token = IMAGE_PROMPT
+        self.video_placeholder_token = VIDEO_PROMPT
+        self.image_token_ids = [tokenizer.convert_tokens_to_ids(token) for token in IMAGE_TOKENS]
+
+    def get_image_tokens(self, image_grid: np.ndarray):
+        resized_h, resized_w, height, width = image_grid
+        if int(height) == 0 or int(width) == 0:
+            per_row = np.full(resized_w, IMAGE_PATCH_TOKEN)
+            use_single_crop_col_tokens = (
+                self.image_use_col_tokens
+                if self.use_single_crop_col_tokens is None
+                else self.use_single_crop_col_tokens
+            )
+            if use_single_crop_col_tokens:
+                per_row = np.concatenate([per_row, [IM_COL_TOKEN]], 0)
+            joint = [
+                [IM_START_TOKEN],
+                np.tile(per_row, [resized_h]),
+                [IM_END_TOKEN],
+            ]
+            return np.concatenate(joint)
+        per_row = np.full(width, IMAGE_PATCH_TOKEN)
+        if self.image_use_col_tokens:
+            per_row = np.concatenate([per_row, [IM_COL_TOKEN]], 0)
+        joint = [
+            [IM_START_TOKEN],
+            np.tile(per_row, [height]),
+            [IM_END_TOKEN],
+        ]
+        per_row = np.full(resized_w, IMAGE_PATCH_TOKEN)
+        use_single_crop_col_tokens = (
+            self.image_use_col_tokens
+            if self.use_single_crop_col_tokens is None
+            else self.use_single_crop_col_tokens
+        )
+        image_start_token = LOW_RES_IMAGE_START_TOKEN if self.use_single_crop_start_token else IM_START_TOKEN
+        if use_single_crop_col_tokens:
+            per_row = np.concatenate([per_row, [IM_COL_TOKEN]], 0)
+        joint = [
+            [image_start_token],
+            np.tile(per_row, [resized_h]),
+            [IM_END_TOKEN],
+        ] + joint
+
+        return np.concatenate(joint)
+
+    def get_video_string(
+        self,
+        video_grid: np.ndarray,
+        timestamps: np.ndarray,
+    ):
+        if self.use_frame_special_tokens:
+            start_token_id = FRAME_START_TOKEN
+            end_token_id = FRAME_END_TOKEN
+        else:
+            start_token_id = IM_START_TOKEN
+            end_token_id = IM_END_TOKEN
+
+        num_frames, h, w = video_grid
+        video_string: str = ""
+        for frame_idx, frame_time in enumerate(timestamps):
+            # `per-frame-compact` time mode
+            prev_space = " " if frame_idx > 0 else ""
+            frame_prefix = prev_space + f"{frame_time:.1f} "  # explicit whitespace before/after image tokens
+
+            video_string += frame_prefix
+            per_row = np.full(w, IMAGE_PATCH_TOKEN)
+            if self.video_use_col_tokens:
+                per_row = np.concatenate([per_row, [IM_COL_TOKEN]], 0)
+            extra_tokens = np.tile(per_row, [h])
+            video_tokens = [
+                [start_token_id],
+                extra_tokens,
+                [end_token_id],
+            ]
+            video_string += "".join(np.concatenate(video_tokens, 0))
+
+        return video_string
+
+    def insert_bos(
+        self,
+        input_ids: np.ndarray,
+        attention_mask: np.ndarray,
+        bos_token_id: int,
+        pad_token_id: int,
+    ):
+        """
+        Args:
+            input_ids: [B, S] array with left padding
+            attention_mask: [B, S] array (0 for pad, 1 for valid)
+            bos_token_id: int
+            pad_token_id: int
+        Returns:
+            input_ids_out: [B, S] or [B, S+1] array with bos inserted if needed
+            attention_mask_out: same shape as input_ids_out
+        """
+
+        need_to_expand = len(input_ids.shape) == 1
+        if need_to_expand:
+            input_ids = input_ids[None, :]
+            attention_mask = attention_mask[None, :]
+
+        B, S = input_ids.shape
+
+        # Handle zero-length sequence
+        if S == 0:
+            new_input_ids = np.full((B, 1), bos_token_id, dtype=input_ids.dtype)
+            new_attention_mask = np.ones((B, 1), dtype=attention_mask.dtype)
+            if need_to_expand:
+                new_input_ids = new_input_ids[0]
+                new_attention_mask = new_attention_mask[0]
+            return new_input_ids, new_attention_mask
+
+        first_valid_index = (attention_mask == 1).argmax(axis=-1)  # [B]
+        bos_already_present = np.all(input_ids[np.arange(B), first_valid_index] == bos_token_id)
+
+        if bos_already_present:
+            if need_to_expand:
+                input_ids = input_ids[0]
+                attention_mask = attention_mask[0]
+            return input_ids, attention_mask
+        else:
+            new_input_ids = np.full((B, S + 1), pad_token_id, dtype=input_ids.dtype)
+            new_attention_mask = np.zeros((B, S + 1), dtype=attention_mask.dtype)
+
+            src_idx = np.tile(np.arange(S), (B, 1))  # [B, S]
+            valid_mask = src_idx >= first_valid_index[:, None]  # [B, S]
+            tgt_idx = src_idx + 1  # shit right
+            batch_idx = np.tile(np.arange(B)[:, None], (1, S))  # [B, S]
+
+            # flatten valid_positions
+            flat_vals = input_ids[valid_mask]
+            flat_batch = batch_idx[valid_mask]
+            flat_tgt = tgt_idx[valid_mask]
+
+            new_input_ids[flat_batch, flat_tgt] = flat_vals
+            new_attention_mask[flat_batch, flat_tgt] = 1
+
+            insert_pos = first_valid_index
+            new_input_ids[np.arange(B), insert_pos] = bos_token_id
+            new_attention_mask[np.arange(B), insert_pos] = 1
+
+            if need_to_expand:
+                new_input_ids = new_input_ids[0]
+                new_attention_mask = new_attention_mask[0]
+
+            return new_input_ids, new_attention_mask
+
+    def __call__(
+        self,
+        text: TextInput | PreTokenizedInput | list[TextInput] | list[PreTokenizedInput] = None,
+        images: ImageInput = None,
+        videos: VideoInput = None,
+        **kwargs: Unpack[MolmoAct2ProcessorKwargs],
+    ) -> BatchFeature:
+        """
+
+        Args:
+            text (`str`, `list[str]`, `list[list[str]]`):
+                The sequence or batch of sequences to be encoded. Each sequence can be a string or a list of strings
+                (pretokenized string). If the sequences are provided as list of strings (pretokenized), you must set
+                `is_split_into_words=True` (to lift the ambiguity with a batch of sequences).
+            images (`PIL.Image.Image`, `np.ndarray`, `torch.Tensor`, `list[PIL.Image.Image]`, `list[np.ndarray]`, `list[torch.Tensor]`):
+                The image or batch of images to be prepared. Each image can be a PIL image, NumPy array or PyTorch
+                tensor. Both channels-first and channels-last formats are supported.
+            videos (`dict[str, Any]` or `list[dict[str, Any]]`):
+                The video or batch of videos to be prepared. Each video can be a dictionary with the following keys:
+                - `"frames"`: `np.ndarray` of shape (T, H, W, 3)
+                - `"timestamps"`: `np.ndarray` of shape (T,)
+                - `"sampled_fps"`: `float` (optional)
+                - `"sampling_augmentation"`: `str` (optional)
+            return_tensors (`str` or [`~utils.TensorType`], *optional*):
+                If set, will return tensors of a particular framework. Acceptable values are:
+                - `'tf'`: Return TensorFlow `tf.constant` objects.
+                - `'pt'`: Return PyTorch `torch.Tensor` objects.
+                - `'np'`: Return NumPy `np.ndarray` objects.
+                - `'jax'`: Return JAX `jnp.ndarray` objects.
+
+        Returns:
+            `BatchFeature`: A [`BatchFeature`] with the following fields:
+            - **input_ids** -- List of token ids to be fed to a model. Returned when `text` is not `None`.
+            - **attention_mask** -- List of indices specifying which tokens should be attended to by the model (when
+              `return_attention_mask=True` or if *"attention_mask"* is in `self.model_input_names` and if `text` is not `None`).
+            - **pixel_values** -- Pixel values to be fed to a model. Returned when `images` is not `None`.
+            - **image_token_pooling** -- Indices of the patches in `image_grids` to pool for each token in `image_tokens`.
+              Returned when `images` is not `None`.
+            - **image_grids** -- Grids of images. Returned when `images` is not `None`.
+            - **image_num_crops** -- Number of crops for each image. Returned when `images` is not `None`.
+            - **pixel_values_videos** -- Pixel values of videos to be fed to a model. Returned when `videos` is not `None`.
+            - **video_token_pooling** -- Indices of the patches in `video_grids` to pool for each token in `video_tokens`.
+              Returned when `videos` is not `None`.
+            - **video_grids** -- Grids of videos. Returned when `videos` is not `None`.
+        """
+
+        output_kwargs = self._merge_kwargs(
+            MolmoAct2ProcessorKwargs,
+            tokenizer_init_kwargs=self.tokenizer.init_kwargs,
+            **kwargs,
+        )
+
+        if images is not None:
+            image_inputs = self.image_processor(images, **output_kwargs["images_kwargs"])
+            image_grids = image_inputs["image_grids"]
+        else:
+            image_inputs = {}
+            image_grids = None
+
+        if videos is not None:
+            videos_inputs = self.video_processor(videos=videos, **output_kwargs["videos_kwargs"])
+            video_grids = videos_inputs["video_grids"]
+            # If user has not requested video metadata, pop it
+            if "return_metadata" not in kwargs:
+                video_metadata = videos_inputs.pop("video_metadata")
+            else:
+                video_metadata = videos_inputs["video_metadata"]
+        else:
+            videos_inputs = {}
+            video_grids = None
+
+        if not isinstance(text, list):
+            text = [text]
+
+        text = text.copy()  # below lines change text in-place
+
+        if image_grids is not None:
+            index = 0
+            for i in range(len(text)):
+                num_images = text[i].count(self.image_placeholder_token)
+                image_grids_i = image_grids[index : index + num_images]
+                for image_grid in image_grids_i:
+                    image_tokens = self.get_image_tokens(image_grid)
+                    image_string = "".join(image_tokens)
+                    text[i] = text[i].replace(self.image_placeholder_token, image_string, 1)
+                index += num_images
+
+        if video_grids is not None:
+            index = 0
+            for i in range(len(text)):
+                num_videos = text[i].count(self.video_placeholder_token)
+                assert num_videos in {0, 1}, "At most one video is supported for now"
+                video_grids_i = video_grids[index : index + num_videos]
+                metadata_i = video_metadata[index : index + num_videos]
+                for video_grid, metadata in zip(video_grids_i, metadata_i):
+                    video_string = self.get_video_string(
+                        video_grid,
+                        metadata.timestamps,
+                    )
+                    text[i] = text[i].replace(self.video_placeholder_token, video_string, 1)
+                index += num_videos
+
+        return_tensors = output_kwargs["text_kwargs"].pop("return_tensors", None)
+        return_mm_token_type_ids = output_kwargs["text_kwargs"].pop("return_mm_token_type_ids", False)
+        text_inputs = self.tokenizer(text, **output_kwargs["text_kwargs"])
+
+        input_ids = text_inputs["input_ids"]
+        attention_mask = text_inputs["attention_mask"]
+
+        input_ids = np.array(input_ids)
+        attention_mask = np.array(attention_mask)
+
+        bos = self.tokenizer.bos_token_id or self.tokenizer.eos_token_id
+        input_ids, attention_mask = self.insert_bos(
+            input_ids, attention_mask, bos, self.tokenizer.pad_token_id
+        )
+
+        if return_mm_token_type_ids:
+            image_tokens = np.array(self.image_token_ids).astype(input_ids.dtype)
+            token_type_ids = np.any(input_ids[:, :, None] == image_tokens[None, None, :], axis=-1)
+            text_inputs["token_type_ids"] = token_type_ids.tolist()
+
+        text_inputs["input_ids"] = input_ids.tolist()
+        text_inputs["attention_mask"] = attention_mask.tolist()
+
+        return BatchFeature(
+            data={**text_inputs, **image_inputs, **videos_inputs},
+            tensor_type=return_tensors,
+        )
+
+    def post_process_image_text_to_text(
+        self, generated_outputs, skip_special_tokens=True, clean_up_tokenization_spaces=False, **kwargs
+    ):
+        """
+        Post-process the output of the model to decode the text.
+
+        Args:
+            generated_outputs (`torch.Tensor` or `np.ndarray`):
+                The output of the model `generate` function. The output is expected to be a tensor of shape `(batch_size, sequence_length)`
+                or `(sequence_length,)`.
+            skip_special_tokens (`bool`, *optional*, defaults to `True`):
+                Whether or not to remove special tokens in the output. Argument passed to the tokenizer's `batch_decode` method.
+            clean_up_tokenization_spaces (`bool`, *optional*, defaults to `False`):
+                Whether or not to clean up the tokenization spaces. Argument passed to the tokenizer's `batch_decode` method.
+            **kwargs:
+                Additional arguments to be passed to the tokenizer's `batch_decode method`.
+
+        Returns:
+            `list[str]`: The decoded text.
+        """
+        return self.tokenizer.batch_decode(
+            generated_outputs,
+            skip_special_tokens=skip_special_tokens,
+            clean_up_tokenization_spaces=clean_up_tokenization_spaces,
+            **kwargs,
+        )
+
+
+MolmoAct2Processor.register_for_auto_class()

src/lerobot/policies/molmoact2/hf_model/video_processing_molmoact2.py

@@ -0,0 +1,997 @@
+#!/usr/bin/env python
+
+# Copyright 2026 The Allen Institute for Artificial Intelligence and The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+# ruff: noqa
+
+"""Video processor class for MolmoAct2"""
+
+from functools import partial
+import os
+import warnings
+from contextlib import redirect_stdout
+from io import BytesIO
+from urllib.parse import urlparse
+from typing import Optional, Union
+from collections.abc import Callable
+
+import numpy as np
+import requests
+import einops
+import torch
+import torchvision.transforms
+
+from transformers.image_utils import (
+    IMAGENET_STANDARD_MEAN,
+    IMAGENET_STANDARD_STD,
+    ImageInput,
+    PILImageResampling,
+    SizeDict,
+    validate_kwargs,
+)
+from transformers.video_utils import (
+    VideoInput,
+    is_valid_video,
+    make_batched_videos,
+    make_batched_metadata,
+    VideoMetadata,
+)
+from transformers.processing_utils import Unpack, VideosKwargs
+from transformers.video_processing_utils import BaseVideoProcessor
+from transformers.utils import logging
+from transformers.feature_extraction_utils import BatchFeature
+from transformers.utils import (
+    is_av_available,
+    is_decord_available,
+    is_torchcodec_available,
+    is_yt_dlp_available,
+    TensorType,
+    logging,
+    to_numpy,
+)
+
+
+logger = logging.get_logger(__name__)
+
+MAX_VIDEO_FPS = 8
+
+
+def normalize_image(
+    image: np.ndarray,
+    image_mean: list[float],
+    image_std: list[float],
+) -> np.ndarray:
+    if np.allclose(image_mean, [0.5, 0.5, 0.5]) and np.allclose(image_std, [0.5, 0.5, 0.5]):
+        return image * np.asarray(2.0, dtype=np.float32) - np.asarray(1.0, dtype=np.float32)
+    image -= np.array(image_mean, dtype=np.float32)[None, None, :]
+    image /= np.array(image_std, dtype=np.float32)[None, None, :]
+    return image
+
+
+def resize_image(
+    image: np.ndarray,
+    desired_output_size: list[int],
+    resample: PILImageResampling,
+) -> np.ndarray:
+    if len(image.shape) == 3:
+        is_video = False
+        image = torch.permute(torch.from_numpy(image), [2, 0, 1])
+    else:
+        is_video = True
+        image = torch.permute(torch.from_numpy(image), [0, 3, 1, 2])
+    dtype = image.dtype
+    if torch.is_floating_point(image):
+        in_min = 0.0
+        in_max = 1.0
+        resized = torchvision.transforms.Resize(
+            desired_output_size,
+            resample,
+            antialias=False,
+        )(image)
+        resized = torch.clip(resized, 0.0, 1.0).to(dtype)
+    else:
+        assert image.dtype == torch.uint8, "SigLIP expects float images or uint8 images, but got {}".format(
+            image.dtype
+        )
+        in_min = 0.0
+        in_max = 255.0
+        resized = torchvision.transforms.Resize(
+            desired_output_size,
+            resample,
+            antialias=False,
+        )(image)
+        resized = torch.clip(resized, 0, 255).to(dtype)
+
+    resized = resized.to(torch.float32)
+    resized = (resized - in_min) / (in_max - in_min)
+
+    if is_video:
+        resized = torch.permute(resized, [0, 2, 3, 1]).numpy()
+    else:
+        resized = torch.permute(resized, [1, 2, 0]).numpy()
+
+    return resized
+
+
+def build_resized_image(
+    image: np.ndarray,
+    base_image_input_size: list[int],
+    resample: PILImageResampling,
+    image_mean: list[float],
+    image_std: list[float],
+    image_patch_size: int,
+) -> tuple[np.ndarray, np.ndarray]:
+    resized = resize_image(
+        image,
+        base_image_input_size,
+        resample,
+    )
+    resized = normalize_image(resized, image_mean, image_std)
+    if len(resized.shape) == 3:
+        resized = np.expand_dims(resized, 0)
+    crop_patch_w = base_image_input_size[1] // image_patch_size
+    crop_patch_h = base_image_input_size[0] // image_patch_size
+    resize_idx = np.arange(crop_patch_w * crop_patch_h).reshape([crop_patch_h, crop_patch_w])
+    return resized, resize_idx
+
+
+def batch_pixels_to_patches(array: np.ndarray, patch_size: int) -> np.ndarray:
+    """Reshape images of [n_images, h, w, 3] -> [n_images, n_patches, pixels_per_patch]"""
+    if len(array.shape) == 3:
+        n_crops, h, w = array.shape
+        h_patches = h // patch_size
+        w_patches = w // patch_size
+        array = np.reshape(array, [n_crops, h_patches, patch_size, w_patches, patch_size])
+        array = np.transpose(array, [0, 1, 3, 2, 4])
+        array = np.reshape(array, [n_crops, h_patches * w_patches, patch_size * patch_size])
+        return array
+    else:
+        n_crops, h, w, c = array.shape
+        h_patches = h // patch_size
+        w_patches = w // patch_size
+        array = np.reshape(array, [n_crops, h_patches, patch_size, w_patches, patch_size, c])
+        array = np.transpose(array, [0, 1, 3, 2, 4, 5])
+        array = np.reshape(array, [n_crops, h_patches * w_patches, patch_size * patch_size * c])
+        return array
+
+
+def arange_for_pooling(
+    idx_arr: np.ndarray,
+    pool_h: int,
+    pool_w: int,
+) -> np.ndarray:
+    h_pad = pool_h * ((idx_arr.shape[0] + pool_h - 1) // pool_h) - idx_arr.shape[0]
+    w_pad = pool_w * ((idx_arr.shape[1] + pool_w - 1) // pool_w) - idx_arr.shape[1]
+    idx_arr = np.pad(
+        idx_arr,
+        [[h_pad // 2, (h_pad + 1) // 2], [w_pad // 2, (w_pad + 1) // 2]],
+        mode="constant",
+        constant_values=-1,
+    )
+    return einops.rearrange(idx_arr, "(h dh) (w dw) -> h w (dh dw)", dh=pool_h, dw=pool_w)
+
+
+def image_to_patches_and_grids(
+    image: ImageInput,
+    base_image_input_size: list[int],
+    resample: PILImageResampling,
+    image_mean: list[float],
+    image_std: list[float],
+    image_patch_size: int,
+    image_pooling_w: int,
+    image_pooling_h: int,
+) -> tuple[np.ndarray, np.ndarray, np.ndarray]:
+    """
+    :return image_grids, the shape of each image after pooling
+    :return crops, the image crops to processes with the ViT
+    :return pooled_patch_idx, for each patch_id tokens in `image_tokens`, the indices of the
+                                patches in `crops` to pool for that token, masked with -1
+    """
+    if isinstance(base_image_input_size, int):
+        base_image_input_size = (base_image_input_size, base_image_input_size)
+
+    pooling_w = image_pooling_w
+    pooling_h = image_pooling_h
+
+    resized, resize_idx = build_resized_image(
+        image,
+        base_image_input_size,
+        resample,
+        image_mean,
+        image_std,
+        image_patch_size,
+    )
+    pooling_idx = arange_for_pooling(resize_idx, pooling_h, pooling_w)
+    h, w = pooling_idx.shape[:2]
+    pooling_idx = pooling_idx.reshape([-1, pooling_h * pooling_w])
+    image_grid = [h, w]
+    return (
+        image_grid,
+        batch_pixels_to_patches(resized, image_patch_size),
+        pooling_idx,
+    )
+
+
+def get_candidate_target_fps(
+    video_fps: int | float,
+    sampling_fps: int | float,
+    max_fps: int | float = MAX_VIDEO_FPS,
+) -> list[float]:
+    """
+    Return the subset of `video_fps` factors that remain multiples of `sampling_fps`.
+
+    Examples:
+        >>> get_candidate_target_fps(video_fps=6, sampling_fps=2)
+        [2, 6]
+        >>> get_candidate_target_fps(video_fps=5, sampling_fps=1)
+        [1, 5]
+        >>> get_candidate_target_fps(video_fps=2, sampling_fps=2)
+        [2]
+        >>> get_candidate_target_fps(video_fps=5, sampling_fps=2)
+        Traceback (most recent call last):
+            ...
+        ValueError: sampling_fps=2 must divide video_fps=5 to produce consistent frame steps.
+    """
+    video_fps = int(video_fps)
+    sampling_fps = int(sampling_fps)
+    max_fps = int(max_fps)
+
+    if sampling_fps is None:
+        raise ValueError("sampling_fps must be provided")
+    if video_fps <= 0 or sampling_fps <= 0:
+        raise ValueError(f"video_fps and sampling_fps must be positive (got {video_fps}, {sampling_fps})")
+    if video_fps % sampling_fps != 0:
+        raise ValueError(f"sampling_fps={sampling_fps} must divide video_fps={video_fps}.")
+
+    candidates = []
+    for candidate in range(sampling_fps, video_fps + 1, sampling_fps):
+        if candidate > max_fps:
+            break
+        if video_fps % candidate == 0:
+            candidates.append(float(candidate))
+
+    return candidates
+
+
+def read_video_decord(
+    video_path,
+    sample_timestamps_fn: Callable,
+    **kwargs,
+) -> np.ndarray:
+    """
+    Decode a video using the Decord backend.
+
+    Args:
+        video_path (`str`):
+            Path to the video file.
+        sample_timestamps_fn (`Callable`):
+            A callable function that will return timestamps at which the video should be sampled.
+
+    Returns:
+        tuple[`np.array`, `VideoMetadata`]: A tuple containing:
+            - Numpy array of frames in RGB (shape: [num_frames, height, width, 3]).
+            - `VideoMetadata` object.
+    """
+    # Lazy import from decord
+    import importlib
+
+    decord = importlib.import_module("decord")
+
+    vr = decord.VideoReader(uri=video_path, ctx=decord.cpu(0))  # decord has problems with gpu
+    video_fps = vr.get_avg_fps()
+    total_num_frames = len(vr)
+    time_stamps = vr.get_frame_timestamp(list(range(len(vr))))
+    duration = time_stamps[-1][1] - time_stamps[0][0]
+
+    metadata = VideoMetadata(
+        total_num_frames=int(total_num_frames),
+        fps=float(video_fps),
+        duration=float(duration),
+        video_backend="decord",
+    )
+
+    target_timestamps = sample_timestamps_fn(metadata=metadata, **kwargs)
+    target_timestamps = np.array(target_timestamps)
+    offset = time_stamps[0, 0]
+
+    ix = np.searchsorted(time_stamps[:, 1], target_timestamps + offset, side="right")
+    ix = np.minimum(ix, len(time_stamps) - 1)
+
+    video = vr.get_batch(ix).asnumpy()
+    metadata.update(
+        {
+            "frames_indices": target_timestamps * video_fps,
+            "height": video.shape[1],
+            "width": video.shape[2],
+        }
+    )
+    return video, metadata
+
+
+def read_video_torchcodec(
+    video_path,
+    sample_timestamps_fn: Callable,
+    **kwargs,
+) -> np.ndarray:
+    """
+    Decode a video using torchcodec decoder.
+
+    Args:
+        video_path (`str`):
+            Path to the video file.
+        sample_timestamps_fn (`Callable`):
+            A callable function that will return timestamps at which the video should be sampled.
+
+    Returns:
+        tuple[`np.array`, `VideoMetadata`]: A tuple containing:
+            - Numpy array of frames in RGB (shape: [num_frames, height, width, 3]).
+            - `VideoMetadata` object.
+    """
+    # Lazy import torchcodec
+    import importlib
+
+    torchcodec = importlib.import_module("torchcodec")
+
+    decoder = torchcodec.decoders.VideoDecoder(
+        video_path,
+        # Interestingly `exact` mode takes less than approximate when we load the whole video
+        seek_mode="exact",
+        # Allow FFmpeg decide on the number of threads for efficiency
+        num_ffmpeg_threads=0,
+    )
+    # If the first frame starts at > 0, we effectively clip the video starting at that time
+    # since (most) video players would also skip to that time
+    time_offset = decoder.metadata.begin_stream_seconds_from_content
+    # Note this duration does assume we started playing at `time_offset`
+    duration = decoder.metadata.duration_seconds
+
+    metadata = VideoMetadata(
+        total_num_frames=decoder.metadata.num_frames,
+        fps=decoder.metadata.average_fps,
+        duration=duration,
+        video_backend="torchcodec",
+        height=decoder.metadata.height,
+        width=decoder.metadata.width,
+    )
+
+    target_timestamps = sample_timestamps_fn(metadata=metadata, **kwargs)
+
+    # Floating point/rounding issues might cause `target_timestamps` to be very slightly
+    # out-of-bounds, to handle this we sanity check then clip them
+    assert all(x >= 0 for x in target_timestamps)
+    assert all(x < duration + 1e-6 for x in target_timestamps)
+    # 1e-6 padding since torchcodec can throw out-of-bounds errors even if you ask for the
+    # exact boundary value, we should still get the first/last frame anyway
+    max_timestamp = decoder.metadata.end_stream_seconds_from_content - 1e-6
+    min_timestamp = decoder.metadata.begin_stream_seconds_from_content + 1e-6
+    # Note we avoid using numpy ops here to reduce floating precision issues
+    timestamps = [x + time_offset for x in target_timestamps]
+    timestamps = [max(min_timestamp, min(max_timestamp, x)) for x in timestamps]
+
+    video = (
+        decoder.get_frames_played_at(timestamps).data.numpy().transpose(0, 2, 3, 1)
+    )  # Convert to THWC format
+    target_timestamps = np.array(target_timestamps)
+    metadata.frames_indices = target_timestamps * metadata.fps
+
+    return video, metadata
+
+
+def read_video_pyav(
+    video_path,
+    sample_timestamps_fn: Callable,
+    **kwargs,
+) -> np.ndarray:
+    """
+    Decode a video using the PyAV backend.
+
+    Args:
+        video_path (`str`):
+            Path to the video file.
+        sample_timestamps_fn (`Callable`):
+            A callable function that will return timestamps at which the video should be sampled.
+
+    Returns:
+        tuple[`np.array`, `VideoMetadata`]: A tuple containing:
+            - Numpy array of frames in RGB (shape: [num_frames, height, width, 3]).
+            - `VideoMetadata` object.
+    """
+    # Lazy import torchcodec
+    import importlib
+
+    av = importlib.import_module("av")
+
+    with av.open(video_path) as container:
+        video_stream = container.streams.video[0]
+        fps = video_stream.average_rate or video_stream.guessed_rate
+        it = container.decode(video=0)
+        frames = list(it)
+
+        stream = container.streams.video[0]
+        start = frames[0].pts * stream.time_base
+        container_end = stream.duration
+        if container_end is not None:
+            container_end *= stream.time_base
+        if container_end is None or container_end < frames[-1].pts:
+            # Some problem with stream duration, so use the frame PTS directly
+            # and guess the duration of the last frame
+            end = frames[-1].pts * stream.time_base + 1 / fps
+        else:
+            end = container_end
+        duration = float(end - start)
+
+        metadata = VideoMetadata(
+            total_num_frames=len(frames),
+            fps=float(fps),
+            duration=float(duration),
+            video_backend="pyav",
+            height=video_stream.height,
+            width=video_stream.width,
+        )
+
+        target_timestamps = sample_timestamps_fn(metadata=metadata, **kwargs)
+        offset = float(start)
+
+        target_timestamps = np.array(target_timestamps)
+        end_time_stamps = np.array([float(frame.pts * stream.time_base) for frame in frames[1:]] + [duration])
+        indices = np.searchsorted(end_time_stamps, target_timestamps + offset, side="right")
+        indices = np.minimum(indices, len(end_time_stamps) - 1)
+
+        video = np.stack(
+            [frames[i].to_ndarray(format="rgb24", channel_last=True) for i in indices],
+            axis=0,
+        )
+
+        metadata.frames_indices = target_timestamps * fps
+
+        return video, metadata
+
+
+VIDEO_DECODERS = {
+    "decord": read_video_decord,
+    "torchcodec": read_video_torchcodec,
+    "pyav": read_video_pyav,
+}
+
+
+def load_video(
+    video: VideoInput,
+    backend: str = "decord",
+    sample_timestamps_fn: Callable | None = None,
+    **kwargs,
+):
+    """
+    Loads `video` to a numpy array.
+
+    Args:
+        video (`VideoInput`):
+            The video to convert to the numpy array format. Can be a link to video or local path.
+        backend (`str`, *optional*, defaults to `"decord"`):
+            The backend to use when loading the video. Can be any of ["decord", "pyav", ""torchcodec"]. Defaults to "decord".
+        sample_timestamps_fn (`Callable`):
+            A callable function that will return timestamps at which the video should be sampled.
+    """
+
+    # Early exit if provided an array or `PIL` frames
+    if not isinstance(video, str):
+        metadata = [None] * len(video)
+        return video, metadata
+
+    if urlparse(video).netloc in ["www.youtube.com", "youtube.com"]:
+        if not is_yt_dlp_available():
+            raise ImportError("To load a video from YouTube url you have  to install `yt_dlp` first.")
+        # Lazy import from yt_dlp
+        import importlib
+
+        yt_dlp = importlib.import_module("yt_dlp")
+
+        buffer = BytesIO()
+        with redirect_stdout(buffer), yt_dlp.YoutubeDL() as f:
+            f.download([video])
+        bytes_obj = buffer.getvalue()
+        file_obj = BytesIO(bytes_obj)
+    elif video.startswith("http://") or video.startswith("https://"):
+        file_obj = BytesIO(requests.get(video, timeout=10).content)
+    elif os.path.isfile(video):
+        file_obj = video
+    else:
+        raise TypeError(
+            "Incorrect format used for video. Should be an url linking to an video or a local path."
+        )
+
+    # can also load with decord, but not cv2/torchvision
+    # both will fail in case of url links
+    video_is_url = video.startswith("http://") or video.startswith("https://")
+    if video_is_url and backend == "opencv":
+        raise ValueError("If you are trying to load a video from URL, you cannot use 'opencv' as backend")
+
+    if (
+        (not is_decord_available() and backend == "decord")
+        or (not is_torchcodec_available() and backend == "torchcodec")
+        or (not is_av_available() and backend == "pyav")
+    ):
+        raise ImportError(
+            f"You chose backend={backend} for loading the video but the required library is not found in your environment "
+            f"Make sure to install {backend} before loading the video."
+        )
+
+    video_decoder = VIDEO_DECODERS[backend]
+    video, metadata = video_decoder(file_obj, sample_timestamps_fn, **kwargs)
+    return video, metadata
+
+
+def get_target_fps(
+    video_fps: float,
+    max_frames: int,
+    total_frames: int,
+    frame_sample_mode: str,
+    candidate_target_fps: tuple[float],
+) -> float:
+    """
+    Get the target fps that best spans the video and has the most frames sampled
+    """
+    num_frames_sampled = 0
+    selected_target_fps = None
+    for target_fps in candidate_target_fps:
+        step_size = max(int(video_fps / target_fps), 1)
+        num_frames_sampled_at_fps = int(total_frames / step_size)
+        if num_frames_sampled == 0:
+            if "uniform" in frame_sample_mode:
+                if num_frames_sampled_at_fps > max_frames:
+                    break
+            selected_target_fps = target_fps
+            num_frames_sampled = num_frames_sampled_at_fps
+
+        else:
+            # the candidate sampling fps increases so frame count can't decrease
+            assert num_frames_sampled <= num_frames_sampled_at_fps
+            if num_frames_sampled_at_fps > max_frames:
+                # choose the sampling fps that spans the video
+                continue
+
+            elif num_frames_sampled_at_fps > num_frames_sampled:
+                # both are less than max_frames, choose the one with higher density of frames sampled
+                selected_target_fps = target_fps
+                num_frames_sampled = num_frames_sampled_at_fps
+    return selected_target_fps
+
+
+def get_frame_times_and_chosen_fps(selected_target_fps, total_frames, max_frames, video_fps):
+    if selected_target_fps is None:
+        frame_indices = np.linspace(0, total_frames, max_frames, endpoint=False, dtype=int)
+    else:
+        step_size = max(int(video_fps / selected_target_fps), 1)
+        frame_indices = np.arange(0, total_frames, step_size)
+    if len(frame_indices) > max_frames:
+        frame_indices = frame_indices[:max_frames]
+    return selected_target_fps, frame_indices
+
+
+class MolmoAct2VideoProcessorKwargs(VideosKwargs, total=False):
+    patch_size: int | None
+    pooling_size: list[int] | None
+    frame_sample_mode: str | None
+    max_fps: int | None
+    sampling_fps: int | None
+
+
+class MolmoAct2VideoProcessor(BaseVideoProcessor):
+    resample = PILImageResampling.BILINEAR
+    size = {"height": 378, "width": 378}
+    image_mean = IMAGENET_STANDARD_MEAN
+    image_std = IMAGENET_STANDARD_STD
+    do_resize = True
+    do_rescale = True
+    do_normalize = True
+    do_convert_rgb = True
+    patch_size = 14
+    pooling_size = [3, 3]
+    do_sample_frames = True
+    frame_sample_mode = "uniform_last_frame"
+    max_fps = 2
+    sampling_fps = 2
+    valid_kwargs = MolmoAct2VideoProcessorKwargs
+    model_input_names = ["pixel_values_videos", "video_token_pooling", "video_grids"]
+
+    def __init__(self, **kwargs: Unpack[MolmoAct2VideoProcessorKwargs]):
+        super().__init__(**kwargs)
+        if self.size is not None and (
+            self.size.get("height", None) is None or self.size.get("width", None) is None
+        ):
+            raise ValueError("size must contain 'height' and 'width' keys.")
+
+    def _further_process_kwargs(
+        self,
+        size: SizeDict | None = None,
+        **kwargs,
+    ) -> dict:
+        """
+        Update kwargs that need further processing before being validated
+        Can be overridden by subclasses to customize the processing of kwargs.
+        """
+        if size is not None and ("height" not in size or "width" not in size):
+            raise ValueError("size must contain 'height' and 'width' keys.")
+
+        return super()._further_process_kwargs(size=size, **kwargs)
+
+    def sample_times(
+        self,
+        metadata: VideoMetadata,
+        frame_sample_mode: str,
+        num_frames: int,
+        max_fps: int | None = None,
+        sampling_fps: int | None = None,
+        **kwargs,
+    ) -> np.ndarray:
+        """
+        Time-based sampling if an array video is passed
+        Args:
+            metadata (`VideoMetadata`):
+                Metadata of the video containing information about total duration, fps and total number of frames.
+            frame_sample_mode (`str`, *optional*):
+                Mode to sample frames. Defaults to `self.frame_sample_mode`.
+            num_frames (`int`, *optional*):
+                Maximum number of frames to sample. Defaults to `self.num_frames`.
+            man_fps (`int`, *optional*):
+                Maximum frames per second to sample.
+            sampling_fps (`int`, *optional*):
+                Sampling frames per second. Defaults to `self.sampling_fps`.
+                Used when `frame_sample_mode` is `"fps"`.
+        """
+        frame_sample_mode = frame_sample_mode or self.frame_sample_mode
+        num_frames = num_frames or self.num_frames
+        sampling_fps = sampling_fps or self.sampling_fps
+
+        duration = metadata.duration or metadata.total_num_frames / metadata.fps
+        if frame_sample_mode == "fps":
+            candidate_target_fps = get_candidate_target_fps(metadata.fps, sampling_fps)
+            # Try larger and larger FPSs until we hit one that can't span the video
+            target_fps = candidate_target_fps[0]
+            for candidate_fps in candidate_target_fps[1:]:
+                if num_frames / candidate_fps < duration:
+                    break
+                target_fps = candidate_fps
+            times = np.arange(0, num_frames) / target_fps
+            times = times[times < duration]
+            return times
+        elif frame_sample_mode == "uniform_last_frame":
+            if max_fps is not None:
+                max_duration = (num_frames - 1) / max_fps  # -1 to include the last frame
+                if max_duration < duration:
+                    times = np.linspace(0, duration, num=num_frames, endpoint=True, dtype=np.float64)
+                else:
+                    times = np.arange(0.0, stop=duration, step=1 / max_fps)
+                    times = np.concatenate([times, [duration]], axis=0)
+                    assert len(times) <= num_frames
+            else:
+                times = np.linspace(0, duration, num=num_frames, endpoint=True, dtype=np.float64)
+            return times
+        else:
+            raise NotImplementedError(frame_sample_mode)
+
+    def sample_frames(
+        self,
+        metadata: VideoMetadata,
+        frame_sample_mode: str | None = None,
+        num_frames: int | None = None,
+        max_fps: int | None = None,
+        sampling_fps: int | None = None,
+        **kwargs,
+    ) -> np.ndarray:
+        """
+        Frame-based sampling if an array video is passed
+        Args:
+            metadata (`VideoMetadata`):
+                Metadata of the video containing information about total duration, fps and total number of frames.
+            frame_sample_mode (`str`, *optional*):
+                Mode to sample frames. Defaults to `self.frame_sample_mode`.
+            num_frames (`int`, *optional*):
+                Maximum number of frames to sample. Defaults to `self.num_frames`.
+            max_fps (`int`, *optional*):
+                Maximum frames per second to sample.
+            sampling_fps (`int`, *optional*):
+                Sampling frames per second. Defaults to `self.sampling_fps`.
+                Used when `frame_sample_mode` is `"fps"`.
+        """
+        frame_sample_mode = frame_sample_mode or self.frame_sample_mode
+        num_frames = num_frames or self.num_frames
+        sampling_fps = sampling_fps or self.sampling_fps
+
+        total_num_frames = metadata.total_num_frames
+        if frame_sample_mode == "uniform_last_frame" and max_fps is not None:
+            duration = total_num_frames / metadata.fps
+            if total_num_frames <= 2:
+                return np.arange(total_num_frames).astype(int)
+            if duration > (num_frames - 1) / max_fps:  # -1 to include the last frame
+                # uniform fallback
+                indices = np.linspace(
+                    0,
+                    total_num_frames - 1,
+                    num=min(num_frames, total_num_frames),
+                    endpoint=True,
+                ).astype(int)
+                return indices
+            else:
+                float_indices = np.arange(
+                    0.0,
+                    stop=total_num_frames - 1,
+                    step=float(metadata.fps / max_fps),
+                )
+                if np.round(float_indices[-1]) != total_num_frames - 1:
+                    float_indices = np.concatenate([float_indices, [total_num_frames - 1]], axis=0)
+                indices = np.round(float_indices).astype(int)
+                assert indices[-1] < total_num_frames
+                assert len(float_indices) <= num_frames
+                return indices
+        elif frame_sample_mode == "uniform_last_frame":
+            indices = np.linspace(
+                0,
+                total_num_frames - 1,
+                num=min(num_frames, total_num_frames),
+                endpoint=True,
+            ).astype(int)
+            return indices
+        elif frame_sample_mode == "fps":
+            candidate_target_fps = get_candidate_target_fps(metadata.fps, sampling_fps)
+            selected_target_fps = get_target_fps(
+                metadata.fps,
+                num_frames,
+                total_num_frames,
+                frame_sample_mode,
+                candidate_target_fps,
+            )
+            _, indices = get_frame_times_and_chosen_fps(
+                selected_target_fps,
+                total_num_frames,
+                num_frames,
+                metadata.fps,
+            )
+            return indices
+        else:
+            raise NotImplementedError(frame_sample_mode)
+
+    def fetch_videos(self, video_url_or_urls: str | list[str] | list[list[str]], sample_timestamps_fn=None):
+        """
+        Convert a single or a list of urls into the corresponding `np.array` objects.
+
+        If a single url is passed, the return value will be a single object. If a list is passed a list of objects is
+        returned.
+        """
+        if (not is_decord_available()) and (not is_torchcodec_available()) and (not is_av_available()):
+            raise ImportError(
+                "MolmoAct2VideoProcessor requires `decord`, `torchcodec`, or `av` to be installed."
+            )
+
+        if is_decord_available():
+            backend = "decord"
+        elif is_torchcodec_available():
+            warnings.warn(
+                "`decord` is not installed and cannot be used to decode the video by default. "
+                "Falling back to `torchcodec`."
+            )
+            backend = "torchcodec"
+        else:
+            warnings.warn(
+                "`decord` is not installed and cannot be used to decode the video by default. "
+                "Falling back to `PyAV`."
+            )
+            backend = "pyav"
+
+        if isinstance(video_url_or_urls, list):
+            return list(
+                zip(
+                    *[
+                        self.fetch_videos(x, sample_timestamps_fn=sample_timestamps_fn)
+                        for x in video_url_or_urls
+                    ]
+                )
+            )
+        else:
+            return load_video(video_url_or_urls, backend=backend, sample_timestamps_fn=sample_timestamps_fn)
+
+    def _decode_and_sample_videos(
+        self,
+        videos: VideoInput,
+        video_metadata: VideoMetadata | dict,
+        do_sample_frames: bool | None = None,
+        sample_indices_fn: Callable | None = None,
+        sample_timestamps_fn: Callable | None = None,
+    ):
+        """
+        Decode input videos and sample frames if needed.
+        """
+        videos = make_batched_videos(videos)
+        video_metadata = make_batched_metadata(videos, video_metadata=video_metadata)
+
+        # Framed-based sampling if an array video is passed
+        # Otherwise, time-based sampling with decoding
+        if is_valid_video(videos[0]) and do_sample_frames:
+            assert video_metadata[0].fps is not None, "FPS must be provided for video input"
+            sampled_videos = []
+            sampled_metadata = []
+            for video, metadata in zip(videos, video_metadata):
+                indices = sample_indices_fn(metadata=metadata)
+                metadata.frames_indices = indices
+                sampled_videos.append(video[indices])
+                sampled_metadata.append(metadata)
+            videos = sampled_videos
+            video_metadata = sampled_metadata
+        elif not is_valid_video(videos[0]):
+            if sample_indices_fn is None:
+                logger.warning(
+                    "do_sample_frames is False, but video array is not provided: "
+                    "Will decode the video and sample frames using MolmoAct2's default sampling mode"
+                )
+            if isinstance(videos[0], list):
+                raise ValueError("A list of images is not supported for video input!")
+            else:
+                videos, video_metadata = self.fetch_videos(videos, sample_timestamps_fn=sample_timestamps_fn)
+
+        return videos, video_metadata
+
+    def _prepare_input_videos(
+        self,
+        videos: VideoInput,
+        **kwargs,
+    ) -> list[np.ndarray]:
+        processed_videos = [to_numpy(video) for video in videos]
+        return processed_videos
+
+    def preprocess(
+        self,
+        videos: VideoInput,
+        **kwargs: Unpack[MolmoAct2VideoProcessorKwargs],
+    ) -> BatchFeature:
+        validate_kwargs(
+            captured_kwargs=kwargs.keys(),
+            valid_processor_keys=list(self.valid_kwargs.__annotations__.keys()) + ["return_tensors"],
+        )
+
+        # Set default kwargs from self. This ensures that if a kwarg is not provided
+        # by the user, it gets its default value from the instance, or is set to None.
+        for kwarg_name in self.valid_kwargs.__annotations__:
+            kwargs.setdefault(kwarg_name, getattr(self, kwarg_name, None))
+
+        do_sample_frames = kwargs.pop("do_sample_frames")
+        video_metadata = kwargs.pop("video_metadata")
+
+        sample_indices_fn = partial(self.sample_frames, **kwargs) if do_sample_frames else None
+        sample_timestamps_fn = partial(self.sample_times, **kwargs)
+        videos, video_metadata = self._decode_and_sample_videos(
+            videos,
+            video_metadata=video_metadata,
+            do_sample_frames=do_sample_frames,
+            sample_indices_fn=sample_indices_fn,
+            sample_timestamps_fn=sample_timestamps_fn,
+        )
+        videos = self._prepare_input_videos(videos=videos)
+
+        kwargs = self._further_process_kwargs(**kwargs)
+
+        return_metadata = kwargs.pop("return_metadata")
+        preprocessed_videos = self._preprocess(videos=videos, **kwargs)
+        if return_metadata:
+            preprocessed_videos["video_metadata"] = video_metadata
+        return preprocessed_videos
+
+    def _preprocess(
+        self,
+        videos: list[np.ndarray],
+        size: SizeDict | None = None,
+        resample: PILImageResampling | None = None,
+        image_mean: float | list[float] | None = None,
+        image_std: float | list[float] | None = None,
+        do_convert_rgb: bool | None = None,
+        patch_size: int | None = None,
+        pooling_size: list[int] | None = None,
+        return_tensors: str | TensorType | None = None,
+        **kwargs,
+    ) -> BatchFeature:
+        """
+        Preprocess a video for the model.
+        Args:
+            videos (`VideoInput`):
+                Video to preprocess.
+            size (`SizeDict`, *optional*, defaults to `self.size`):
+                Size of the image after resizing.
+            resample (`PILImageResampling`, *optional*, defaults to `self.resample`):
+                Resampling filter to use when resizing the image. This can be one of the enum `PILImageResampling`. Only
+                has an effect if `do_resize` is set to `True`.
+            image_mean (`float` or `list[float]`, *optional*, defaults to `self.image_mean`):
+                Image mean to use for normalization. Only has an effect if `do_normalize` is set to `True`.
+            image_std (`float` or `list[float]`, *optional*, defaults to `self.image_std`):
+                Image standard deviation to use for normalization. Only has an effect if `do_normalize` is set to
+                `True`.
+            do_convert_rgb (`bool`, *optional*, defaults to `self.do_convert_rgb`):
+                Whether to convert the image to RGB.
+            patch_size (`int`, *optional*, defaults to `self.patch_size`):
+                The spatial patch size of the vision encoder.
+            pooling_size (`list[int]`, *optional*, defaults to `self.pooling_size`):
+                The pooling size of the vision adapter.
+            return_tensors (`str` or `TensorType`, *optional*):
+                The type of tensors to return. Can be one of:
+                - Unset: Return a list of `np.ndarray`.
+                - `TensorType.TENSORFLOW` or `'tf'`: Return a batch of type `tf.Tensor`.
+                - `TensorType.PYTORCH` or `'pt'`: Return a batch of type `torch.Tensor`.
+                - `TensorType.NUMPY` or `'np'`: Return a batch of type `np.ndarray`.
+                - `TensorType.JAX` or `'jax'`: Return a batch of type `jax.numpy.ndarray`.
+
+        Returns:
+            A `BatchFeature` containing the following keys:
+                - `pixel_values_videos`: The preprocessed videos.
+                - `video_token_pooling`: The indices of the patches in `crops` to pool for each token in `video_tokens`.
+                - `video_grids`: The video grids.
+        """
+        if size.height is None or size.width is None:
+            raise ValueError("size must contain 'height' and 'width' keys.")
+
+        base_image_input_size = [size.height, size.width]
+
+        resample = resample or self.resample
+        image_mean = image_mean or self.image_mean
+        image_std = image_std or self.image_std
+        do_convert_rgb = do_convert_rgb or self.do_convert_rgb
+
+        patch_size = patch_size or self.patch_size
+        pooling_size = pooling_size or self.pooling_size
+
+        image_pooling_h, image_pooling_w = pooling_size
+
+        batch_grids = []
+        batch_crops = []
+        batch_pooled_patches_idx = []
+
+        for video in videos:
+            all_crops = []
+            pooled_patches_idx = []
+
+            for frame in video:
+                image_grid, crops, pooled_idx = image_to_patches_and_grids(
+                    frame,
+                    base_image_input_size,
+                    resample,
+                    image_mean,
+                    image_std,
+                    patch_size,
+                    image_pooling_w,
+                    image_pooling_h,
+                )
+                offset = sum(np.prod(x.shape[:2]) for x in all_crops)
+                pooled_idx_with_offset = np.where(pooled_idx >= 0, pooled_idx + offset, pooled_idx)
+                pooled_patches_idx.append(pooled_idx_with_offset)
+                all_crops.append(crops)
+
+            video_grid = np.array([len(video), image_grid[0], image_grid[1]])
+            all_crops = np.concatenate(all_crops, 0)
+            pooled_patches_idx = np.concatenate(pooled_patches_idx, 0)
+
+            batch_grids.append(video_grid)
+            batch_crops.append(all_crops)
+            batch_pooled_patches_idx.append(pooled_patches_idx)
+
+        video_grids = np.stack(batch_grids, 0)
+        pixel_values_videos = np.concatenate(batch_crops, 0)
+        video_token_pooling = np.concatenate(batch_pooled_patches_idx, 0)
+
+        data = dict(
+            pixel_values_videos=pixel_values_videos,
+            video_token_pooling=video_token_pooling,
+            video_grids=video_grids,
+        )
+
+        return BatchFeature(data, tensor_type=return_tensors)
+
+
+MolmoAct2VideoProcessor.register_for_auto_class()

src/lerobot/policies/pi0/modeling_pi0.py

@@ -15,7 +15,6 @@
 # limitations under the License.
 
 import builtins
-import copy
 import logging
 import math
 from collections import deque
@@ -30,6 +29,7 @@ from lerobot.utils.import_utils import _transformers_available, require_package
 
 # Conditional import for type checking and lazy loading
 if TYPE_CHECKING or _transformers_available:
+    from transformers.cache_utils import DynamicCache
     from transformers.models.auto import CONFIG_MAPPING
     from transformers.models.gemma import modeling_gemma
 
@@ -41,6 +41,7 @@ if TYPE_CHECKING or _transformers_available:
     )
 else:
     CONFIG_MAPPING = None
+    DynamicCache = None
     modeling_gemma = None
     PiGemmaForCausalLM = None
     _gated_residual = None
@@ -141,6 +142,15 @@ def make_att_2d_masks(pad_masks, att_masks):  # see openpi `make_att_2d_masks` (
     return att_2d_masks & pad_2d_masks
 
 
+def clone_past_key_values(past_key_values):
+    """Clone the DynamicCache returned by prefix prefill for compiled denoising."""
+    return DynamicCache(
+        tuple(
+            (keys.clone(), values.clone(), sliding_window) for keys, values, sliding_window in past_key_values
+        )
+    )
+
+
 def pad_vector(vector, new_dim):
     """Pad the last dimension of a vector to new_dim with zeros.
 
@@ -227,16 +237,13 @@ def resize_with_pad_torch(  # see openpi `resize_with_pad_torch` (exact copy)
 
 
 # Define the complete layer computation function for gradient checkpointing
-def compute_layer_complete(
-    layer_idx, inputs_embeds, attention_mask, position_ids, adarms_cond, paligemma, gemma_expert
-):
-    models = [paligemma.model.language_model, gemma_expert.model]
+def compute_layer_complete(inputs_embeds, attention_mask, position_ids, adarms_cond, layers, rotary_emb):
     query_states = []
     key_states = []
     value_states = []
     gates = []
     for i, hidden_states in enumerate(inputs_embeds):
-        layer = models[i].layers[layer_idx]
+        layer = layers[i]
         hidden_states, gate = layernorm_forward(layer.input_layernorm, hidden_states, adarms_cond[i])
         gates.append(gate)
         input_shape = hidden_states.shape[:-1]
@@ -258,15 +265,16 @@ def compute_layer_complete(
         device=query_states.device,
         dtype=query_states.dtype,
     )
-    cos, sin = paligemma.model.language_model.rotary_emb(dummy_tensor, position_ids)
+    cos, sin = rotary_emb(dummy_tensor, position_ids)
     query_states, key_states = modeling_gemma.apply_rotary_pos_emb(
         query_states, key_states, cos, sin, unsqueeze_dim=1
     )
     batch_size = query_states.shape[0]
-    scaling = paligemma.model.language_model.layers[layer_idx].self_attn.scaling
+    paligemma_layer = layers[0]
+    scaling = paligemma_layer.self_attn.scaling
     # Attention computation
     att_output, _ = modeling_gemma.eager_attention_forward(
-        paligemma.model.language_model.layers[layer_idx].self_attn,
+        paligemma_layer.self_attn,
         query_states,
         key_states,
         value_states,
@@ -274,13 +282,13 @@ def compute_layer_complete(
         scaling,
     )
     # Get head_dim from the current layer, not from the model
-    head_dim = paligemma.model.language_model.layers[layer_idx].self_attn.head_dim
+    head_dim = paligemma_layer.self_attn.head_dim
     att_output = att_output.reshape(batch_size, -1, 1 * 8 * head_dim)
     # Process layer outputs
     outputs_embeds = []
     start_pos = 0
     for i, hidden_states in enumerate(inputs_embeds):
-        layer = models[i].layers[layer_idx]
+        layer = layers[i]
         end_pos = start_pos + hidden_states.shape[1]
         if att_output.dtype != layer.self_attn.o_proj.weight.dtype:
             att_output = att_output.to(layer.self_attn.o_proj.weight.dtype)
@@ -488,8 +496,9 @@ class PaliGemmaWithExpertModel(
             prefix_output = None
             prefix_past_key_values = None
         else:
-            models = [self.paligemma.model.language_model, self.gemma_expert.model]
-            num_layers = self.paligemma.config.text_config.num_hidden_layers
+            paligemma_layers = self.paligemma.model.language_model.layers
+            gemma_expert_layers = self.gemma_expert.model.layers
+            rotary_emb = self.paligemma.model.language_model.rotary_emb
 
             # Check if gradient checkpointing is enabled for any of the models
             use_gradient_checkpointing = (
@@ -499,36 +508,39 @@ class PaliGemmaWithExpertModel(
             ) or (hasattr(self, "gradient_checkpointing") and self.gradient_checkpointing and self.training)
 
             # Process all layers with gradient checkpointing if enabled
-            for layer_idx in range(num_layers):
+            for layers in zip(paligemma_layers, gemma_expert_layers, strict=True):
                 if use_gradient_checkpointing:
                     inputs_embeds = torch.utils.checkpoint.checkpoint(
                         compute_layer_complete,
-                        layer_idx,
                         inputs_embeds,
                         attention_mask,
                         position_ids,
                         adarms_cond,
                         use_reentrant=False,
                         preserve_rng_state=False,
-                        paligemma=self.paligemma,
-                        gemma_expert=self.gemma_expert,
+                        layers=layers,
+                        rotary_emb=rotary_emb,
                     )
                 else:
                     inputs_embeds = compute_layer_complete(
-                        layer_idx,
                         inputs_embeds,
                         attention_mask,
                         position_ids,
                         adarms_cond,
-                        paligemma=self.paligemma,
-                        gemma_expert=self.gemma_expert,
+                        layers=layers,
+                        rotary_emb=rotary_emb,
                     )
 
             # final norm
+            final_norms = (
+                self.paligemma.model.language_model.norm,
+                self.gemma_expert.model.norm,
+            )
+
             def compute_final_norms(inputs_embeds, adarms_cond):
                 outputs_embeds = []
                 for i, hidden_states in enumerate(inputs_embeds):
-                    out_emb, _ = layernorm_forward(models[i].norm, hidden_states, adarms_cond[i])
+                    out_emb, _ = layernorm_forward(final_norms[i], hidden_states, adarms_cond[i])
                     outputs_embeds.append(out_emb)
                 return outputs_embeds
 
@@ -907,7 +919,7 @@ class PI0Pytorch(nn.Module):  # see openpi `PI0Pytorch`
         full_att_2d_masks_4d = self._prepare_attention_masks_4d(full_att_2d_masks)
         self.paligemma_with_expert.gemma_expert.model.config._attn_implementation = "eager"  # noqa: SLF001
 
-        past_key_values = copy.deepcopy(past_key_values)
+        past_key_values = clone_past_key_values(past_key_values)
         outputs_embeds, _ = self.paligemma_with_expert.forward(
             attention_mask=full_att_2d_masks_4d,
             position_ids=position_ids,

src/lerobot/policies/pi05/modeling_pi05.py

@@ -15,7 +15,6 @@
 # limitations under the License.
 
 import builtins
-import copy
 import logging
 import math
 from collections import deque
@@ -30,6 +29,7 @@ from lerobot.utils.import_utils import _transformers_available, require_package
 
 # Conditional import for type checking and lazy loading
 if TYPE_CHECKING or _transformers_available:
+    from transformers.cache_utils import DynamicCache
     from transformers.models.auto import CONFIG_MAPPING
     from transformers.models.gemma import modeling_gemma
 
@@ -41,6 +41,7 @@ if TYPE_CHECKING or _transformers_available:
     )
 else:
     CONFIG_MAPPING = None
+    DynamicCache = None
     modeling_gemma = None
     PiGemmaForCausalLM = None
     _gated_residual = None
@@ -138,6 +139,15 @@ def make_att_2d_masks(pad_masks, att_masks):  # see openpi `make_att_2d_masks` (
     return att_2d_masks & pad_2d_masks
 
 
+def clone_past_key_values(past_key_values):
+    """Clone the DynamicCache returned by prefix prefill for compiled denoising."""
+    return DynamicCache(
+        tuple(
+            (keys.clone(), values.clone(), sliding_window) for keys, values, sliding_window in past_key_values
+        )
+    )
+
+
 def pad_vector(vector, new_dim):
     """Pad the last dimension of a vector to new_dim with zeros.
 
@@ -224,16 +234,13 @@ def resize_with_pad_torch(  # see openpi `resize_with_pad_torch` (exact copy)
 
 
 # Define the complete layer computation function for gradient checkpointing
-def compute_layer_complete(
-    layer_idx, inputs_embeds, attention_mask, position_ids, adarms_cond, paligemma, gemma_expert
-):
-    models = [paligemma.model.language_model, gemma_expert.model]
+def compute_layer_complete(inputs_embeds, attention_mask, position_ids, adarms_cond, layers, rotary_emb):
     query_states = []
     key_states = []
     value_states = []
     gates = []
     for i, hidden_states in enumerate(inputs_embeds):
-        layer = models[i].layers[layer_idx]
+        layer = layers[i]
         hidden_states, gate = layernorm_forward(layer.input_layernorm, hidden_states, adarms_cond[i])
         gates.append(gate)
         input_shape = hidden_states.shape[:-1]
@@ -255,15 +262,16 @@ def compute_layer_complete(
         device=query_states.device,
         dtype=query_states.dtype,
     )
-    cos, sin = paligemma.model.language_model.rotary_emb(dummy_tensor, position_ids)
+    cos, sin = rotary_emb(dummy_tensor, position_ids)
     query_states, key_states = modeling_gemma.apply_rotary_pos_emb(
         query_states, key_states, cos, sin, unsqueeze_dim=1
     )
     batch_size = query_states.shape[0]
-    scaling = paligemma.model.language_model.layers[layer_idx].self_attn.scaling
+    paligemma_layer = layers[0]
+    scaling = paligemma_layer.self_attn.scaling
     # Attention computation
     att_output, _ = modeling_gemma.eager_attention_forward(
-        paligemma.model.language_model.layers[layer_idx].self_attn,
+        paligemma_layer.self_attn,
         query_states,
         key_states,
         value_states,
@@ -271,13 +279,13 @@ def compute_layer_complete(
         scaling,
     )
     # Get head_dim from the current layer, not from the model
-    head_dim = paligemma.model.language_model.layers[layer_idx].self_attn.head_dim
+    head_dim = paligemma_layer.self_attn.head_dim
     att_output = att_output.reshape(batch_size, -1, 1 * 8 * head_dim)
     # Process layer outputs
     outputs_embeds = []
     start_pos = 0
     for i, hidden_states in enumerate(inputs_embeds):
-        layer = models[i].layers[layer_idx]
+        layer = layers[i]
         end_pos = start_pos + hidden_states.shape[1]
         if att_output.dtype != layer.self_attn.o_proj.weight.dtype:
             att_output = att_output.to(layer.self_attn.o_proj.weight.dtype)
@@ -485,8 +493,9 @@ class PaliGemmaWithExpertModel(
             prefix_output = None
             prefix_past_key_values = None
         else:
-            models = [self.paligemma.model.language_model, self.gemma_expert.model]
-            num_layers = self.paligemma.config.text_config.num_hidden_layers
+            paligemma_layers = self.paligemma.model.language_model.layers
+            gemma_expert_layers = self.gemma_expert.model.layers
+            rotary_emb = self.paligemma.model.language_model.rotary_emb
 
             # Check if gradient checkpointing is enabled for any of the models
             use_gradient_checkpointing = (
@@ -496,36 +505,39 @@ class PaliGemmaWithExpertModel(
             ) or (hasattr(self, "gradient_checkpointing") and self.gradient_checkpointing and self.training)
 
             # Process all layers with gradient checkpointing if enabled
-            for layer_idx in range(num_layers):
+            for layers in zip(paligemma_layers, gemma_expert_layers, strict=True):
                 if use_gradient_checkpointing:
                     inputs_embeds = torch.utils.checkpoint.checkpoint(
                         compute_layer_complete,
-                        layer_idx,
                         inputs_embeds,
                         attention_mask,
                         position_ids,
                         adarms_cond,
                         use_reentrant=False,
                         preserve_rng_state=False,
-                        paligemma=self.paligemma,
-                        gemma_expert=self.gemma_expert,
+                        layers=layers,
+                        rotary_emb=rotary_emb,
                     )
                 else:
                     inputs_embeds = compute_layer_complete(
-                        layer_idx,
                         inputs_embeds,
                         attention_mask,
                         position_ids,
                         adarms_cond,
-                        paligemma=self.paligemma,
-                        gemma_expert=self.gemma_expert,
+                        layers=layers,
+                        rotary_emb=rotary_emb,
                     )
 
             # final norm
+            final_norms = (
+                self.paligemma.model.language_model.norm,
+                self.gemma_expert.model.norm,
+            )
+
             def compute_final_norms(inputs_embeds, adarms_cond):
                 outputs_embeds = []
                 for i, hidden_states in enumerate(inputs_embeds):
-                    out_emb, _ = layernorm_forward(models[i].norm, hidden_states, adarms_cond[i])
+                    out_emb, _ = layernorm_forward(final_norms[i], hidden_states, adarms_cond[i])
                     outputs_embeds.append(out_emb)
                 return outputs_embeds
 
@@ -880,7 +892,7 @@ class PI05Pytorch(nn.Module):  # see openpi `PI0Pytorch`
         full_att_2d_masks_4d = self._prepare_attention_masks_4d(full_att_2d_masks)
         self.paligemma_with_expert.gemma_expert.model.config._attn_implementation = "eager"  # noqa: SLF001
 
-        past_key_values = copy.deepcopy(past_key_values)
+        past_key_values = clone_past_key_values(past_key_values)
         outputs_embeds, _ = self.paligemma_with_expert.forward(
             attention_mask=full_att_2d_masks_4d,
             position_ids=position_ids,

src/lerobot/policies/pi052/__init__.py

@@ -0,0 +1,42 @@
+# Copyright 2026 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""π0.5 v2 — full reproduction of the π0.5 paper's hierarchical
+inference recipe on lerobot.
+
+Extends :class:`lerobot.policies.pi05.PI05Policy` with:
+
+* recipe-driven training (PR 1's :class:`RenderMessagesStep`),
+* PaliGemma ``lm_head`` cross-entropy on supervised subtask spans
+  (the "high-level subtask prediction" of the paper, §IV.D),
+* AR text generation at inference (:meth:`PI052Policy.select_message`),
+* per-component prompt dropout (Pi 0.7 §V.E) for regularising the
+  text head against missing context at inference.
+
+See ``src/lerobot/configs/recipes/subtasks_vqa.yaml`` for the
+canonical training recipe and
+``examples/training/pi052_hirobot.slurm`` for the launcher.
+"""
+
+from .configuration_pi052 import PI052Config
+from .modeling_pi052 import PI052Policy
+from .processor_pi052 import make_pi052_pre_post_processors
+from .text_processor_pi052 import PI052TextTokenizerStep
+
+__all__ = [
+    "PI052Config",
+    "PI052Policy",
+    "PI052TextTokenizerStep",
+    "make_pi052_pre_post_processors",
+]

src/lerobot/policies/pi052/configuration_pi052.py

@@ -0,0 +1,235 @@
+# Copyright 2026 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""π0.5 v2 (with text head) — reproduction of the π0.5 paper's
+hierarchical inference recipe.
+
+Same architecture as the existing ``PI05Policy`` (PaliGemma 2B VLM +
+~300M Gemma action expert, joint training with FAST tokens during
+pre-train and flow matching during post-train), but with the
+PaliGemma ``lm_head`` re-enabled so the same model can be supervised
+to predict both:
+
+  * **subtask strings** at the high level (cross-entropy on the LM
+    head), and
+  * **action chunks** at the low level (flow matching on the
+    action-expert tokens).
+
+This is the dual-head co-training pattern from the paper:
+
+    L = H(x, f_θ_text) + α * ‖ω - a - f_θ_action(a_τ, o, ℓ)‖²
+
+with α = 10.0 per § IV.D of arxiv:2504.16054. The π0.5 model splits
+inference into a text-prediction step followed by an action-prediction
+step, which the multi-rate ``PI052Runtime`` (in
+``lerobot.policies.pi052.inference``) drives at separate rates.
+"""
+
+from dataclasses import dataclass
+
+from lerobot.configs import PreTrainedConfig
+from lerobot.optim.optimizers import AdamWConfig
+
+from ..pi05.configuration_pi05 import PI05Config
+
+
+@PreTrainedConfig.register_subclass("pi052")
+@dataclass
+class PI052Config(PI05Config):
+    """π0.5 with the PaliGemma LM head re-enabled for subtask prediction.
+
+    Recipe-driven dual-head training: the flow head supervises actions,
+    the LM head supervises subtask / plan / memory / VQA text. The
+    flow:text loss split is the milder 5:1 (see ``flow_loss_weight``).
+    """
+
+    # Recipe / language stack ---------------------------------------------
+    recipe_path: str | None = "recipes/subtasks_vqa.yaml"
+    """Path (absolute or relative to ``src/lerobot/configs/``) to a
+    ``TrainingRecipe`` YAML. Defaults to the canonical Hi-Robot blend
+    shipped alongside this policy. Set to ``None`` to disable recipe
+    rendering and fall back to π0.5's single-task ``Task: ... Action:``
+    prompt path (unannotated datasets keep working that way)."""
+
+    apply_chat_template: bool = False
+    """PaliGemma is *not* chat-pretrained — its tokenizer doesn't ship a
+    chat template, so we don't apply one. The recipe renderer's output
+    is concatenated as a plain prefix + assistant suffix instead,
+    mirroring how the π0.5 paper's high-level inference samples text
+    auto-regressively after the prefix."""
+
+    # Loss weights --------------------------------------------------------
+    # Paper §IV.D uses α=10 between the flow and text terms, assuming
+    # text is a rare auxiliary task. With the recipe stack the flow-only
+    # `low_level` branch fires on a large share of samples, so α=10
+    # swamps the LM head and collapses generation into degenerate
+    # repetition. We use the milder 5:1 split here.
+    text_loss_weight: float = 1.0
+    """Weight on the LM-head cross-entropy term. Set to ``0`` to disable
+    text training entirely (reverts to flow-only / π0.5 behaviour)."""
+
+    flow_loss_weight: float = 5.0
+    """Weight on the action-expert flow-matching term. ``5.0`` — a milder
+    flow:text split than the paper's α=10, since the flow-only
+    ``low_level`` recipe already gives the action expert frequent
+    gradient. Lower it further if the LM head still underfits."""
+
+    # Backbone training ---------------------------------------------------
+    unfreeze_lm_head: bool = True
+    """Whether to keep the PaliGemma ``lm_head`` unfrozen for fine-tuning.
+    The existing ``PI05Policy`` zeroes / freezes the head on load
+    because it never reads from it. Must be ``True`` for π0.5-style
+    hierarchical inference."""
+
+    # Per-component prompt dropout (Pi0.7 §V.E) ---------------------------
+    # Randomly drop non-target context messages so the LM head learns
+    # to handle missing /
+    # stale plan / memory at inference. Defaults to 0.0 so behaviour
+    # is identical until explicitly enabled.
+    plan_dropout_prob: float = 0.0
+    memory_dropout_prob: float = 0.0
+    subtask_dropout_prob: float = 0.0
+
+    # FAST discrete-action supervision — paper §III.B-C ------------------
+    # When enabled, actions are *also* tokenised via the FAST tokenizer
+    # ("physical-intelligence/fast") and supervised with cross-entropy
+    # on the PaliGemma LM head — exactly as in the paper's pre-training
+    # objective (Eq. 1 mixes FAST CE + flow MSE + subtask CE). The
+    # ActionTokenizerProcessorStep is wired into the preprocessor
+    # pipeline when this flag is set; the loss is computed in
+    # PI052Policy.forward.
+    enable_fast_action_loss: bool = True
+    """If True, tokenise actions with the FAST tokenizer and add a
+    cross-entropy loss on the LM head. On by default to match the
+    π0.5 paper's three-loss objective (text CE + FAST CE + flow MSE,
+    §III.B-C Eq. 1). Set to False if you only want the
+    post-training-style flow + text recipe."""
+
+    action_tokenizer_name: str = "physical-intelligence/fast"
+    """HF identifier for the FAST action tokenizer."""
+
+    max_action_tokens: int = 256
+    """Maximum number of FAST tokens per action chunk."""
+
+    fast_skip_tokens: int = 128
+    """Number of low-vocab tokens the FAST tokenizer skips to avoid
+    collisions with PaliGemma's text vocabulary."""
+
+    fast_action_loss_weight: float = 1.0
+    """Weight on the FAST-action-token CE loss. Paper §III.C uses 1.0."""
+
+    auto_fit_fast_tokenizer: bool = False
+    """If True, the processor factory checks ``fast_tokenizer_cache_dir``
+    for a previously-fitted tokenizer keyed on ``(dataset_repo_id,
+    base_tokenizer_name, fit_samples)``. On cache miss, it loads
+    ``action_tokenizer_name`` as a base, samples
+    ``fast_tokenizer_fit_samples`` action chunks from the dataset, runs
+    ``.fit()``, saves the result, and uses *that* fitted path as the
+    actual tokenizer. Pertsch et al. 2025 (FAST paper [64], π0.5 §III.C)
+    explicitly recommend per-dataset fitting for best compression.
+
+    Off by default because the fit requires a separate pre-training
+    pass over the dataset (~1-2 min on a medium dataset) and depends
+    on the FAST tokenizer snapshot having a ``.fit()`` method. Opt in
+    when you want paper-faithful compression; leave off to fall back
+    on the universal ``physical-intelligence/fast`` codebook."""
+
+    fast_tokenizer_cache_dir: str = "~/.cache/lerobot/fast_tokenizers"
+    """Where fitted FAST tokenizers are stored. ``~`` expands."""
+
+    fast_tokenizer_fit_samples: int = 1024
+    """Number of action chunks to sample for the fit. The FAST paper uses
+    a few thousand; 1024 is a reasonable default for medium datasets."""
+
+    # Knowledge insulation — paper §III.B --------------------------------
+    # When enabled, gradients from the action expert's flow loss are
+    # blocked from flowing back into the VLM's K/V projections. This
+    # prevents the action loss from over-fitting the language backbone
+    # to robot-specific features. Implemented in ``modeling_pi052`` as
+    # a per-instance monkey-patch on ``paligemma_with_expert.forward``
+    # that splits queries into VLM and action halves and ``.detach()``-s
+    # the VLM K/V tensors used in the action-half's attention.
+    knowledge_insulation: bool = False
+    """If True, route every transformer layer through the KI
+    attention path that blocks action→VLM gradient flow on K/V."""
+
+    # Learning-rate defaults --------------------------------------------
+    # pi052 inherits π0.5's openpi-validated optimizer config (peak LR
+    # 2.5e-5, cosine→2.5e-6, 1k warmup, AdamW (0.9, 0.95), wd=0.01,
+    # grad_clip=1.0). The only place pi052 needs to diverge from pi05
+    # is the LM-head LR multiplier: pi05 has no text supervision so the
+    # head doesn't get gradients; pi052 always has text supervision
+    # (subtask / memory / VQA) via the recipe, and under KI the LM head
+    # only sees gradients on ~30–45% of the batch (the text-CE mask
+    # share of the recipe). Under aggressive cosine decay this is too
+    # weak to keep the head pinned, so it drifts back toward PaliGemma's
+    # pretrained ``<loc>`` first-token bias. 5x is the documented fix
+    # (see ``PI05Config.lm_head_lr_scale`` docstring); the wiring is
+    # already in ``PI05Policy.get_optim_params`` — it splits the LM head
+    # + tied ``embed_tokens`` into their own param group while sharing
+    # the same cosine lambda, so the 5x ratio is preserved across decay.
+    lm_head_lr_scale: float = 5.0
+
+    # PaLM-style z-loss on text CE. Penalises the log-partition function
+    # ``z = log Σ exp(logits)`` drifting away from zero — without it, large-
+    # vocab models (PaliGemma is 257k) can let ``logsumexp`` grow unbounded
+    # while CE stays low, because a uniform additive logit bias cancels in
+    # softmax. PaLM appendix B / Chinchilla report z-loss is essential for
+    # stable large-vocab CE; it especially helps under ``lm_head_lr_scale=
+    # 5.0`` which amplifies drift risk on the LM head. ``1e-4`` is the
+    # commonly cited weight; set 0 to disable entirely.
+    text_ce_z_loss_weight: float = 1e-4
+
+    # Liger Triton kernels (rope + geglu + layer_norm) are now patched
+    # unconditionally at model build time — see ``_enable_hf_kernels``
+    # in ``modeling_pi052``. The patch is process-global, idempotent
+    # and degrades gracefully if ``liger-kernel`` is missing. Measured
+    # at -4.5% step time on H100 (bench job 22161421); peak memory
+    # unchanged. ``fused_linear_cross_entropy`` ships separately via
+    # ``_shifted_lin_ce`` / ``_fast_lin_ce``.
+    use_hf_kernels: bool = True
+    """Deprecated. Liger HF kernels are patched unconditionally by
+    ``_enable_hf_kernels`` — this field is retained as a no-op for
+    backward compatibility with checkpoints saved before commit
+    d70c8104 (which still serialize ``use_hf_kernels: true`` into
+    ``config.json``). Loading those configs would otherwise raise
+    ``DecodingError: The fields use_hf_kernels are not valid for
+    PI052Config`` (job 22164492). Remove in a future major bump."""
+
+    # Optimizer foreach/fused. pi052 carries these locally because the shared
+    # PI05Config (kept identical to upstream main) does not define them; the
+    # checkpoints we train serialize both keys into config.json, so they must
+    # be valid PI052Config fields and flow into the AdamW preset below.
+    optimizer_foreach: bool | None = False
+    optimizer_fused: bool | None = True
+
+    def get_optimizer_preset(self) -> AdamWConfig:
+        return AdamWConfig(
+            lr=self.optimizer_lr,
+            betas=self.optimizer_betas,
+            eps=self.optimizer_eps,
+            weight_decay=self.optimizer_weight_decay,
+            grad_clip_norm=self.optimizer_grad_clip_norm,
+            foreach=self.optimizer_foreach,
+            fused=self.optimizer_fused,
+        )
+
+    def __post_init__(self) -> None:
+        super().__post_init__()
+        # Backbone needs gradients flowing through the text head when
+        # we're training it. Override the π0.5 default
+        # (``train_expert_only=True``) unless the user explicitly opts
+        # out of text training via ``text_loss_weight=0``.
+        if self.text_loss_weight > 0 and self.unfreeze_lm_head:
+            self.train_expert_only = False

src/lerobot/policies/pi052/fit_fast_tokenizer.py

@@ -0,0 +1,304 @@
+# Copyright 2026 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""Dataset-specific FAST action tokenizer fitting.
+
+The published ``physical-intelligence/fast`` tokenizer is a *universal*
+codebook fitted on a heterogeneous mix of robot datasets. Per Pertsch
+et al. 2025 (the FAST paper, [64] in the π0.5 paper) and §III.C of
+π0.5 itself, the recommended practice is to **finetune the tokenizer on
+your specific dataset's action distribution** before training the
+policy — same way one would adapt a language tokenizer to a domain
+corpus. Without this finetune step, action sequences from your robot
+may require more tokens per chunk than necessary, lowering effective
+compression and slowing convergence of the action-CE loss.
+
+This module provides a single utility, :func:`fit_fast_tokenizer`,
+that does the finetune. The training entry point invokes it
+automatically when the policy's ``enable_fast_action_loss`` and
+``auto_fit_fast_tokenizer`` flags are both ``True`` and no cached
+fitted tokenizer is found at ``fast_tokenizer_cache_dir``.
+
+The fitted tokenizer is saved to
+``{cache_dir}/{dataset_hash}_{base_hash}/`` so successive training
+runs over the same dataset re-use it.
+"""
+
+from __future__ import annotations
+
+import hashlib
+import logging
+import os
+import time
+from pathlib import Path
+
+import numpy as np
+
+logger = logging.getLogger(__name__)
+
+# Marker file the cache-hit check looks for. ``ProcessorMixin.save_pretrained``
+# writes ``processor_config.json`` (NOT ``preprocessor_config.json`` —
+# that's the image / feature-extractor convention). Centralised here so
+# the cache-hit check and the rank-N readiness wait agree on the same
+# sentinel.
+_CACHE_SENTINEL = "processor_config.json"
+
+
+def _dataset_signature(
+    dataset_repo_id: str,
+    base_tokenizer_name: str,
+    n_samples: int,
+    chunk_size: int,
+) -> str:
+    """Deterministic short hash for naming the cache directory.
+
+    Keys on (dataset, base tokenizer, sample count, chunk size) so any
+    of those changing re-runs the fit. ``chunk_size`` matters because
+    the tokenizer is fit on chunks of that length.
+    """
+    h = hashlib.sha256()
+    h.update(dataset_repo_id.encode("utf-8"))
+    h.update(b"\0")
+    h.update(base_tokenizer_name.encode("utf-8"))
+    h.update(b"\0")
+    h.update(str(n_samples).encode("utf-8"))
+    h.update(b"\0")
+    h.update(str(chunk_size).encode("utf-8"))
+    return h.hexdigest()[:16]
+
+
+def fit_fast_tokenizer(
+    *,
+    dataset_repo_id: str,
+    cache_dir: str | Path,
+    base_tokenizer_name: str = "physical-intelligence/fast",
+    n_samples: int = 1024,
+    chunk_size: int = 50,
+    seed: int = 42,
+) -> str:
+    """Fit a FAST tokenizer on a LeRobot dataset's action distribution.
+
+    Args:
+        dataset_repo_id: HF Hub repo id of the LeRobotDataset to fit on.
+        cache_dir: Directory under which to save (and look up) fitted
+            tokenizers. The actual save path is
+            ``{cache_dir}/{signature}``.
+        base_tokenizer_name: HF identifier for the base FAST tokenizer
+            to finetune from. ``physical-intelligence/fast`` is the
+            universal one.
+        n_samples: Number of action chunks to sample for the fit. The
+            FAST paper uses a few thousand; ``1024`` is a good default
+            for medium datasets.
+        chunk_size: Length of each action chunk (matches
+            ``policy.chunk_size``). The FAST tokenizer is fit on
+            sequences of this length.
+        seed: RNG seed for sample selection.
+
+    Returns:
+        The local path to the fitted tokenizer. Passed directly to
+        ``--policy.action_tokenizer_name`` for the training run.
+
+    Raises:
+        ImportError: If the ``transformers`` library doesn't expose
+            ``AutoProcessor`` or the FAST tokenizer doesn't have a
+            ``.fit()`` method (then you're on an older FAST snapshot —
+            update to the current published model).
+        FileNotFoundError: If the dataset can't be loaded.
+    """
+    cache_dir = Path(cache_dir)
+    sig = _dataset_signature(dataset_repo_id, base_tokenizer_name, n_samples, chunk_size)
+    out_dir = cache_dir / sig
+
+    if out_dir.exists() and (out_dir / _CACHE_SENTINEL).exists():
+        logger.info(
+            "FAST tokenizer cache hit: %s — re-using fitted tokenizer for "
+            "dataset=%s base=%s n_samples=%d",
+            out_dir, dataset_repo_id, base_tokenizer_name, n_samples,
+        )
+        return str(out_dir)
+
+    # DDP-safe fit: only the (local) main process actually fits + saves;
+    # other ranks poll the cache sentinel until the leader is done.
+    # Without this guard, all N ranks fit concurrently and race on
+    # ``save_pretrained`` + ``AutoProcessor.from_pretrained`` (the latter
+    # copies ``processing_action_tokenizer.py`` into ``HF_MODULES_CACHE``
+    # and compiles a ``.pyc`` — concurrent writers occasionally produce
+    # a stale / partial ``.pyc`` and the subsequent ``from .. import
+    # UniversalActionProcessor`` raises ``AttributeError``.
+    is_leader = (
+        int(os.environ.get("RANK", "0")) == 0
+        and int(os.environ.get("LOCAL_RANK", "0")) == 0
+    )
+    if not is_leader:
+        timeout_s = 1800.0  # 30 min — covers ~1024-sample fits on cold caches
+        start = time.monotonic()
+        while not (out_dir / _CACHE_SENTINEL).exists():
+            if time.monotonic() - start > timeout_s:
+                raise RuntimeError(
+                    f"FAST tokenizer fit: non-leader rank timed out after "
+                    f"{timeout_s:.0f}s waiting for {out_dir / _CACHE_SENTINEL}. "
+                    "Leader rank likely crashed during the fit."
+                )
+            time.sleep(2.0)
+        logger.info("FAST tokenizer ready (leader populated cache): %s", out_dir)
+        return str(out_dir)
+
+    logger.info(
+        "FAST tokenizer cache miss — fitting on dataset=%s "
+        "base=%s n_samples=%d chunk_size=%d → %s",
+        dataset_repo_id, base_tokenizer_name, n_samples, chunk_size, out_dir,
+    )
+
+    from transformers import AutoProcessor  # noqa: PLC0415
+
+    from lerobot.datasets.lerobot_dataset import LeRobotDataset  # noqa: PLC0415
+
+    # Stream a single episode's worth of action chunks at a time so
+    # we don't blow memory on huge datasets. Random episode +
+    # random start offset gives a reasonable spread.
+    #
+    # Actions are read straight from the underlying HF dataset's
+    # ``action`` *column* — never via ``ds[i]``. ``ds[i]`` builds a full
+    # training item (delta-timestamp expansion + video decode + image
+    # transforms); a single bad video frame would then throw and, since
+    # the failure was swallowed at debug level, silently starve the fit
+    # of every chunk. The action column carries no video, so reading it
+    # directly is both faster and immune to decode errors.
+    rng = np.random.default_rng(seed)
+    actions_buf: list[np.ndarray] = []
+
+    # Resolve the dataset's data parquet shards directly, sidestepping
+    # ``LeRobotDataset(repo_id, episodes=[N])`` which on v3-format
+    # datasets routes through HF datasets'' split lookup and raises
+    # ``ValueError: Instruction "train" corresponds to no data!`` for
+    # every episode (job 22182985 looped through 13,293 skipped episodes
+    # for ~2.5 h before NCCL killed it). Reading the ``action`` column
+    # straight from the parquet shards is also faster: each per-episode
+    # ``LeRobotDataset`` instantiation re-parses every meta file.
+    from huggingface_hub import snapshot_download  # noqa: PLC0415
+    import pyarrow as _pa  # noqa: PLC0415
+    import pyarrow.parquet as _pq  # noqa: PLC0415
+
+    snap = Path(snapshot_download(repo_id=dataset_repo_id, repo_type="dataset"))
+    data_files = sorted((snap / "data").glob("chunk-*/file-*.parquet"))
+    if not data_files:
+        raise RuntimeError(
+            f"FAST fit: no ``data/chunk-*/file-*.parquet`` shards found under {snap!s}."
+        )
+
+    # Read just the (episode_index, action) columns once across all
+    # shards. This is the same pattern used elsewhere in the codebase
+    # for whole-dataset audits and stays under ~2 GB even on 32 k-episode
+    # / 29 M-frame datasets because the action column is a fixed-length
+    # float vector.
+    tables = [_pq.read_table(f, columns=["episode_index", "action"]) for f in data_files]
+    table = _pa.concat_tables(tables)
+    eps = table["episode_index"].to_numpy()
+    acts_col = table["action"]
+    # ``action`` may be a fixed-shape ListArray or a 2-D NumericArray;
+    # ``to_numpy(zero_copy_only=False)`` produces an object array of
+    # 1-D NumPy actions either way, which we stack into (N, D).
+    try:
+        acts = np.stack(acts_col.to_numpy(zero_copy_only=False)).astype(np.float32)
+    except Exception:  # noqa: BLE001
+        # Fallback path for nested-list types: flatten via to_pylist().
+        acts = np.asarray(acts_col.to_pylist(), dtype=np.float32)
+    if acts.ndim != 2:
+        raise RuntimeError(
+            f"FAST fit: expected ``action`` rows to be 1-D vectors; got shape {acts.shape}."
+        )
+
+    # Episode index → slice (start, stop) into ``acts`` along axis 0.
+    # ``eps`` is monotonically increasing within each parquet shard but
+    # we make no assumption across shards — sort once and group.
+    order = np.argsort(eps, kind="stable")
+    eps_sorted = eps[order]
+    boundaries = np.searchsorted(eps_sorted, np.arange(int(eps_sorted.max()) + 2))
+    ep_to_slice: dict[int, tuple[int, int]] = {
+        int(ep): (int(boundaries[ep]), int(boundaries[ep + 1]))
+        for ep in range(len(boundaries) - 1)
+        if boundaries[ep] < boundaries[ep + 1]
+    }
+    num_episodes = len(ep_to_slice)
+    # ``acts`` is in original (un-sorted-by-episode) row order; reorder
+    # so per-episode slices are contiguous.
+    acts = acts[order]
+
+    samples_per_episode = max(1, n_samples // max(num_episodes, 1))
+    collected = 0
+    eps_visited = 0
+    short_episodes = 0
+    ep_indices = list(ep_to_slice.keys())
+    for ep_idx in rng.permutation(ep_indices):
+        if collected >= n_samples:
+            break
+        start, stop = ep_to_slice[int(ep_idx)]
+        ep_actions = acts[start:stop]
+        if ep_actions.shape[0] < chunk_size:
+            short_episodes += 1
+            continue
+        starts = rng.integers(0, ep_actions.shape[0] - chunk_size + 1, size=samples_per_episode)
+        for s in starts:
+            actions_buf.append(ep_actions[int(s) : int(s) + chunk_size])
+            collected += 1
+            if collected >= n_samples:
+                break
+        eps_visited += 1
+
+    if not actions_buf:
+        raise RuntimeError(
+            f"FAST fit collected zero action chunks from {dataset_repo_id!r}: "
+            f"all {num_episodes} episodes were shorter than chunk_size="
+            f"{chunk_size} ({short_episodes} too short) or had an unreadable "
+            "``action`` column. Lower ``chunk_size`` to match your episode "
+            "lengths."
+        )
+
+    actions = np.stack(actions_buf, axis=0).astype(np.float32)  # (N, H, D)
+    logger.info(
+        "FAST fit: collected %d chunks of shape %s from %d episodes",
+        actions.shape[0], actions.shape[1:], eps_visited,
+    )
+
+    # Quantile-normalise per dimension before fitting.
+    #
+    # The FAST tokenizer DCT-transforms actions, scales by ``scale`` and
+    # rounds to integer tokens; the integer *range* must fit the
+    # codebook (vocab_size, default 1024). Raw motor units (e.g. encoder
+    # ticks) blow that range up — hence "Vocab size 1024 is too small".
+    # More importantly, at training time ``ActionTokenizerProcessorStep``
+    # runs *after* the QUANTILES ``NormalizerProcessorStep``, so it
+    # encodes normalised actions. Fitting on raw actions would mismatch
+    # that space. We replicate QUANTILES normalisation here (per-dim
+    # [q01, q99] → [-1, 1], clipped) so the fit and the training-time
+    # encode see the same distribution.
+    flat = actions.reshape(-1, actions.shape[-1])
+    q01 = np.quantile(flat, 0.01, axis=0)
+    q99 = np.quantile(flat, 0.99, axis=0)
+    span = np.where((q99 - q01) > 1e-6, q99 - q01, 1.0)
+    actions = np.clip((actions - q01) / span * 2.0 - 1.0, -1.0, 1.0).astype(np.float32)
+
+    base = AutoProcessor.from_pretrained(base_tokenizer_name, trust_remote_code=True)
+    if not hasattr(base, "fit"):
+        raise ImportError(
+            f"Base FAST tokenizer {base_tokenizer_name!r} has no ``.fit()`` "
+            "method — your transformers / model snapshot is too old. Update "
+            "to the current ``physical-intelligence/fast`` revision."
+        )
+
+    fitted = base.fit(actions)
+    out_dir.mkdir(parents=True, exist_ok=True)
+    fitted.save_pretrained(str(out_dir))
+    logger.info("FAST fit: saved fitted tokenizer to %s", out_dir)
+    return str(out_dir)

src/lerobot/policies/pi052/inference/__init__.py

@@ -0,0 +1,73 @@
+# Copyright 2026 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""PI052 inference / runtime orchestration.
+
+Multi-rate runtime that mirrors the recipe-time training shape:
+
+  low_level_execution        → LowLevelForward + DispatchAction (high Hz)
+  high_level_subtask         → HighLevelSubtaskFwd (~1 Hz)
+  memory_update              → MemoryUpdateFwd (event: subtask_change)
+  user_interjection_response → UserInterjectionFwd (event: stdin)
+  ask_vqa_*                  → AskVQAFwd (event: stdin question)
+  speech tool calls          → DispatchToolCalls (event: tool_call_pending)
+
+The CLI ``lerobot-pi052-runtime`` builds a ``PI052Runtime`` and calls
+``run()``.
+"""
+
+from .repl import StdinReader
+from .runtime import PI052Runtime
+from .runtime_state import initial_runtime_state, push_log, set_if_changed, take_event
+from .steps import (
+    AskVQAFwd,
+    DispatchAction,
+    DispatchToolCalls,
+    HighLevelSubtaskFwd,
+    InferenceStep,
+    LowLevelForward,
+    MemoryUpdateFwd,
+    UserInterjectionFwd,
+)
+from .triggers import EventTrigger, HzTrigger, Tick, TickClock, Trigger
+from .ui import make_state_panel, print_robot_lines, print_user_line
+
+__all__ = [
+    # runtime
+    "PI052Runtime",
+    "StdinReader",
+    # state helpers
+    "initial_runtime_state",
+    "push_log",
+    "set_if_changed",
+    "take_event",
+    # triggers
+    "Trigger",
+    "Tick",
+    "TickClock",
+    "HzTrigger",
+    "EventTrigger",
+    # steps
+    "InferenceStep",
+    "LowLevelForward",
+    "DispatchAction",
+    "HighLevelSubtaskFwd",
+    "MemoryUpdateFwd",
+    "UserInterjectionFwd",
+    "AskVQAFwd",
+    "DispatchToolCalls",
+    # UI
+    "make_state_panel",
+    "print_robot_lines",
+    "print_user_line",
+]

src/lerobot/policies/pi052/inference/repl.py

@@ -0,0 +1,105 @@
+# Copyright 2026 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""Stdin REPL event collector for the PI052 runtime.
+
+Reads non-blocking stdin lines, classifies each one heuristically:
+
+  "stop" / "quit" / "exit"               → state["stop"] = True
+  "/action" / "/pause"                    → set state["mode"]
+  ends with "?"                           → user_vqa_query event
+  starts with "task:" or first line       → set runtime task
+  anything else                           → user_interjection event
+
+Plugged into the runtime via ``event_collector=StdinReader().poll``.
+
+Note: the shipped CLI (``lerobot-pi052-runtime``) drives stdin
+directly in its REPL / autonomous loops and does *not* wire this
+collector; it's kept as the documented embedding hook and for tests.
+"""
+
+from __future__ import annotations
+
+import select
+import sys
+from dataclasses import dataclass, field
+from typing import Any
+
+
+@dataclass
+class StdinReader:
+    """Non-blocking stdin line collector for the runtime loop."""
+
+    prompt: str = "> "
+    _seen_first_line: bool = field(default=False, init=False)
+    _prompted: bool = field(default=False, init=False)
+
+    def poll(self, state: dict[str, Any]) -> None:
+        """Drain pending stdin lines into runtime events."""
+        # Print the input prompt once on every fresh tick if we don't
+        # already have a pending line; matches the expected REPL feel.
+        if not self._prompted:
+            print(self.prompt, end="", flush=True)
+            self._prompted = True
+
+        # ``select`` with timeout=0 makes this non-blocking. Only works
+        # for actual TTY / pipe stdins; CI / scripted runs hit EOF.
+        try:
+            ready, _, _ = select.select([sys.stdin], [], [], 0)
+        except (ValueError, OSError):
+            return
+        if not ready:
+            return
+
+        line = sys.stdin.readline()
+        if not line:  # EOF
+            state["stop"] = True
+            return
+        line = line.strip()
+        self._prompted = False  # we'll re-prompt next tick
+        if not line:
+            return
+
+        lower = line.lower()
+        if lower in {"stop", "quit", "exit"}:
+            state["stop"] = True
+            return
+
+        # Slash commands flip the run mode. ``/pause`` stops the action
+        # loop (the action steps gate on ``state["mode"]``); ``/action``
+        # resumes it.
+        if lower.split(" ", 1)[0] in {"/action", "/act", "/run"}:
+            state["mode"] = "action"
+            return
+        if lower in {"/pause", "/p"}:
+            state["mode"] = "paused"
+            queue = state.get("action_queue")
+            if hasattr(queue, "clear"):
+                queue.clear()
+            return
+
+        # First non-control line sets the task if no task is active.
+        if not state.get("task"):
+            task = line[5:].strip() if lower.startswith("task:") else line
+            state["task"] = task
+            print(f"[pi052] Task: {task}", flush=True)
+            self._seen_first_line = True
+            return
+
+        # Question → VQA; statement → interjection.
+        if lower.endswith("?"):
+            state["recent_vqa_query"] = line
+            state.setdefault("events_this_tick", []).append("user_vqa_query")
+        else:
+            state["recent_interjection"] = line
+            state.setdefault("events_this_tick", []).append("user_interjection")

src/lerobot/policies/pi052/inference/runtime.py

@@ -0,0 +1,205 @@
+# Copyright 2026 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""PI052 runtime loop.
+
+Threads the multi-rate inference pipeline together with a stdin REPL
+event collector, drives ticks through :class:`TickClock`, and prints
+state-change updates to the user.
+"""
+
+from __future__ import annotations
+
+import logging
+from collections import deque
+from dataclasses import dataclass, field
+from typing import Any, Callable
+
+from .runtime_state import initial_runtime_state, push_log
+from .steps import (
+    AskVQAFwd,
+    DispatchAction,
+    DispatchToolCalls,
+    HighLevelSubtaskFwd,
+    InferenceStep,
+    LowLevelForward,
+    MemoryUpdateFwd,
+)
+from .triggers import EventTrigger, HzTrigger, TickClock
+
+logger = logging.getLogger(__name__)
+
+
+@dataclass
+class PI052Runtime:
+    """Compose the inference pipeline and drive it tick-by-tick."""
+
+    policy: Any
+    tools: dict[str, Any] = field(default_factory=dict)
+    """Name → tool-instance dict, e.g. ``{"say": SayTool(...)}``. Read
+    from :func:`lerobot.tools.get_tools(meta)` when wiring the
+    runtime."""
+    observation_provider: Callable[[], dict | None] | None = None
+    """Closure returning the current preprocessed observation batch.
+    ``None`` for dry-run / language-only sessions."""
+    robot_executor: Callable[[Any], None] | None = None
+    """Closure that takes one action chunk and forwards it to the
+    robot. ``None`` for dry-run."""
+    event_collector: Callable[[dict], None] | None = None
+    """Per-tick hook that polls external sources (stdin, network) and
+    appends event names to ``state["events_this_tick"]``."""
+    chunk_hz: float = 4.0
+    ctrl_hz: float = 50.0
+    high_level_hz: float = 1.0
+    max_rate_hz: float = 50.0
+
+    pipeline: list[InferenceStep] = field(init=False)
+    state: dict[str, Any] = field(init=False)
+    _stop: bool = field(default=False, init=False)
+
+    def __post_init__(self) -> None:
+        # Subtask + memory + VQA configuration. Pipeline:
+        #
+        #   HighLevelSubtaskFwd → generate the next subtask via the LM
+        #                         head at ~``high_level_hz``; writes
+        #                         ``current_subtask`` and emits
+        #                         ``subtask_change`` on a transition.
+        #   MemoryUpdateFwd     → on ``subtask_change``, refresh
+        #                         ``current_memory`` from the
+        #                         ``memory_update`` head.
+        #   AskVQAFwd           → answer camera-grounded stdin questions.
+        #   LowLevelForward     → action chunk conditioned on the
+        #                         generated ``current_subtask``.
+        #   DispatchAction      → drain the chunk to the robot.
+        #   DispatchToolCalls   → fire any pending tool calls.
+        #
+        # Order matters: ``HighLevelSubtaskFwd`` must run before
+        # ``MemoryUpdateFwd`` so the event is visible the same tick, and
+        # both must run before ``LowLevelForward`` (which is gated on
+        # "action queue empty") so the chunk consumes the freshest
+        # subtask. ``UserInterjectionFwd`` is still importable but
+        # disabled until plan generation is wired in.
+        self.pipeline = [
+            HighLevelSubtaskFwd(
+                trigger=HzTrigger(self.high_level_hz),
+                policy=self.policy,
+                observation_provider=self.observation_provider,
+            ),
+            # Listens for the ``subtask_change`` event raised by
+            # ``HighLevelSubtaskFwd`` and refreshes ``current_memory``.
+            MemoryUpdateFwd(
+                trigger=EventTrigger("subtask_change"),
+                policy=self.policy,
+                observation_provider=self.observation_provider,
+            ),
+            AskVQAFwd(
+                policy=self.policy,
+                observation_provider=self.observation_provider,
+            ),
+            LowLevelForward(
+                trigger=HzTrigger(self.chunk_hz),
+                policy=self.policy,
+                observation_provider=self.observation_provider,
+            ),
+            DispatchAction(
+                trigger=HzTrigger(self.ctrl_hz),
+                robot_executor=self.robot_executor,
+            ),
+            DispatchToolCalls(tools=self.tools),
+        ]
+        self.state = initial_runtime_state()
+
+    # ------------------------------------------------------------------
+    # Lifecycle
+    # ------------------------------------------------------------------
+
+    def set_task(self, task: str) -> None:
+        """Set or replace the active task. Logged for the REPL."""
+        self.state["task"] = task
+        push_log(self.state, f"Task: {task}")
+
+    def stop(self) -> None:
+        self._stop = True
+
+    def run(self, *, max_ticks: int | None = None) -> None:
+        """Main loop. Returns when ``stop()`` is called or after
+        ``max_ticks`` ticks (useful for tests / dry-run)."""
+        clock = TickClock(max_rate_hz=self.max_rate_hz)
+        while not self._stop:
+            tick = clock.advance()
+            self.state["_tick"] = tick
+            self.state["events_this_tick"] = []
+            self.state["log_lines"] = []
+
+            if self.event_collector is not None:
+                self.event_collector(self.state)
+            if self.state.get("stop"):
+                self._stop = True
+                break
+
+            for step in self.pipeline:
+                self.state = step(self.state)
+
+            self._flush_logs()
+            if max_ticks is not None and tick.index >= max_ticks:
+                break
+
+        self._on_shutdown()
+
+    # ------------------------------------------------------------------
+    # REPL helper: drive one full pipeline pass and return its logs
+    # ------------------------------------------------------------------
+
+    def step_once(self) -> list[str]:
+        """Run one tick of the pipeline and return the log lines.
+
+        Used by the interactive REPL: instead of a background thread,
+        the CLI drives ticks synchronously after each user input. Logs
+        are returned (not printed) so the caller can route them into
+        the rich-Live chat scrollback.
+        """
+        from .triggers import Tick  # noqa: PLC0415
+
+        # Synthesize a tick. We don't need the real wall-clock pacing
+        # here — the REPL drives the runtime, not vice versa — but
+        # ``HzTrigger`` uses ``tick.monotonic_seconds`` to gate, so we
+        # bump it generously so every Hz-triggered step considers
+        # itself due.
+        import time as _time  # noqa: PLC0415
+
+        prev_index = self.state.get("_tick").index if isinstance(self.state.get("_tick"), Tick) else 0
+        self.state["_tick"] = Tick(index=prev_index + 1, monotonic_seconds=_time.monotonic())
+        self.state["log_lines"] = []
+        # ``events_this_tick`` is set up by the caller before
+        # ``step_once`` (the REPL pushes user-driven events first).
+        self.state.setdefault("events_this_tick", [])
+
+        for step in self.pipeline:
+            self.state = step(self.state)
+
+        return list(self.state.get("log_lines") or [])
+
+    # ------------------------------------------------------------------
+    # I/O
+    # ------------------------------------------------------------------
+
+    def _flush_logs(self) -> None:
+        for line in self.state.get("log_lines") or []:
+            print(f"[pi052] {line}", flush=True)
+
+    def _on_shutdown(self) -> None:
+        # Drain any queued action chunks safely.
+        queue = self.state.get("action_queue")
+        if isinstance(queue, deque):
+            queue.clear()
+        print("[pi052] runtime stopped", flush=True)

src/lerobot/policies/pi052/inference/runtime_state.py

@@ -0,0 +1,95 @@
+# Copyright 2026 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""Runtime state passed between inference steps each tick.
+
+The runtime threads a single dict through the pipeline; this module
+documents the shape and provides factories. We use a plain ``dict``
+rather than a frozen dataclass because steps freely add and remove
+keys (``events_this_tick``, ``messages_pending``, ``tool_calls_pending``,
+…) and dataclass field churn would just get in the way.
+
+Stable keys (read by multiple steps):
+
+  task          str             the current top-level task
+  current_plan  str | None      latest plan emitted by the planner
+  current_subtask str | None    latest subtask the policy is executing
+  current_memory str | None     latest compressed memory
+  recent_interjection str | None  most recent user interjection text (consumed)
+
+  action_queue  collections.deque[Tensor]  pending action chunks
+  tool_calls_pending list[dict]  parsed but not-yet-dispatched tool calls
+
+  events_this_tick list[str]    triggers consumed this tick
+  _tick         Tick            current tick (set by the loop)
+
+  mode          str             "action" (run the robot) | "paused"
+                                 (action loop stopped — robot holds)
+
+  log_lines     list[str]       human-readable status lines printed each tick
+"""
+
+from __future__ import annotations
+
+from collections import deque
+from typing import Any
+
+
+def initial_runtime_state(task: str | None = None) -> dict[str, Any]:
+    """Build a fresh runtime state dict with sensible defaults."""
+    return {
+        "task": task,
+        "current_plan": None,
+        "current_subtask": None,
+        "current_memory": None,
+        "recent_interjection": None,
+        "action_queue": deque(),
+        "tool_calls_pending": [],
+        "events_this_tick": [],
+        "log_lines": [],
+        "mode": "action",
+        "stop": False,
+    }
+
+
+def take_event(state: dict[str, Any], event_name: str) -> bool:
+    """Pop ``event_name`` from ``events_this_tick`` if present.
+
+    Steps that consume an event call this so the same event doesn't
+    re-fire on a sibling step within the same tick.
+    """
+    events: list[str] = state.get("events_this_tick") or []
+    if event_name in events:
+        events.remove(event_name)
+        return True
+    return False
+
+
+def push_log(state: dict[str, Any], line: str) -> None:
+    """Append ``line`` to the per-tick log buffer; the runtime prints
+    it at the end of the tick."""
+    state.setdefault("log_lines", []).append(line)
+
+
+def set_if_changed(state: dict[str, Any], key: str, value: Any, label: str | None = None) -> bool:
+    """Update ``state[key]`` and log a diff line if the value changed.
+
+    Returns ``True`` if the value actually changed.
+    """
+    prev = state.get(key)
+    if prev == value:
+        return False
+    state[key] = value
+    if label is not None:
+        push_log(state, f"  {label}: {value}")
+    return True

src/lerobot/policies/pi052/inference/steps.py

@@ -0,0 +1,955 @@
+# Copyright 2026 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""Inference steps for the PI052 multi-rate runtime.
+
+Each step is a tiny class with a ``trigger`` and an ``__call__(state)``;
+the runtime applies them in order each tick. When a step's trigger
+doesn't fire, the step is a no-op and the runtime moves on.
+
+Stream-to-step mapping mirrors the ``subtasks_vqa.yaml`` recipe:
+
+* ``LowLevelForward``        — calls ``policy.select_action`` for the
+                                action chunk; trained by
+                                ``low_level_execution``
+* ``EnqueueChunk``           — pushes the chunk to ``action_queue``
+* ``DispatchAction``         — pops one action per control tick and
+                                forwards to the robot
+* ``HighLevelSubtaskFwd``    — calls ``policy.select_message`` for the
+                                next subtask; trained by
+                                ``high_level_subtask``
+* ``MemoryUpdateFwd``        — fires on subtask boundary; trained by
+                                ``memory_update``
+* ``UserInterjectionFwd``    — fires on stdin interjection; trained by
+                                ``user_interjection_response``
+* ``AskVQAFwd``              — fires on stdin question; trained by
+                                ``ask_vqa_*``
+* ``DispatchToolCalls``      — pops ``tool_calls_pending`` and calls
+                                the matching ``Tool`` instance
+"""
+
+from __future__ import annotations
+
+import logging
+import re
+from dataclasses import dataclass, field
+from typing import Any
+
+from .runtime_state import push_log, set_if_changed, take_event
+from .triggers import EventTrigger, HzTrigger, Trigger
+
+logger = logging.getLogger(__name__)
+
+
+# ---------------------------------------------------------------------------
+# Step base + runner
+# ---------------------------------------------------------------------------
+
+
+@dataclass
+class InferenceStep:
+    """A trigger-gated callable. Subclasses override :meth:`run`."""
+
+    trigger: Trigger
+
+    def __call__(self, state: dict[str, Any]) -> dict[str, Any]:
+        if not self.trigger.should_fire(state["_tick"], state):
+            return state
+        return self.run(state) or state
+
+    def run(self, state: dict[str, Any]) -> dict[str, Any] | None:  # pragma: no cover
+        raise NotImplementedError
+
+
+# ---------------------------------------------------------------------------
+# Low-level (action) path
+# ---------------------------------------------------------------------------
+
+
+@dataclass
+class LowLevelForward(InferenceStep):
+    """Run the policy's action head and produce one action chunk."""
+
+    policy: Any = None
+    observation_provider: Any = None
+    """Callable ``() -> dict``: returns the current observation batch
+    (already preprocessed). Typically wraps the robot's camera /
+    proprio reads. ``None`` in dry-run mode → step skips."""
+
+    trigger: Trigger = field(default_factory=lambda: HzTrigger(hz=4.0))
+
+    def run(self, state: dict[str, Any]) -> dict[str, Any] | None:
+        if self.policy is None or self.observation_provider is None:
+            return None
+        # ``/vlm`` mode pauses the whole action loop so the robot holds
+        # position while the operator probes the VLM with VQA.
+        if state.get("mode", "action") != "action":
+            return None
+        if not state.get("task"):
+            return None
+
+        # PI052 produces *action chunks* (typically 50 steps via
+        # flow-matching). Every step gets dispatched to the robot;
+        # popping one per dispatch tick is essentially free. Only
+        # generate a new chunk once the previous one has fully
+        # drained — this is the canonical "sense → think → act"
+        # loop. Refreshing while a chunk is still queued causes the
+        # new chunk to "telescope" past the old one (planned from an
+        # observation that's already 25+ steps stale by the time it
+        # starts dispatching).
+        queue = state.setdefault("action_queue", [])
+        if len(queue) > 0:
+            return None
+
+        observation = self.observation_provider()
+        if observation is None:
+            return None
+
+        # The action expert is conditioned on the SUBTASK generated by
+        # the high-level loop (``HighLevelSubtaskFwd`` runs earlier in
+        # the pipeline and writes ``current_subtask``). Matches the
+        # training-time ``low_level_execution`` recipe — ``user(${subtask})``.
+        # Falls back to the task string only on the very first frame,
+        # before the high-level loop has produced a subtask.
+        subtask = state.get("current_subtask") or state.get("task") or ""
+        ctx = [{"role": "user", "content": subtask}]
+        # ``add_generation_prompt=False`` to match the training-time
+        # prefix shape: at training the action expert sees the rendered
+        # user turn ending at ``<|im_end|>`` (no trailing
+        # ``<|im_start|>assistant\n``). Passing True here would append
+        # extra role-marker tokens the action expert never saw during
+        # training.
+        text_batch = _build_text_batch(self.policy, ctx, add_generation_prompt=False)
+        from lerobot.utils.constants import (  # noqa: PLC0415
+            OBS_LANGUAGE_ATTENTION_MASK,
+            OBS_LANGUAGE_TOKENS,
+        )
+
+        observation = dict(observation)
+        observation[OBS_LANGUAGE_TOKENS] = text_batch["lang_tokens"]
+        observation[OBS_LANGUAGE_ATTENTION_MASK] = text_batch["lang_masks"]
+
+        try:
+            # ``predict_action_chunk`` returns the *full* chunk shape
+            # ``(batch, n_action_steps, action_dim)``. Enqueue every
+            # step so DispatchAction at ctrl_hz can drain them
+            # smoothly until the next refresh.
+            chunk = self.policy.predict_action_chunk(observation)
+        except Exception as exc:  # noqa: BLE001
+            logger.warning(
+                "predict_action_chunk failed: %s",
+                exc,
+                exc_info=logger.isEnabledFor(logging.DEBUG),
+            )
+            push_log(
+                state,
+                f"  [warn] predict_action_chunk failed: "
+                f"{type(exc).__name__}: {exc}",
+            )
+            return None
+
+        # ``chunk`` shape: ``(batch, n_action_steps, action_dim)``. Push
+        # each step as a ``(1, action_dim)`` tensor so the existing
+        # action executor's batch-squeeze logic works unchanged.
+        if chunk.ndim == 3:
+            chunk_iter = chunk[0]  # ``(n_action_steps, action_dim)``
+        elif chunk.ndim == 2:
+            chunk_iter = chunk
+        else:
+            chunk_iter = chunk.unsqueeze(0)
+
+        for step in chunk_iter:
+            queue.append(step.unsqueeze(0))
+        state["last_chunk_size"] = int(chunk_iter.shape[0])
+        return None
+
+
+@dataclass
+class DispatchAction(InferenceStep):
+    """Pop one action per tick and hand it to the robot.
+
+    In dry-run mode (``robot_executor=None``) the step still pops the
+    queue so it doesn't grow unbounded — the popped tensor is logged
+    instead of executed.
+
+    Wall-clock catch-up: the action queue represents an open-loop
+    trajectory at a fixed step rate (``trigger.hz`` ≈ ``ctrl_hz``).
+    When the main loop stalls — e.g. an LLM call for the high-level
+    subtask blocks for ~2 s on MPS — the dispatch trigger fires only
+    once over that whole interval. Naively popping a single entry per
+    fire makes the robot lag further and further behind the planned
+    timeline, and a 50-step chunk would take ~125 s to drain instead
+    of ~1.7 s. Track real elapsed time between dispatches and pop
+    ``round(elapsed * hz)`` entries, sending the most recent one. The
+    skipped intermediate joint targets are stale anyway — the dynamixel
+    will smooth toward the latest goal position.
+    """
+
+    robot_executor: Any = None
+    trigger: Trigger = field(default_factory=lambda: HzTrigger(hz=50.0))
+    _last_dispatch_t: float | None = field(default=None, init=False)
+
+    def run(self, state: dict[str, Any]) -> dict[str, Any] | None:
+        import time as _time  # noqa: PLC0415
+
+        # ``/vlm`` mode pauses dispatch — the robot holds its last
+        # commanded position while the operator runs VQA.
+        if state.get("mode", "action") != "action":
+            self._last_dispatch_t = None
+            return None
+
+        queue = state.get("action_queue")
+        if not queue:
+            # Reset wall-clock anchor when the queue is empty so the
+            # next chunk doesn't see a huge fake "elapsed" window.
+            self._last_dispatch_t = None
+            return None
+
+        now = _time.monotonic()
+        hz = getattr(self.trigger, "hz", 30.0)
+        if self._last_dispatch_t is None or hz <= 0:
+            n_to_pop = 1
+        else:
+            elapsed = now - self._last_dispatch_t
+            # ``max(1, ...)`` so we always pop at least one when the
+            # trigger fires; ``min(len(queue), ...)`` so we don't run
+            # off the end of the chunk.
+            n_to_pop = max(1, min(len(queue), int(round(elapsed * hz))))
+        self._last_dispatch_t = now
+
+        # Drain ``n_to_pop`` stale entries, keep only the latest as the
+        # action actually sent. The intermediate joint targets would
+        # all be ~10–30 ms apart in chunk time — the robot can't track
+        # them individually anyway when the host loop is slow.
+        latest = None
+        for _ in range(n_to_pop):
+            if not queue:
+                break
+            latest = queue.popleft() if hasattr(queue, "popleft") else queue.pop(0)
+            state["actions_dispatched"] = state.get("actions_dispatched", 0) + 1
+
+        if latest is not None and self.robot_executor is not None:
+            self.robot_executor(latest)
+        return None
+
+
+# ---------------------------------------------------------------------------
+# High-level (text) paths — all use policy.select_message
+# ---------------------------------------------------------------------------
+
+
+_LOC_TOKENIZER_CACHE: dict[str, Any] = {}
+
+
+def _get_loc_tokenizer(tok_name: str, auto_tokenizer_cls: Any, register_loc_fn: Any) -> Any:
+    """Return a loc-token-registered tokenizer, loading from disk only once.
+
+    ``AutoTokenizer.from_pretrained`` + loc-token registration is expensive and
+    the result is immutable, so cache per ``tok_name``.
+    """
+    tokenizer = _LOC_TOKENIZER_CACHE.get(tok_name)
+    if tokenizer is None:
+        tokenizer = register_loc_fn(auto_tokenizer_cls.from_pretrained(tok_name))
+        _LOC_TOKENIZER_CACHE[tok_name] = tokenizer
+    return tokenizer
+
+
+def _build_text_batch(
+    policy: Any,
+    prompt_messages: list[dict[str, Any]],
+    *,
+    add_generation_prompt: bool = True,
+) -> dict[str, Any]:
+    """Tokenize chat messages into the batch ``select_message`` expects.
+
+    PI052's backbone (PaliGemma) ships no chat template, so we train on
+    a plain role-prefixed concatenation built by
+    ``PI052TextTokenizerStep``. We reuse that exact formatter so the
+    inference prefix matches training; ``add_generation_prompt`` appends
+    the bare ``Assistant: `` header the LM head continues from.
+    """
+    import torch  # noqa: PLC0415
+    from transformers import AutoTokenizer  # noqa: PLC0415
+
+    from lerobot.policies.pi052.text_processor_pi052 import (  # noqa: PLC0415
+        _flatten_say_tool_calls,
+        _format_messages,
+        _strip_blocks,
+        register_paligemma_loc_tokens,
+    )
+
+    tok_name = (
+        getattr(policy.config, "tokenizer_name", None) or "google/paligemma-3b-pt-224"
+    )
+    # Register PaliGemma's <locDDDD> tokens so inference encoding /
+    # decoding sees them as single vocab ids — must match training.
+    # The tokenizer is read-only after registration, so cache it: rebuilding it
+    # from disk on every call dominated eval runtime (this runs twice per env
+    # per replan — subtask gen + action prompt).
+    tokenizer = _get_loc_tokenizer(tok_name, AutoTokenizer, register_paligemma_loc_tokens)
+
+    messages = [_strip_blocks(_flatten_say_tool_calls(m)) for m in prompt_messages]
+    prompt, _spans = _format_messages(messages)
+    if add_generation_prompt:
+        prompt = prompt + "Assistant: "
+
+    encoded = tokenizer(prompt, return_tensors="pt")
+    ids = encoded["input_ids"]
+    attn = encoded.get("attention_mask")
+    if attn is None and tokenizer.pad_token_id is not None:
+        attn = ids != tokenizer.pad_token_id
+    if attn is not None and hasattr(attn, "dtype") and attn.dtype != torch.bool:
+        attn = attn.bool()
+
+    # Move tokens onto the policy's device — otherwise prefix embedding
+    # raises a device-mismatch on every forward (CPU tensor vs MPS / CUDA
+    # model), which the caller's broad except would swallow silently.
+    device = getattr(getattr(policy, "config", None), "device", None)
+    if device is not None:
+        try:
+            ids = ids.to(device)
+            if attn is not None and hasattr(attn, "to"):
+                attn = attn.to(device)
+        except Exception as exc:  # noqa: BLE001
+            logger.debug("could not move pi052 lang tokens to %s: %s", device, exc)
+    return {"lang_tokens": ids, "lang_masks": attn, "tokenizer": tokenizer}
+
+
+def _strip_recipe_keys(m: dict[str, Any]) -> dict[str, Any]:
+    new = dict(m)
+    new.pop("stream", None)
+    new.pop("target", None)
+    return new
+
+
+@dataclass
+class HighLevelSubtaskFwd(InferenceStep):
+    """At ~1 Hz, ask the policy for the next subtask.
+
+    Mirrors the ``high_level_subtask`` recipe layout exactly:
+
+        user:   "${task}\\nPlan: ${plan}\\nMemory: ${memory}"
+        user:   "Current subtask: ${subtask}"        (if subtask present)
+        ↓ generate ↓
+        assistant: <next subtask>
+    """
+
+    policy: Any = None
+    observation_provider: Any = None
+    """Same shape as ``LowLevelForward.observation_provider``. When
+    set, the resulting observation is merged into ``select_message``'s
+    batch so text generation runs against real video + state."""
+
+    trigger: Trigger = field(default_factory=lambda: HzTrigger(hz=1.0))
+
+    def run(self, state: dict[str, Any]) -> dict[str, Any] | None:
+        if self.policy is None or not state.get("task"):
+            return None
+        # ``/vlm`` mode pauses subtask generation along with the rest of
+        # the action loop.
+        if state.get("mode", "action") != "action":
+            return None
+        # Gate to chunk boundaries: only generate a fresh subtask when
+        # the action queue is empty (i.e. right before LowLevelForward
+        # refreshes the chunk). ``select_message`` takes ~2 s on MPS,
+        # and running it every loop iteration starves DispatchAction
+        # at ctrl_hz=30 — the queue drains at ~0.4 actions/sec instead
+        # of 30/sec and the robot barely moves. Tying it to the same
+        # "queue empty" condition as the chunk refresh produces a
+        # clean sense → think → act cycle.
+        #
+        # Rearm the trigger when skipping so a low-hz schedule
+        # (e.g. ``--high_level_hz=0.2`` = once per 5 s) doesn't lose
+        # the slot: the trigger fires once on the timer but the brief
+        # queue-empty window almost never coincides, so without rearm
+        # HL would effectively never run.
+        queue = state.get("action_queue") or []
+        if len(queue) > 0:
+            if hasattr(self.trigger, "rearm"):
+                self.trigger.rearm()
+            return None
+        # Per-chunk-boundary throttle: at each "queue empty" moment we
+        # increment a counter; subtask gen only fires once the counter
+        # reaches ``subtask_chunks_per_gen``. Lets the operator run e.g.
+        # 5 action chunks per subtask-gen so the LM head doesn't churn
+        # every 1.7 s (a fresh subtask while the previous one is still
+        # being executed is wasted compute *and* causes the action
+        # expert's flow trajectory to be re-planned mid-grasp).
+        chunks_per_gen = max(1, int(state.get("subtask_chunks_per_gen", 1) or 1))
+        # Initialise so the first chunk boundary fires immediately
+        # (counter starts at chunks_per_gen, decrements per skip,
+        # generates and resets when it hits 0).
+        if "_hl_chunks_until_gen" not in state:
+            state["_hl_chunks_until_gen"] = 0
+        if state["_hl_chunks_until_gen"] > 0:
+            state["_hl_chunks_until_gen"] -= 1
+            if hasattr(self.trigger, "rearm"):
+                self.trigger.rearm()
+            return None
+        state["_hl_chunks_until_gen"] = chunks_per_gen - 1
+        ctx = _msgs_for_subtask(state)
+        observation = _maybe_observation(self.observation_provider)
+        # Default: greedy argmax, no min_new_tokens, no special-token
+        # suppression — matches training. Operator can override via
+        # ``--text_min_new_tokens=N --text_temperature=T --text_top_p=P``
+        # on the CLI; useful for under-trained checkpoints whose LM
+        # head still favours EOS at position 0 (pre-trained chat
+        # backbone's short-turn prior hasn't been fully overridden
+        # by the fine-tuning supervision yet).
+        msg = _generate_with_policy(
+            self.policy,
+            ctx,
+            observation=observation,
+            state=state,
+            label="subtask gen",
+            min_new_tokens=int(state.get("text_gen_min_new_tokens") or 0),
+            temperature=float(state.get("text_gen_temperature") or 0.0),
+            top_p=float(state.get("text_gen_top_p") or 1.0),
+            # Subtasks never legitimately contain PaliGemma ``<loc>``
+            # tokens — suppress them so a checkpoint whose LM head
+            # has drifted toward the pretrained loc-prior falls back
+            # to its (still-correct) text mass.
+            suppress_loc_tokens=True,
+        )
+        # Diagnostics: surface what the model is *actually* producing
+        # at chunk boundaries, even when the output gets rejected or
+        # repeats. Memorisation collapse looks like "same accepted
+        # subtask N times in a row" or "gibberish_count rising while
+        # current_subtask is stuck". The state panel renders these.
+        state["last_subtask_raw"] = msg or ""
+        # Persistent empty completion is its own failure mode (model
+        # immediately EOS-es from the chat-template generation
+        # prompt) — surface it once every N occurrences so the
+        # operator can distinguish "generation failing silently"
+        # from "generating fine but filter rejecting".
+        if not msg:
+            empties = state.get("subtask_empty_count", 0) + 1
+            state["subtask_empty_count"] = empties
+            if empties == 1 or empties % 5 == 0:
+                debug = getattr(self.policy, "_last_select_message_debug", "") or ""
+                if debug:
+                    push_log(
+                        state,
+                        f"  [info] subtask gen empty (×{empties}); {debug}",
+                    )
+                else:
+                    push_log(
+                        state,
+                        f"  [info] subtask gen returned empty (×{empties}) — "
+                        "no tokens generated (head EOS-ing before any "
+                        "non-special token).",
+                    )
+        if msg and _looks_like_gibberish(msg):
+            # Bump a counter so the operator can see the model is
+            # struggling without spamming the log every tick. A first
+            # rejection still logs once so the failure is visible.
+            count = state.get("subtask_gibberish_count", 0) + 1
+            state["subtask_gibberish_count"] = count
+            if count == 1 or count % 30 == 0:
+                push_log(
+                    state,
+                    f"  [info] subtask gen rejected (gibberish ×{count}): {msg[:60]!r}",
+                )
+            return None
+        if msg:
+            prev_subtask = state.get("current_subtask")
+            changed = set_if_changed(state, "current_subtask", msg, label="subtask")
+            if changed:
+                # Stash the just-completed subtask so ``MemoryUpdateFwd``
+                # can drop it into its prompt as ``Completed subtask:``
+                # — the recipe binds ``completed_subtask`` to
+                # ``nth_prev(style=subtask, offset=1)``, i.e. the subtask
+                # that was active *before* the change.
+                if prev_subtask:
+                    state["prior_subtask"] = prev_subtask
+                # Subtask change is a downstream trigger.
+                state.setdefault("events_this_tick", []).append("subtask_change")
+                state["subtask_repeat_count"] = 0
+            else:
+                # Same accepted string regenerated — memorisation tell.
+                # Once this counter climbs past a few, you're seeing
+                # the model unable to move past the current subtask
+                # despite the chunk having drained (visual scene may
+                # have changed but the LM is replaying training
+                # tokens).
+                state["subtask_repeat_count"] = (
+                    state.get("subtask_repeat_count", 0) + 1
+                )
+        # Silently skip empty completions — common when the model
+        # warms up or generates only EOS; logging it every tick at
+        # ctrl_hz is just noise.
+        return None
+
+
+@dataclass
+class MemoryUpdateFwd(InferenceStep):
+    """On subtask boundary, refresh the compressed memory.
+
+    Mirrors the ``memory_update`` recipe layout exactly:
+
+        user:      "${task}"
+        assistant: "Previous memory: ${prior_memory}"   (if prior memory)
+        user:      "Completed subtask: ${completed_subtask}"  (if subtask)
+        ↓ generate ↓
+        assistant: <new memory>
+    """
+
+    policy: Any = None
+    observation_provider: Any = None
+    trigger: Trigger = field(default_factory=lambda: EventTrigger("subtask_change"))
+
+    def run(self, state: dict[str, Any]) -> dict[str, Any] | None:
+        # Don't consume the event — multiple steps may want to react.
+        if self.policy is None:
+            return None
+        ctx = _msgs_for_memory(state)
+        observation = _maybe_observation(self.observation_provider)
+        new_memory = _generate_with_policy(
+            self.policy,
+            ctx,
+            observation=observation,
+            state=state,
+            label="memory gen",
+            suppress_loc_tokens=True,
+        )
+        state["last_memory_raw"] = new_memory or ""
+        if new_memory and _looks_like_gibberish(new_memory):
+            count = state.get("memory_gibberish_count", 0) + 1
+            state["memory_gibberish_count"] = count
+            push_log(
+                state,
+                f"  [info] memory gen rejected (gibberish ×{count}): {new_memory[:60]!r}",
+            )
+            return None
+        if new_memory:
+            set_if_changed(state, "current_memory", new_memory, label="memory")
+        return None
+
+
+@dataclass
+class UserInterjectionFwd(InferenceStep):
+    """On stdin interjection, refresh the plan + emit a paired ``say``.
+
+    Mirrors the ``user_interjection_response`` recipe layout exactly:
+
+        user:      "${task}"
+        assistant: "Previous plan:\\n${prior_plan}"   (if prior plan)
+        user:      "${interjection}"                  (the new utterance)
+        ↓ generate ↓
+        assistant: <plan + <say>...</say>>
+    """
+
+    policy: Any = None
+    observation_provider: Any = None
+    trigger: Trigger = field(default_factory=lambda: EventTrigger("user_interjection"))
+
+    def run(self, state: dict[str, Any]) -> dict[str, Any] | None:
+        if self.policy is None or not take_event(state, "user_interjection"):
+            return None
+        ctx = _msgs_for_interjection(state)
+        observation = _maybe_observation(self.observation_provider)
+        out = _generate_with_policy(
+            self.policy,
+            ctx,
+            observation=observation,
+            state=state,
+            label="plan/say gen",
+            suppress_loc_tokens=True,
+        )
+        if not out:
+            # Don't log every empty completion — happens repeatedly on
+            # MPS during warm-up and floods the panel. The user can
+            # re-trigger by typing again.
+            return None
+        if _looks_like_gibberish(out):
+            count = state.get("plan_gibberish_count", 0) + 1
+            state["plan_gibberish_count"] = count
+            push_log(
+                state,
+                f"  [info] plan/say gen rejected (gibberish ×{count}): {out[:60]!r}",
+            )
+            return None
+        # Heuristic split: model is trained to emit one assistant turn
+        # carrying both plan text AND a `say` tool call. Look for a
+        # "<say>...</say>" or "say(...)" marker; fall back to whole
+        # text → plan, no speech.
+        plan_text, speech_text = _split_plan_and_say(out)
+        if plan_text and _looks_like_gibberish(plan_text):
+            plan_text = ""
+        if plan_text:
+            set_if_changed(state, "current_plan", plan_text, label="plan")
+        if speech_text:
+            push_log(state, f"  speech: {speech_text}")
+            state.setdefault("tool_calls_pending", []).append(
+                {
+                    "type": "function",
+                    "function": {"name": "say", "arguments": {"text": speech_text}},
+                }
+            )
+            state.setdefault("events_this_tick", []).append("tool_call_pending")
+        # Mark interjection consumed.
+        state["recent_interjection"] = None
+        return None
+
+
+@dataclass
+class AskVQAFwd(InferenceStep):
+    """On stdin question, answer a frame-grounded VQA.
+
+    Mirrors the ``ask_vqa_*`` recipe layout exactly: a single user
+    turn carrying just the VQA question, plus the camera image block
+    in training (we drop the image at inference because the dataset's
+    image preprocessing doesn't match SmolVLM's vision tower input).
+
+        user:   <question>
+        ↓ generate ↓
+        assistant: <vqa answer>
+    """
+
+    policy: Any = None
+    observation_provider: Any = None
+    trigger: Trigger = field(default_factory=lambda: EventTrigger("user_vqa_query"))
+
+    def run(self, state: dict[str, Any]) -> dict[str, Any] | None:
+        if self.policy is None or not take_event(state, "user_vqa_query"):
+            return None
+        question = state.get("recent_vqa_query")
+        if not question:
+            return None
+        ctx = _msgs_for_vqa(question)
+        observation = _maybe_observation(self.observation_provider)
+        answer = _generate_with_policy(
+            self.policy,
+            ctx,
+            observation=observation,
+            state=state,
+            label="vqa gen",
+        )
+        # VQA answers are intentionally JSON-like during training, so
+        # ``_looks_like_gibberish`` would false-positive on them. Keep
+        # the answer as-is — the VQA panel line lets the user judge.
+        if answer:
+            push_log(state, f"  vqa: {answer}")
+        state["recent_vqa_query"] = None
+        return None
+
+
+# ---------------------------------------------------------------------------
+# Tool dispatch
+# ---------------------------------------------------------------------------
+
+
+@dataclass
+class DispatchToolCalls(InferenceStep):
+    """Pop ``tool_calls_pending`` and execute them via :data:`TOOL_REGISTRY`."""
+
+    tools: dict[str, Any] = field(default_factory=dict)
+    trigger: Trigger = field(default_factory=lambda: EventTrigger("tool_call_pending"))
+
+    def run(self, state: dict[str, Any]) -> dict[str, Any] | None:
+        take_event(state, "tool_call_pending")
+        pending = state.get("tool_calls_pending") or []
+        for call in pending:
+            try:
+                fn = (call or {}).get("function") or {}
+                name = fn.get("name")
+                args = fn.get("arguments") or {}
+                tool = self.tools.get(name)
+                if tool is None:
+                    push_log(state, f"  [warn] tool {name!r} not registered — skipping call")
+                    continue
+                tool.call(args)
+            except Exception as exc:  # noqa: BLE001
+                push_log(state, f"  [error] tool dispatch failed: {exc}")
+        state["tool_calls_pending"] = []
+        return None
+
+
+# ---------------------------------------------------------------------------
+# Helpers
+# ---------------------------------------------------------------------------
+
+
+def _looks_like_gibberish(text: str) -> bool:
+    """Heuristically detect generation that's clearly off the rails.
+
+    Memorised models can collapse to dominant-mode outputs when the
+    prompt drifts even slightly from training distribution. Reject:
+
+    * empty / whitespace-only
+    * too few alphabetic characters (mostly punctuation)
+    * a single character repeated past the threshold
+    * starts with ``":"`` and contains no letters
+    * too few unique tokens — e.g. ``"the"``, ``"the the the"``,
+      ``"Ass\\n::\\nthe"`` (the collapse seen on real-robot frames
+      where the model emits one or two memorised tokens repeatedly)
+    * chat-template fragment leakage (``Assistant:``, ``User:``,
+      ``Ass\\n``)
+
+    Real subtasks look like ``"close the gripper to grasp the blue
+    cube"`` — multiple unique alphabetic tokens, no role-marker
+    fragments. Anything materially shorter than that is rejected.
+    """
+    if not text or not text.strip():
+        return True
+    stripped = text.strip()
+    alpha = sum(1 for c in stripped if c.isalpha())
+    if alpha < max(3, len(stripped) // 8):
+        return True
+    if stripped.startswith('":') and stripped.count('"') > stripped.count(" "):
+        return True
+    # Single repeating char: e.g. ``""""""``.
+    if len(set(stripped)) <= 2 and len(stripped) > 4:
+        return True
+    # Chat-template fragment leakage — the model emits ``Ass``,
+    # ``Assistant:``, ``User:``, often with extra newlines/colons.
+    # Reject if the cleaned text is mostly role-marker shards.
+    cleaned = stripped.replace("\n", " ").replace(":", " ")
+    for marker in ("Assistant", "User", "Ass "):
+        if marker in cleaned and len(cleaned.split()) < 4:
+            return True
+    tokens = [t for t in cleaned.split() if any(c.isalpha() for c in t)]
+    unique_alpha = {t.lower() for t in tokens}
+    # Short degenerate output — model stuck on ``the`` or a couple of
+    # memorised single-token continuations.
+    if len(unique_alpha) < 3 and len(stripped) < 80:
+        return True
+    # Long repetition collapse — the LM head loops an n-gram for the
+    # whole generation budget ("the arm the arm … the the the the").
+    # Length-independent: many tokens but a tiny unique ratio. The
+    # earlier ``< 80`` check missed these because the looped string
+    # blows well past 80 chars.
+    if len(tokens) >= 8 and len(unique_alpha) <= max(3, len(tokens) // 10):
+        return True
+    return False
+
+
+def _control_context_messages(
+    state: dict[str, Any],
+    *,
+    include_completed: bool = False,
+    extra_user: str | None = None,
+) -> list[dict[str, Any]]:
+    """Build a chat-template-ready prompt from current runtime state.
+
+    Mirrors what ``subtasks_vqa.yaml`` renders into ``${task}\nPlan:
+    ${plan}\nMemory: ${memory}`` for the high-level branches.
+    """
+    # Always emit ``Plan: `` / ``Memory: `` labels — even with empty
+    # values — to mirror the training-time recipe substitution.
+    task = state.get("task") or ""
+    plan = state.get("current_plan") or ""
+    memory = state.get("current_memory") or ""
+    parts = [task, f"Plan: {plan}", f"Memory: {memory}"]
+    if include_completed and state.get("current_subtask"):
+        parts.append(f"Completed subtask: {state['current_subtask']}")
+    head = "\n".join(parts)
+    msgs: list[dict[str, Any]] = [{"role": "user", "content": head}]
+    if extra_user:
+        msgs.append({"role": "user", "content": extra_user})
+    return msgs
+
+
+# ---------------------------------------------------------------------------
+# Per-recipe prompt builders. Each one mirrors a single sub-recipe's
+# message layout in ``subtasks_vqa.yaml`` so the chat-templated
+# prompt at inference matches what the model saw during training.
+# Generic ``_control_context_messages`` is kept around as a fallback
+# for ad-hoc callers but the four high-level steps now use these.
+# ---------------------------------------------------------------------------
+
+
+def _hirobot_user_head(state: dict[str, Any]) -> str:
+    """Build the ``task\\nPlan: …\\nMemory: …`` user content string.
+
+    Mirrors what the recipe renders at training time, where
+    ``language_render._substitute`` substitutes empty strings for
+    missing ``${plan}`` / ``${memory}`` bindings — i.e. the
+    ``Plan: `` / ``Memory: `` prefix labels are *always* in the
+    user turn, even when their values aren't set yet. Skipping them
+    here (the previous behaviour) produced a different prompt shape
+    on early frames before plan / memory are populated and on
+    samples where the dataset has no plan / memory annotation.
+    """
+    task = state.get("task") or ""
+    plan = state.get("current_plan") or ""
+    memory = state.get("current_memory") or ""
+    return f"{task}\nPlan: {plan}\nMemory: {memory}"
+
+
+def _msgs_for_subtask(state: dict[str, Any]) -> list[dict[str, Any]]:
+    """``high_level_subtask`` recipe layout — predict the subtask from the
+    task. The v-current recipe's user turn is just ``${task}`` (plan and
+    memory are not trained), so the inference prompt is the bare task —
+    no ``Plan: `` / ``Memory: `` lines.
+    """
+    return [{"role": "user", "content": state.get("task") or ""}]
+
+
+def _msgs_for_memory(state: dict[str, Any]) -> list[dict[str, Any]]:
+    """Memory-update prompt — mirrors ``memory_update`` recipe layout.
+
+    Recipe layout (``subtask_mem.yaml``):
+
+        user:      "${task}"
+        assistant: "Previous memory: ${prior_memory}"     (if_present prior)
+        user:      "Completed subtask: ${completed}"       (if_present completed)
+        assistant: → predicts new memory
+
+    Fired by ``MemoryUpdateFwd`` on a ``subtask_change`` event:
+    ``state['current_memory']`` is the memory the policy last emitted
+    (= the ``prior_memory`` binding at training), and
+    ``state['prior_subtask']`` is the subtask that just got replaced
+    (= the ``completed_subtask`` binding at training).
+    """
+    msgs: list[dict[str, Any]] = [
+        {"role": "user", "content": state.get("task") or ""},
+    ]
+    prior_memory = state.get("current_memory")
+    if prior_memory:
+        msgs.append(
+            {"role": "assistant", "content": f"Previous memory: {prior_memory}"}
+        )
+    completed_subtask = state.get("prior_subtask")
+    if completed_subtask:
+        msgs.append(
+            {"role": "user", "content": f"Completed subtask: {completed_subtask}"}
+        )
+    return msgs
+
+
+def _msgs_for_interjection(state: dict[str, Any]) -> list[dict[str, Any]]:
+    """``user_interjection_response`` recipe layout."""
+    msgs: list[dict[str, Any]] = [
+        {"role": "user", "content": state.get("task") or ""}
+    ]
+    if state.get("current_plan"):
+        msgs.append(
+            {"role": "assistant", "content": f"Previous plan:\n{state['current_plan']}"}
+        )
+    interjection = state.get("recent_interjection")
+    if interjection:
+        msgs.append({"role": "user", "content": interjection})
+    return msgs
+
+
+def _msgs_for_plan(state: dict[str, Any]) -> list[dict[str, Any]]:
+    """``plan_generation`` recipe layout — bare task → plan.
+
+    The assistant turn is the generation target, so we only render
+    the user turn at inference; the runtime appends the predicted
+    plan after sampling.
+    """
+    return [{"role": "user", "content": state.get("task") or ""}]
+
+
+def _msgs_for_vqa(question: str) -> list[dict[str, Any]]:
+    """``ask_vqa_*`` recipe layout (text-only at inference)."""
+    return [{"role": "user", "content": question}]
+
+
+def _maybe_observation(provider: Any) -> dict | None:
+    """Pull one observation from ``provider`` if it's set, else ``None``.
+
+    Errors from the provider are logged at debug level and swallowed —
+    text generation still runs (in text-only mode) so a flaky frame
+    source doesn't kill the REPL.
+    """
+    if provider is None:
+        return None
+    try:
+        return provider()
+    except Exception as exc:  # noqa: BLE001
+        logger.debug("observation_provider raised %s — falling back to text-only", exc)
+        return None
+
+
+def _generate_with_policy(
+    policy: Any,
+    messages: list[dict[str, Any]],
+    *,
+    observation: dict | None = None,
+    state: dict[str, Any] | None = None,
+    label: str = "select_message",
+    min_new_tokens: int = 0,
+    temperature: float = 0.0,
+    top_p: float = 1.0,
+    suppress_loc_tokens: bool = False,
+) -> str:
+    """Drive ``policy.select_message`` with a chat batch (and optional obs).
+
+    When ``observation`` carries ``observation.images.*`` and
+    ``observation.state``, those are merged into the batch so
+    ``select_message`` runs the same VLM prefix the policy was trained
+    on. Without an observation the runtime falls back to a text-only
+    prompt — the text head still runs, but generations may drift from
+    the training distribution.
+
+    Failures are surfaced both to the module logger (``warning``) and,
+    when ``state`` is given, to the runtime's user-visible log via
+    :func:`push_log`, so the REPL no longer "looks dead" when
+    something goes wrong inside generation.
+    """
+    if not hasattr(policy, "select_message"):
+        if state is not None:
+            push_log(state, f"  [warn] policy has no select_message — skipping {label}")
+        return ""
+    text_batch = _build_text_batch(policy, messages)
+    try:
+        from lerobot.utils.constants import (  # noqa: PLC0415
+            OBS_LANGUAGE_ATTENTION_MASK,
+            OBS_LANGUAGE_TOKENS,
+        )
+
+        batch: dict[str, Any] = {
+            OBS_LANGUAGE_TOKENS: text_batch["lang_tokens"],
+            OBS_LANGUAGE_ATTENTION_MASK: text_batch["lang_masks"],
+        }
+        if observation:
+            for k, v in observation.items():
+                if isinstance(k, str) and k.startswith("observation.") and k not in batch:
+                    batch[k] = v
+        kwargs: dict[str, Any] = {
+            "tokenizer": text_batch["tokenizer"],
+            "min_new_tokens": min_new_tokens,
+            "temperature": temperature,
+            "top_p": top_p,
+        }
+        kwargs["suppress_loc_tokens"] = suppress_loc_tokens
+        return policy.select_message(batch, **kwargs)
+    except Exception as exc:  # noqa: BLE001
+        logger.warning("%s failed: %s", label, exc, exc_info=logger.isEnabledFor(logging.DEBUG))
+        if state is not None:
+            push_log(state, f"  [warn] {label} failed: {type(exc).__name__}: {exc}")
+        return ""
+
+
+_SAY_RE = re.compile(r"<\s*say\s*>(.*?)<\s*/\s*say\s*>", re.IGNORECASE | re.DOTALL)
+
+
+def _split_plan_and_say(text: str) -> tuple[str, str]:
+    """Pull a ``<say>...</say>`` snippet out of ``text``; remainder is plan.
+
+    The training-time tool-call serializer wraps ``say(text="…")`` in a
+    deterministic textual marker so prefix-LM-style training learns to
+    emit it. The runtime parses it back here. If no marker is present,
+    the entire text is treated as plan with no speech.
+    """
+    if not text:
+        return "", ""
+    match = _SAY_RE.search(text)
+    if not match:
+        return text.strip(), ""
+    speech = match.group(1).strip().strip('"').strip("'")
+    plan = (text[: match.start()] + text[match.end() :]).strip()
+    return plan, speech

src/lerobot/policies/pi052/inference/triggers.py

@@ -0,0 +1,134 @@
+# Copyright 2026 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""Trigger primitives for PI052's multi-rate inference runtime.
+
+Mirrors the plan's Section "Runtime orchestration": each
+``InferenceStep`` is gated by a :class:`Trigger` that decides per tick
+whether the step fires. Two trigger flavours cover all the cadences
+the canonical recipe needs:
+
+* :class:`HzTrigger` for periodic beats (action chunks at ~3-5 Hz,
+  high-level subtask generation at ~1 Hz, action dispatch at ~50 Hz)
+* :class:`EventTrigger` for one-shot reactions (subtask boundary →
+  memory update; user interjection → plan refresh; user VQA query →
+  vqa answer; pending tool call → dispatcher)
+
+Triggers are stateless except for ``HzTrigger``'s last-fire timestamp.
+The runtime stores the :class:`Tick` clock as ``state["_tick"]`` so
+every step shares a single time source.
+"""
+
+from __future__ import annotations
+
+import time
+from dataclasses import dataclass, field
+from typing import Any, Protocol
+
+
+@dataclass
+class Tick:
+    """Single tick from :class:`TickClock`. Carries time references the
+    runtime steps consume to gate themselves."""
+
+    index: int
+    """Monotonic counter — increments by one per tick."""
+
+    monotonic_seconds: float
+    """``time.monotonic()`` at the start of this tick."""
+
+
+@dataclass
+class TickClock:
+    """Drives the runtime loop at up to ``max_rate_hz``.
+
+    Sleeps just enough between :meth:`advance` calls to enforce the
+    rate. With ``max_rate_hz=50`` the loop wakes ~every 20ms; the
+    higher-level ``HzTrigger`` slices that timeline into sub-cadences.
+    """
+
+    max_rate_hz: float = 50.0
+    _index: int = field(default=0, init=False)
+    _last_seconds: float | None = field(default=None, init=False)
+
+    def advance(self) -> Tick:
+        period = 1.0 / max(self.max_rate_hz, 0.1)
+        now = time.monotonic()
+        if self._last_seconds is not None:
+            sleep_for = (self._last_seconds + period) - now
+            if sleep_for > 0:
+                time.sleep(sleep_for)
+                now = time.monotonic()
+        self._last_seconds = now
+        self._index += 1
+        return Tick(index=self._index, monotonic_seconds=now)
+
+
+class Trigger(Protocol):
+    """Decide whether the next ``InferenceStep`` should fire."""
+
+    def should_fire(self, tick: Tick, state: dict[str, Any]) -> bool: ...
+
+
+@dataclass
+class HzTrigger:
+    """Fire at most ``hz`` times per second.
+
+    A step that gates further (e.g. ``HighLevelSubtaskFwd`` skipping
+    when the action queue is non-empty) and wants the trigger to
+    retry next tick instead of waiting a full period can call
+    :meth:`rearm` from inside ``run``. Without this, a low-hz trigger
+    (e.g. ``hz=0.2`` = once per 5 s) almost never coincides with the
+    brief queue-empty window and the step never fires at all.
+    """
+
+    hz: float
+    _last_seconds: float | None = field(default=None, init=False)
+
+    def should_fire(self, tick: Tick, state: dict[str, Any]) -> bool:
+        period = 1.0 / max(self.hz, 1e-6)
+        if self._last_seconds is None or (tick.monotonic_seconds - self._last_seconds) >= period:
+            self._last_seconds = tick.monotonic_seconds
+            return True
+        return False
+
+    def rearm(self) -> None:
+        """Mark the trigger as not having fired, so the next tick re-evaluates.
+
+        Used by a step that decided to skip after ``should_fire`` already
+        committed the firing — keeps the cadence honest without losing
+        the slot.
+        """
+        self._last_seconds = None
+
+
+@dataclass
+class EventTrigger:
+    """Fire when ``event_name`` is in ``state["events_this_tick"]``.
+
+    The runtime fills ``events_this_tick`` once per tick from:
+
+    * stdin / network input (``user_interjection``, ``user_vqa_query``,
+      ``stop``)
+    * internal state transitions (``subtask_change``,
+      ``tool_call_pending``)
+
+    The list is consumed (cleared at the end of the tick) so events
+    fire at most once.
+    """
+
+    event_name: str
+
+    def should_fire(self, tick: Tick, state: dict[str, Any]) -> bool:
+        events: list[str] = state.get("events_this_tick") or []
+        return self.event_name in events

src/lerobot/policies/pi052/inference/ui.py

@@ -0,0 +1,127 @@
+# Copyright 2026 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""Rich-based REPL layout for the PI052 runtime.
+
+Two-zone terminal layout:
+
+    [chat scrollback — user messages / robot responses, scrolls naturally]
+
+    ┌── State ──────────────────────────────────────────┐
+    │ task     please clean up the kitchen              │
+    │ subtask  grasp the handle of the sponge           │
+    │ plan     1. grasp sponge  2. wipe  3. tidy        │
+    │ memory   sponge picked up; counter still dirty    │
+    └───────────────────────────────────────────────────┘
+    > _
+
+The state panel re-renders on every state change. Chat lines are
+``console.print``'d above the live region so they accumulate naturally
+in scrollback. Implemented with :class:`rich.live.Live` plus
+:func:`rich.console.Console.input` for the prompt — when an input is
+pending, ``rich.Live`` auto-suspends so the input doesn't fight the
+panel for cursor position.
+"""
+
+from __future__ import annotations
+
+from typing import Any
+
+try:  # rich is optional; only required for the interactive REPL.
+    from rich.console import Console
+    from rich.panel import Panel
+    from rich.table import Table
+    from rich.text import Text
+
+    _HAS_RICH = True
+except ImportError:  # pragma: no cover
+    _HAS_RICH = False
+    Console = Any  # type: ignore[assignment]
+    Panel = Any  # type: ignore[assignment]
+    Table = Any  # type: ignore[assignment]
+    Text = Any  # type: ignore[assignment]
+
+
+_STATE_KEYS = (
+    ("task", "task"),
+    ("current_subtask", "subtask"),
+    ("current_plan", "plan"),
+    ("current_memory", "memory"),
+)
+
+
+def make_state_panel(state: dict[str, Any]) -> Any:
+    """Render the persistent state panel for the live region.
+
+    Returns a :class:`rich.panel.Panel`. Caller passes it to
+    ``Live.update(panel)`` whenever the state changes.
+    """
+    if not _HAS_RICH:
+        raise RuntimeError(
+            "rich is required for the interactive REPL. "
+            "`pip install rich` (it's a transitive dep of lerobot)."
+        )
+    table = Table.grid(padding=(0, 2), expand=True)
+    table.add_column(justify="right", style="dim", no_wrap=True, width=10)
+    table.add_column(justify="left")
+    for key, label in _STATE_KEYS:
+        value = state.get(key)
+        if value is None:
+            rendered = Text("(not set)", style="dim italic")
+        else:
+            rendered = Text(str(value), style="bold")
+        table.add_row(label, rendered)
+    queue = state.get("action_queue")
+    queue_len = len(queue) if hasattr(queue, "__len__") else 0
+    pending = state.get("tool_calls_pending") or []
+    footer = Text.assemble(
+        ("queued actions: ", "dim"),
+        (str(queue_len), "bold cyan"),
+        ("    pending tool calls: ", "dim"),
+        (str(len(pending)), "bold magenta"),
+    )
+    table.add_row("", footer)
+    run_mode = state.get("mode", "action")
+    mode_tag = (
+        "[green]action[/]" if run_mode == "action" else "[yellow]paused[/]"
+    )
+    return Panel(
+        table,
+        title=f"[bold]PI052 state[/] · mode: {mode_tag}",
+        border_style="cyan",
+    )
+
+
+def print_user_line(console: Any, line: str) -> None:
+    """Append a user-typed line to the chat scrollback."""
+    if not _HAS_RICH:
+        print(f"you: {line}", flush=True)
+        return
+    console.print(f"[bold cyan]you:[/] {line}")
+
+
+def print_robot_lines(console: Any, lines: list[str]) -> None:
+    """Append robot/runtime log lines to the chat scrollback."""
+    if not _HAS_RICH:
+        for line in lines:
+            print(f"robot: {line.lstrip()}", flush=True)
+        return
+    for line in lines:
+        # The runtime uses leading whitespace + "label: text"; render
+        # the label in green and the value in default for readability.
+        stripped = line.lstrip()
+        if ":" in stripped:
+            label, _, value = stripped.partition(":")
+            console.print(f"[bold green]robot[/] [dim]({label.strip()})[/] {value.strip()}")
+        else:
+            console.print(f"[bold green]robot:[/] {stripped}")

src/lerobot/policies/pi052/inference/vqa.py

@@ -0,0 +1,423 @@
+# Copyright 2026 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""Interactive VQA for the PI052 runtime.
+
+In ``/vlm`` mode a typed line is treated as a VQA question. This module
+runs the full interactive flow:
+
+  1. pull the current observation and list available cameras,
+  2. ask the operator which camera to ground the question on,
+  3. generate the answer with the VLM conditioned on that one camera,
+  4. parse the JSON answer; if it carries a bounding box (``bbox``) or a
+     point (``keypoint``), draw the overlay on the camera frame, save a
+     PNG to ``./vqa_overlays/`` and auto-open it.
+
+VQA answer schemas mirror the annotation pipeline's ``VQA_ANSWER_SHAPES``
+(see ``lerobot.annotations.steerable_pipeline.validator``):
+
+  * ``bbox``     — ``{"detections": [{"label", "bbox_format": "xyxy",
+                    "bbox": [x1, y1, x2, y2]}, ...]}``
+  * ``keypoint`` — ``{"label", "point_format": "xy", "point": [x, y]}``
+  * ``count`` / ``attribute`` / ``spatial`` — text-only, no overlay.
+"""
+
+from __future__ import annotations
+
+import json
+import logging
+import os
+import re
+import subprocess
+import sys
+import time
+import webbrowser
+from pathlib import Path
+from typing import Any
+
+from .runtime_state import push_log
+
+logger = logging.getLogger(__name__)
+
+_IMAGE_PREFIX = "observation.images."
+
+# PaliGemma detection / pointing vocabulary. PI052 trains spatial VQA
+# answers in this native ``<locNNNN>`` format (index in [0, 1023],
+# normalized to the image axis) instead of pixel-coordinate JSON, so the
+# answer string the runtime parses can be e.g.
+# ``<loc0512><loc0301> blue cube`` (point) or
+# ``<loc0100><loc0080><loc0400><loc0360> blue cube`` (box).
+_LOC_RE = re.compile(r"<loc(\d{1,4})>")
+
+# Iteration order for shape matching — most specific keys first so an
+# answer is classified deterministically.
+_SHAPE_ORDER = ("bbox", "keypoint", "count", "attribute", "spatial")
+
+_BBOX_COLOR = (255, 64, 64)
+_POINT_COLOR = (64, 220, 64)
+
+
+# ---------------------------------------------------------------------------
+# Camera selection
+# ---------------------------------------------------------------------------
+
+
+def available_cameras(observation: dict | None) -> list[str]:
+    """Return the sorted ``observation.images.*`` keys present in ``observation``."""
+    if not observation:
+        return []
+    return sorted(k for k in observation if isinstance(k, str) and k.startswith(_IMAGE_PREFIX))
+
+
+def camera_short_name(camera_key: str) -> str:
+    """Strip the ``observation.images.`` prefix for display."""
+    return camera_key[len(_IMAGE_PREFIX) :] if camera_key.startswith(_IMAGE_PREFIX) else camera_key
+
+
+def prompt_camera_choice(
+    cameras: list[str],
+    *,
+    input_fn: Any = input,
+    print_fn: Any = print,
+) -> str | None:
+    """Ask the operator which camera frame to draw a VQA overlay on.
+
+    Accepts either the menu number or the (short or full) camera name.
+    A single-camera setup auto-selects without prompting. Returns the
+    chosen ``observation.images.*`` key, or ``None`` if the operator
+    cancels / gives an invalid answer.
+    """
+    if not cameras:
+        return None
+    if len(cameras) == 1:
+        return cameras[0]
+    print_fn("Draw the result on which camera?")
+    for i, cam in enumerate(cameras, 1):
+        print_fn(f"  [{i}] {camera_short_name(cam)}")
+    try:
+        raw = str(input_fn("camera> ")).strip()
+    except (EOFError, KeyboardInterrupt):
+        return None
+    if not raw:
+        return cameras[0]
+    if raw.isdigit():
+        idx = int(raw) - 1
+        return cameras[idx] if 0 <= idx < len(cameras) else None
+    for cam in cameras:
+        if raw == cam or raw == camera_short_name(cam):
+            return cam
+    return None
+
+
+# ---------------------------------------------------------------------------
+# Answer parsing
+# ---------------------------------------------------------------------------
+
+
+def _loc_to_norm(idx: int) -> float:
+    """PaliGemma ``<locNNNN>`` index → normalized [0, 1] axis coordinate."""
+    return max(0.0, min(1023.0, float(idx))) / 1023.0
+
+
+def parse_loc_answer(answer: str) -> dict | None:
+    """Parse a PaliGemma ``<loc>``-format spatial VQA answer.
+
+    PI052 trains spatial answers in PaliGemma's native detection
+    vocabulary, label-first: a point is ``<label> <locY><locX>``, a box
+    is ``<label> <locY0><locX0><locY1><locX1>``, and multiple boxes are
+    joined by `` ; `` (e.g. ``cube <loc..><loc..><loc..><loc..> ; box
+    <loc..><loc..><loc..><loc..>``). Loc-first formats are also accepted
+    — this parser strips loc tokens and treats the remainder as the
+    label, so order is irrelevant. Coordinates come back *normalized*
+    ([0, 1]); the overlay denormalizes them against the chosen camera
+    frame's pixel size.
+
+    Returns ``{"kind", "payload", "normalized": True}`` on success
+    (``payload`` mirrors the JSON shapes so the overlay code is shared),
+    or ``None`` when the answer carries no ``<loc>`` tokens.
+    """
+    if not answer or "<loc" not in answer:
+        return None
+    segments = [seg for seg in answer.split(";") if "<loc" in seg]
+    points: list[tuple[float, float, str]] = []
+    boxes: list[tuple[float, float, float, float, str]] = []
+    for seg in segments:
+        locs = [int(m) for m in _LOC_RE.findall(seg)]
+        label = _LOC_RE.sub("", seg).strip()
+        if len(locs) == 2:
+            y, x = (_loc_to_norm(v) for v in locs[:2])
+            points.append((x, y, label))
+        elif len(locs) >= 4:
+            y1, x1, y2, x2 = (_loc_to_norm(v) for v in locs[:4])
+            boxes.append((x1, y1, x2, y2, label))
+    if boxes:
+        detections = [
+            {"label": lbl, "bbox_format": "xyxy", "bbox": [x1, y1, x2, y2]}
+            for (x1, y1, x2, y2, lbl) in boxes
+        ]
+        return {"kind": "bbox", "payload": {"detections": detections}, "normalized": True}
+    if len(points) == 1:
+        x, y, lbl = points[0]
+        return {
+            "kind": "keypoint",
+            "payload": {"label": lbl, "point_format": "xy", "point": [x, y]},
+            "normalized": True,
+        }
+    if points:  # several bare points → treat as detections-as-points
+        detections = [
+            {"label": lbl, "bbox_format": "xyxy", "bbox": [x, y, x, y]} for (x, y, lbl) in points
+        ]
+        return {"kind": "bbox", "payload": {"detections": detections}, "normalized": True}
+    return None
+
+
+def parse_vqa_answer(answer: str) -> dict | None:
+    """Parse a VQA answer string into ``{"kind", "payload"}``.
+
+    ``kind`` is one of the ``VQA_ANSWER_SHAPES`` names (``bbox``,
+    ``keypoint``, ``count``, ``attribute``, ``spatial``) or ``"unknown"``
+    when the JSON doesn't match any known shape. PaliGemma ``<loc>``
+    spatial answers are detected first (PI052 trains them in that native
+    format). Returns ``None`` when the answer is neither ``<loc>`` text
+    nor a parseable JSON object.
+    """
+    if not answer or not answer.strip():
+        return None
+    loc_parsed = parse_loc_answer(answer)
+    if loc_parsed is not None:
+        return loc_parsed
+    try:
+        payload = json.loads(answer)
+    except (ValueError, TypeError):
+        return None
+    if not isinstance(payload, dict):
+        return None
+
+    try:
+        from lerobot.annotations.steerable_pipeline.validator import (  # noqa: PLC0415
+            VQA_ANSWER_SHAPES,
+        )
+
+        shapes = VQA_ANSWER_SHAPES
+    except ImportError:  # pragma: no cover - annotation extra not installed
+        shapes = {
+            "bbox": {"detections"},
+            "keypoint": {"label", "point_format", "point"},
+            "count": {"label", "count"},
+            "attribute": {"label", "attribute", "value"},
+            "spatial": {"subject", "relation", "object"},
+        }
+
+    keys = set(payload)
+    for kind in _SHAPE_ORDER:
+        required = shapes.get(kind)
+        if required and required <= keys:
+            return {"kind": kind, "payload": payload}
+    return {"kind": "unknown", "payload": payload}
+
+
+def answer_has_overlay(parsed: dict | None) -> bool:
+    """True iff ``parsed`` carries drawable spatial coordinates."""
+    return bool(parsed) and parsed.get("kind") in ("bbox", "keypoint")
+
+
+# ---------------------------------------------------------------------------
+# Overlay drawing
+# ---------------------------------------------------------------------------
+
+
+def observation_image_to_pil(image_tensor: Any) -> Any:
+    """Convert an ``observation.images.*`` tensor to a PIL RGB image.
+
+    The runtime observation stores images as ``(1, C, H, W)`` (or
+    ``(C, H, W)``) float tensors in ``[0, 1]``. Reuses
+    ``image_array_to_pil_image`` which handles the CHW→HWC transpose and
+    the float→uint8 scaling.
+    """
+    from lerobot.datasets.image_writer import image_array_to_pil_image  # noqa: PLC0415
+
+    arr = image_tensor
+    if hasattr(arr, "detach"):
+        arr = arr.detach().cpu()
+    if hasattr(arr, "numpy"):
+        arr = arr.numpy()
+    while arr.ndim > 3:  # drop leading batch dim(s)
+        arr = arr[0]
+    return image_array_to_pil_image(arr).convert("RGB")
+
+
+def draw_vqa_overlay(image: Any, parsed: dict) -> Any:
+    """Draw ``bbox`` / ``keypoint`` answers onto a copy of ``image``.
+
+    Non-spatial answers (``count`` / ``attribute`` / ``spatial`` /
+    ``unknown``) are returned as an unmodified copy. When ``parsed`` has
+    ``normalized=True`` (PaliGemma ``<loc>`` answers) the [0, 1]
+    coordinates are scaled to the image's pixel size.
+    """
+    from PIL import ImageDraw  # noqa: PLC0415
+
+    img = image.convert("RGB").copy()
+    kind = parsed.get("kind")
+    payload = parsed.get("payload") or {}
+    draw = ImageDraw.Draw(img)
+    w, h = img.size
+    sx, sy = (w, h) if parsed.get("normalized") else (1, 1)
+
+    if kind == "bbox":
+        for det in payload.get("detections") or []:
+            if not isinstance(det, dict):
+                continue
+            box = det.get("bbox")
+            if not (isinstance(box, list | tuple) and len(box) == 4):
+                continue
+            try:
+                x1, y1, x2, y2 = (float(v) for v in box)
+            except (TypeError, ValueError):
+                continue
+            x1, x2 = x1 * sx, x2 * sx
+            y1, y2 = y1 * sy, y2 * sy
+            draw.rectangle([x1, y1, x2, y2], outline=_BBOX_COLOR, width=3)
+            label = str(det.get("label", "")).strip()
+            if label:
+                draw.text((x1 + 3, max(0.0, y1 - 12)), label, fill=_BBOX_COLOR)
+    elif kind == "keypoint":
+        point = payload.get("point")
+        if isinstance(point, list | tuple) and len(point) == 2:
+            try:
+                x, y = float(point[0]) * sx, float(point[1]) * sy
+            except (TypeError, ValueError):
+                return img
+            r = 6
+            draw.ellipse([x - r, y - r, x + r, y + r], outline=_POINT_COLOR, width=3)
+            draw.line([x - 2 * r, y, x + 2 * r, y], fill=_POINT_COLOR, width=2)
+            draw.line([x, y - 2 * r, x, y + 2 * r], fill=_POINT_COLOR, width=2)
+            label = str(payload.get("label", "")).strip()
+            if label:
+                draw.text((x + r + 3, y - r), label, fill=_POINT_COLOR)
+    return img
+
+
+def _open_file(path: Path) -> None:
+    """Best-effort open ``path`` in the OS default viewer."""
+    try:
+        if sys.platform == "darwin":
+            subprocess.run(["open", str(path)], check=False)
+        elif sys.platform.startswith("linux"):
+            subprocess.run(["xdg-open", str(path)], check=False)
+        elif os.name == "nt":
+            os.startfile(str(path))  # type: ignore[attr-defined]  # noqa: S606
+        else:  # pragma: no cover - exotic platform
+            webbrowser.open(path.resolve().as_uri())
+    except Exception as exc:  # noqa: BLE001
+        logger.debug("could not auto-open %s: %s", path, exc)
+
+
+def save_and_open_overlay(image: Any, out_dir: str | Path = "./vqa_overlays") -> Path:
+    """Save ``image`` as a timestamped PNG under ``out_dir`` and auto-open it."""
+    out = Path(out_dir)
+    out.mkdir(parents=True, exist_ok=True)
+    path = out / f"vqa_{int(time.time() * 1000)}.png"
+    image.save(path)
+    _open_file(path)
+    return path
+
+
+# ---------------------------------------------------------------------------
+# Orchestrator
+# ---------------------------------------------------------------------------
+
+
+def handle_vqa_query(
+    *,
+    policy: Any,
+    observation_provider: Any,
+    question: str,
+    state: dict[str, Any],
+    input_fn: Any = input,
+    print_fn: Any = print,
+) -> None:
+    """Run one interactive VQA question end to end.
+
+    Called synchronously from the input layer while the runtime is in
+    ``/question`` mode (the action loop is gated off, so the policy is
+    not in concurrent use). Progress is reported via both
+    :func:`push_log` (REPL panel scrollback) and ``print_fn`` (direct
+    stdout) — in autonomous question mode the panel redraw is suspended,
+    so the direct print is what the operator actually sees.
+    """
+    from .steps import _generate_with_policy, _msgs_for_vqa  # noqa: PLC0415
+
+    def report(line: str) -> None:
+        """Surface a line both to the panel scrollback and to stdout."""
+        push_log(state, line)
+        try:
+            print_fn(line)
+        except Exception:  # noqa: BLE001
+            pass
+
+    if policy is None or not hasattr(policy, "select_message"):
+        report("  [warn] vqa: policy has no select_message — skipping")
+        return
+
+    observation: dict | None = None
+    if observation_provider is not None:
+        try:
+            observation = observation_provider()
+        except Exception as exc:  # noqa: BLE001
+            logger.debug("observation_provider raised %s", exc)
+
+    # Feed the FULL observation (every camera + state) to the VLM. The
+    # ``ask_vqa_*`` recipes look single-camera, but the image *block* is
+    # stripped before tokenization — the actual frames reach the model
+    # via PI052's ``OBS_IMAGES_*`` channels, and ``embed_prefix``
+    # consumes *all* ``config.image_features`` regardless of which
+    # camera the sub-recipe was tagged for. So the model always sees
+    # every camera; the operator never has to name one to ask.
+    answer = _generate_with_policy(
+        policy,
+        _msgs_for_vqa(question),
+        observation=observation,
+        state=state,
+        label="vqa gen",
+    )
+    if not answer:
+        report("  [info] vqa gen returned empty")
+        return
+    report(f"  vqa: {answer}")
+
+    parsed = parse_vqa_answer(answer)
+    if not answer_has_overlay(parsed):
+        if parsed is None:
+            report("  [info] vqa answer is not JSON — no overlay")
+        return
+
+    # The answer carries a bounding box / point. Its pixel coordinates
+    # are camera-specific and the text answer doesn't say which camera,
+    # so ask the operator *now* — only when there is actually something
+    # to draw — which camera frame to render the overlay on.
+    cameras = available_cameras(observation)
+    if observation is None or not cameras:
+        report("  [info] no camera image — cannot draw overlay")
+        return
+    chosen = prompt_camera_choice(cameras, input_fn=input_fn, print_fn=print_fn)
+    if chosen is None:
+        report("  [info] overlay skipped — no camera selected")
+        return
+    try:
+        pil = observation_image_to_pil(observation[chosen])
+        overlay = draw_vqa_overlay(pil, parsed)
+        path = save_and_open_overlay(overlay)
+        report(f"  vqa overlay ({camera_short_name(chosen)}) saved: {path}")
+    except Exception as exc:  # noqa: BLE001
+        logger.warning("vqa overlay failed: %s", exc, exc_info=logger.isEnabledFor(logging.DEBUG))
+        report(f"  [warn] vqa overlay failed: {type(exc).__name__}: {exc}")