refactor(pi052): load native base checkpoint

Keep the parity tests compatible with the repository's current Ruff hooks and remove the requested migration warning from the docs.

Co-authored-by: Cursor <cursoragent@cursor.com>

AGENT_GUIDE.md

@@ -61,20 +61,16 @@ Full details in [`docs/source/so101.mdx`](./docs/source/so101.mdx) and [`docs/so
 **4.1 Install**
 
 ```bash
-# uv (recommended — see AGENTS.md and CLAUDE.md)
+pip install 'lerobot[feetech]'              # SO-100/SO-101 motor stack
-uv sync --locked --extra feetech          # SO-100/SO-101 motor stack
+# pip install 'lerobot[all]'                # everything
-# uv sync --locked --extra all            # everything
+# pip install 'lerobot[aloha,pusht]'        # specific features
-# uv sync --locked --extra smolvla        # add SmolVLA deps
+# pip install 'lerobot[smolvla]'            # add SmolVLA deps
-
-# pip (alternative, e.g. when not working from source)
-# pip install 'lerobot[feetech]'
-# pip install 'lerobot[all]'
-# pip install 'lerobot[smolvla]'
-
 git lfs install && git lfs pull
-hf auth login                             # required to push datasets/policies
+hf auth login                               # required to push datasets/policies
 ```
 
+Contributors can alternatively use `uv sync --locked --extra feetech` (see `AGENTS.md`).
+
 **4.2 Find USB ports** — run once per arm, unplug when prompted.
 
 ```bash

README.md

@@ -101,13 +101,13 @@ lerobot-train \
   --dataset.repo_id=lerobot/aloha_mobile_cabinet
 ```
 
-| Category                   | Models                                                                                                                                                                                                                                                                                                                                                                                     |
+| Category                   | Models                                                                                                                                                                                                                                                                                                                                                                                                                       |
-| -------------------------- | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ |
+| -------------------------- | ---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
-| **Imitation Learning**     | [ACT](./docs/source/policy_act_README.md), [Diffusion](./docs/source/policy_diffusion_README.md), [VQ-BeT](./docs/source/policy_vqbet_README.md), [Multitask DiT Policy](./docs/source/policy_multi_task_dit_README.md)                                                                                                                                                                    |
+| **Imitation Learning**     | [ACT](./docs/source/policy_act_README.md), [Diffusion](./docs/source/policy_diffusion_README.md), [VQ-BeT](./docs/source/policy_vqbet_README.md), [Multitask DiT Policy](./docs/source/policy_multi_task_dit_README.md)                                                                                                                                                                                                      |
-| **Reinforcement Learning** | [HIL-SERL](./docs/source/hilserl.mdx), [TDMPC](./docs/source/policy_tdmpc_README.md) & QC-FQL (coming soon)                                                                                                                                                                                                                                                                                |
+| **Reinforcement Learning** | [HIL-SERL](./docs/source/hilserl.mdx), [TDMPC](./docs/source/policy_tdmpc_README.md) & QC-FQL (coming soon)                                                                                                                                                                                                                                                                                                                  |
-| **VLAs Models**            | [Pi0](./docs/source/pi0.mdx), [Pi0Fast](./docs/source/pi0fast.mdx), [Pi0.5](./docs/source/pi05.mdx), [GR00T N1.7](./docs/source/policy_groot_README.md), [SmolVLA](./docs/source/policy_smolvla_README.md), [XVLA](./docs/source/xvla.mdx), [EO-1](./docs/source/eo1.mdx), [MolmoAct2](./docs/source/molmoact2.mdx), [WALL-OSS](./docs/source/walloss.mdx), [EVO1](./docs/source/evo1.mdx) |
+| **VLAs Models**            | [Pi0](./docs/source/pi0.mdx), [Pi0Fast](./docs/source/pi0fast.mdx), [Pi0.5](./docs/source/pi05.mdx), [Pi052](./docs/source/pi052.mdx), [GR00T N1.7](./docs/source/policy_groot_README.md), [SmolVLA](./docs/source/policy_smolvla_README.md), [XVLA](./docs/source/xvla.mdx), [EO-1](./docs/source/eo1.mdx), [MolmoAct2](./docs/source/molmoact2.mdx), [WALL-OSS](./docs/source/walloss.mdx), [EVO1](./docs/source/evo1.mdx) |
-| **World Models**           | [VLA-JEPA](./docs/source/vla_jepa.mdx), [LingBot-VA](./docs/source/lingbot_va.mdx), [FastWAM](./docs/source/fastwam.mdx)                                                                                                                                                                                                                                                                   |
+| **World Models**           | [VLA-JEPA](./docs/source/vla_jepa.mdx), [LingBot-VA](./docs/source/lingbot_va.mdx), [FastWAM](./docs/source/fastwam.mdx)                                                                                                                                                                                                                                                                                                     |
-| **Reward Models**          | [SARM](./docs/source/sarm.mdx), [TOPReward](./docs/source/topreward.mdx), [Robometer](./docs/source/robometer.mdx)                                                                                                                                                                                                                                                                         |
+| **Reward Models**          | [SARM](./docs/source/sarm.mdx), [TOPReward](./docs/source/topreward.mdx), [Robometer](./docs/source/robometer.mdx)                                                                                                                                                                                                                                                                                                           |
 
 Similarly to the hardware, you can easily implement your own policy & leverage LeRobot's data collection, training, and visualization tools, and share your model to the HF Hub.
 

docker/Dockerfile.internal

@@ -68,16 +68,17 @@ ENV HOME=/home/user_lerobot \
 # issues with MuJoCo and OpenGL drivers.
 RUN uv venv --python python${PYTHON_VERSION}
 
-# Install third-party dependencies separately for layer caching
+# Install Python dependencies for caching
 COPY --chown=user_lerobot:user_lerobot setup.py pyproject.toml uv.lock README.md MANIFEST.in ./
-RUN uv sync --locked --extra all --no-install-project --no-cache
+COPY --chown=user_lerobot:user_lerobot src/ src/
+
+RUN uv sync --locked --extra all --no-cache
 
 RUN chmod +x /lerobot/.venv/lib/python${PYTHON_VERSION}/site-packages/triton/backends/nvidia/bin/ptxas
 
-# Copy the application source code and install the local project
+# Copy the rest of the application source code
 # Make sure to have the git-LFS files for testing
 COPY --chown=user_lerobot:user_lerobot . .
-RUN uv sync --locked --extra all --no-cache
 
 # Set the default command
 CMD ["/bin/bash"]

docker/Dockerfile.user

@@ -60,14 +60,15 @@ ENV HOME=/home/user_lerobot \
 # run other Python projects in the same container without dependency conflicts.
 RUN uv venv
 
-# Install third-party dependencies separately for layer caching
+# Install Python dependencies for caching
 COPY --chown=user_lerobot:user_lerobot setup.py pyproject.toml uv.lock README.md MANIFEST.in ./
-RUN uv sync --locked --extra all --no-install-project --no-cache
+COPY --chown=user_lerobot:user_lerobot src/ src/
 
-# Copy the application code and install the local project
+RUN uv sync --locked --extra all --no-cache
+
+# Copy the rest of the application code
 # Make sure to have the git-LFS files for testing
 COPY --chown=user_lerobot:user_lerobot . .
-RUN uv sync --locked --extra all --no-cache
 
 # Set the default command
 CMD ["/bin/bash"]

docs/source/_toctree.yml

@@ -63,6 +63,8 @@
     title: π₀-FAST (Pi0Fast)
   - local: pi05
     title: π₀.₅ (Pi05)
+  - local: pi052
+    title: π₀.₅ with language supervision (Pi052)
   - local: molmoact2
     title: MolmoAct2
   - local: vla_jepa

docs/source/backwardcomp.mdx

@@ -58,7 +58,7 @@ final_action = postprocessor(action)
 
 ## Hardware API redesign
 
-PR [#777](https://github.com/huggingface/lerobot/pull/777) improves the LeRobot calibration but is **not backward-compatible**. Below is an overview of what changed and how you can continue to work with datasets created before this pull request.
+PR [#777](https://github.com/huggingface/lerobot/pull/777) improves the LeRobot calibration but is **not backward-compatible**. Below is a overview of what changed and how you can continue to work with datasets created before this pull request.
 
 ### What changed?
 
@@ -129,8 +129,8 @@ python examples/backward_compatibility/replay.py \
 
 Policies output actions in the same format as the datasets (`torch.Tensors`). Therefore, the same transformations should be applied.
 
-To find these transformations, we recommend first replaying an episode of the dataset your policy was trained on using the section above.
+To find these transformations, we recommend to first try and and replay an episode of the dataset your policy was trained on using the section above.
-Then, add these same transformations to your inference script (shown here in the `record.py` script):
+Then, add these same transformations on your inference script (shown here in the `record.py` script):
 
 ```diff
 action_values = predict_action(

docs/source/bring_your_own_policies.mdx

@@ -191,6 +191,162 @@ def make_my_policy_pre_post_processors(
 
 ---
 
+## Adding high- and low-level language control
+
+The policy API above is sufficient for training and standard evaluation. To use a language-conditioned policy with interactive `lerobot-rollout`, also register a runtime adapter. The adapter keeps policy-specific prompting and tokenization out of the generic control loop.
+
+The runtime supports two policy shapes:
+
+| Policy shape     | Behavior                                                                | Adapter                                        |
+| ---------------- | ----------------------------------------------------------------------- | ---------------------------------------------- |
+| Low-level / flat | The operator's task or subtask directly conditions action prediction.   | Reuse `DirectTaskPolicyAdapter`.               |
+| High + low level | The policy generates subtasks or memory, then conditions actions on it. | Subclass `BaseLanguageAdapter`, as PI052 does. |
+
+During a rollout, `RuntimeState` stores the high-level task and the active language context:
+
+```text
+task ──> adapter.generate_text("subtask", ...) ──> state.language_context["subtask"]
+                                                       │
+observation ──> processors ──> adapter.select_action() ─┴─> action chunk ──> robot
+```
+
+The generic runtime handles generation frequency, pause/resume, prompt replacement, action queues, and dispatch. The adapter only translates between that runtime contract and your policy.
+
+### Low-level policies
+
+If your policy already consumes the live task through its normal preprocessor and implements `predict_action_chunk`, register the shared direct adapter. PI0.5 and MolmoAct2 use this path:
+
+```python
+# src/lerobot/runtime/registry.py
+_ADAPTERS = {
+    # ...
+    "my_policy": "lerobot.runtime.adapter:DirectTaskPolicyAdapter",
+}
+```
+
+Run it with direct-subtask mode so the operator supplies the instruction used by the action policy:
+
+```bash
+lerobot-rollout \
+    --language \
+    --policy.path=user/my_policy_checkpoint \
+    --robot.type=so101_follower \
+    --robot.port=/dev/ttyACM0 \
+    --direct_subtask
+```
+
+The rollout context builds the observation batch with the current instruction before `DirectTaskPolicyAdapter` calls `policy.predict_action_chunk(observation)`. No text-generation method is required.
+
+### Hierarchical policies
+
+For a policy that generates language and actions, subclass [`BaseLanguageAdapter`](https://github.com/huggingface/lerobot/blob/main/src/lerobot/runtime/adapter.py) and implement two methods:
+
+- `generate_text(kind, observation, state, user_text=None) -> str` generates a `subtask`, `memory`, or interjection response.
+- `select_action(observation, state)` builds the low-level prompt from the active context and returns an action chunk.
+
+This abbreviated adapter follows [`PI052PolicyAdapter`](https://github.com/huggingface/lerobot/blob/main/src/lerobot/policies/pi052/inference/pi052_adapter.py):
+
+```python
+# inference/my_policy_adapter.py
+from typing import Any
+
+from lerobot.runtime import RuntimeState
+from lerobot.runtime.adapter import BaseLanguageAdapter
+from lerobot.utils.constants import (
+    OBS_LANGUAGE_ATTENTION_MASK,
+    OBS_LANGUAGE_TOKENS,
+)
+
+
+class MyPolicyAdapter(BaseLanguageAdapter):
+    def select_action(self, observation: dict[str, Any], state: RuntimeState):
+        instruction = state.language_context.get("subtask") or state.task or ""
+        tokens, attention_mask = tokenize_instruction(instruction)
+
+        batch = dict(observation)
+        batch[OBS_LANGUAGE_TOKENS] = tokens
+        batch[OBS_LANGUAGE_ATTENTION_MASK] = attention_mask
+        return self.policy.predict_action_chunk(batch)
+
+    def generate_text(
+        self,
+        kind: str,
+        observation: dict[str, Any] | None,
+        state: RuntimeState,
+        user_text: str | None = None,
+    ) -> str:
+        messages = self.build_messages(kind, state, user_text)
+        batch, tokenizer = tokenize_messages(messages, observation)
+        return self.policy.select_message(
+            batch,
+            tokenizer=tokenizer,
+            min_new_tokens=self.gen.min_new_tokens,
+            temperature=self.gen.temperature,
+            top_p=self.gen.top_p,
+        )
+
+    def build_messages(
+        self, kind: str, state: RuntimeState, user_text: str | None
+    ) -> list[dict[str, str]]:
+        if kind == "subtask":
+            return [{"role": "user", "content": state.task or ""}]
+        if kind == "memory":
+            return [
+                {"role": "user", "content": state.task or ""},
+                {
+                    "role": "user",
+                    "content": f"Completed subtask: {state.extra.get('prior_subtask', '')}",
+                },
+            ]
+        if kind == "interjection":
+            return [
+                {"role": "user", "content": state.task or ""},
+                {"role": "user", "content": user_text or ""},
+            ]
+        raise ValueError(f"Unsupported text kind: {kind}")
+```
+
+`tokenize_instruction` and `tokenize_messages` are policy-specific helpers. They must reproduce the prompt format used during training; PI052, for example, adds the discretized robot state to its low-level subtask prompt and uses the same PaliGemma formatting for `select_message`.
+
+`BaseLanguageAdapter` provides the default hierarchy: regenerate a subtask at action-chunk boundaries, update memory when the subtask changes, and handle user interjections. Override `_regenerate_context` only if your policy uses a different hierarchy.
+
+Register the adapter with a lazy import so importing LeRobot does not load the model or its optional dependencies:
+
+```python
+# src/lerobot/runtime/registry.py
+_ADAPTERS = {
+    # ...
+    "my_policy": "lerobot.policies.my_policy.inference.my_policy_adapter:MyPolicyAdapter",
+}
+```
+
+The key must match the policy's registered type. Once registered, the same checkpoint works through the shared entry point:
+
+```bash
+lerobot-rollout \
+    --language \
+    --policy.path=user/my_hierarchical_checkpoint \
+    --robot.type=so101_follower \
+    --robot.port=/dev/ttyACM0 \
+    --task="put the cup in the sink"
+```
+
+For RoboCasa-compatible policies, replace the robot arguments with `--sim --sim.task=<task>`. Without `--direct_subtask`, the adapter generates the low-level subtask; with it, the operator bypasses high-level generation and supplies each subtask.
+
+### Keep training and deployment aligned
+
+The adapter is intentionally small, but its prompts are part of the model contract:
+
+- Use the same tokenizer, role formatting, special tokens, image ordering, and state encoding as training.
+- Condition `select_action` on `state.language_context["subtask"]`, falling back to `state.task` for direct or not-yet-generated prompts.
+- Return a full action chunk from `select_action`; the runtime handles control-rate dispatch.
+- Keep optional model dependencies inside lazy imports.
+- Test adapter selection, generated-message routing, action-batch construction, and direct-subtask behavior with a lightweight fake policy.
+
+PI052 is the complete in-tree reference: its [processor](https://github.com/huggingface/lerobot/blob/main/src/lerobot/policies/pi052/processor_pi052.py) renders the training recipe, its [policy](https://github.com/huggingface/lerobot/blob/main/src/lerobot/policies/pi052/modeling_pi052.py) exposes text and action generation, and its [adapter](https://github.com/huggingface/lerobot/blob/main/src/lerobot/policies/pi052/inference/pi052_adapter.py) reconstructs those same prompts at deployment.
+
+---
+
 ## Path A: Out-of-tree plugin
 
 The fastest way to ship a policy: package it as a standalone Python distribution and install it alongside LeRobot. No PR required, you own the release cycle, and you can publish to PyPI under your own namespace.

docs/source/cameras.mdx

@@ -136,10 +136,6 @@ config = RealSenseCameraConfig(
     height=480,
     color_mode=ColorMode.RGB,
     use_depth=True,
-    # Optional fixed color controls. Omit them to leave the current sensor settings unchanged.
-    exposure=120,
-    gain=64,
-    white_balance=4600,
     rotation=Cv2Rotation.NO_ROTATION
 )
 
@@ -158,15 +154,6 @@ finally:
 ```
 <!-- prettier-ignore-end -->
 
-Manual color controls disable the corresponding automatic exposure or white-balance mode. Their
-supported ranges vary by camera model; an invalid value raises an error at connection time that
-includes the range reported by the sensor. Requesting an unsupported control also raises an error.
-Omitted controls leave the sensor's existing automatic or manual setting unchanged. These options
-require `use_rgb=True`.
-
-On the RealSense D405, the color stream is provided by the Stereo Module, so changing manual
-exposure or gain also affects the depth stream.
-
 </hfoption>
 </hfoptions>
 

docs/source/env_processor.mdx

@@ -88,6 +88,20 @@ policy_preprocessor = NormalizerProcessorStep(stats=dataset_stats)
 
 The same policy can work with different environment processors, and the same environment processor can work with different policies:
 
+````python
+# Use SmolVLA policy with LIBERO environment
+# Use SmolVLA policy with LIBERO environment
+libero_preprocessor, libero_postprocessor = make_env_pre_post_processors(
+    env_cfg=libero_cfg,
+    policy_cfg=smolvla_cfg,
+)
+smolvla_preprocessor, smolvla_postprocessor = make_pre_post_processors(smolvla_cfg)
+# Or use ACT policy with the same LIBERO environment
+libero_preprocessor, libero_postprocessor = make_env_pre_post_processors(
+    env_cfg=libero_cfg,
+    policy_cfg=act_cfg,
+)
+act_preprocessor, act_postprocessor = make_pre_post_processors(act_cfg)
 ```python
 # Use SmolVLA policy with LIBERO environment
 libero_preprocessor, libero_postprocessor = make_env_pre_post_processors(
@@ -102,7 +116,6 @@ libero_preprocessor, libero_postprocessor = make_env_pre_post_processors(
     policy_cfg=act_cfg,
 )
 act_preprocessor, act_postprocessor = make_pre_post_processors(act_cfg)
-```
 
 ### 3. **Easier Experimentation**
 
@@ -132,7 +145,7 @@ class LiberoVelocityProcessorStep(ObservationProcessorStep):
         state = torch.cat([eef_pos, eef_axisangle, eef_vel,
                           gripper_pos, gripper_vel], dim=-1)  # 14D
         return state
-```
+````
 
 ### 4. **Cleaner Environment Code**
 

docs/source/feetech.mdx

@@ -40,10 +40,10 @@ This tutorial guides you through updating the firmware of Feetech motors using t
 For each motor you want to update:
 
 1. **Select the motor** from the list by clicking on it
-2. **Click the Upgrade tab**:
+2. **Click on Upgrade tab**:
-3. **Click the Online button**:
+3. **Click on Online button**:
-   - If a potential firmware update is found, it will be displayed in the box
+   - If an potential firmware update is found, it will be displayed in the box
-4. **Click the Upgrade button**:
+4. **Click on Upgrade button**:
    - The update progress will be displayed
 
 ## Step 6: Verify Update

docs/source/hope_jr.mdx

@@ -211,7 +211,7 @@ Record, Replay and Train with Hope-JR is still experimental.
 
 ### Record
 
-This step records the dataset, which can be seen as an example [here](https://huggingface.co/datasets/nepyope/hand_record_test_with_video_data).
+This step records the dataset, which can be seen as an example [here](https://huggingface.co/datasets/nepyope/hand_record_test_with_video_data/settings).
 
 ```bash
 lerobot-record \

docs/source/inference.mdx

@@ -1,6 +1,6 @@
 # Policy Deployment (lerobot-rollout)
 
-`lerobot-rollout` is the single CLI for deploying trained policies on real robots. It supports multiple execution strategies and inference backends, from quick evaluation to continuous recording and human-in-the-loop data collection.
+`lerobot-rollout` is the single CLI for deploying trained policies on real robots or in an interactive simulator. It supports multiple execution strategies and inference backends, from quick evaluation to continuous recording, language-driven control, and human-in-the-loop data collection.
 
 ## Quick Start
 
@@ -197,6 +197,52 @@ Teleop is optional — if omitted the robot holds its position during the reset
 
 ---
 
+## Interactive language control
+
+Language-conditioned policies can expose a high-level text head in addition to
+their action head. Add `--language` to open-prompt one of these policies on a
+real robot. Language-only flags such as `--direct_subtask` select this mode
+automatically.
+
+MolmoAct2 has no high-level planner, so use direct-subtask mode and type each
+next low-level instruction yourself:
+
+```bash
+lerobot-rollout \
+    --policy.path=lerobot/MolmoAct2-SO100_101-LeRobot \
+    --policy.device=cuda \
+    --robot.type=so101_follower \
+    --robot.port=/dev/ttyACM1 \
+    --robot.cameras='{"cam0":{"type":"opencv","index_or_path":"/dev/video0","width":640,"height":480,"fps":30,"fourcc":"MJPG","backend":200},"cam1":{"type":"opencv","index_or_path":"/dev/video2","width":640,"height":480,"fps":30,"fourcc":"MJPG","backend":200}}' \
+    --direct_subtask \
+    --robot.max_relative_target='{"shoulder_pan":5,"shoulder_lift":5,"elbow_flex":5,"wrist_flex":5,"wrist_roll":5,"gripper":5}'
+```
+
+The robot starts paused. Type a subtask, then use `/resume` and `/pause` to
+control action dispatch. Check the workspace and motion limits before resuming.
+Without `--direct_subtask`, a policy such as PI052 generates its active subtask
+from the high-level `--task` itself.
+
+RoboCasa uses the same runtime and processor path. `--sim` selects it
+automatically, so no robot configuration is needed:
+
+```bash
+MUJOCO_GL=egl lerobot-rollout \
+    --policy.path=lerobot/pi052_robocasa \
+    --sim --sim.task=CloseFridge --sim.split=pretrain \
+    --task="close the fridge" \
+    --disable_memory \
+    --sim.render_size=384 \
+    --sim.views=robot0_agentview_left,robot0_eye_in_hand,robot0_agentview_right \
+    --mode=action --ctrl_hz=20
+```
+
+Open `http://localhost:8010` for the live simulator view. Add
+`--sim.direct_subtask` to bypass the language planner and make each typed prompt
+the action policy's current subtask.
+
+---
+
 ## Inference Backends
 
 Select a backend with `--inference.type=<name>`. All strategies work with both backends.

docs/source/integrate_hardware.mdx

@@ -18,7 +18,7 @@ If you're using Feetech or Dynamixel motors, LeRobot provides built-in bus inter
 - [`DynamixelMotorsBus`](https://github.com/huggingface/lerobot/blob/main/src/lerobot/motors/dynamixel/dynamixel.py) – for controlling Dynamixel servos
 
 Please refer to the [`MotorsBus`](https://github.com/huggingface/lerobot/blob/main/src/lerobot/motors/motors_bus.py) abstract class to learn about its API.
-For a good example of how it can be used, you can have a look at our own [SO101 follower implementation](https://github.com/huggingface/lerobot/blob/main/src/lerobot/robots/so_follower/so_follower.py)
+For a good example of how it can be used, you can have a look at our own [SO101 follower implementation](https://github.com/huggingface/lerobot/blob/main/src/lerobot/robots/so_follower/so101_follower/so101_follower.py)
 
 Use these if compatible. Otherwise, you'll need to find or write a Python interface (not covered in this tutorial):
 

docs/source/language_and_recipes.mdx

@@ -141,6 +141,17 @@ 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/subtask_mem.yaml` trains the compact core blend — high-level subtask prediction, low-level execution, and memory. `recipes/subtask_mem_vqa_speech.yaml` is the fuller variant that also adds VQA and spoken interjection responses.
+
+A message recipe with a supervised assistant turn on the `low_level` stream trains
+the π0.5 paper's joint sequence instead of a blend: the target span gets text CE
+while also conditioning the action losses in the same forward.
+`recipes/subtask_joint.yaml` is the provided example; pair it with
+`--policy.joint_subtask_conditioning=true` at inference.
+
 ## Graceful absence
 
 If both language columns are missing, `None`, or empty, `RenderMessagesStep` is a no-op.

docs/source/lekiwi.mdx

@@ -51,7 +51,7 @@ In addition to these instructions, you need to install the Feetech SDK & ZeroMQ
 pip install -e ".[lekiwi]"
 ```
 
-Great 🤗! You are now done installing LeRobot, and we can begin assembling the SO100/SO101 arms and the mobile base 🤖.
+Great :hugs:! You are now done installing LeRobot, and we can begin assembling the SO100/SO101 arms and the mobile base :robot:.
 Every time you now want to use LeRobot, you can go to the `~/lerobot` folder where we installed LeRobot and run one of the commands.
 
 # Step-by-Step Assembly Instructions

docs/source/pi052.mdx

@@ -0,0 +1,274 @@
+# π₀.₅ with language supervision (Pi052)
+
+Pi052 extends [Pi05](./pi05) with a trainable PaliGemma language head and a
+runtime that alternates language generation with action generation. A single
+checkpoint can predict a low-level subtask, optionally update memory or answer
+visual questions, and condition its flow-matching action expert on that text.
+
+Use Pi05 when you only need task-conditioned actions. Use Pi052 when the policy
+must generate or consume intermediate language during a rollout.
+
+## How Pi052 differs from Pi05
+
+| Capability          | Pi05                                                   | Pi052                                                                             |
+| ------------------- | ------------------------------------------------------ | --------------------------------------------------------------------------------- |
+| Action model        | PaliGemma vision-language prefix + Gemma action expert | Same base architecture                                                            |
+| Language head       | Not trained for runtime generation                     | Re-enabled and trained with text cross-entropy                                    |
+| Action conditioning | Episode task                                           | Active low-level subtask plus normalized robot state                              |
+| Training targets    | Flow-matching actions                                  | Flow actions, recipe-selected text, and optional FAST action tokens               |
+| Dataset requirement | Standard images, state, actions, and task              | The same fields plus language annotations for every language capability you train |
+| Rollout             | Direct task-to-action policy                           | Hierarchical task → subtask → action loop, with optional memory and VQA           |
+
+Pi052 can initialize from a Pi05 checkpoint. The policy architecture remains
+compatible, while Pi052 builds its own processors so recipe labels and FAST
+labels are not silently replaced by the Pi05 processor stack.
+
+## Install
+
+Install LeRobot with the PI dependencies:
+
+```bash
+git clone https://github.com/huggingface/lerobot.git
+cd lerobot
+python -m venv .venv
+source .venv/bin/activate
+pip install -e ".[pi]"
+```
+
+The `pi` extra includes the PaliGemma/FAST dependencies. Install
+`liger-kernel` for the supported fused training kernels; optional FlashRT
+backends also require the Hugging Face `kernels` package and a supported CUDA
+GPU.
+
+## Prepare language-annotated data
+
+Pi052 does not infer supervised subtasks from a normal LeRobot dataset during
+training. The dataset must contain the language targets used by the selected
+recipe in the optional `language_persistent` and `language_events` columns.
+
+At minimum, annotate a continuous `subtask` timeline so each training frame has
+an active low-level instruction. Add `memory`, VQA, interjections, and speech
+annotations only if the recipe trains those capabilities.
+
+The provided recipes are:
+
+| Recipe                                | Required annotations                                                    | Trains                                                             |
+| ------------------------------------- | ----------------------------------------------------------------------- | ------------------------------------------------------------------ |
+| `recipes/subtask.yaml`                | `subtask`                                                               | Subtask prediction and subtask-conditioned actions                 |
+| `recipes/subtask_joint.yaml`          | `subtask`                                                               | Paper-style joint sequence: subtask text and actions in one sample |
+| `recipes/subtask_mem.yaml`            | `subtask`, `memory`                                                     | Subtasks, actions, and memory updates                              |
+| `recipes/subtask_mem_vqa_speech.yaml` | `subtask`, `memory`, `vqa`; interjection/speech rows for those branches | Subtasks, actions, memory, VQA, and spoken replies                 |
+
+The blend recipes factorize training into separate high-level (task → subtask)
+and low-level (subtask → actions) samples, matching how inference decomposes
+π(a|o, subtask)·π(subtask|o, task). `recipes/subtask_joint.yaml` instead uses
+the π0.5 paper's single-sequence layout — the supervised subtask span is
+attended causally and conditions the FAST and flow losses in the same forward.
+Checkpoints trained with the joint recipe must set
+`--policy.joint_subtask_conditioning=true` at inference so the flow prefix
+rebuilds the same layout (task turn with state, then the generated subtask as a
+causal assistant turn); leave it `false` for the blend recipes.
+
+Use `lerobot-annotate` to generate these columns. The repository includes a
+Hugging Face Jobs launcher that you can edit for your source and destination
+datasets. For a local annotation run, first install
+`pip install -e ".[annotations]"`:
+
+```bash
+HF_TOKEN=hf_... uv run python examples/annotations/run_hf_job.py
+```
+
+Before a long training run, inspect several episodes and verify that subtasks
+are temporally correct and cover the full demonstration. See
+[Annotation Pipeline](./annotation_pipeline) for generation and validation, and
+[Language Columns and Recipes](./language_and_recipes) for the schema and
+recipe resolver.
+
+<Tip>
+  If a dataset has no language columns, recipe rendering becomes a no-op and
+  Pi052 falls back to the plain Pi05 prompt path. This is useful for
+  compatibility but does not train the language planner.
+</Tip>
+
+## Train Pi052
+
+This example initializes Pi052 from the native Pi052 initialization checkpoint
+and trains the default subtask-and-memory recipe:
+
+```bash
+lerobot-train \
+    --dataset.repo_id=${HF_USER}/my_language_annotated_dataset \
+    --policy.type=pi052 \
+    --policy.pretrained_path=lerobot/pi052_base \
+    --policy.recipe_path=recipes/subtask_mem.yaml \
+    --policy.dtype=bfloat16 \
+    --policy.device=cuda \
+    --policy.freeze_vision_encoder=false \
+    --policy.gradient_checkpointing=true \
+    --batch_size=8 \
+    --steps=30000 \
+    --output_dir=outputs/pi052 \
+    --job_name=pi052 \
+    --wandb.enable=true
+```
+
+For subtask-only data, change the recipe to `recipes/subtask.yaml` and disable
+memory during rollout. Start with a small run and confirm that W&B examples show
+the expected prompt, text target, and action endpoints before scaling up.
+
+### Main training controls
+
+| Option                              |                    Default | Purpose                                                             |
+| ----------------------------------- | -------------------------: | ------------------------------------------------------------------- |
+| `policy.recipe_path`                | `recipes/subtask_mem.yaml` | Selects the language/action objective mixture                       |
+| `policy.text_loss_weight`           |                      `1.0` | Language-head cross-entropy weight; `0` disables text training      |
+| `policy.flow_loss_weight`           |                     `10.0` | Continuous action flow-loss weight                                  |
+| `policy.enable_fast_action_loss`    |                     `true` | Adds discrete FAST action-token supervision                         |
+| `policy.fast_action_loss_weight`    |                      `1.0` | FAST cross-entropy weight                                           |
+| `policy.knowledge_insulation`       |                     `true` | Blocks action-loss gradients through the VLM K/V path               |
+| `policy.flow_num_repeats`           |                        `5` | Reuses one VLM prefix for independent denoising targets             |
+| `policy.lm_head_lr_scale`           |                      `1.0` | Scales language-head learning rate; `1.0` uses the base rate        |
+| `policy.fast_skip_tokens`           |                     `1152` | FAST id offset; skips `<seg>`+`<loc>` so VQA and FAST never collide |
+| `policy.joint_subtask_conditioning` |                    `false` | Rebuilds the joint-sequence prefix at inference (see recipes)       |
+
+`fast_skip_tokens=1152` places FAST codes below PaliGemma's `<loc>` range.
+openpi's pi0-FAST convention is `128` (FAST occupies the `<loc>` ids); use that
+value only to stay weight-compatible with checkpoints trained that way, and
+avoid combining it with the VQA recipe, whose `<loc>` targets would share
+embedding rows with FAST codes.
+
+The loss weights are starting points, not dataset-independent constants. Track
+flow loss and text/FAST losses separately, and inspect generated subtasks rather
+than selecting a checkpoint from total loss alone.
+
+### Dataset-specific FAST tokenizer
+
+The universal FAST tokenizer works out of the box. For a large or
+embodiment-specific dataset, Pi052 can fit and cache a tokenizer on normalized
+actions before training:
+
+```bash
+lerobot-train \
+    ... \
+    --policy.auto_fit_fast_tokenizer=true \
+    --policy.fast_tokenizer_fit_samples=4096
+```
+
+The fit runs once per dataset/tokenizer configuration. Keep
+`auto_fit_fast_tokenizer=false` when you do not want the extra preprocessing
+pass.
+
+## Training performance
+
+Pi052 uses optimized training paths by default:
+
+- batches repeated flow targets and suffix projections instead of replaying
+  small operations in Python;
+- caches constant action masks and computes RoPE positions once per forward;
+- selects the text/FAST cross-entropy implementation from target shape and
+  sparsity;
+- skips the mathematically dead VLM/vision backward on knowledge-insulated,
+  flow-only batches;
+- uses native non-reentrant SigLIP layer checkpointing when gradient
+  checkpointing is enabled; and
+- retains the Liger RoPE/GeGLU kernels while avoiding the slower LayerNorm
+  patch at SigLIP shapes.
+
+Optional training backends are disabled by default:
+
+| Option                                 | When to try it                                                                                    |
+| -------------------------------------- | ------------------------------------------------------------------------------------------------- |
+| `policy.use_flashrt_adarms=true`       | Fused adaptive RMSNorm and gated residuals on supported CUDA GPUs                                 |
+| `policy.use_compiled_text_ce=true`     | Compiled materialized-logit CE buckets                                                            |
+| `policy.use_compiled_vision=true`      | Compiled vision only when the vision pass has no gradients                                        |
+| `policy.use_flex_attention=true`       | Profiled CUDA setups with knowledge insulation and `flow_num_repeats > 1`; otherwise SDPA is used |
+| `policy.use_manual_attention=true`     | Explicitly profiled shapes where materialized attention is faster                                 |
+| `policy.manual_attention_scope=action` | Restricts manual attention to action queries                                                      |
+
+Do not enable every backend blindly. Flex and manual attention are mutually
+exclusive, and attention/AdaRMS alternatives require knowledge insulation.
+The benchmark-best configuration used compiled text CE and FlashRT AdaRMS,
+with Flex/manual attention and compiled vision disabled.
+
+### Reported training benchmarks
+
+These benchmarks measure complete optimizer steps with three real camera
+inputs, BF16 transformer/action execution, FP32 vision, fused AdamW, and no
+video decoding or network I/O. Results vary with GPU, batch shape, annotation
+mixture, and checkpointing:
+
+| Workload                   |     RTX PRO 6000 Blackwell |                 A100 80 GB |
+| -------------------------- | -------------------------: | -------------------------: |
+| Full flow + text, batch 1  | 4.75× vs checkpointing off | 3.33× vs checkpointing off |
+| Full flow + text, batch 8  | 2.16× vs checkpointing off | 1.66× vs checkpointing off |
+| Full flow + text, batch 64 |  1.24× vs checkpointing on |  1.15× vs checkpointing on |
+| Flow-only, batch 1         | 3.70× vs checkpointing off | 3.58× vs checkpointing off |
+| Flow-only, batch 64        |  3.76× vs checkpointing on |  3.61× vs checkpointing on |
+
+On those 80 GB GPUs, full training was fastest without gradient checkpointing
+through batch 8, then required checkpointing at batch 16 and above. Treat that
+as a tuning rule to test on your hardware, not a universal threshold. Flow-only
+means both text and FAST supervision are disabled; it is useful for action-only
+ablation or post-training but does not learn the language runtime.
+
+## Inference performance
+
+Pi052 has two inference loops, and both avoid repeatedly encoding the expensive
+multimodal prefix:
+
+1. **Action denoising** encodes the image/language prefix once, reuses its KV
+   cache across flow steps, precomputes the timestep schedule on-device, and
+   crops temporary suffix K/V instead of cloning the prefix cache.
+2. **Language decoding** uses autoregressive KV caching, so each new token only
+   processes the sampled token against cached image/language keys instead of
+   rerunning the full prefix.
+
+The runtime also runs language and actions at different rates. Increase
+`--subtask_chunks_per_gen` when a subtask remains valid across several action
+chunks, lower `--high_level_hz`, or use `--direct_subtask` to bypass language
+generation entirely. These settings reduce compute but also slow replanning.
+
+`--fp8` enables the optional FlashRT inference MLP swap on supported CUDA GPUs.
+It calibrates on the first observation and falls back to BF16 when unavailable;
+because FP8 can change outputs slightly, validate task success before using it
+for production rollouts.
+
+## Run a checkpoint
+
+RoboCasa:
+
+```bash
+MUJOCO_GL=egl lerobot-rollout \
+    --policy.path=lerobot/pi052_robocasa \
+    --sim --sim.task=CloseFridge --sim.split=pretrain \
+    --task="close the fridge" \
+    --disable_memory \
+    --sim.render_size=384 \
+    --sim.views=robot0_agentview_left,robot0_eye_in_hand,robot0_agentview_right \
+    --mode=action --ctrl_hz=20
+```
+
+Open `http://localhost:8010` for the live view. Without
+`--sim.direct_subtask`, Pi052 generates the low-level subtask; with it, each
+prompt becomes the action policy's subtask directly.
+
+The same runtime supports real robots. See [Interactive language
+control](./inference#interactive-language-control) for the real-arm command,
+safety behavior, and runtime controls.
+
+## Troubleshooting
+
+- **No text loss or generated subtasks:** confirm the selected recipe can bind
+  the annotations on sampled frames and that `policy.text_loss_weight > 0`.
+- **Subtasks look plausible but actions fail:** verify subtask boundaries,
+  normalized state/action statistics, and that low-level recipe samples are
+  present.
+- **Text collapses to repeated or location tokens:** inspect text-target
+  coverage, language-head learning rate, and the balance between flow, FAST,
+  and text losses.
+- **Out of memory:** reduce batch size first, then enable gradient
+  checkpointing. Do not enable compiled or alternative attention backends
+  without profiling their memory on your camera count.
+- **Slow rollout:** separate action latency from language latency, then tune
+  `--subtask_chunks_per_gen`, `--high_level_hz`, and the number of flow
+  inference steps.

docs/source/pi0fast.mdx

@@ -109,15 +109,21 @@ lerobot-train \
 
 ### Key Training Parameters
 
-| Parameter                              | Description                                        | Default                         |
+| Parameter                               | Description                                        | Default                         |
-| -------------------------------------- | -------------------------------------------------- | ------------------------------- |
+| --------------------------------------- | -------------------------------------------------- | ------------------------------- |
-| `--policy.gradient_checkpointing=true` | Reduces memory usage significantly during training | `false`                         |
+| `--policy.gradient_checkpointing=true`  | Reduces memory usage significantly during training | `false`                         |
-| `--policy.dtype=bfloat16`              | Use mixed precision training for efficiency        | `float32`                       |
+| `--policy.dtype=bfloat16`               | Use mixed precision training for efficiency        | `float32`                       |
-| `--policy.chunk_size`                  | Number of action steps to predict (action horizon) | `50`                            |
+| `--policy.chunk_size`                   | Number of action steps to predict (action horizon) | `50`                            |
-| `--policy.n_action_steps`              | Number of action steps to execute                  | `50`                            |
+| `--policy.n_action_steps`               | Number of decoded action steps to execute          | `50`                            |
-| `--policy.max_action_tokens`           | Maximum number of FAST tokens per action chunk     | `256`                           |
+| `--policy.max_action_tokens`            | Maximum number of FAST tokens per action chunk     | `256`                           |
-| `--policy.action_tokenizer_name`       | FAST tokenizer to use                              | `lerobot/fast-action-tokenizer` |
+| `--policy.action_tokenizer_name`        | FAST tokenizer to use                              | `lerobot/fast-action-tokenizer` |
-| `--policy.compile_model=true`          | Enable torch.compile for faster training           | `false`                         |
+| `--policy.auto_fit_fast_tokenizer=true` | Fit and cache a tokenizer for the training dataset | `false`                         |
+| `--policy.compile_model=true`           | Enable torch.compile for faster training           | `false`                         |
+
+Set `--policy.auto_fit_fast_tokenizer=true` to sample action chunks from the
+training dataset and cache a fitted tokenizer under
+`~/.cache/lerobot/fast_tokenizers`. This also works when fine-tuning with
+`--policy.path`; leave it disabled to retain the checkpoint's tokenizer.
 
 ## Inference
 
@@ -151,6 +157,9 @@ actions = policy.predict_action_chunk(batch)
 
 The model takes images, text instructions, and robot state as input, and outputs discrete FAST tokens that are decoded back to continuous actions.
 
+PI0-FAST always decodes a complete `chunk_size` action chunk. `n_action_steps` controls only
+how many actions from that chunk are executed before the policy predicts again.
+
 ## Configuration Options
 
 | Parameter            | Description                                     | Default    |
@@ -174,7 +183,7 @@ The model takes images, text instructions, and robot state as input, and outputs
 
 ## Reproducing π₀Fast results
 
-We reproduce the results of π₀Fast on the LIBERO benchmark using the LeRobot implementation. We take the LeRobot PiFast base model [lerobot/pi0fast-base](https://huggingface.co/lerobot/pi0fast-base) and finetune for an additional 40k steps in bfloat16, with batch size of 256 on 8 H100 GPUs using the [HuggingFace LIBERO dataset](https://huggingface.co/datasets/HuggingFaceVLA/libero).
+We reproduce the results of π₀Fast on the LIBERO benchmark using the LeRobot implementation. We take the LeRobot PiFast base model [lerobot/pi0fast-base](https://huggingface.co/lerobot/pi0fast-base) and finetune for an additional 40kk steps in bfloat16, with batch size of 256 on 8 H100 GPUs using the [HuggingFace LIBERO dataset](https://huggingface.co/datasets/HuggingFaceVLA/libero).
 
 The finetuned model can be found here:
 

docs/source/processors_robots_teleop.mdx

@@ -22,7 +22,7 @@ With processors, you choose the learning features you want to use for your polic
 ## Three pipelines
 
 We often compose three pipelines. Depending on your setup, some can be empty if action and observation spaces already match.
-Each of these pipelines handles different conversions between different action and observation spaces. Below is a quick explanation of each pipeline.
+Each of these pipelines handle different conversions between different action and observation spaces. Below is a quick explanation of each pipeline.
 
 1. Pipeline 1: Teleop action space → dataset action space (phone pose → EE targets)
 2. Pipeline 2: Dataset action space → robot command space (EE targets → joints)
@@ -74,15 +74,15 @@ In the phone to SO-100 follower examples we use the following adapters:
 - `robot_action_to_transition`: transforms the teleop action dict to a pipeline transition.
 - `transition_to_robot_action`: transforms the pipeline transition to a robot action dict.
 - `observation_to_transition`: transforms the robot observation dict to a pipeline transition.
-- `transition_to_observation`: transforms the pipeline transition to an observation dict.
+- `transition_to_observation`: transforms the pipeline transition to a observation dict.
 
-Check out [src/lerobot/processor/converters.py](https://github.com/huggingface/lerobot/blob/main/src/lerobot/processor/converters.py) for more details.
+Checkout [src/lerobot/processor/converters.py](https://github.com/huggingface/lerobot/blob/main/src/lerobot/processor/converters.py) for more details.
 
 ## Dataset feature contracts
 
 Dataset features are determined by the keys saved in the dataset. Each step can declare what features it modifies in a contract called `transform_features(...)`. Once you build a processor, the processor can then aggregate all of these features with `aggregate_pipeline_dataset_features()` and merge multiple feature dicts with `combine_feature_dicts(...)`.
 
-Below is an example of how we declare features with the `transform_features` method in the phone to SO-100 follower examples:
+Below is and example of how we declare features with the `transform_features` method in the phone to SO-100 follower examples:
 
 ```python
     def transform_features(

docs/source/rtc.mdx

@@ -57,7 +57,7 @@ policy_cfg.rtc_config = RTCConfig(
 policy = PI0Policy.from_pretrained("lerobot/pi0_base", policy_cfg=policy_cfg, device="cuda")
 
 # Now use predict_action_chunk with RTC parameters
-inference_delay = 4  # How many steps of inference latency, this value should be calculated based on the inference latency of the policy
+inference_delay = 4  # How many steps of inference latency, this values should be calculated based on the inference latency of the policy
 
 # Initialize the action queue
 action_queue = ActionQueue(policy_cfg.rtc_config)
@@ -100,7 +100,7 @@ Typical values: 8-12 steps
 RTCConfig(execution_horizon=10)
 ```
 
-**`max_guidance_weight`**: How strongly to enforce consistency with the previous chunk. This is a hyperparameter that can be tuned to balance the smoothness of the transitions and the reactivity of the policy. For 10 steps flow matching (SmolVLA, Pi0, Pi0.5), a value of 10.0 is an optimal value.
+**`max_guidance_weight`**: How strongly to enforce consistency with the previous chunk. This is a hyperparameter that can be tuned to balance the smoothness of the transitions and the reactivity of the policy. For 10 steps flow matching (SmolVLA, Pi0, Pi0.5), a value of 10.0 is a optimal value.
 
 **`prefix_attention_schedule`**: How to weight consistency across the overlap region.
 

docs/source/smolvla.mdx

@@ -93,7 +93,7 @@ lerobot-train --help
 
 ## Evaluate the finetuned model and run it in real-time
 
-Similarly for when recording an episode, it is recommended that you are logged in to the HuggingFace Hub. You can follow the corresponding steps: [Record a dataset](./il_robots#record-a-dataset).
+Similarly for when recording an episode, it is recommended that you are logged in to the HuggingFace Hub. You can follow the corresponding steps: [Record a dataset](./il_robots).
 Once you are logged in, you can run inference in your setup by doing:
 
 ```bash

docs/source/unitree_g1.mdx

@@ -8,15 +8,6 @@
 
 The Unitree G1 humanoid is now supported in LeRobot! You can teleoperate, train locomanipulation policies, test in sim, and more. Both 29 and 23 DoF variants are supported.
 
-<Tip>
-  **New: SONIC whole-body control.** The `SonicWholeBodyController` runs NVIDIA's
-  [GEAR-SONIC](https://huggingface.co/nvidia/GEAR-SONIC) decoder on the G1, turning a
-  64-D latent motion token into full-body joint targets at 50 Hz. This lets you drive
-  the robot from a VLA policy trained on SONIC motion tokens (token in → whole-body
-  motion out) with `lerobot-rollout`, in sim or on the physical robot. See
-  [Whole-body control with SONIC](#whole-body-control-with-sonic) below.
-</Tip>
-
 ---
 
 ## Part 1: Getting Started
@@ -68,7 +59,7 @@ lerobot-teleoperate \
   --robot.controller=GrootLocomotionController
 ```
 
-This will launch a [MuJoCo sim instance](https://huggingface.co/lerobot/unitree-g1-mujoco/tree/main) for the G1. You can connect a gamepad to your machine before launching in order to control the robot's locomotion in sim. We support [HolosomaLocomotionController](https://github.com/amazon-far/holosoma), [GrootLocomotionController](https://github.com/NVlabs/GR00T-WholeBodyControl), and [SonicWholeBodyController](https://huggingface.co/nvidia/GEAR-SONIC) via `--robot.controller`.
+This will launch a [MuJoCo sim instance](https://huggingface.co/lerobot/unitree-g1-mujoco/tree/main) for the G1. You can connect a gamepad to your machine before launching in order to control the robot's locomotion in sim. We support both [HolosomaLocomotionController](https://github.com/amazon-far/holosoma) and [GrootLocomotionController](https://github.com/NVlabs/GR00T-WholeBodyControl) via `--robot.controller`.
 
 - Press `9` to release the robot
 - Press `7` / `8` to increase / decrease waist height
@@ -299,52 +290,6 @@ lerobot-rollout \
 
 ---
 
-## Whole-body control with SONIC
-
-The `SonicWholeBodyController` runs NVIDIA's [GEAR-SONIC](https://huggingface.co/nvidia/GEAR-SONIC)
-decoder on the G1. Each 50 Hz tick it consumes a **64-D latent motion token** and emits
-full-body joint targets — the encoder is bypassed, so a policy feeds tokens in and the
-decoder turns them into motion. Before the first token arrives the controller holds a
-neutral (idle) pose.
-
-This makes the G1 drivable by a VLA policy trained to output SONIC motion tokens (token
-as both `observation.state` and `action`, e.g. [`nepyope/sonic_walk`](https://huggingface.co/nepyope/sonic_walk))
-using the standard `lerobot-rollout`. The controller always runs **onboard** the robot;
-the laptop is a thin client that streams tokens and receives camera frames over ZMQ.
-
-**On the robot** — start the server in handshake mode so it instantiates and runs the
-controller onboard against local DDS at full rate:
-
-```bash
-cd ~/lerobot
-python src/lerobot/robots/unitree_g1/run_g1_server.py --handshake --camera
-```
-
-**From your laptop** — run the token policy; selecting `--robot.controller=SonicWholeBodyController`
-implicitly switches the robot to the 64-D latent-token action/observation interface:
-
-```bash
-lerobot-rollout \
-  --policy.path=nepyope/sonic_walk \
-  --policy.device=cuda \
-  --robot.type=unitree_g1 \
-  --robot.is_simulation=false \
-  --robot.robot_ip=<ROBOT_IP> \
-  --robot.controller=SonicWholeBodyController \
-  --robot.cameras='{"ego_view": {"type": "zmq", "server_address": "<ROBOT_IP>", "port": 5555, "camera_name": "head_camera", "width": 640, "height": 480, "fps": 30}}' \
-  --task="walk back and forth" \
-  --duration=1000 \
-  --fps=30
-```
-
-<Tip>
-  SONIC is a token-only decoder in LeRobot: the only input path is the 64-D latent
-  vector. To train your own token policy, expose the 64-D token as the action (a config
-  choice, e.g. `pi05` with a 64-D action dim) — no policy code changes are needed.
-</Tip>
-
----
-
 ## Additional Resources
 
 - [Unitree SDK Documentation](https://github.com/unitreerobotics/unitree_sdk2_python)

docs/source/using_dataset_tools.mdx

@@ -50,11 +50,11 @@ lerobot-edit-dataset \
 Divide a dataset into multiple subsets.
 
 ```bash
-# Split by fractions (e.g. 60% train, 20% val, 20% test)
+# Split by fractions (e.g. 80% train, 20% test, 20% val)
 lerobot-edit-dataset \
     --repo_id lerobot/pusht \
     --operation.type split \
-    --operation.splits '{"train": 0.6, "val": 0.2, "test": 0.2}'
+    --operation.splits '{"train": 0.8, "test": 0.2, "val": 0.2}'
 
 # Split by specific episode indices
 lerobot-edit-dataset \

pyproject.toml

@@ -150,6 +150,7 @@ pygame-dep = ["pygame>=2.5.1,<2.7.0"]
 # There is no cmeel-urdfdom 5.x; <5 selects the 4.x ABI the placo/pin wheels are built against.
 placo-dep = ["placo>=0.9.6,<0.9.16", "cmeel-urdfdom>=4,<5", "cmeel-tinyxml2<11"]
 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,<2.0.0", "protobuf>=6.31.1,<8.0.0"]
 accelerate-dep = ["accelerate>=1.14.0,<2.0.0"]
 can-dep = ["python-can>=4.2.0,<5.0.0"]
@@ -212,7 +213,7 @@ 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", "lerobot[accelerate-dep]"]
 multi_task_dit = ["lerobot[transformers-dep]", "lerobot[diffusers-dep]"]
@@ -494,19 +495,6 @@ ignore_errors = true
 module = "lerobot.envs.*"
 ignore_errors = false
 
-[[tool.mypy.overrides]]
-module = "lerobot.annotations.*"
-ignore_errors = false
-disallow_untyped_defs = true
-disallow_incomplete_defs = true
-check_untyped_defs = true
-
-[[tool.mypy.overrides]]
-module = "lerobot.transforms.*"
-ignore_errors = false
-disallow_untyped_defs = true
-disallow_incomplete_defs = true
-check_untyped_defs = true
 
 # [[tool.mypy.overrides]]
 # module = "lerobot.utils.*"

src/lerobot/cameras/opencv/camera_opencv.py

@@ -120,22 +120,14 @@ class OpenCVCamera(Camera):
         self.rotation: int | None = get_cv2_rotation(config.rotation)
         self.backend: int = config.backend
 
-        self.capture_width: int | None = None
+        if self.height and self.width:
-        self.capture_height: int | None = None
+            self.capture_width, self.capture_height = self.width, self.height
-        self._reset_connection_settings()
+            if self.rotation in [cv2.ROTATE_90_CLOCKWISE, cv2.ROTATE_90_COUNTERCLOCKWISE]:
+                self.capture_width, self.capture_height = self.height, self.width
 
     def __str__(self) -> str:
         return f"{self.__class__.__name__}({self.index_or_path})"
 
-    def _reset_connection_settings(self) -> None:
-        """Restore settings that may have been auto-detected during a failed connection."""
-        self.fps = self.config.fps
-        self.width = self.config.width
-        self.height = self.config.height
-        self.capture_width, self.capture_height = self.width, self.height
-        if self.rotation in [cv2.ROTATE_90_CLOCKWISE, cv2.ROTATE_90_COUNTERCLOCKWISE]:
-            self.capture_width, self.capture_height = self.height, self.width
-
     @property
     def is_connected(self) -> bool:
         """Checks if the camera is currently connected and opened."""
@@ -172,25 +164,17 @@ class OpenCVCamera(Camera):
                 f"Failed to open {self}.Run `lerobot-find-cameras opencv` to find available cameras."
             )
 
-        try:
+        self._configure_capture_settings()
-            self._configure_capture_settings()
+        self._start_read_thread()
-            self._start_read_thread()
 
-            if warmup and self.warmup_s > 0:
+        if warmup and self.warmup_s > 0:
-                start_time = time.time()
+            start_time = time.time()
-                while time.time() - start_time < self.warmup_s:
+            while time.time() - start_time < self.warmup_s:
-                    self.async_read(timeout_ms=self.warmup_s * 1000)
+                self.async_read(timeout_ms=self.warmup_s * 1000)
-                    time.sleep(0.1)
+                time.sleep(0.1)
-                with self.frame_lock:
+            with self.frame_lock:
-                    if self.latest_frame is None:
+                if self.latest_frame is None:
-                        raise ConnectionError(f"{self} failed to capture frames during warmup.")
+                    raise ConnectionError(f"{self} failed to capture frames during warmup.")
-        except BaseException:
-            try:
-                self._cleanup_resources()
-            except Exception:
-                logger.exception(f"Failed to fully clean up {self} after connect() failed.")
-            self._reset_connection_settings()
-            raise
 
         logger.info(f"{self} connected.")
 
@@ -328,36 +312,32 @@ class OpenCVCamera(Camera):
 
         for target in targets_to_scan:
             camera = cv2.VideoCapture(target)
-            try:
+            if camera.isOpened():
-                if camera.isOpened():
+                default_width = int(camera.get(cv2.CAP_PROP_FRAME_WIDTH))
-                    default_width = int(camera.get(cv2.CAP_PROP_FRAME_WIDTH))
+                default_height = int(camera.get(cv2.CAP_PROP_FRAME_HEIGHT))
-                    default_height = int(camera.get(cv2.CAP_PROP_FRAME_HEIGHT))
+                default_fps = camera.get(cv2.CAP_PROP_FPS)
-                    default_fps = camera.get(cv2.CAP_PROP_FPS)
+                default_format = camera.get(cv2.CAP_PROP_FORMAT)
-                    default_format = camera.get(cv2.CAP_PROP_FORMAT)
 
-                    # Get FOURCC code and convert to string
+                # Get FOURCC code and convert to string
-                    default_fourcc_code = camera.get(cv2.CAP_PROP_FOURCC)
+                default_fourcc_code = camera.get(cv2.CAP_PROP_FOURCC)
-                    default_fourcc_code_int = int(default_fourcc_code)
+                default_fourcc_code_int = int(default_fourcc_code)
-                    default_fourcc = "".join(
+                default_fourcc = "".join([chr((default_fourcc_code_int >> 8 * i) & 0xFF) for i in range(4)])
-                        [chr((default_fourcc_code_int >> 8 * i) & 0xFF) for i in range(4)]
-                    )
 
-                    camera_info = {
+                camera_info = {
-                        "name": f"OpenCV Camera @ {target}",
+                    "name": f"OpenCV Camera @ {target}",
-                        "type": "OpenCV",
+                    "type": "OpenCV",
-                        "id": target,
+                    "id": target,
-                        "backend_api": camera.getBackendName(),
+                    "backend_api": camera.getBackendName(),
-                        "default_stream_profile": {
+                    "default_stream_profile": {
-                            "format": default_format,
+                        "format": default_format,
-                            "fourcc": default_fourcc,
+                        "fourcc": default_fourcc,
-                            "width": default_width,
+                        "width": default_width,
-                            "height": default_height,
+                        "height": default_height,
-                            "fps": default_fps,
+                        "fps": default_fps,
-                        },
+                    },
-                    }
+                }
 
-                    found_cameras_info.append(camera_info)
+                found_cameras_info.append(camera_info)
-            finally:
                 camera.release()
 
         return found_cameras_info
@@ -516,26 +496,6 @@ class OpenCVCamera(Camera):
             self.latest_timestamp = None
             self.new_frame_event.clear()
 
-    def _cleanup_resources(self) -> None:
-        """Stop background reads and release the capture, including after partial setup."""
-        read_thread = self.thread
-        videocapture = self.videocapture
-
-        try:
-            self._stop_read_thread()
-        finally:
-            self.videocapture = None
-            try:
-                if videocapture is not None:
-                    videocapture.release()
-            finally:
-                # Releasing the device may unblock a hardware read that outlived
-                # the first bounded join in _stop_read_thread().
-                if read_thread is not None and read_thread.is_alive():
-                    read_thread.join(timeout=2.0)
-                    if read_thread.is_alive():  # pragma: no cover
-                        logger.warning(f"{self} read thread remained alive after releasing the capture.")
-
     @check_if_not_connected
     def async_read(self, timeout_ms: float = 200) -> NDArray[Any]:
         """
@@ -626,6 +586,16 @@ class OpenCVCamera(Camera):
         if not self.is_connected and self.thread is None:
             raise DeviceNotConnectedError(f"{self} not connected.")
 
-        self._cleanup_resources()
+        if self.thread is not None:
+            self._stop_read_thread()
+
+        if self.videocapture is not None:
+            self.videocapture.release()
+            self.videocapture = None
+
+        with self.frame_lock:
+            self.latest_frame = None
+            self.latest_timestamp = None
+            self.new_frame_event.clear()
 
         logger.info(f"{self} disconnected.")

src/lerobot/cameras/realsense/camera_realsense.py

@@ -121,9 +121,6 @@ class RealSenseCamera(Camera):
 
         self.config = config
 
-        self.width: int | None = config.width
-        self.height: int | None = config.height
-
         if config.serial_number_or_name.isdigit():
             self.serial_number = config.serial_number_or_name
         else:
@@ -134,9 +131,6 @@ class RealSenseCamera(Camera):
         self.use_rgb = config.use_rgb
         self.use_depth = config.use_depth
         self.warmup_s = config.warmup_s
-        self.exposure: int | None = config.exposure
-        self.gain: int | None = config.gain
-        self.white_balance: int | None = config.white_balance
 
         self.rs_pipeline: rs.pipeline | None = None
         self.rs_profile: rs.pipeline_profile | None = None
@@ -151,23 +145,14 @@ class RealSenseCamera(Camera):
 
         self.rotation: int | None = get_cv2_rotation(config.rotation)
 
-        self.capture_width: int | None = None
+        if self.height and self.width:
-        self.capture_height: int | None = None
+            self.capture_width, self.capture_height = self.width, self.height
-        self._reset_connection_settings()
+            if self.rotation in [cv2.ROTATE_90_CLOCKWISE, cv2.ROTATE_90_COUNTERCLOCKWISE]:
+                self.capture_width, self.capture_height = self.height, self.width
 
     def __str__(self) -> str:
         return f"{self.__class__.__name__}({self.serial_number})"
 
-    def _reset_connection_settings(self) -> None:
-        """Restore settings that may have been auto-detected during a failed connection."""
-        self.fps = self.config.fps
-        self.width = self.config.width
-        self.height = self.config.height
-        self.warmup_s = self.config.warmup_s
-        self.capture_width, self.capture_height = self.width, self.height
-        if self.rotation in [cv2.ROTATE_90_CLOCKWISE, cv2.ROTATE_90_COUNTERCLOCKWISE]:
-            self.capture_width, self.capture_height = self.height, self.width
-
     @property
     def is_connected(self) -> bool:
         """Checks if the camera pipeline is started and streams are active."""
@@ -187,8 +172,7 @@ class RealSenseCamera(Camera):
 
         Raises:
             DeviceAlreadyConnectedError: If the camera is already connected.
-            ValueError: If the configuration is invalid, a requested sensor option is unsupported,
+            ValueError: If the configuration is invalid (e.g., missing serial/name, name not unique).
-                or a requested sensor value is invalid.
             ConnectionError: If the camera is found but fails to start the pipeline or no RealSense devices are detected at all.
             RuntimeError: If the pipeline starts but fails to apply requested settings.
         """
@@ -206,31 +190,22 @@ class RealSenseCamera(Camera):
                 f"Failed to open {self}.Run `lerobot-find-cameras realsense` to find available cameras."
             ) from e
 
-        try:
+        self._configure_capture_settings()
-            self._configure_capture_settings()
+        self._start_read_thread()
-            self._configure_sensor_options()
-            self._start_read_thread()
 
-            # NOTE(Steven/Caroline): Enforcing at least one second of warmup as RS cameras need a bit of time before the first read. If we don't wait, the first read from the warmup will raise.
+        # NOTE(Steven/Caroline): Enforcing at least one second of warmup as RS cameras need a bit of time before the first read. If we don't wait, the first read from the warmup will raise.
-            self.warmup_s = max(self.warmup_s, 1)
+        self.warmup_s = max(self.warmup_s, 1)
 
-            warmup_read = self.async_read if self.use_rgb else self.async_read_depth
+        warmup_read = self.async_read if self.use_rgb else self.async_read_depth
-            start_time = time.time()
+        start_time = time.time()
-            while time.time() - start_time < self.warmup_s:
+        while time.time() - start_time < self.warmup_s:
-                warmup_read(timeout_ms=self.warmup_s * 1000)
+            warmup_read(timeout_ms=self.warmup_s * 1000)
-                time.sleep(0.1)
+            time.sleep(0.1)
-            with self.frame_lock:
+        with self.frame_lock:
-                if (self.use_rgb and self.latest_color_frame is None) or (
+            if (self.use_rgb and self.latest_color_frame is None) or (
-                    self.use_depth and self.latest_depth_frame is None
+                self.use_depth and self.latest_depth_frame is None
-                ):
+            ):
-                    raise ConnectionError(f"{self} failed to capture frames during warmup.")
+                raise ConnectionError(f"{self} failed to capture frames during warmup.")
-        except BaseException:
-            try:
-                self._cleanup_resources()
-            except Exception:
-                logger.exception(f"Failed to fully clean up {self} after connect() failed.")
-            self._reset_connection_settings()
-            raise
 
         logger.info(f"{self} connected.")
 
@@ -364,111 +339,6 @@ class RealSenseCamera(Camera):
         self.new_frame_event.clear()
         return self._async_read(timeout_ms=10000, read_depth=read_depth)
 
-    def _get_color_sensor(self) -> "rs.sensor":
-        """Returns the sensor that controls the color stream.
-
-        Most RealSense cameras expose "RGB Camera" for color. The D405 has no
-        separate RGB module — its color stream comes from "Stereo Module".
-        We try RGB Camera first, then fall back to Stereo Module.
-        """
-        if self.rs_profile is None:
-            raise RuntimeError(f"{self}: rs_profile must be initialized before use.")
-
-        device = self.rs_profile.get_device()
-        sensors = {s.get_info(rs.camera_info.name): s for s in device.query_sensors()}
-
-        for name in ("RGB Camera", "Stereo Module"):
-            if name in sensors:
-                return sensors[name]
-
-        available = list(sensors.keys())
-        raise RuntimeError(f"{self}: no color sensor found. Available sensors: {available}")
-
-    def _set_sensor_option(self, sensor: "rs.sensor", option: "rs.option", value: float, label: str) -> None:
-        """Sets a sensor option, re-raising range errors with actionable diagnostics."""
-        try:
-            sensor.set_option(option, value)
-        except Exception as e:
-            range_info = ""
-            try:
-                option_range = sensor.get_option_range(option)
-                range_info = (
-                    f" (supported range: min={option_range.min}, max={option_range.max}, "
-                    f"step={option_range.step}, default={option_range.default})"
-                )
-            except Exception:
-                range_info = " (option range unavailable)"
-            raise ValueError(
-                f"{self}: failed to set {label} to {value}{range_info}. Original error: {e}"
-            ) from e
-
-    def _configure_sensor_options(self) -> None:
-        """Applies manual sensor options (exposure, gain, white balance) to the color sensor.
-
-        When exposure or gain is set, auto-exposure is disabled first. When white_balance
-        is set, auto white balance is disabled first. An omitted option is left unchanged,
-        and configuration is skipped entirely if all options are omitted.
-
-        Raises:
-            ValueError: If the sensor does not support a requested option or a requested
-                value is invalid. Invalid-value errors include the option name, requested
-                value, and supported range when available.
-        """
-        if self.exposure is None and self.gain is None and self.white_balance is None:
-            return
-
-        color_sensor = self._get_color_sensor()
-
-        requested_options = (
-            (rs.option.exposure, self.exposure, "exposure"),
-            (rs.option.gain, self.gain, "gain"),
-            (rs.option.white_balance, self.white_balance, "white balance"),
-        )
-        unsupported_options = [
-            label
-            for option, value, label in requested_options
-            if value is not None and not color_sensor.supports(option)
-        ]
-        if unsupported_options:
-            raise ValueError(
-                f"{self}: color sensor does not support requested manual options: {unsupported_options}."
-            )
-
-        manual_exposure_requested = self.exposure is not None or self.gain is not None
-        if manual_exposure_requested:
-            if color_sensor.supports(rs.option.enable_auto_exposure):
-                self._set_sensor_option(color_sensor, rs.option.enable_auto_exposure, 0, "auto-exposure")
-                logger.info(f"{self} auto-exposure disabled.")
-            else:
-                logger.warning(
-                    f"{self} sensor does not support disabling auto-exposure; "
-                    "applying manual exposure/gain directly."
-                )
-
-        if self.exposure is not None:
-            self._set_sensor_option(color_sensor, rs.option.exposure, self.exposure, "exposure")
-            logger.info(f"{self} exposure set to {self.exposure}.")
-
-        if self.gain is not None:
-            self._set_sensor_option(color_sensor, rs.option.gain, self.gain, "gain")
-            logger.info(f"{self} gain set to {self.gain}.")
-
-        if self.white_balance is not None:
-            if color_sensor.supports(rs.option.enable_auto_white_balance):
-                self._set_sensor_option(
-                    color_sensor, rs.option.enable_auto_white_balance, 0, "auto white balance"
-                )
-                logger.info(f"{self} auto white balance disabled.")
-            else:
-                logger.warning(
-                    f"{self} sensor does not support disabling auto white balance; "
-                    "applying manual white balance directly."
-                )
-            self._set_sensor_option(
-                color_sensor, rs.option.white_balance, self.white_balance, "white balance"
-            )
-            logger.info(f"{self} white balance set to {self.white_balance}.")
-
     @check_if_not_connected
     def read_depth(self, timeout_ms: int = 200) -> NDArray[Any]:
         """
@@ -671,27 +541,6 @@ class RealSenseCamera(Camera):
             self.latest_timestamp = None
             self.new_frame_event.clear()
 
-    def _cleanup_resources(self) -> None:
-        """Stop background reads and stop the pipeline, including after partial setup."""
-        read_thread = self.thread
-        rs_pipeline = self.rs_pipeline
-
-        try:
-            self._stop_read_thread()
-        finally:
-            self.rs_pipeline = None
-            self.rs_profile = None
-            try:
-                if rs_pipeline is not None:
-                    rs_pipeline.stop()
-            finally:
-                # Stopping the pipeline may unblock a hardware read that outlived
-                # the first bounded join in _stop_read_thread().
-                if read_thread is not None and read_thread.is_alive():
-                    read_thread.join(timeout=2.0)
-                    if read_thread.is_alive():  # pragma: no cover
-                        logger.warning(f"{self} read thread remained alive after stopping the pipeline.")
-
     def _async_read(self, timeout_ms: float, read_depth: bool = False) -> NDArray[Any]:
         """Shared helper for :meth:`async_read`/:meth:`async_read_depth`: return the latest buffered frame."""
         if self.thread is None or not self.thread.is_alive():
@@ -835,5 +684,18 @@ class RealSenseCamera(Camera):
                 f"Attempted to disconnect {self}, but it appears already disconnected."
             )
 
-        self._cleanup_resources()
+        if self.thread is not None:
+            self._stop_read_thread()
+
+        if self.rs_pipeline is not None:
+            self.rs_pipeline.stop()
+            self.rs_pipeline = None
+            self.rs_profile = None
+
+        with self.frame_lock:
+            self.latest_color_frame = None
+            self.latest_depth_frame = None
+            self.latest_timestamp = None
+            self.new_frame_event.clear()
+
         logger.info(f"{self} disconnected.")

src/lerobot/cameras/realsense/configuration_realsense.py

@@ -46,17 +46,6 @@ class RealSenseCameraConfig(CameraConfig):
         use_depth: Whether to enable depth stream. Defaults to False.
         rotation: Image rotation setting (0°, 90°, 180°, or 270°). Defaults to no rotation.
         warmup_s: Time reading frames before returning from connect (in seconds)
-        exposure: Manual exposure value for the color sensor. When set, auto-exposure is
-            disabled and this fixed value is used. Valid ranges are camera-model specific
-            and reported if the value is rejected. Defaults to None (leave unchanged).
-        gain: Manual gain value for the color sensor. When set, auto-exposure is disabled
-            and this fixed gain is used, which also freezes exposure at its current value
-            when no exposure is configured. Valid ranges are camera-model specific and
-            reported if the value is rejected. Defaults to None (leave unchanged).
-        white_balance: Manual white balance value for the color sensor. When set, auto
-            white balance is disabled and this fixed value is used. Valid ranges are
-            camera-model specific and reported if the value is rejected. Defaults to None
-            (leave unchanged).
 
     Note:
         - Either name or serial_number must be specified.
@@ -72,9 +61,6 @@ class RealSenseCameraConfig(CameraConfig):
     use_depth: bool = False
     rotation: Cv2Rotation = Cv2Rotation.NO_ROTATION
     warmup_s: int = 1
-    exposure: int | None = None
-    gain: int | None = None
-    white_balance: int | None = None
 
     def __post_init__(self) -> None:
         self.color_mode = ColorMode(self.color_mode)
@@ -83,18 +69,6 @@ class RealSenseCameraConfig(CameraConfig):
         if not self.use_rgb and not self.use_depth:
             raise ValueError("At least one of `use_rgb` or `use_depth` must be enabled.")
 
-        manual_color_options = {
-            "exposure": self.exposure,
-            "gain": self.gain,
-            "white_balance": self.white_balance,
-        }
-        configured_color_options = [name for name, value in manual_color_options.items() if value is not None]
-        if configured_color_options and not self.use_rgb:
-            raise ValueError(
-                "Manual color sensor options require `use_rgb=True`. "
-                f"Configured options: {configured_color_options}."
-            )
-
         values = (self.fps, self.width, self.height)
         if any(v is not None for v in values) and any(v is None for v in values):
             raise ValueError(

src/lerobot/configs/dataset.py

@@ -71,19 +71,13 @@ class DatasetRecordConfig:
     # Number of threads per encoder instance. None = auto (codec default).
     # Lower values reduce CPU usage, maps to 'lp' (via svtav1-params) for libsvtav1 and 'threads' for h264/hevc..
     encoder_threads: int | None = None
-    # Skip appending the date-time tag to repo_id, keeping the user-provided name as-is
-    # (e.g. self-managed versioned names intended for a later `lerobot-edit-dataset merge`).
-    no_stamp: bool = False
 
     def stamp_repo_id(self) -> None:
         """Append a date-time tag to ``repo_id`` so each recording session gets a unique name.
 
         Must be called explicitly at dataset *creation* time — not on resume,
         where the existing ``repo_id`` (already stamped) must be preserved.
-        No-op when ``no_stamp`` is set, preserving a user-managed ``repo_id``.
         """
-        if self.no_stamp:
-            return
         if self.repo_id:
             timestamp = datetime.now().strftime("%Y%m%d_%H%M%S")
             self.repo_id = f"{self.repo_id}_{timestamp}"

src/lerobot/configs/default.py

@@ -33,6 +33,8 @@ class DatasetConfig:
     # looked up under $HF_LEROBOT_HOME/repo_id and Hub downloads use a revision-safe cache under $HF_LEROBOT_HOME/hub.
     root: str | None = None
     episodes: list[int] | None = None
+    # Episode indices to drop (e.g. corrupt or heterogeneous ones). Applied on top of `episodes`.
+    exclude_episodes: list[int] | None = None
     image_transforms: ImageTransformsConfig = field(default_factory=ImageTransformsConfig)
     revision: str | None = None
     use_imagenet_stats: bool = True
@@ -62,6 +64,10 @@ class DatasetConfig:
             if len(self.episodes) != len(set(self.episodes)):
                 duplicates = sorted({ep for ep in self.episodes if self.episodes.count(ep) > 1})
                 raise ValueError(f"Episode indices contain duplicates: {duplicates}")
+        if self.exclude_episodes is not None and any(ep < 0 for ep in self.exclude_episodes):
+            raise ValueError(
+                f"exclude_episodes must be non-negative, got: {[ep for ep in self.exclude_episodes if ep < 0]}"
+            )
 
 
 @dataclass

src/lerobot/configs/recipe.py

@@ -78,7 +78,7 @@ class MessageTurn:
             raise ValueError(f"Unsupported message stream: {self.stream!r}")
         if self.content is None and self.tool_calls_from is None:
             raise ValueError("MessageTurn.content is required unless tool_calls_from is set.")
-        if self.content is not None and not isinstance(self.content, (str, list)):
+        if self.content is not None and not isinstance(self.content, str | list):
             raise TypeError("MessageTurn.content must be a string, a list of HF-style blocks, or None.")
         if isinstance(self.content, list):
             for block in self.content:
@@ -147,7 +147,7 @@ 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."""
+        """Validate bindings and require text or low-level action supervision."""
         assert self.messages is not None
         known_bindings = set(DEFAULT_BINDINGS) | set(self.bindings or {}) | {"task"}
 
@@ -156,8 +156,14 @@ class TrainingRecipe:
             if missing:
                 raise ValueError(f"MessageTurn references unknown binding(s): {sorted(missing)}")
 
-        if not any(turn.target for turn in self.messages):
+        has_target = any(turn.target for turn in self.messages)
-            raise ValueError("Message recipes must contain at least one target turn.")
+        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.yaml

@@ -0,0 +1,16 @@
+# Predicts subtasks from tasks and trains subtask-conditioned action flow without memory or plans.
+# Requires `subtask` annotations; samples with missing `if_present` bindings do not render.
+
+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: 0.70
+    messages:
+      # The low-level stream trains action flow on the generated or annotated subtask.
+      - {role: user, content: "${subtask}", stream: low_level, if_present: subtask}

src/lerobot/configs/recipes/subtask_joint.yaml

@@ -0,0 +1,13 @@
+# Paper-style joint sequence (pi0.5 §IV-B): one sample supervises the subtask
+# text with CE and, because the assistant turn is part of the prefix, conditions
+# the FAST and flow action losses on the same annotated subtask in one forward.
+# The supervised span is attended causally; the action losses see task + subtask.
+#
+# Pair with `--policy.joint_subtask_conditioning=true` at inference so the flow
+# prefix reproduces this layout (task turn with state + causal generated subtask).
+# Samples without a `subtask` annotation fall back to a plain task-prompt
+# low-level sample via `if_present`.
+
+messages:
+  - {role: user, content: "${task}", stream: low_level}
+  - {role: assistant, content: "${subtask}", stream: low_level, target: true, if_present: subtask}

src/lerobot/configs/recipes/subtask_mem.yaml

@@ -0,0 +1,30 @@
+# Trains subtask prediction, subtask-conditioned action flow, and memory updates without plans.
+# Requires `subtask` and `memory`; missing `if_present` bindings skip the affected sub-recipe.
+
+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: 0.60
+    messages:
+      # The low-level stream trains action flow on the generated or annotated subtask.
+      - {role: user, content: "${subtask}", stream: low_level, if_present: subtask}
+
+  memory_update:
+    # `active_at` densifies sparse boundaries while preserving the prior-memory/subtask mapping.
+    # Inference controls update timing through `subtask_change` events.
+    weight: 0.15
+    bindings:
+      prior_memory: "nth_prev(style=memory, offset=1)"
+      current_memory: "active_at(t, style=memory)"
+      completed_subtask: "nth_prev(style=subtask, offset=1)"
+    messages:
+      - {role: user, content: "${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_speech.yaml

@@ -0,0 +1,70 @@
+# Adds memory, spoken interjection responses, and camera-grounded VQA to subtask/action training.
+# Missing optional annotations skip only their sub-recipe; `say` tool calls tokenize as `<say>...</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 low-level stream trains action flow on the generated or annotated subtask.
+      - {role: user, content: "${subtask}", stream: low_level, if_present: subtask}
+
+  memory_update:
+    # `active_at` densifies sparse boundaries while preserving the prior-memory/subtask mapping.
+    # Inference controls update timing through `subtask_change` events.
+    weight: 0.10
+    bindings:
+      prior_memory: "nth_prev(style=memory, offset=1)"
+      current_memory: "active_at(t, style=memory)"
+      completed_subtask: "nth_prev(style=subtask, offset=1)"
+    messages:
+      - {role: user, content: "${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}
+      # The assistant target is a `say` tool call flattened to a `<say>...</say>` marker.
+      - {role: assistant, stream: high_level, target: true, if_present: speech, tool_calls_from: speech}
+
+  # Each camera uses a separate VQA sub-recipe for view-specific binding.
+  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/datasets/aggregate.py

@@ -19,7 +19,6 @@ import copy
 import logging
 import shutil
 from pathlib import Path
-from typing import Any, NotRequired, TypedDict
 
 import datasets
 import pandas as pd
@@ -50,32 +49,8 @@ from .utils import (
 )
 from .video_utils import concatenate_video_files, get_video_duration_in_s
 
-logger = logging.getLogger(__name__)
 
-type FeatureDict = dict[str, dict[str, Any]]
+def merge_video_feature_info_for_aggregate(all_metadata: list[LeRobotDatasetMetadata]) -> dict[str, dict]:
-type ChunkFile = tuple[int, int]
-
-
-class IndexState(TypedDict):
-    chunk: int
-    file: int
-    src_to_dst: NotRequired[dict[ChunkFile, ChunkFile]]
-
-
-class VideoIndex(TypedDict):
-    chunk: int
-    file: int
-    latest_duration: float
-    episode_duration: float
-    src_to_offset: NotRequired[dict[ChunkFile, float]]
-    src_to_dst: NotRequired[dict[ChunkFile, ChunkFile]]
-    dst_file_durations: NotRequired[dict[ChunkFile, float]]
-
-
-type VideoIndexState = dict[str, VideoIndex]
-
-
-def merge_video_feature_info_for_aggregate(all_metadata: list[LeRobotDatasetMetadata]) -> FeatureDict:
     """Create a merged video feature info dictionary for aggregation. The video encoder info is merged field-by-field: each key is kept only when every source agrees; otherwise that key is set to ``null`` (or ``{}`` for ``video.extra_options``) and a warning is logged.
 
     Args:
@@ -84,14 +59,14 @@ def merge_video_feature_info_for_aggregate(all_metadata: list[LeRobotDatasetMeta
     Returns:
         dict: A dictionary of merged video feature info.
     """
-    merged_info: FeatureDict = copy.deepcopy(all_metadata[0].features)
+    merged_info = copy.deepcopy(all_metadata[0].features)
     video_keys = [k for k in merged_info if merged_info[k].get("dtype") == "video"]
 
     for vk in video_keys:
         video_infos = [m.features.get(vk, {}).get("info") or {} for m in all_metadata]
         base_video_info = video_infos[0]
 
-        merged_encoder_info: dict[str, Any] = {}
+        merged_encoder_info: dict = {}
         fallback_keys: list[str] = []
         for info_key in VIDEO_ENCODER_INFO_KEYS:
             values = [info.get(info_key, None) for info in video_infos]
@@ -105,7 +80,7 @@ def merge_video_feature_info_for_aggregate(all_metadata: list[LeRobotDatasetMeta
                 merged_encoder_info[info_key] = {} if info_key == "video.extra_options" else None
 
         if fallback_keys:
-            logger.warning(
+            logging.warning(
                 f"Merging heterogeneous or incomplete video encoder metadata for feature {vk}. "
                 f"Setting these keys to null: {fallback_keys}.",
             )
@@ -117,7 +92,7 @@ def merge_video_feature_info_for_aggregate(all_metadata: list[LeRobotDatasetMeta
     return merged_info
 
 
-def validate_all_metadata(all_metadata: list[LeRobotDatasetMetadata]) -> tuple[int, str | None, FeatureDict]:
+def validate_all_metadata(all_metadata: list[LeRobotDatasetMetadata]):
     """Validates that all dataset metadata have consistent properties.
 
     Ensures all datasets have the same fps, robot_type, and features to guarantee
@@ -154,9 +129,7 @@ def validate_all_metadata(all_metadata: list[LeRobotDatasetMetadata]) -> tuple[i
     return fps, robot_type, features
 
 
-def update_data_df(
+def update_data_df(df, src_meta, dst_meta):
-    df: pd.DataFrame, src_meta: LeRobotDatasetMetadata, dst_meta: LeRobotDatasetMetadata
-) -> pd.DataFrame:
     """Updates a data DataFrame with new indices and task mappings for aggregation.
 
     Adjusts episode indices, frame indices, and task indices to account for
@@ -181,12 +154,12 @@ def update_data_df(
 
 
 def update_meta_data(
-    df: pd.DataFrame,
+    df,
-    dst_meta: LeRobotDatasetMetadata,
+    dst_meta,
-    meta_idx: IndexState,
+    meta_idx,
-    data_idx: IndexState,
+    data_idx,
-    videos_idx: VideoIndexState,
+    videos_idx,
-) -> pd.DataFrame:
+):
     """Updates metadata DataFrame with new chunk, file, and timestamp indices.
 
     Adjusts all indices and timestamps to account for previously aggregated
@@ -316,7 +289,7 @@ def aggregate_datasets(
     chunk_size: int | None = None,
     concatenate_videos: bool = True,
     concatenate_data: bool = True,
-) -> None:
+):
     """Aggregates multiple LeRobot datasets into a single unified dataset.
 
     This is the main function that orchestrates the aggregation process by:
@@ -336,7 +309,7 @@ def aggregate_datasets(
         concatenate_videos: When False, keep one mp4 per source file instead of packing into shards.
         concatenate_data: When False, keep one parquet per source file instead of packing into shards.
     """
-    logger.info("Start aggregate_datasets")
+    logging.info("Start aggregate_datasets")
 
     if data_files_size_in_mb is None:
         data_files_size_in_mb = DEFAULT_DATA_FILE_SIZE_IN_MB
@@ -368,15 +341,15 @@ def aggregate_datasets(
         video_files_size_in_mb=video_files_size_in_mb,
     )
 
-    logger.info("Find all tasks")
+    logging.info("Find all tasks")
     unique_tasks = pd.concat([m.tasks for m in all_metadata]).index.unique()
     dst_meta.tasks = pd.DataFrame(
         {"task_index": range(len(unique_tasks))}, index=pd.Index(unique_tasks, name="task")
     )
 
-    meta_idx: IndexState = {"chunk": 0, "file": 0}
+    meta_idx = {"chunk": 0, "file": 0}
-    data_idx: IndexState = {"chunk": 0, "file": 0}
+    data_idx = {"chunk": 0, "file": 0}
-    videos_idx: VideoIndexState = {
+    videos_idx = {
         key: {"chunk": 0, "file": 0, "latest_duration": 0, "episode_duration": 0} for key in video_keys
     }
 
@@ -400,17 +373,12 @@ def aggregate_datasets(
         dst_meta.info.total_frames += src_meta.total_frames
 
     finalize_aggregation(dst_meta, all_metadata)
-    logger.info("Aggregation complete.")
+    logging.info("Aggregation complete.")
 
 
 def aggregate_videos(
-    src_meta: LeRobotDatasetMetadata,
+    src_meta, dst_meta, videos_idx, video_files_size_in_mb, chunk_size, concatenate_videos=True
-    dst_meta: LeRobotDatasetMetadata,
+):
-    videos_idx: VideoIndexState,
-    video_files_size_in_mb: float,
-    chunk_size: int,
-    concatenate_videos: bool = True,
-) -> VideoIndexState:
     """Aggregates video chunks from a source dataset into the destination dataset.
 
     Handles video file concatenation and rotation based on file size limits.
@@ -438,16 +406,15 @@ def aggregate_videos(
             videos_idx[key]["dst_file_durations"] = {}
 
     for key, video_idx in videos_idx.items():
-        unique_chunk_file_pairs: list[ChunkFile] = sorted(
+        unique_chunk_file_pairs = {
-            {
+            (chunk, file)
-                (chunk, file)
+            for chunk, file in zip(
-                for chunk, file in zip(
+                src_meta.episodes[f"videos/{key}/chunk_index"],
-                    src_meta.episodes[f"videos/{key}/chunk_index"],
+                src_meta.episodes[f"videos/{key}/file_index"],
-                    src_meta.episodes[f"videos/{key}/file_index"],
+                strict=False,
-                    strict=False,
+            )
-                )
+        }
-            }
+        unique_chunk_file_pairs = sorted(unique_chunk_file_pairs)
-        )
 
         chunk_idx = video_idx["chunk"]
         file_idx = video_idx["file"]
@@ -522,14 +489,7 @@ def aggregate_videos(
     return videos_idx
 
 
-def aggregate_data(
+def aggregate_data(src_meta, dst_meta, data_idx, data_files_size_in_mb, chunk_size, concatenate_data=True):
-    src_meta: LeRobotDatasetMetadata,
-    dst_meta: LeRobotDatasetMetadata,
-    data_idx: IndexState,
-    data_files_size_in_mb: float,
-    chunk_size: int,
-    concatenate_data: bool = True,
-) -> IndexState:
     """Aggregates data chunks from a source dataset into the destination dataset.
 
     Reads source data files, updates indices to match the aggregated dataset,
@@ -550,16 +510,14 @@ def aggregate_data(
     Returns:
         dict: Updated data_idx with current chunk and file indices.
     """
-    unique_chunk_file_ids: list[ChunkFile] = sorted(
+    unique_chunk_file_ids = {
-        {
+        (c, f)
-            (c, f)
+        for c, f in zip(
-            for c, f in zip(
+            src_meta.episodes["data/chunk_index"], src_meta.episodes["data/file_index"], strict=False
-                src_meta.episodes["data/chunk_index"],
+        )
-                src_meta.episodes["data/file_index"],
+    }
-                strict=False,
+
-            )
+    unique_chunk_file_ids = sorted(unique_chunk_file_ids)
-        }
-    )
     contains_images = len(dst_meta.image_keys) > 0
 
     # retrieve features schema for proper image typing in parquet
@@ -567,7 +525,7 @@ def aggregate_data(
 
     # Track source to destination file mapping for metadata update
     # This is critical for handling datasets that are already results of a merge
-    src_to_dst: dict[ChunkFile, ChunkFile] = {}
+    src_to_dst: dict[tuple[int, int], tuple[int, int]] = {}
 
     for src_chunk_idx, src_file_idx in unique_chunk_file_ids:
         src_path = src_meta.root / DEFAULT_DATA_PATH.format(
@@ -606,13 +564,7 @@ def aggregate_data(
     return data_idx
 
 
-def aggregate_metadata(
+def aggregate_metadata(src_meta, dst_meta, meta_idx, data_idx, videos_idx):
-    src_meta: LeRobotDatasetMetadata,
-    dst_meta: LeRobotDatasetMetadata,
-    meta_idx: IndexState,
-    data_idx: IndexState,
-    videos_idx: VideoIndexState,
-) -> IndexState:
     """Aggregates metadata from a source dataset into the destination dataset.
 
     Reads source metadata files, updates all indices and timestamps,
@@ -628,16 +580,16 @@ def aggregate_metadata(
     Returns:
         dict: Updated meta_idx with current chunk and file indices.
     """
-    chunk_file_ids: list[ChunkFile] = sorted(
+    chunk_file_ids = {
-        {
+        (c, f)
-            (c, f)
+        for c, f in zip(
-            for c, f in zip(
+            src_meta.episodes["meta/episodes/chunk_index"],
-                src_meta.episodes["meta/episodes/chunk_index"],
+            src_meta.episodes["meta/episodes/file_index"],
-                src_meta.episodes["meta/episodes/file_index"],
+            strict=False,
-                strict=False,
+        )
-            )
+    }
-        }
+
-    )
+    chunk_file_ids = sorted(chunk_file_ids)
     for chunk_idx, file_idx in chunk_file_ids:
         src_path = src_meta.root / DEFAULT_EPISODES_PATH.format(chunk_index=chunk_idx, file_index=file_idx)
         df = pd.read_parquet(src_path)
@@ -670,16 +622,16 @@ def aggregate_metadata(
 def append_or_create_parquet_file(
     df: pd.DataFrame,
     src_path: Path,
-    idx: IndexState,
+    idx: dict[str, int],
     max_mb: float,
     chunk_size: int,
     default_path: str,
     contains_images: bool = False,
-    aggr_root: Path | None = None,
+    aggr_root: Path = None,
     hf_features: datasets.Features | None = None,
     concatenate: bool = True,
     one_row_group_per_episode: bool = False,
-) -> tuple[IndexState, ChunkFile]:
+) -> tuple[dict[str, int], tuple[int, int]]:
     """Appends data to an existing parquet file or creates a new one based on size constraints.
 
     Manages file rotation when size limits are exceeded to prevent individual files
@@ -702,13 +654,7 @@ def append_or_create_parquet_file(
     Returns:
         tuple: (updated_idx, (dst_chunk, dst_file)) where updated_idx is the index dict
                and (dst_chunk, dst_file) is the actual destination file the data was written to.
-
-    Raises:
-        ValueError: If aggr_root is not provided.
     """
-    if aggr_root is None:
-        raise ValueError("aggr_root must be provided.")
-
     dst_chunk, dst_file = idx["chunk"], idx["file"]
     dst_path = aggr_root / default_path.format(chunk_index=dst_chunk, file_index=dst_file)
 
@@ -752,9 +698,7 @@ def append_or_create_parquet_file(
     return idx, (dst_chunk, dst_file)
 
 
-def finalize_aggregation(
+def finalize_aggregation(aggr_meta, all_metadata):
-    aggr_meta: LeRobotDatasetMetadata, all_metadata: list[LeRobotDatasetMetadata]
-) -> None:
     """Finalizes the dataset aggregation by writing summary files and statistics.
 
     Writes the tasks file, info file with total counts and splits, and
@@ -764,16 +708,16 @@ def finalize_aggregation(
         aggr_meta: Aggregated dataset metadata.
         all_metadata: List of all source dataset metadata objects.
     """
-    logger.info("write tasks")
+    logging.info("write tasks")
     write_tasks(aggr_meta.tasks, aggr_meta.root)
 
-    logger.info("write info")
+    logging.info("write info")
     aggr_meta.info.total_tasks = len(aggr_meta.tasks)
     aggr_meta.info.total_episodes = sum(m.total_episodes for m in all_metadata)
     aggr_meta.info.total_frames = sum(m.total_frames for m in all_metadata)
     aggr_meta.info.splits = {"train": f"0:{sum(m.total_episodes for m in all_metadata)}"}
     write_info(aggr_meta.info, aggr_meta.root)
 
-    logger.info("write stats")
+    logging.info("write stats")
     aggr_meta.stats = aggregate_stats([m.stats for m in all_metadata])
     write_stats(aggr_meta.stats, aggr_meta.root)

src/lerobot/datasets/dataset_metadata.py

@@ -188,8 +188,8 @@ class LeRobotDatasetMetadata:
     def _load_metadata(self):
         self.info = load_info(self.root)
         check_version_compatibility(self.repo_id, self._version, CODEBASE_VERSION)
-        self.tasks = load_tasks(self.root) if self.total_tasks > 0 else None
+        self.tasks = load_tasks(self.root)
-        self.episodes = load_episodes(self.root) if self.total_episodes > 0 else None
+        self.episodes = load_episodes(self.root)
         self.stats = load_stats(self.root)
 
     def ensure_readable(self) -> None:

src/lerobot/datasets/dataset_reader.py

@@ -163,10 +163,40 @@ 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)
+        # Annotated datasets may have language columns absent from metadata.
+        # Extend the schema before the strict Parquet cast.
+        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:
+        """Register language columns found in Parquet but missing from metadata."""
+        # Leave empty datasets to fail through the normal loading path.
+        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/factory.py

@@ -66,6 +66,17 @@ def resolve_delta_timestamps(
     return delta_timestamps
 
 
+def _resolve_episodes(
+    episodes: list[int] | None, exclude_episodes: list[int] | None, total_episodes: int
+) -> list[int] | None:
+    """Apply an episode exclusion list on top of an optional allowlist."""
+    if not exclude_episodes:
+        return episodes
+    base = episodes if episodes is not None else list(range(total_episodes))
+    excluded = set(exclude_episodes)
+    return [episode for episode in base if episode not in excluded]
+
+
 def make_dataset(cfg: TrainPipelineConfig) -> LeRobotDataset | MultiLeRobotDataset:
     """Handles the logic of setting up delta timestamps and image transforms before creating a dataset.
 
@@ -87,11 +98,14 @@ def make_dataset(cfg: TrainPipelineConfig) -> LeRobotDataset | MultiLeRobotDatas
             cfg.dataset.repo_id, root=cfg.dataset.root, revision=cfg.dataset.revision
         )
         delta_timestamps = resolve_delta_timestamps(cfg.trainable_config, ds_meta)
+        episodes = _resolve_episodes(
+            cfg.dataset.episodes, cfg.dataset.exclude_episodes, ds_meta.total_episodes
+        )
         if not cfg.dataset.streaming:
             dataset = LeRobotDataset(
                 cfg.dataset.repo_id,
                 root=cfg.dataset.root,
-                episodes=cfg.dataset.episodes,
+                episodes=episodes,
                 delta_timestamps=delta_timestamps,
                 image_transforms=image_transforms,
                 revision=cfg.dataset.revision,
@@ -104,7 +118,7 @@ def make_dataset(cfg: TrainPipelineConfig) -> LeRobotDataset | MultiLeRobotDatas
             dataset = StreamingLeRobotDataset(
                 cfg.dataset.repo_id,
                 root=cfg.dataset.root,
-                episodes=cfg.dataset.episodes,
+                episodes=episodes,
                 delta_timestamps=delta_timestamps,
                 image_transforms=image_transforms,
                 revision=cfg.dataset.revision,

src/lerobot/datasets/language_render.py

@@ -162,14 +162,28 @@ def render_sample(
     task: str | None = None,
     dataset_ctx: Any | None = None,
 ) -> RenderedMessages | None:
-    """Render the chat-style messages for a single dataset sample.
+    """Resolve one sample's bindings and render its message recipe.
 
-    Resolves the recipe's bindings against ``persistent`` and ``events`` rows
+    Returns ``None`` when no text or low-level action supervision applies.
-    at frame timestamp ``t``, then expands the recipe's message templates.
-    Returns ``None`` if the resolved sample contains no target message.
     """
     persistent_rows = _normalize_rows(persistent or [])
     event_rows = _normalize_rows(events or [])
+
+    # Route sparse VQA frames to a matching view-specific component before weighted selection.
+    # This avoids dropping annotated frames or selecting VQA without annotations.
+    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 +197,55 @@ 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 a matching VQA component, or return ``None`` for normal selection.
+
+    Multiple matching views are selected 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]
+
+    # Choose among matching cameras by relative weight, or uniformly when all weights are zero.
+    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 +409,9 @@ def _render_message_recipe(
         if turn.target:
             target_indices.append(message_idx)
 
-    if not target_indices:
+    # Keep samples with either text targets or low-level action supervision.
+    has_low_level = any(stream == "low_level" for stream in streams)
+    if not target_indices and not has_low_level:
         return None
 
     rendered = {
@@ -403,14 +468,12 @@ def _validate_rendered(rendered: RenderedMessages) -> None:
 
     if len(streams) != len(messages):
         raise ValueError("message_streams must be aligned with messages.")
-    if not target_indices:
+    # Require text or low-level action supervision.
-        raise ValueError("Rendered samples must contain at least one target message.")
+    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.")
-    # ``stream`` is enforced non-None at MessageTurn construction time
-    # (see ``MessageTurn.__post_init__``), so a missing stream here would
-    # mean the dataclass invariant was bypassed; no need to re-check.
 
 
 def _nth_relative(

src/lerobot/envs/configs.py

@@ -560,7 +560,13 @@ class RoboCasaEnv(EnvConfig):
             kwargs["split"] = self.split
         return kwargs
 
-    def create_envs(self, n_envs: int, use_async_envs: bool = False):
+    def create_envs(
+        self,
+        n_envs: int,
+        use_async_envs: bool = False,
+        terminate_on_success: bool = True,
+        horizon: int | None = None,
+    ):
         from .robocasa import create_robocasa_envs
 
         if self.task is None:
@@ -574,6 +580,8 @@ class RoboCasaEnv(EnvConfig):
             env_cls=env_cls,
             episode_length=self.episode_length,
             obj_registries=tuple(self.obj_registries),
+            terminate_on_success=terminate_on_success,
+            horizon=horizon,
         )
 
 

src/lerobot/envs/libero.py

@@ -384,12 +384,7 @@ class LiberoEnv(gym.Env):
 
     def close(self):
         if self._env is not None:
-            try:
+            self._env.close()
-                self._env.close()
-            finally:
-                # LIBERO deletes its inner env on close, so this wrapper must
-                # be recreated before the next reset.
-                self._env = None
 
 
 def _make_env_fns(

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)
+OBS_STATE_DIM = 16  # ee_pos_rel(3) + ee_quat_rel(4) + base_pos(3) + base_quat(4) + gripper_qpos(2)
-ACTION_DIM = 12  # base_motion(4) + control_mode(1) + ee_pos(3) + ee_rot(3) + gripper(1)
+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.end_effector_position": flat_action[0:3],
-        "action.control_mode": flat_action[4:5],
+        "action.end_effector_rotation": flat_action[3:6],
-        "action.end_effector_position": flat_action[5:8],
+        "action.gripper_close": flat_action[6:7],
-        "action.end_effector_rotation": flat_action[8:11],
+        "action.base_motion": flat_action[7:11],
-        "action.gripper_close": flat_action[11:12],
+        "action.control_mode": flat_action[11:12],
     }
 
 
@@ -136,9 +137,16 @@ class RoboCasaEnv(gym.Env):
         episode_length: int | None = None,
         obj_registries: Sequence[str] = DEFAULT_OBJ_REGISTRIES,
         episode_index: int = 0,
+        terminate_on_success: bool = True,
+        horizon: int | None = None,
     ):
         super().__init__()
         self.task = task
+        # When False, a task-success does NOT end/reset the episode — used by the
+        # interactive sim so one kitchen persists across sequential prompts.
+        self.terminate_on_success = terminate_on_success
+        # Underlying robosuite horizon (steps before truncation). None -> default.
+        self.horizon = horizon
         self.obs_type = obs_type
         self.render_mode = render_mode
         self.observation_width = observation_width
@@ -210,12 +218,16 @@ class RoboCasaEnv(gym.Env):
         # (only None/"all"/"pretrain"/"target" are valid). Always pass a
         # valid value so we don't hit that default. Extra kwargs are
         # forwarded to the underlying kitchen env via create_env/robosuite.make.
+        extra_kwargs: dict[str, Any] = {}
+        if self.horizon is not None:
+            extra_kwargs["horizon"] = int(self.horizon)
         self._env = RoboCasaGymEnv(
             env_name=self.task,
             camera_widths=self.observation_width,
             camera_heights=self.observation_height,
             split=self.split if self.split is not None else "all",
             obj_registries=self.obj_registries,
+            **extra_kwargs,
         )
 
         ep_meta = self._env.env.get_ep_meta()
@@ -230,12 +242,14 @@ 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,
@@ -280,7 +294,7 @@ class RoboCasaEnv(gym.Env):
         raw_obs, reward, done, truncated, info = self._env.step(action_dict)
 
         is_success = bool(info.get("success", False))
-        terminated = done or is_success
+        terminated = done or (is_success and self.terminate_on_success)
         info.update({"task": self.task, "done": done, "is_success": is_success})
 
         observation = self._format_raw_obs(raw_obs)
@@ -313,6 +327,8 @@ def _make_env_fns(
     split: str | None,
     episode_length: int | None,
     obj_registries: Sequence[str],
+    terminate_on_success: bool = True,
+    horizon: int | None = None,
 ) -> list[Callable[[], RoboCasaEnv]]:
     """Build n_envs factory callables for a single task.
 
@@ -335,6 +351,8 @@ def _make_env_fns(
             episode_length=episode_length,
             obj_registries=obj_registries,
             episode_index=episode_index,
+            terminate_on_success=terminate_on_success,
+            horizon=horizon,
         )
 
     return [partial(_make_env, i) for i in range(n_envs)]
@@ -348,6 +366,8 @@ def create_robocasa_envs(
     env_cls: Callable[[Sequence[Callable[[], Any]]], Any] | None = None,
     episode_length: int | None = None,
     obj_registries: Sequence[str] = DEFAULT_OBJ_REGISTRIES,
+    terminate_on_success: bool = True,
+    horizon: int | None = None,
 ) -> dict[str, dict[int, Any]]:
     """Create vectorized RoboCasa365 environments with a consistent return shape.
 
@@ -409,6 +429,8 @@ def create_robocasa_envs(
             split=split,
             episode_length=episode_length,
             obj_registries=obj_registries,
+            terminate_on_success=terminate_on_success,
+            horizon=horizon,
         )
 
         if is_async:

src/lerobot/envs/robotwin.py

@@ -384,9 +384,7 @@ class RoboTwinEnv(gym.Env):
 
         self._env: Any | None = None  # deferred — created on first reset() inside worker
         self._step_count: int = 0
-        self._black_frame: np.ndarray = np.zeros(
+        self._black_frame = np.zeros((self.observation_height, self.observation_width, 3), dtype=np.uint8)
-            (self.observation_height, self.observation_width, 3), dtype=np.uint8
-        )
 
         image_spaces = {
             cam: spaces.Box(

src/lerobot/envs/vlabench.py

@@ -373,7 +373,7 @@ class VLABenchEnv(gym.Env):
 
         if action.shape[0] != 7:
             # Unknown layout — fall back to zero-pad so the sim doesn't crash.
-            padded: np.ndarray = np.zeros(ctrl_dim, dtype=np.float64)
+            padded = np.zeros(ctrl_dim, dtype=np.float64)
             padded[: min(action.shape[0], ctrl_dim)] = action[:ctrl_dim]
             return padded
 

src/lerobot/motors/dynamixel/tables.py

@@ -122,9 +122,6 @@ MODEL_ENCODING_TABLE = {
     "xm430-w350": X_SERIES_ENCODINGS_TABLE,
     "xm540-w270": X_SERIES_ENCODINGS_TABLE,
     "xc430-w150": X_SERIES_ENCODINGS_TABLE,
-    "xh540-w150": X_SERIES_ENCODINGS_TABLE,
-    "xc330-t288": X_SERIES_ENCODINGS_TABLE,
-    "xc330-t181": X_SERIES_ENCODINGS_TABLE,
 }
 
 # {model: model_resolution}
@@ -137,9 +134,6 @@ MODEL_RESOLUTION = {
     "xm430-w350": 4096,
     "xm540-w270": 4096,
     "xc430-w150": 4096,
-    "xh540-w150": 4096,
-    "xc330-t288": 4096,
-    "xc330-t181": 4096,
 }
 
 # {model: model_number}
@@ -151,9 +145,6 @@ MODEL_NUMBER_TABLE = {
     "xm430-w350": 1020,
     "xm540-w270": 1120,
     "xc430-w150": 1070,
-    "xh540-w150": 1110,
-    "xc330-t288": 1220,
-    "xc330-t181": 1210,
 }
 
 # {model: available_operating_modes}
@@ -165,9 +156,6 @@ MODEL_OPERATING_MODES = {
     "xm430-w350": [0, 1, 3, 4, 5, 16],
     "xm540-w270": [0, 1, 3, 4, 5, 16],
     "xc430-w150": [1, 3, 4, 16],
-    "xh540-w150": [0, 1, 3, 4, 5, 16],
-    "xc330-t288": [0, 1, 3, 4, 5, 16],
-    "xc330-t181": [0, 1, 3, 4, 5, 16],
 }
 
 MODEL_CONTROL_TABLE = {
@@ -178,9 +166,6 @@ MODEL_CONTROL_TABLE = {
     "xm430-w350": X_SERIES_CONTROL_TABLE,
     "xm540-w270": X_SERIES_CONTROL_TABLE,
     "xc430-w150": X_SERIES_CONTROL_TABLE,
-    "xh540-w150": X_SERIES_CONTROL_TABLE,
-    "xc330-t288": X_SERIES_CONTROL_TABLE,
-    "xc330-t181": X_SERIES_CONTROL_TABLE,
 }
 
 MODEL_BAUDRATE_TABLE = {
@@ -191,9 +176,6 @@ MODEL_BAUDRATE_TABLE = {
     "xm430-w350": X_SERIES_BAUDRATE_TABLE,
     "xm540-w270": X_SERIES_BAUDRATE_TABLE,
     "xc430-w150": X_SERIES_BAUDRATE_TABLE,
-    "xh540-w150": X_SERIES_BAUDRATE_TABLE,
-    "xc330-t288": X_SERIES_BAUDRATE_TABLE,
-    "xc330-t181": X_SERIES_BAUDRATE_TABLE,
 }
 
 AVAILABLE_BAUDRATES = [

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

@@ -28,6 +28,7 @@ from .multi_task_dit.configuration_multi_task_dit import MultiTaskDiTConfig as M
 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
@@ -56,6 +57,7 @@ __all__ = [
     "PI0Config",
     "PI0FastConfig",
     "PI05Config",
+    "PI052Config",
     "SmolVLAConfig",
     "TDMPCConfig",
     "VLAJEPAConfig",

src/lerobot/policies/evo1/processor_evo1.py

@@ -302,33 +302,6 @@ def _pad_evo1_stats(
     return padded_stats
 
 
-def _refresh_evo1_normalization_steps(
-    config: Evo1Config,
-    preprocessor: PolicyProcessorPipeline,
-    postprocessor: PolicyProcessorPipeline,
-) -> None:
-    """Re-pad checkpoint-loaded (un)normalizer stats/features to EVO1's fixed widths.
-
-    Loading a checkpoint injects the raw dataset stats (unpadded to max_state_dim/max_action_dim)
-    into the (un)normalizer via the generic override path in make_pre_post_processors. Those stats
-    and their declared features must be re-padded/reshaped to EVO1's fixed widths, otherwise
-    normalization fails against the padded state/action tensors (e.g. state padded to 24 vs. 8-dim
-    LIBERO stats). Padding is a no-op when stats are already at the target width.
-    """
-    normalization_features = _evo1_normalization_features(config)
-    action_features = _evo1_action_features(config)
-    for step in preprocessor.steps:
-        if isinstance(step, NormalizerProcessorStep):
-            step.features = normalization_features
-            step.stats = _pad_evo1_stats(config, step.stats)
-            step.to(device=step.device, dtype=step.dtype)
-    for step in postprocessor.steps:
-        if isinstance(step, UnnormalizerProcessorStep):
-            step.features = action_features
-            step.stats = _pad_evo1_stats(config, step.stats)
-            step.to(device=step.device, dtype=step.dtype)
-
-
 def reconcile_evo1_processors(
     config: Evo1Config,
     preprocessor: PolicyProcessorPipeline,
@@ -336,19 +309,16 @@ def reconcile_evo1_processors(
 ) -> tuple[PolicyProcessorPipeline, PolicyProcessorPipeline]:
     """Reconcile checkpoint-loaded pipelines with the current EVO1 config.
 
-    Three things cannot be restored from a serialized pipeline alone: the EVO1 batch converter
+    Two things cannot be restored from a serialized pipeline alone: the EVO1 batch converter
-    (converters are plain functions and are never serialized), eval-time CLI overrides of the
+    (converters are plain functions and are never serialized), and eval-time CLI overrides of the
-    action postprocessing flags (`postprocess_action_dim`, `binarize_gripper`, `gripper_*`), and the
+    action postprocessing flags (`postprocess_action_dim`, `binarize_gripper`, `gripper_*`). This
-    (un)normalizer stats/features when the generic override path injects raw, unpadded dataset
+    restores the converter and rebuilds the action step from the current config so those overrides
-    stats. This restores the converter, re-pads the normalization stats to EVO1's fixed widths, and
+    take effect.
-    rebuilds the action step from the current config so those overrides take effect.
     """
     # Pipelines reloaded from a checkpoint come back with the default batch converter, which drops
     # non-observation extras (embodiment_id, state_mask, custom task fields) needed by EVO1.
     preprocessor.to_transition = evo1_batch_to_transition
 
-    _refresh_evo1_normalization_steps(config, preprocessor, postprocessor)
-
     action_step = Evo1ActionProcessorStep(
         action_dim=_evo1_action_dim(config),
         binarize_gripper=config.binarize_gripper,

src/lerobot/policies/factory.py

@@ -44,19 +44,12 @@ from lerobot.utils.constants import (
     POLICY_PREPROCESSOR_DEFAULT_NAME,
 )
 from lerobot.utils.feature_utils import dataset_to_policy_features
-from lerobot.utils.import_utils import _peft_available, require_package
 
 from .evo1.configuration_evo1 import Evo1Config
 from .groot.configuration_groot import GrootConfig
 from .pretrained import PreTrainedPolicy
 from .utils import validate_visual_features_consistency
 
-if TYPE_CHECKING or _peft_available:
-    from peft import PeftConfig, PeftModel
-else:
-    PeftConfig = None
-    PeftModel = None
-
 
 def _reconnect_relative_absolute_steps(
     preprocessor: PolicyProcessorPipeline, postprocessor: PolicyProcessorPipeline
@@ -144,6 +137,12 @@ 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
+    # Dataset source used by policies that optionally fit processor artifacts.
+    dataset_repo_id: str | None
+    dataset_root: str | None
+    dataset_revision: str | None
+    dataset_episodes: list[int] | None
+    dataset_exclude_episodes: list[int] | None
     dataset_meta: Any | None
 
 
@@ -178,6 +177,10 @@ def make_pre_post_processors(
         ValueError: If no processor factory exists for the given policy configuration type.
     """
     if pretrained_path:
+        # Register the PI052-only stateful tokenizer step before deserializing its pipeline.
+        if policy_cfg.type == "pi052":
+            from .pi052 import processor_pi052 as _processor_pi052  # noqa: F401
+
         if isinstance(policy_cfg, GrootConfig):
             from .groot.processor_groot import make_groot_pre_post_processors_from_pretrained
 
@@ -197,12 +200,29 @@ def make_pre_post_processors(
                 ),
             )
 
+        preprocessor_overrides = dict(kwargs.get("preprocessor_overrides") or {})
+        if policy_cfg.type == "pi0_fast" and getattr(policy_cfg, "auto_fit_fast_tokenizer", False):
+            from .pi052.fit_fast_tokenizer import resolve_fast_tokenizer
+
+            fitted_tokenizer = resolve_fast_tokenizer(
+                policy_cfg,
+                kwargs.get("dataset_repo_id"),
+                kwargs.get("dataset_root"),
+                kwargs.get("dataset_stats"),
+                kwargs.get("dataset_revision"),
+                kwargs.get("dataset_episodes"),
+                kwargs.get("dataset_exclude_episodes"),
+            )
+            tokenizer_overrides = dict(preprocessor_overrides.get("action_tokenizer_processor") or {})
+            tokenizer_overrides["action_tokenizer_name"] = fitted_tokenizer
+            preprocessor_overrides["action_tokenizer_processor"] = tokenizer_overrides
+
         preprocessor = PolicyProcessorPipeline.from_pretrained(
             pretrained_model_name_or_path=pretrained_path,
             config_filename=kwargs.get(
                 "preprocessor_config_filename", f"{POLICY_PREPROCESSOR_DEFAULT_NAME}.json"
             ),
-            overrides=kwargs.get("preprocessor_overrides", {}),
+            overrides=preprocessor_overrides,
             to_transition=batch_to_transition,
             to_output=transition_to_batch,
             revision=pretrained_revision,
@@ -234,6 +254,11 @@ def make_pre_post_processors(
         config=policy_cfg,
         dataset_stats=kwargs.get("dataset_stats"),
         dataset_meta=kwargs.get("dataset_meta"),
+        dataset_repo_id=kwargs.get("dataset_repo_id"),
+        dataset_root=kwargs.get("dataset_root"),
+        dataset_revision=kwargs.get("dataset_revision"),
+        episodes=kwargs.get("dataset_episodes"),
+        exclude_episodes=kwargs.get("dataset_exclude_episodes"),
     )
 
 
@@ -341,15 +366,12 @@ def make_policy(
         # Load a pretrained PEFT model on top of the policy. The pretrained path points to the folder/repo
         # of the adapter and the adapter's config contains the path to the base policy. So we need the
         # adapter config first, then load the correct policy and then apply PEFT.
-        require_package("peft", extra="peft")
+        from peft import PeftConfig, PeftModel
 
         logging.info("Loading policy's PEFT adapter.")
 
         peft_pretrained_path = str(cfg.pretrained_path)
-        peft_config = PeftConfig.from_pretrained(
+        peft_config = PeftConfig.from_pretrained(peft_pretrained_path)
-            peft_pretrained_path,
-            revision=cfg.pretrained_revision,
-        )
 
         kwargs["pretrained_name_or_path"] = peft_config.base_model_name_or_path
         if not kwargs["pretrained_name_or_path"]:
@@ -360,14 +382,9 @@ def make_policy(
                 "the adapter was trained."
             )
 
-        kwargs["revision"] = peft_config.revision
         policy = policy_cls.from_pretrained(**kwargs)
         policy = PeftModel.from_pretrained(
-            policy,
+            policy, peft_pretrained_path, config=peft_config, is_trainable=True
-            peft_pretrained_path,
-            config=peft_config,
-            revision=cfg.pretrained_revision,
-            is_trainable=True,
         )
 
     else:
@@ -439,6 +456,7 @@ def _make_processors_from_policy_config(
     config: PreTrainedConfig,
     dataset_stats: dict[str, dict[str, torch.Tensor]] | None = None,
     dataset_meta: Any | None = None,
+    **optional_kwargs: Any,
 ) -> tuple[Any, Any]:
     """Create pre- and post-processors from a policy configuration using dynamic imports.
 
@@ -474,7 +492,9 @@ def _make_processors_from_policy_config(
     function = getattr(module, function_name, None)
     if function is None:
         raise ValueError(f"Processor for policy type '{policy_type}' is not implemented.")
+    parameters = inspect.signature(function).parameters
     call_kwargs: dict[str, Any] = {"dataset_stats": dataset_stats}
-    if "dataset_meta" in inspect.signature(function).parameters:
+    if "dataset_meta" in parameters:
         call_kwargs["dataset_meta"] = dataset_meta
+    call_kwargs.update({name: value for name, value in optional_kwargs.items() if name in parameters})
     return function(config, **call_kwargs)

src/lerobot/policies/gaussian_actor/configuration_gaussian_actor.py

@@ -37,19 +37,13 @@ def is_image_feature(key: str) -> bool:
 @dataclass
 class ConcurrencyConfig:
     """Configuration for the concurrency of the actor and learner.
-
     Possible values are:
     - "threads": Use threads for the actor and learner.
     - "processes": Use processes for the actor and learner.
-
-    ``multiprocessing_context`` selects the process-wide start method when
-    processes are used. Set it to ``None`` to preserve Python's default or a
-    method already selected by the embedding application.
     """
 
     actor: str = "threads"
     learner: str = "threads"
-    multiprocessing_context: str | None = "spawn"
 
 
 @dataclass

src/lerobot/policies/molmoact2/modeling_molmoact2.py

@@ -43,22 +43,11 @@ from torch.distributions import Beta
 
 from lerobot.policies.pretrained import PreTrainedPolicy
 from lerobot.utils.constants import ACTION
-from lerobot.utils.import_utils import (
+from lerobot.utils.import_utils import _scipy_available, _transformers_available, require_package
-    _peft_available,
-    _scipy_available,
-    _transformers_available,
-    require_package,
-)
 
 from ..rtc.modeling_rtc import RTCProcessor
 from .configuration_molmoact2 import MolmoAct2Config
 
-if TYPE_CHECKING or _peft_available:
-    from peft import LoraConfig, get_peft_model
-else:
-    LoraConfig = None
-    get_peft_model = None
-
 logger = logging.getLogger(__name__)
 
 
@@ -1742,11 +1731,13 @@ class MolmoAct2Policy(PreTrainedPolicy):
 
     def _build_inner_lora_config(self):
         require_package("peft", extra="molmoact2")
+        from peft import LoraConfig
 
         return LoraConfig(**self._get_inner_peft_targets())
 
     def _apply_lora_adapters(self) -> None:
         require_package("peft", extra="molmoact2")
+        from peft import get_peft_model
 
         peft_config = self._build_inner_lora_config()
         self._validate_peft_config(peft_config)

src/lerobot/policies/pi05/modeling_pi05.py

@@ -22,6 +22,7 @@ from typing import TYPE_CHECKING, Literal, TypedDict, Unpack
 
 import torch
 import torch.nn.functional as F  # noqa: N812
+from safetensors.torch import load_file
 from torch import Tensor, nn
 
 from lerobot.utils.import_utils import _transformers_available, require_package
@@ -30,6 +31,7 @@ from lerobot.utils.import_utils import _transformers_available, require_package
 if TYPE_CHECKING or _transformers_available:
     from transformers.models.auto import CONFIG_MAPPING
     from transformers.models.gemma import modeling_gemma
+    from transformers.utils import cached_file
 
     from ..pi_gemma import (
         PaliGemmaForConditionalGenerationWithPiGemma,
@@ -44,20 +46,21 @@ else:
     _gated_residual = None
     layernorm_forward = None
     PaliGemmaForConditionalGenerationWithPiGemma = None
+    cached_file = None
 from lerobot.configs import PreTrainedConfig
 from lerobot.utils.constants import (
     ACTION,
     OBS_LANGUAGE_ATTENTION_MASK,
     OBS_LANGUAGE_TOKENS,
+    OPENPI_ATTENTION_MASK_VALUE,
 )
 
-from ..common.flow_matching import euler_integrate, sample_noise, sample_time_beta
+from ..common.flow_matching import sample_noise, sample_time_beta
 from ..common.vla_utils import (
     clone_past_key_values,
     create_sinusoidal_pos_embedding,
     make_att_2d_masks,
     pad_vector,
-    prepare_attention_masks_4d,
     resize_with_pad_torch,
 )
 from ..pretrained import PreTrainedPolicy, T
@@ -71,6 +74,9 @@ class ActionSelectKwargs(TypedDict, total=False):
     execution_horizon: int | None
 
 
+_SAFETENSORS_FILE = "model.safetensors"
+
+
 # Define the complete layer computation function for gradient checkpointing
 def compute_layer_complete(inputs_embeds, attention_mask, position_ids, adarms_cond, layers, rotary_emb):
     query_states = []
@@ -401,6 +407,12 @@ class PaliGemmaWithExpertModel(
 class PI05Pytorch(nn.Module):  # see openpi `PI0Pytorch`
     """Core PI05 PyTorch model."""
 
+    use_hf_vision_checkpointing_api = False
+    checkpoint_vision_embeddings = True
+    use_typed_attention_masks = False
+    use_on_device_suffix_mask = False
+    precompute_denoise_times = False
+
     def __init__(self, config: PI05Config, rtc_processor: RTCProcessor | None = None):
         super().__init__()
         self.config = config
@@ -444,7 +456,11 @@ class PI05Pytorch(nn.Module):  # see openpi `PI0Pytorch`
         """Enable gradient checkpointing for memory optimization."""
         self.gradient_checkpointing_enabled = True
         self.paligemma_with_expert.paligemma.model.language_model.gradient_checkpointing = True
-        self.paligemma_with_expert.paligemma.model.vision_tower.gradient_checkpointing = True
+        vision_tower = self.paligemma_with_expert.paligemma.model.vision_tower
+        if self.use_hf_vision_checkpointing_api:
+            vision_tower.gradient_checkpointing_enable(gradient_checkpointing_kwargs={"use_reentrant": False})
+        else:
+            vision_tower.gradient_checkpointing = True
         self.paligemma_with_expert.gemma_expert.model.gradient_checkpointing = True
         logging.info("Enabled gradient checkpointing for PI05Pytorch model")
 
@@ -452,7 +468,11 @@ class PI05Pytorch(nn.Module):  # see openpi `PI0Pytorch`
         """Disable gradient checkpointing."""
         self.gradient_checkpointing_enabled = False
         self.paligemma_with_expert.paligemma.model.language_model.gradient_checkpointing = False
-        self.paligemma_with_expert.paligemma.model.vision_tower.gradient_checkpointing = False
+        vision_tower = self.paligemma_with_expert.paligemma.model.vision_tower
+        if self.use_hf_vision_checkpointing_api:
+            vision_tower.gradient_checkpointing_disable()
+        else:
+            vision_tower.gradient_checkpointing = False
         self.paligemma_with_expert.gemma_expert.model.gradient_checkpointing = False
         logging.info("Disabled gradient checkpointing for PI05Pytorch model")
 
@@ -467,6 +487,14 @@ class PI05Pytorch(nn.Module):  # see openpi `PI0Pytorch`
             )
         return func(*args, **kwargs)
 
+    def _prepare_attention_masks_4d(self, att_2d_masks, dtype=None):
+        """Helper method to prepare 4D attention masks for transformer."""
+        att_2d_masks_4d = att_2d_masks[:, None, :, :]
+        result = torch.where(att_2d_masks_4d, 0.0, OPENPI_ATTENTION_MASK_VALUE)
+        if dtype is not None:
+            result = result.to(dtype=dtype)
+        return result
+
     def sample_noise(self, shape, device):
         return sample_noise(shape, device)
 
@@ -488,13 +516,16 @@ class PI05Pytorch(nn.Module):  # see openpi `PI0Pytorch`
         pad_masks = []
         att_masks = []
 
-        # Process images
+        if self.checkpoint_vision_embeddings:
-        for img, img_mask in zip(images, img_masks, strict=True):
 
-            def image_embed_func(img):
+            def embed_image(img):
-                return self.paligemma_with_expert.embed_image(img)
+                return self._apply_checkpoint(self.paligemma_with_expert.embed_image, img)
 
-            img_emb = self._apply_checkpoint(image_embed_func, img)
+            img_embs = [embed_image(img) for img in images]
+        else:
+            img_embs = [self.paligemma_with_expert.embed_image(img) for img in images]
+
+        for img_emb, img_mask in zip(img_embs, img_masks, strict=True):
             bsize, num_img_embs = img_emb.shape[:2]
 
             embs.append(img_emb)
@@ -556,8 +587,15 @@ class PI05Pytorch(nn.Module):  # see openpi `PI0Pytorch`
 
         # Set attention masks so that image, language and state inputs do not attend to action tokens
         att_masks += [1] + ([0] * (self.config.chunk_size - 1))
-        att_masks = torch.tensor(att_masks, dtype=action_emb.dtype, device=action_emb.device)
+
-        att_masks = att_masks[None, :].expand(bsize, len(att_masks))
+        if self.use_on_device_suffix_mask:
+            n = len(att_masks)
+            att_masks = torch.zeros(n, dtype=action_emb.dtype, device=action_emb.device)
+            att_masks[0] = 1
+            att_masks = att_masks[None, :].expand(bsize, n)
+        else:
+            att_masks = torch.tensor(att_masks, dtype=action_emb.dtype, device=action_emb.device)
+            att_masks = att_masks[None, :].expand(bsize, len(att_masks))
 
         return action_emb, pad_masks, att_masks, adarms_cond
 
@@ -583,7 +621,7 @@ class PI05Pytorch(nn.Module):  # see openpi `PI0Pytorch`
         att_2d_masks = make_att_2d_masks(pad_masks, att_masks)
         position_ids = torch.cumsum(pad_masks, dim=1) - 1
 
-        att_2d_masks_4d = prepare_attention_masks_4d(att_2d_masks)
+        att_2d_masks_4d = self._prepare_attention_masks_4d(att_2d_masks)
 
         def forward_func(prefix_embs, suffix_embs, att_2d_masks_4d, position_ids, adarms_cond):
             (_, suffix_out), _ = self.paligemma_with_expert.forward(
@@ -641,7 +679,8 @@ class PI05Pytorch(nn.Module):  # see openpi `PI0Pytorch`
         prefix_att_2d_masks = make_att_2d_masks(prefix_pad_masks, prefix_att_masks)
         prefix_position_ids = torch.cumsum(prefix_pad_masks, dim=1) - 1
 
-        prefix_att_2d_masks_4d = prepare_attention_masks_4d(prefix_att_2d_masks)
+        mask_dtype = prefix_embs.dtype if self.use_typed_attention_masks else None
+        prefix_att_2d_masks_4d = self._prepare_attention_masks_4d(prefix_att_2d_masks, dtype=mask_dtype)
         self.paligemma_with_expert.paligemma.model.language_model.config._attn_implementation = "eager"  # noqa: SLF001
 
         _, past_key_values = self.paligemma_with_expert.forward(
@@ -652,21 +691,52 @@ class PI05Pytorch(nn.Module):  # see openpi `PI0Pytorch`
             use_cache=True,
         )
 
-        return euler_integrate(
+        dt = -1.0 / num_steps
-            lambda input_x_t, current_timestep: self.denoise_step(
+
-                prefix_pad_masks=prefix_pad_masks,
+        times = None
-                past_key_values=past_key_values,
+        if self.precompute_denoise_times:
-                x_t=input_x_t,
+            times = torch.tensor(
-                timestep=current_timestep,
+                [1.0 + step * dt for step in range(num_steps)], dtype=torch.float32, device=device
-            ),
+            )
-            noise,
+
-            num_steps,
+        x_t = noise
-            rtc_processor=self.rtc_processor,
+        for step in range(num_steps):
-            rtc_enabled=self._rtc_enabled(),
+            time = 1.0 + step * dt
-            inference_delay=kwargs.get("inference_delay"),
+            if times is None:
-            prev_chunk_left_over=kwargs.get("prev_chunk_left_over"),
+                time_tensor = torch.tensor(time, dtype=torch.float32, device=device).expand(bsize)
-            execution_horizon=kwargs.get("execution_horizon"),
+            else:
-        )
+                time_tensor = times[step].expand(bsize)
+
+            def denoise_step_partial_call(input_x_t, current_timestep=time_tensor):
+                return self.denoise_step(
+                    prefix_pad_masks=prefix_pad_masks,
+                    past_key_values=past_key_values,
+                    x_t=input_x_t,
+                    timestep=current_timestep,
+                )
+
+            if self._rtc_enabled():
+                inference_delay = kwargs.get("inference_delay")
+                prev_chunk_left_over = kwargs.get("prev_chunk_left_over")
+                execution_horizon = kwargs.get("execution_horizon")
+
+                v_t = self.rtc_processor.denoise_step(
+                    x_t=x_t,
+                    prev_chunk_left_over=prev_chunk_left_over,
+                    inference_delay=inference_delay,
+                    time=time,
+                    original_denoise_step_partial=denoise_step_partial_call,
+                    execution_horizon=execution_horizon,
+                )
+            else:
+                v_t = denoise_step_partial_call(x_t)
+
+            x_t = x_t + dt * v_t
+
+            if self.rtc_processor is not None and self.rtc_processor.is_debug_enabled():
+                self.rtc_processor.track(time=time, x_t=x_t, v_t=v_t)
+
+        return x_t
 
     def denoise_step(
         self,
@@ -689,7 +759,7 @@ class PI05Pytorch(nn.Module):  # see openpi `PI0Pytorch`
         prefix_offsets = torch.sum(prefix_pad_masks, dim=-1)[:, None]
         position_ids = prefix_offsets + torch.cumsum(suffix_pad_masks, dim=1) - 1
 
-        full_att_2d_masks_4d = prepare_attention_masks_4d(full_att_2d_masks)
+        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 = clone_past_key_values(past_key_values)
@@ -713,6 +783,10 @@ class PI05Policy(PreTrainedPolicy):
 
     config_class = PI05Config
     name = "pi05"
+    model_class = PI05Pytorch
+    eval_after_pretrained_load = False
+    show_openpi_disclaimer = True
+    use_native_pretrained_loader = False
 
     def __init__(
         self,
@@ -730,7 +804,7 @@ class PI05Policy(PreTrainedPolicy):
 
         # Initialize the core PI05 model
         self.init_rtc_processor()
-        self.model = PI05Pytorch(config, rtc_processor=self.rtc_processor)
+        self.model = self.model_class(config, rtc_processor=self.rtc_processor)
 
         # Enable gradient checkpointing if requested
         if config.gradient_checkpointing:
@@ -756,16 +830,31 @@ class PI05Policy(PreTrainedPolicy):
         strict: bool = True,
         **kwargs,
     ) -> T:
-        """Override the from_pretrained method to handle key remapping and display important disclaimer."""
+        """Load a native LeRobot checkpoint or convert the PI05 base checkpoint."""
-        print(
+        if cls.use_native_pretrained_loader:
-            "The PI05 model is a direct port of the OpenPI implementation. \n"
+            return super().from_pretrained(
-            "This implementation follows the original OpenPI structure for compatibility. \n"
+                pretrained_name_or_path,
-            "Original implementation: https://github.com/Physical-Intelligence/openpi"
+                config=config,
-        )
+                force_download=force_download,
+                resume_download=resume_download,
+                proxies=proxies,
+                token=token,
+                cache_dir=cache_dir,
+                local_files_only=local_files_only,
+                revision=revision,
+                strict=strict,
+                **kwargs,
+            )
+
+        if cls.show_openpi_disclaimer:
+            print(
+                "The PI05 model is a direct port of the OpenPI implementation. \n"
+                "This implementation follows the original OpenPI structure for compatibility. \n"
+                "Original implementation: https://github.com/Physical-Intelligence/openpi"
+            )
         if pretrained_name_or_path is None:
             raise ValueError("pretrained_name_or_path is required")
 
-        # Use provided config if available, otherwise create default config
         if config is None:
             config = PreTrainedConfig.from_pretrained(
                 pretrained_name_or_path=pretrained_name_or_path,
@@ -779,85 +868,41 @@ class PI05Policy(PreTrainedPolicy):
                 **kwargs,
             )
 
-        # Initialize model without loading weights
-        # Check if dataset_stats were provided in kwargs
         model = cls(config, **kwargs)
+        model_id = str(pretrained_name_or_path)
+        resolved_file = cached_file(
+            model_id,
+            _SAFETENSORS_FILE,
+            _raise_exceptions_for_missing_entries=False,
+            force_download=force_download,
+            resume_download=resume_download,
+            proxies=proxies,
+            token=token,
+            cache_dir=cache_dir,
+            local_files_only=local_files_only,
+            revision=revision,
+        )
+        if resolved_file is None:
+            raise FileNotFoundError(f"No {_SAFETENSORS_FILE} found in {model_id!r}.")
 
-        # Load state dict (expects keys with "model." prefix)
+        fixed_state_dict = model._fix_pytorch_state_dict_keys(load_file(resolved_file), model.config)
-        try:
+        remapped_state_dict = {
-            print(f"Loading model from: {pretrained_name_or_path}")
+            key if key.startswith("model.") else f"model.{key}": value
-            try:
+            for key, value in fixed_state_dict.items()
-                from transformers.utils import cached_file
+        }
-
+        remapped_state_dict = model._prepare_pretrained_state_dict(remapped_state_dict)
-                resolved_file = cached_file(
+        missing_keys, unexpected_keys = model.load_state_dict(remapped_state_dict, strict=strict)
-                    pretrained_name_or_path,
+        if missing_keys:
-                    "model.safetensors",
+            logging.warning("Missing %s checkpoint keys: %s", cls.name, missing_keys)
-                    cache_dir=kwargs.get("cache_dir"),
+        if unexpected_keys:
-                    force_download=kwargs.get("force_download", False),
+            logging.warning("Unexpected %s checkpoint keys: %s", cls.name, unexpected_keys)
-                    resume_download=kwargs.get("resume_download"),
+        if model.eval_after_pretrained_load:
-                    proxies=kwargs.get("proxies"),
+            model.eval()
-                    token=kwargs.get("token"),
-                    revision=kwargs.get("revision"),
-                    local_files_only=kwargs.get("local_files_only", False),
-                )
-                from safetensors.torch import load_file
-
-                original_state_dict = load_file(resolved_file)
-                print("✓ Loaded state dict from model.safetensors")
-            except Exception as e:
-                print(f"Could not load state dict from remote files: {e}")
-                print("Returning model without loading pretrained weights")
-                return model
-
-            # First, fix any key differences (see openpi model.py, _fix_pytorch_state_dict_keys)
-            fixed_state_dict = model._fix_pytorch_state_dict_keys(original_state_dict, model.config)
-
-            # Then add "model." prefix for all keys that don't already have it
-            remapped_state_dict = {}
-            remap_count = 0
-
-            for key, value in fixed_state_dict.items():
-                if not key.startswith("model."):
-                    new_key = f"model.{key}"
-                    remapped_state_dict[new_key] = value
-                    remap_count += 1
-                else:
-                    remapped_state_dict[key] = value
-
-            if remap_count > 0:
-                print(f"Remapped {remap_count} state dict keys")
-
-            # Load the remapped state dict into the model
-            missing_keys, unexpected_keys = model.load_state_dict(remapped_state_dict, strict=strict)
-
-            if missing_keys:
-                print(f"Missing keys when loading state dict: {len(missing_keys)} keys")
-                if len(missing_keys) <= 5:
-                    for key in missing_keys:
-                        print(f"  - {key}")
-                else:
-                    for key in missing_keys[:5]:
-                        print(f"  - {key}")
-                    print(f"  ... and {len(missing_keys) - 5} more")
-
-            if unexpected_keys:
-                print(f"Unexpected keys when loading state dict: {len(unexpected_keys)} keys")
-                if len(unexpected_keys) <= 5:
-                    for key in unexpected_keys:
-                        print(f"  - {key}")
-                else:
-                    for key in unexpected_keys[:5]:
-                        print(f"  - {key}")
-                    print(f"  ... and {len(unexpected_keys) - 5} more")
-
-            if not missing_keys and not unexpected_keys:
-                print("All keys loaded successfully!")
-
-        except Exception as e:
-            print(f"Warning: Could not load state dict: {e}")
-
         return model
 
+    def _prepare_pretrained_state_dict(self, state_dict: dict[str, Tensor]) -> dict[str, Tensor]:
+        return state_dict
+
     def _fix_pytorch_state_dict_keys(
         self, state_dict, model_config
     ):  # see openpi `BaseModelConfig, _fix_pytorch_state_dict_keys`
@@ -1028,12 +1073,16 @@ class PI05Policy(PreTrainedPolicy):
 
         # Action queue logic for n_action_steps > 1
         if len(self._action_queue) == 0:
-            actions = self.predict_action_chunk(batch)[:, : self.config.n_action_steps]
+            action_batch = self._prepare_action_batch(batch)
+            actions = self.predict_action_chunk(action_batch)[:, : self.config.n_action_steps]
             # Transpose to get shape (n_action_steps, batch_size, action_dim)
             self._action_queue.extend(actions.transpose(0, 1))
 
         return self._action_queue.popleft()
 
+    def _prepare_action_batch(self, batch: dict[str, Tensor]) -> dict[str, Tensor]:
+        return batch
+
     @torch.no_grad()
     def predict_action_chunk(self, batch: dict[str, Tensor], **kwargs: Unpack[ActionSelectKwargs]) -> Tensor:
         """Predict a chunk of actions given environment observations."""

src/lerobot/policies/pi052/__init__.py

@@ -1,6 +1,4 @@
-#!/usr/bin/env python
+# Copyright 2026 The HuggingFace Inc. team. All rights reserved.
-
-# Copyright 2025 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.
@@ -14,14 +12,8 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-"""Unitree G1 locomotion controllers (Groot, Holosoma, SONIC)."""
+"""PI052 configuration; model and processors are imported lazily by their factories."""
 
-from .gr00t_locomotion import GrootLocomotionController
+from .configuration_pi052 import PI052Config
-from .holosoma_locomotion import HolosomaLocomotionController
-from .sonic_whole_body import SonicWholeBodyController
 
-__all__ = [
+__all__ = ["PI052Config"]
-    "GrootLocomotionController",
-    "HolosomaLocomotionController",
-    "SonicWholeBodyController",
-]

src/lerobot/policies/pi052/configuration_pi052.py

@@ -0,0 +1,170 @@
+# 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.
+
+"""PI0.5 with hierarchical text generation and flow-matched actions."""
+
+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):
+    """PI0.5 with recipe-driven text and action supervision."""
+
+    # Recipe / language stack ---------------------------------------------
+    recipe_path: str | None = "recipes/subtask_mem.yaml"
+    """Recipe path, or ``None`` for the plain PI0.5 prompt."""
+
+    apply_chat_template: bool = False
+    """Apply the tokenizer's chat template."""
+
+    # Balance frequent recipe text supervision against the paper's α=10 flow weight.
+    text_loss_weight: float = 1.0
+    """Text cross-entropy weight; ``0`` disables it."""
+
+    flow_loss_weight: float = 10.0
+    """Flow-matching loss weight."""
+
+    # Backbone training ---------------------------------------------------
+    unfreeze_lm_head: bool = True
+    """Train PaliGemma's language head."""
+
+    # Optional context dropout improves tolerance to missing or stale language state.
+    plan_dropout_prob: float = 0.0
+    memory_dropout_prob: float = 0.0
+    subtask_dropout_prob: float = 0.0
+
+    # FAST adds discrete-action CE to the text and flow objectives from paper §III.B-C.
+    enable_fast_action_loss: bool = True
+    """Add FAST action-token cross-entropy."""
+
+    action_tokenizer_name: str = "physical-intelligence/fast"
+    """FAST tokenizer identifier."""
+
+    max_action_tokens: int = 256
+    """Maximum FAST tokens per action chunk."""
+
+    fast_skip_tokens: int = 1152
+    """Reserved vocabulary IDs skipped by FAST token mapping."""
+
+    fast_action_loss_weight: float = 1.0
+    """FAST action-token loss weight."""
+
+    subtask_replan_steps: int = 0
+    """Steps between subtask generations; non-positive replans every chunk."""
+
+    joint_subtask_conditioning: bool = False
+    """Condition actions on the task and generated subtask."""
+
+    auto_fit_fast_tokenizer: bool = False
+    """Fit and cache a dataset-specific FAST tokenizer."""
+
+    fast_tokenizer_cache_dir: str = "~/.cache/lerobot/fast_tokenizers"
+    """Cache directory for fitted FAST tokenizers."""
+
+    fast_tokenizer_fit_samples: int = 1024
+    """Action chunks sampled for tokenizer fitting."""
+
+    fast_tokenizer_validation_samples: int = 256
+    """Held-out chunks used for tokenizer validation."""
+
+    fast_tokenizer_max_reconstruction_rmse: float = 0.10
+    """Maximum validation reconstruction RMSE."""
+
+    fast_tokenizer_max_dim_rmse: float = 0.20
+    """Maximum per-dimension validation RMSE."""
+
+    # Knowledge insulation detaches VLM K/V from action-loss gradients (paper §III.B).
+    knowledge_insulation: bool = True
+    """Detach VLM keys and values from action-loss gradients."""
+
+    # Optional training backends. Defaults preserve the eager/SDPA path.
+    use_flashrt_adarms: bool = False
+    """Use FlashRT adaptive RMSNorm kernels."""
+
+    use_compiled_text_ce: bool = False
+    """Compile text and FAST cross-entropy."""
+
+    use_compiled_vision: bool = False
+    """Compile the SigLIP vision tower."""
+
+    use_flex_attention: bool = False
+    """Use FlexAttention for knowledge insulation."""
+
+    use_manual_attention: bool = False
+    """Use manual attention for profiled KI shapes."""
+
+    manual_attention_scope: str = "all"
+    """Manual-attention scope: ``all`` or ``action``."""
+
+    # Scale language-head updates relative to the base optimizer schedule.
+    lm_head_lr_scale: float = 1.0
+
+    # Scale backbone and action-expert optimizer groups independently.
+    backbone_lr_scale: float = 1.0
+    action_expert_lr_scale: float = 1.0
+
+    # Reuse each VLM prefix across independent denoising draws; 1 restores single-draw flow.
+    flow_num_repeats: int = 5
+
+    # PaLM-style z-loss stabilizes large-vocabulary CE; 0 disables it.
+    text_ce_z_loss_weight: float = 1e-4
+
+    use_flashrt_fp8_mlp: bool = False
+    """Use calibrated FlashRT FP8 MLP kernels."""
+
+    # Keep serialized PI052 AdamW options local because PI05Config lacks them.
+    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__()
+        if self.enable_fast_action_loss and not self.recipe_path:
+            raise ValueError("PI052 FAST action loss requires recipe_path to build action supervision.")
+        if self.text_loss_weight > 0 and self.unfreeze_lm_head:
+            self.train_expert_only = False
+        if self.flow_num_repeats < 1:
+            raise ValueError(f"flow_num_repeats must be >= 1, got {self.flow_num_repeats}")
+        if self.fast_tokenizer_validation_samples < 1:
+            raise ValueError("fast_tokenizer_validation_samples must be >= 1")
+        if self.fast_tokenizer_max_reconstruction_rmse <= 0 or self.fast_tokenizer_max_dim_rmse <= 0:
+            raise ValueError("FAST tokenizer reconstruction thresholds must be positive")
+        if self.manual_attention_scope not in {"all", "action"}:
+            raise ValueError(
+                f"manual_attention_scope must be 'all' or 'action', got {self.manual_attention_scope!r}"
+            )
+        if self.use_flex_attention and self.use_manual_attention:
+            raise ValueError("use_flex_attention and use_manual_attention are mutually exclusive")
+        if self.use_flex_attention and self.flow_num_repeats == 1:
+            raise ValueError("use_flex_attention requires flow_num_repeats > 1")
+        if not self.knowledge_insulation and (
+            self.use_flex_attention or self.use_manual_attention or self.use_flashrt_adarms
+        ):
+            raise ValueError("KI attention and AdaRMS optimizations require knowledge_insulation=True")

src/lerobot/policies/pi052/fit_fast_tokenizer.py

@@ -0,0 +1,522 @@
+# 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.
+
+"""Fit and cache a FAST tokenizer for a dataset's action distribution.
+
+Training invokes this automatically when FAST loss and automatic fitting are enabled.
+"""
+
+from __future__ import annotations
+
+import hashlib
+import json
+import logging
+import os
+import shutil
+import time
+from pathlib import Path
+from typing import Any
+
+import numpy as np
+
+logger = logging.getLogger(__name__)
+
+# ``ProcessorMixin.save_pretrained`` writes this shared cache sentinel.
+_CACHE_SENTINEL = "processor_config.json"
+
+
+def _is_global_leader() -> bool:
+    return int(os.environ.get("RANK", "0")) == 0
+
+
+def _jsonable(value: Any) -> Any:
+    if hasattr(value, "detach"):
+        value = value.detach().cpu().numpy()
+    if isinstance(value, np.ndarray):
+        return value.tolist()
+    if isinstance(value, dict):
+        return {key: _jsonable(item) for key, item in sorted(value.items())}
+    if isinstance(value, (list, tuple)):
+        return [_jsonable(item) for item in value]
+    return value
+
+
+def _dataset_signature(
+    dataset_repo_id: str,
+    base_tokenizer_name: str,
+    n_samples: int,
+    chunk_size: int,
+    normalization_mode: str,
+    dataset_revision: str | None = None,
+    episodes: list[int] | None = None,
+    exclude_episodes: list[int] | None = None,
+    action_stats: dict | None = None,
+    use_relative_actions: bool = False,
+    relative_action_mask: list[bool] | None = None,
+    validation_samples: int = 256,
+    max_reconstruction_rmse: float = 0.10,
+    max_dim_rmse: float = 0.20,
+) -> str:
+    """Hash every input that changes the fitted action distribution."""
+    payload = {
+        "dataset_repo_id": dataset_repo_id,
+        "dataset_revision": dataset_revision,
+        "base_tokenizer_name": base_tokenizer_name,
+        "n_samples": n_samples,
+        "chunk_size": chunk_size,
+        "normalization_mode": normalization_mode,
+        "episodes": episodes,
+        "exclude_episodes": exclude_episodes,
+        "action_stats": action_stats,
+        "use_relative_actions": use_relative_actions,
+        "relative_action_mask": relative_action_mask,
+        "validation_samples": validation_samples,
+        "max_reconstruction_rmse": max_reconstruction_rmse,
+        "max_dim_rmse": max_dim_rmse,
+    }
+    encoded = json.dumps(_jsonable(payload), sort_keys=True, separators=(",", ":")).encode()
+    return hashlib.sha256(encoded).hexdigest()[:16]
+
+
+def _select_episode_indices(
+    available_episodes: list[int],
+    episodes: list[int] | None,
+    exclude_episodes: list[int] | None,
+) -> list[int]:
+    allowed = set(episodes) if episodes is not None else set(available_episodes)
+    excluded = set(exclude_episodes or [])
+    return [episode for episode in available_episodes if episode in allowed and episode not in excluded]
+
+
+def _apply_relative_actions(
+    actions: np.ndarray,
+    states: np.ndarray,
+    relative_action_mask: list[bool] | None,
+) -> np.ndarray:
+    """Match RelativeActionsProcessorStep before tokenizer fitting."""
+    action_dim = actions.shape[-1]
+    mask = list(relative_action_mask) if relative_action_mask is not None else [True] * action_dim
+    if len(mask) < action_dim:
+        mask.extend([True] * (action_dim - len(mask)))
+    mask_array = np.asarray(mask[:action_dim], dtype=np.float32)
+    relative = actions.copy()
+    relative -= states[:, None, :action_dim] * mask_array
+    return relative
+
+
+def _normalize_actions(
+    actions: np.ndarray,
+    normalization_mode: str,
+    action_stats: dict | None = None,
+) -> np.ndarray:
+    """Match the action normalization applied by the training preprocessor."""
+    mode = getattr(normalization_mode, "value", normalization_mode).upper()
+    flat = actions.reshape(-1, actions.shape[-1])
+    stats = action_stats or {}
+
+    def stat(name: str, fallback) -> np.ndarray:
+        value = stats.get(name)
+        if value is None:
+            value = fallback()
+        if hasattr(value, "detach"):
+            value = value.detach().cpu().numpy()
+        return np.asarray(value, dtype=np.float32)
+
+    if mode == "IDENTITY":
+        return actions
+    if mode == "MEAN_STD":
+        mean = stat("mean", lambda: flat.mean(axis=0))
+        std = stat("std", lambda: flat.std(axis=0))
+        return ((actions - mean) / np.where(std == 0, 1e-8, std)).astype(np.float32)
+    if mode in {"QUANTILES", "QUANTILE10"}:
+        low_name, high_name, low_q, high_q = (
+            ("q01", "q99", 0.01, 0.99) if mode == "QUANTILES" else ("q10", "q90", 0.10, 0.90)
+        )
+        low = stat(low_name, lambda: np.quantile(flat, low_q, axis=0))
+        high = stat(high_name, lambda: np.quantile(flat, high_q, axis=0))
+    elif mode == "MIN_MAX":
+        low = stat("min", lambda: flat.min(axis=0))
+        high = stat("max", lambda: flat.max(axis=0))
+    else:
+        raise ValueError(f"Unsupported FAST tokenizer normalization mode: {mode}")
+
+    return (2.0 * (actions - low) / np.where(high == low, 1e-8, high - low) - 1.0).astype(np.float32)
+
+
+def _validate_fast_reconstruction(
+    tokenizer: Any,
+    actions: np.ndarray,
+    max_reconstruction_rmse: float,
+    max_dim_rmse: float,
+) -> tuple[dict[str, Any], np.ndarray]:
+    """Decode held-out chunks and reject tokenizers with excessive quantization error."""
+    decoded = np.asarray(tokenizer.decode(tokenizer(actions)), dtype=np.float32)
+    if decoded.shape != actions.shape:
+        raise RuntimeError(
+            f"FAST tokenizer reconstruction shape mismatch: expected {actions.shape}, got {decoded.shape}."
+        )
+    if not np.isfinite(decoded).all():
+        raise RuntimeError("FAST tokenizer reconstruction contains non-finite values.")
+
+    squared_error = np.square(decoded - actions)
+    rmse = float(np.sqrt(squared_error.mean()))
+    dim_rmse = np.sqrt(squared_error.mean(axis=(0, 1)))
+    nonconstant_dims = np.ptp(actions, axis=(0, 1)) > 1e-8
+    max_observed_dim_rmse = float(dim_rmse[nonconstant_dims].max(initial=0.0))
+    report = {
+        "num_validation_chunks": int(actions.shape[0]),
+        "reconstruction_rmse": rmse,
+        "max_dim_rmse": max_observed_dim_rmse,
+        "dim_rmse": dim_rmse.tolist(),
+        "max_reconstruction_rmse": max_reconstruction_rmse,
+        "max_allowed_dim_rmse": max_dim_rmse,
+    }
+    if rmse > max_reconstruction_rmse or max_observed_dim_rmse > max_dim_rmse:
+        raise RuntimeError(
+            "FAST tokenizer reconstruction error exceeds the configured limit: "
+            f"rmse={rmse:.4f} (max {max_reconstruction_rmse:.4f}), "
+            f"max_dim_rmse={max_observed_dim_rmse:.4f} (max {max_dim_rmse:.4f})."
+        )
+    return report, decoded
+
+
+def _load_fast_fitter(base_tokenizer_name: str) -> Any:
+    """Load FAST's fitting implementation without requiring its universal BPE weights."""
+    from transformers import AutoProcessor  # noqa: PLC0415
+
+    try:
+        return AutoProcessor.from_pretrained(base_tokenizer_name, trust_remote_code=True)
+    except ValueError as error:
+        if base_tokenizer_name != "physical-intelligence/fast":
+            raise
+        logger.warning(
+            "Could not load the universal FAST tokenizer backend; loading its fitting class directly: %s",
+            error,
+        )
+        from transformers.dynamic_module_utils import get_class_from_dynamic_module  # noqa: PLC0415
+
+        return get_class_from_dynamic_module(
+            "processing_action_tokenizer.UniversalActionProcessor",
+            base_tokenizer_name,
+        )
+
+
+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,
+    dataset_root: str | Path | None = None,
+    dataset_revision: str | None = None,
+    episodes: list[int] | None = None,
+    exclude_episodes: list[int] | None = None,
+    normalization_mode: str = "QUANTILES",
+    action_stats: dict | None = None,
+    use_relative_actions: bool = False,
+    relative_action_mask: list[bool] | None = None,
+    validation_samples: int = 256,
+    max_reconstruction_rmse: float = 0.10,
+    max_dim_rmse: float = 0.20,
+) -> 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)
+    normalization_mode = getattr(normalization_mode, "value", normalization_mode).upper()
+    sig = _dataset_signature(
+        dataset_repo_id,
+        base_tokenizer_name,
+        n_samples,
+        chunk_size,
+        normalization_mode,
+        dataset_revision,
+        episodes,
+        exclude_episodes,
+        action_stats,
+        use_relative_actions,
+        relative_action_mask,
+        validation_samples,
+        max_reconstruction_rmse,
+        max_dim_rmse,
+    )
+    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)
+
+    # One global rank populates the shared cache; every other rank waits for the atomic publish.
+    is_leader = _is_global_leader()
+    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,
+    )
+
+    # Read action columns directly to avoid video decoding and bound memory to sampled episodes.
+    rng = np.random.default_rng(seed)
+    actions_buf: list[np.ndarray] = []
+
+    # Read v3 parquet shards directly to avoid split lookup failures and repeated metadata parsing.
+    import pyarrow as _pa  # noqa: PLC0415
+    import pyarrow.parquet as _pq  # noqa: PLC0415
+
+    if dataset_root is not None:
+        snap = Path(dataset_root)
+    else:
+        from huggingface_hub import snapshot_download  # noqa: PLC0415
+
+        snap = Path(
+            snapshot_download(repo_id=dataset_repo_id, repo_type="dataset", revision=dataset_revision)
+        )
+    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}.")
+
+    columns = ["episode_index", "action"]
+    if use_relative_actions:
+        columns.append("observation.state")
+    tables = [_pq.read_table(f, columns=columns) for f in data_files]
+    table = _pa.concat_tables(tables)
+    eps = table["episode_index"].to_numpy()
+    acts_col = table["action"]
+    # Normalize Arrow action representations into an (N, D) array.
+    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}.")
+    states = None
+    if use_relative_actions:
+        try:
+            states = np.stack(table["observation.state"].to_numpy(zero_copy_only=False)).astype(np.float32)
+        except Exception:  # noqa: BLE001
+            states = np.asarray(table["observation.state"].to_pylist(), dtype=np.float32)
+        if states.ndim != 2:
+            raise RuntimeError(
+                f"FAST fit: expected ``observation.state`` rows to be 1-D vectors; got {states.shape}."
+            )
+
+    # Sort once because episode order is only guaranteed within each shard.
+    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]
+    if states is not None:
+        states = states[order]
+
+    ep_indices = _select_episode_indices(list(ep_to_slice), episodes, exclude_episodes)
+    if not ep_indices:
+        raise RuntimeError("FAST fit: episode selection is empty after applying exclusions.")
+    total_samples = n_samples + validation_samples
+    samples_per_episode = max(1, (total_samples + len(ep_indices) - 1) // len(ep_indices))
+    collected = 0
+    eps_visited = 0
+    short_episodes = 0
+    states_buf: list[np.ndarray] = []
+    for ep_idx in rng.permutation(ep_indices):
+        if collected >= total_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])
+            if states is not None:
+                states_buf.append(states[start + int(s)])
+            collected += 1
+            if collected >= total_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)
+    if states is not None:
+        actions = _apply_relative_actions(actions, np.stack(states_buf), relative_action_mask)
+    logger.info(
+        "FAST fit: collected %d chunks of shape %s from %d episodes",
+        actions.shape[0],
+        actions.shape[1:],
+        eps_visited,
+    )
+
+    actions = _normalize_actions(actions, normalization_mode, action_stats)
+
+    base = _load_fast_fitter(base_tokenizer_name)
+    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."
+        )
+
+    if actions.shape[0] < total_samples:
+        raise RuntimeError(
+            f"FAST fit collected {actions.shape[0]} chunks, but {total_samples} are required "
+            f"for {n_samples} fit and {validation_samples} validation chunks."
+        )
+    fit_actions = actions[:n_samples]
+    validation_actions = actions[n_samples:total_samples]
+    fitted = base.fit(fit_actions)
+    validation_report, decoded_actions = _validate_fast_reconstruction(
+        fitted,
+        validation_actions,
+        max_reconstruction_rmse,
+        max_dim_rmse,
+    )
+    cache_dir.mkdir(parents=True, exist_ok=True)
+    staging_dir = cache_dir / f".{sig}.tmp-{os.getpid()}"
+    shutil.rmtree(staging_dir, ignore_errors=True)
+    fitted.save_pretrained(str(staging_dir))
+    (staging_dir / "reconstruction_validation.json").write_text(
+        json.dumps(validation_report, indent=2) + "\n"
+    )
+    np.savez_compressed(
+        staging_dir / "reconstruction_examples.npz",
+        original=validation_actions[:8],
+        decoded=decoded_actions[:8],
+    )
+    if out_dir.exists():
+        shutil.rmtree(out_dir)
+    staging_dir.replace(out_dir)
+    logger.info("FAST fit: saved fitted tokenizer to %s", out_dir)
+    return str(out_dir)
+
+
+def resolve_fast_tokenizer(
+    config: Any,
+    dataset_repo_id: str | None,
+    dataset_root: str | Path | None = None,
+    dataset_stats: dict | None = None,
+    dataset_revision: str | None = None,
+    episodes: list[int] | None = None,
+    exclude_episodes: list[int] | None = None,
+) -> str:
+    """Return the configured tokenizer, fitting a cached dataset-specific one when requested."""
+    if not getattr(config, "auto_fit_fast_tokenizer", False) or dataset_repo_id is None:
+        return config.action_tokenizer_name
+
+    relative_action_mask = None
+    if getattr(config, "use_relative_actions", False):
+        action_names = getattr(config, "action_feature_names", None)
+        exclude_tokens = [
+            str(name).lower() for name in getattr(config, "relative_exclude_joints", []) if name
+        ]
+        if action_names is not None and exclude_tokens:
+            relative_action_mask = [
+                not any(token == str(name).lower() or token in str(name).lower() for token in exclude_tokens)
+                for name in action_names
+            ]
+
+    fit_kwargs = {
+        "dataset_repo_id": dataset_repo_id,
+        "cache_dir": Path(config.fast_tokenizer_cache_dir).expanduser(),
+        "base_tokenizer_name": config.action_tokenizer_name,
+        "n_samples": config.fast_tokenizer_fit_samples,
+        "chunk_size": config.chunk_size,
+        "dataset_root": dataset_root,
+        "dataset_revision": dataset_revision,
+        "episodes": episodes,
+        "exclude_episodes": exclude_episodes,
+        "normalization_mode": config.normalization_mapping.get("ACTION", "QUANTILES"),
+        "action_stats": (dataset_stats or {}).get("action"),
+        "use_relative_actions": getattr(config, "use_relative_actions", False),
+        "relative_action_mask": relative_action_mask,
+    }
+    validation_fields = {
+        "validation_samples": "fast_tokenizer_validation_samples",
+        "max_reconstruction_rmse": "fast_tokenizer_max_reconstruction_rmse",
+        "max_dim_rmse": "fast_tokenizer_max_dim_rmse",
+    }
+    fit_kwargs.update(
+        {
+            argument: getattr(config, attribute)
+            for argument, attribute in validation_fields.items()
+            if hasattr(config, attribute)
+        }
+    )
+    return fit_fast_tokenizer(**fit_kwargs)

src/lerobot/policies/pi052/flashrt_fp8.py

@@ -0,0 +1,263 @@
+# 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.
+
+"""Optional FlashRT FP8 MLP kernels with one-pass calibration and BF16 fallback."""
+
+from __future__ import annotations
+
+import logging
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F  # noqa: N812
+
+logger = logging.getLogger(__name__)
+
+_FP8_MAX = 448.0
+
+
+def _roundtrip_fp8(x: torch.Tensor, scale: torch.Tensor) -> torch.Tensor:
+    """Quantize->dequantize an activation through FP8 E4M3 at ``scale`` (f32)."""
+    q = torch.clamp(x.float() / scale.float(), -_FP8_MAX, _FP8_MAX).to(torch.float8_e4m3fn)
+    return q.float() * scale.float()
+
+
+_SWIGLU_REPO = "flashrt/flashrt-fp8-swiglu-ffn"
+_GELU_REPO = "flashrt/flashrt-fp8-ffn"
+_GEMM_REPO = "flashrt/flashrt-gemm-epilogues"
+
+
+def _get_kernel(repo: str):
+    """Load a cached FlashRT Hub package."""
+    from kernels import get_kernel
+
+    return get_kernel(repo, version=1)
+
+
+def _quantize_fp8(weight: torch.Tensor) -> tuple[torch.Tensor, torch.Tensor]:
+    scale = max(weight.detach().float().abs().max().item(), 1e-12) / _FP8_MAX
+    fp8 = torch.clamp(weight.float() / scale, -_FP8_MAX, _FP8_MAX).to(torch.float8_e4m3fn)
+    return fp8.contiguous(), torch.tensor([scale], dtype=torch.float32)
+
+
+def _static_scale(amax: float, safety: float) -> torch.Tensor:
+    return torch.tensor([max(amax, 1e-12) / _FP8_MAX * safety], dtype=torch.float32)
+
+
+class _FlashRTGeGLU(nn.Module):
+    """FP8 Gemma GeGLU MLP."""
+
+    def __init__(self, mlp, in_amax, hid_amax, ffn_ops, quant_ops, safety, fuse_weight=None):
+        super().__init__()
+        self.ffn_ops = ffn_ops
+        self.quant_ops = quant_ops
+        self.in_features = mlp.gate_proj.weight.shape[1]
+        device = mlp.gate_proj.weight.device
+        gate_up = torch.cat([mlp.gate_proj.weight, mlp.up_proj.weight], dim=0).float()
+        # Fold fixed RMSNorm weights into GEMM; adaptive norms use identity scaling.
+        if fuse_weight is not None:
+            f = 1.0 + fuse_weight.detach().float()
+            gate_up = gate_up * f[None, :]
+            channel_scale = (1.0 / f).to(torch.bfloat16)
+        else:
+            channel_scale = torch.ones(self.in_features, dtype=torch.bfloat16)
+        gate_up_fp8, gate_up_scale = _quantize_fp8(gate_up)
+        down_fp8, down_scale = _quantize_fp8(mlp.down_proj.weight)
+        self.register_buffer("gate_up_fp8", gate_up_fp8.to(device))
+        self.register_buffer("down_fp8", down_fp8.to(device))
+        self.register_buffer("gate_up_scale", gate_up_scale.to(device))
+        self.register_buffer("down_scale", down_scale.to(device))
+        self.register_buffer("input_scale", _static_scale(in_amax, safety).to(device))
+        self.register_buffer("hidden_scale", _static_scale(hid_amax, safety).to(device))
+        self.register_buffer("channel_scale", channel_scale.to(device))
+        self.safety = safety
+        self.calibrating = False
+        self._ia = 0.0
+        self._ha = 0.0
+
+    def _calibrate_step(self, x):
+        # Track input and hidden maxima on live FP8-propagated activations.
+        flat = x.reshape(-1, self.in_features).to(torch.bfloat16)
+        xq = flat.float() * self.channel_scale.float()
+        self._ia = max(self._ia, xq.abs().max().item())
+        self.input_scale.copy_(_static_scale(self._ia, self.safety).to(self.input_scale.device))
+        xdq = _roundtrip_fp8(xq, self.input_scale)
+        wdq = self.gate_up_fp8.float() * self.gate_up_scale.float()
+        gate, up = (xdq @ wdq.t()).chunk(2, dim=-1)
+        hidden = F.gelu(gate, approximate="tanh") * up
+        self._ha = max(self._ha, hidden.abs().max().item())
+        self.hidden_scale.copy_(_static_scale(self._ha, self.safety).to(self.hidden_scale.device))
+
+    def forward(self, x):
+        if self.calibrating:
+            self._calibrate_step(x)
+        shape = x.shape
+        flat = x.reshape(-1, self.in_features).to(torch.bfloat16)
+        x_fp8 = self.quant_ops.channel_scale_quantize_fp8_static_bf16(
+            flat, self.channel_scale, self.input_scale
+        )
+        out = self.ffn_ops.fp8_geglu_mlp_bf16(
+            x_fp8,
+            self.gate_up_fp8,
+            self.down_fp8,
+            self.input_scale,
+            self.gate_up_scale,
+            self.hidden_scale,
+            self.down_scale,
+        )
+        return out.reshape(shape)
+
+
+class _FlashRTGeluMLP(nn.Module):
+    """FP8 SigLIP GELU MLP."""
+
+    def __init__(self, mlp, in_amax, hid_amax, ffn_ops, quant_ops, safety):
+        super().__init__()
+        self.ffn_ops = ffn_ops
+        self.quant_ops = quant_ops
+        self.in_features = mlp.fc1.weight.shape[1]
+        self.out_features = mlp.fc2.weight.shape[0]
+        device = mlp.fc1.weight.device
+        up_fp8, up_scale = _quantize_fp8(mlp.fc1.weight)
+        down_fp8, down_scale = _quantize_fp8(mlp.fc2.weight)
+        self.register_buffer("up_fp8", up_fp8.to(device))
+        self.register_buffer("down_fp8", down_fp8.to(device))
+        self.register_buffer("up_scale", up_scale.to(device))
+        self.register_buffer("down_scale", down_scale.to(device))
+        self.register_buffer("up_bias", mlp.fc1.bias.detach().to(torch.bfloat16))
+        self.register_buffer("down_bias", mlp.fc2.bias.detach().to(torch.bfloat16))
+        self.register_buffer("input_scale", _static_scale(in_amax, safety).to(device))
+        self.register_buffer("hidden_scale", _static_scale(hid_amax, safety).to(device))
+        self.register_buffer(
+            "channel_scale", torch.ones(self.in_features, device=device, dtype=torch.bfloat16)
+        )
+        self.safety = safety
+        self.calibrating = False
+        self._ia = 0.0
+        self._ha = 0.0
+
+    def _calibrate_step(self, x):
+        flat = x.reshape(-1, self.in_features).to(torch.bfloat16)
+        self._ia = max(self._ia, flat.float().abs().max().item())
+        self.input_scale.copy_(_static_scale(self._ia, self.safety).to(self.input_scale.device))
+        xdq = _roundtrip_fp8(flat.float(), self.input_scale)
+        hid = (xdq @ (self.up_fp8.float() * self.up_scale.float()).t()) + self.up_bias.float()
+        hid = F.gelu(hid, approximate="tanh")
+        self._ha = max(self._ha, hid.abs().max().item())
+        self.hidden_scale.copy_(_static_scale(self._ha, self.safety).to(self.hidden_scale.device))
+
+    def forward(self, x):
+        if self.calibrating:
+            self._calibrate_step(x)
+        shape = x.shape
+        dtype = x.dtype
+        flat = x.reshape(-1, self.in_features).to(torch.bfloat16)
+        x_fp8 = self.quant_ops.channel_scale_quantize_fp8_static_bf16(
+            flat, self.channel_scale, self.input_scale
+        )
+        out = self.ffn_ops.fp8_gelu_mlp_bf16(
+            x_fp8,
+            self.up_fp8,
+            self.up_bias,
+            self.down_fp8,
+            self.down_bias,
+            self.input_scale,
+            self.up_scale,
+            self.hidden_scale,
+            self.down_scale,
+        )
+        return out.reshape(*shape[:-1], self.out_features).to(dtype)
+
+
+def _siglip_mlps(model) -> list:
+    tower = model.paligemma_with_expert.paligemma.model.vision_tower
+    return [m for _, m in tower.named_modules() if type(m).__name__ == "SiglipMLP"]
+
+
+def _run_forward(policy, batches) -> None:
+    """Run eager action prediction so calibration reaches Python module forwards."""
+    model = policy.model
+    saved = {name: vars(model).pop(name) for name in ("sample_actions", "forward") if name in vars(model)}
+    with torch.inference_mode():
+        for batch in batches:
+            policy.predict_action_chunk(
+                {k: (v.clone() if torch.is_tensor(v) else v) for k, v in batch.items()}
+            )
+    torch.cuda.synchronize()
+    vars(model).update(saved)
+
+
+def _fixed_norm_weight(norm):
+    """Return a fixed RMSNorm fold weight, or ``None`` for adaptive norms."""
+    return norm.weight if getattr(norm, "dense", None) is None else None
+
+
+def _fp8_supported(device) -> bool:
+    """Return whether the device supports FP8 E4M3 tensor cores (CUDA SM >= 8.9)."""
+    if device.type != "cuda" or not torch.cuda.is_available():
+        return False
+    major, minor = torch.cuda.get_device_capability(device)
+    return (major, minor) >= (8, 9)
+
+
+def apply_fp8_mlp(policy, batch, *, safety: float = 1.05) -> bool:
+    """Replace Gemma and SigLIP MLPs with FlashRT FP8 kernels calibrated on the supplied batch.
+
+    Returns ``False`` without modifying BF16 execution when FP8 or its kernels are unavailable.
+    """
+    device = next(policy.parameters()).device
+    if not _fp8_supported(device):
+        logger.warning(
+            "PI052: device %s has no FP8 (E4M3) support (needs CUDA SM>=8.9); keeping BF16.",
+            device,
+        )
+        return False
+    batches = batch if isinstance(batch, (list, tuple)) else [batch]
+    try:
+        ffn_ops = _get_kernel(_SWIGLU_REPO)
+        gelu_ops = _get_kernel(_GELU_REPO)
+        quant_ops = _get_kernel(_GEMM_REPO)
+    except Exception as exc:  # noqa: BLE001
+        logger.warning("PI052: FlashRT FP8 kernels unavailable (%s); keeping BF16.", exc)
+        return False
+
+    model = policy.model
+    calibrating = []
+
+    gemma_layers = list(model.paligemma_with_expert.gemma_expert.model.layers) + list(
+        model.paligemma_with_expert.paligemma.model.language_model.layers
+    )
+    for layer in gemma_layers:
+        fw = _fixed_norm_weight(layer.post_attention_layernorm)
+        layer.mlp = _FlashRTGeGLU(layer.mlp, 1.0, 1.0, ffn_ops, quant_ops, safety, fuse_weight=fw).to(device)
+        calibrating.append(layer.mlp)
+
+    siglip = _siglip_mlps(model)
+    for mlp_parent in model.paligemma_with_expert.paligemma.model.vision_tower.vision_model.encoder.layers:
+        mlp_parent.mlp = _FlashRTGeluMLP(mlp_parent.mlp, 1.0, 1.0, gelu_ops, quant_ops, safety).to(device)
+        calibrating.append(mlp_parent.mlp)
+
+    # Calibrate every swapped module in one FP8-propagated forward.
+    for m in calibrating:
+        m.calibrating = True
+    _run_forward(policy, batches)
+    for m in calibrating:
+        m.calibrating = False
+
+    logger.info(
+        "PI052: FlashRT FP8 enabled (%d Gemma + %d SigLIP MLPs).",
+        len(gemma_layers),
+        len(siglip),
+    )
+    return True

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

@@ -0,0 +1,19 @@
+# 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 adapter for the policy-agnostic language runtime."""
+
+from .pi052_adapter import PI052PolicyAdapter
+
+__all__ = ["PI052PolicyAdapter"]

src/lerobot/policies/pi052/inference/pi052_adapter.py

@@ -0,0 +1,254 @@
+# 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 actions and text generation for the generic language runtime."""
+
+from __future__ import annotations
+
+import logging
+from typing import Any
+
+from lerobot.runtime import RuntimeState
+from lerobot.runtime.adapter import BaseLanguageAdapter
+
+logger = logging.getLogger(__name__)
+
+_LOC_TOKENIZER_CACHE: dict[str, Any] = {}
+
+
+class PI052PolicyAdapter(BaseLanguageAdapter):
+    """Runtime bridge for PI052 policies."""
+
+    def select_action(self, observation: dict[str, Any], state: RuntimeState) -> Any:
+        import torch  # noqa: PLC0415
+
+        from lerobot.utils.constants import (  # noqa: PLC0415
+            OBS_LANGUAGE_ATTENTION_MASK,
+            OBS_LANGUAGE_TOKENS,
+            OBS_STATE,
+        )
+
+        subtask = state.language_context.get("subtask") or state.task or ""
+        # Match the training prompt by conditioning on both subtask and discretized state.
+        state_str = None
+        obs_state = observation.get(OBS_STATE)
+        if isinstance(obs_state, torch.Tensor) and obs_state.numel() > 0:
+            from lerobot.policies.pi052.text_processor_pi052 import discretize_state_str  # noqa: PLC0415
+
+            state_row = obs_state[0] if obs_state.ndim > 1 else obs_state
+            state_str = discretize_state_str(state_row)
+
+        batch = dict(observation)
+        if getattr(self.policy.config, "joint_subtask_conditioning", False):
+            # Joint sequences keep the task turn (with state) and render the
+            # subtask as a causal assistant turn, exactly as trained.
+            from transformers import AutoTokenizer  # noqa: PLC0415
+
+            from lerobot.policies.pi052.text_processor_pi052 import (  # noqa: PLC0415
+                encode_prompt_with_targets,
+                register_paligemma_loc_tokens,
+            )
+            from lerobot.utils.constants import OBS_LANGUAGE_CAUSAL_MARKS  # noqa: PLC0415
+
+            task = state.task or ""
+            task_content = task if state_str is None else f"{task}, State: {state_str};"
+            tok_name = getattr(self.policy.config, "tokenizer_name", None) or "google/paligemma-3b-pt-224"
+            tokenizer = _get_loc_tokenizer(tok_name, AutoTokenizer, register_paligemma_loc_tokens)
+            ids, attn, marks = encode_prompt_with_targets(
+                tokenizer,
+                [
+                    {"role": "user", "content": task_content},
+                    {"role": "assistant", "content": subtask},
+                ],
+                target_indices=[1],
+            )
+            device = getattr(self.policy.config, "device", None)
+            if device is not None:
+                ids, attn, marks = ids.to(device), attn.to(device), marks.to(device)
+            batch[OBS_LANGUAGE_TOKENS] = ids
+            batch[OBS_LANGUAGE_ATTENTION_MASK] = attn
+            batch[OBS_LANGUAGE_CAUSAL_MARKS] = marks
+        else:
+            content = subtask if state_str is None else f"{subtask}, State: {state_str};"
+            text_batch = _build_text_batch(
+                self.policy,
+                [{"role": "user", "content": content}],
+                add_generation_prompt=False,
+            )
+            batch[OBS_LANGUAGE_TOKENS] = text_batch["lang_tokens"]
+            batch[OBS_LANGUAGE_ATTENTION_MASK] = text_batch["lang_masks"]
+        return self.policy.predict_action_chunk(batch)
+
+    def generate_text(
+        self,
+        kind: str,
+        observation: dict[str, Any] | None,
+        state: RuntimeState,
+        user_text: str | None = None,
+    ) -> str:
+        messages = self.build_messages(kind, state, user_text=user_text)
+        if kind == "subtask" and getattr(self.policy.config, "joint_subtask_conditioning", False):
+            # Joint samples carry state on the task turn, so the subtask must be
+            # generated from the same state-bearing prompt.
+            import torch  # noqa: PLC0415
+
+            from lerobot.policies.pi052.text_processor_pi052 import discretize_state_str  # noqa: PLC0415
+            from lerobot.utils.constants import OBS_STATE  # noqa: PLC0415
+
+            obs_state = (observation or {}).get(OBS_STATE)
+            if isinstance(obs_state, torch.Tensor) and obs_state.numel() > 0:
+                state_row = obs_state[0] if obs_state.ndim > 1 else obs_state
+                for m in reversed(messages):
+                    if m.get("role") == "user":
+                        m["content"] = f"{m.get('content', '')}, State: {discretize_state_str(state_row)};"
+                        break
+        return _generate_with_policy(
+            self.policy,
+            messages,
+            observation=observation,
+            state=state,
+            label=f"{kind} gen",
+            min_new_tokens=self.gen.min_new_tokens,
+            temperature=self.gen.temperature,
+            top_p=self.gen.top_p,
+            suppress_loc_tokens=True,  # all runtime text is prose; never emit <loc>
+        )
+
+    def build_messages(
+        self,
+        kind: str,
+        state: RuntimeState,
+        *,
+        user_text: str | None = None,
+    ) -> list[dict[str, Any]]:
+        if kind in ("subtask", "plan"):
+            return [{"role": "user", "content": state.task or ""}]
+        if kind == "memory":
+            messages = [{"role": "user", "content": state.task or ""}]
+            if state.language_context.get("memory"):
+                messages.append(
+                    {"role": "assistant", "content": f"Previous memory: {state.language_context['memory']}"}
+                )
+            if state.extra.get("prior_subtask"):
+                messages.append(
+                    {"role": "user", "content": f"Completed subtask: {state.extra['prior_subtask']}"}
+                )
+            return messages
+        if kind == "interjection":
+            messages = [{"role": "user", "content": state.task or ""}]
+            if state.language_context.get("plan"):
+                messages.append(
+                    {"role": "assistant", "content": f"Previous plan:\n{state.language_context['plan']}"}
+                )
+            if user_text:
+                messages.append({"role": "user", "content": user_text})
+            return messages
+        raise ValueError(f"Unknown PI052 text kind: {kind}")
+
+
+def _get_loc_tokenizer(tok_name: str, auto_tokenizer_cls: Any, register_loc_fn: Any) -> Any:
+    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]:
+    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"
+    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:
+        # No trailing space: SentencePiece folds it into the first target token
+        # ("▁move"), so a space-suffixed prefill ends in a lone "▁" the model
+        # never saw at this position during training.
+        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()
+
+    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 _generate_with_policy(
+    policy: Any,
+    messages: list[dict[str, Any]],
+    *,
+    observation: dict[str, Any] | None = None,
+    state: RuntimeState | 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:
+    if not hasattr(policy, "select_message"):
+        if state is not None:
+            state.log(f"  [warn] policy has no select_message — skipping {label}")
+        return ""
+    text_batch = _build_text_batch(policy, messages)
+    try:
+        from lerobot.utils.constants import OBS_LANGUAGE_ATTENTION_MASK, OBS_LANGUAGE_TOKENS  # noqa: PLC0415
+
+        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
+        return policy.select_message(
+            batch,
+            tokenizer=text_batch["tokenizer"],
+            min_new_tokens=min_new_tokens,
+            temperature=temperature,
+            top_p=top_p,
+            suppress_loc_tokens=suppress_loc_tokens,
+        )
+    except Exception as exc:  # noqa: BLE001
+        logger.warning("%s failed: %s", label, exc, exc_info=logger.isEnabledFor(logging.DEBUG))
+        if state is not None:
+            state.log(f"  [warn] {label} failed: {type(exc).__name__}: {exc}")
+        return ""

src/lerobot/policies/pi052/processor_pi052.py

@@ -0,0 +1,164 @@
+# 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 processor factory with optional recipe rendering and text tokenization.
+
+Without a recipe it delegates to the standard PI0.5 pipeline.
+"""
+
+from __future__ import annotations
+
+from pathlib import Path
+from typing import Any
+
+import torch
+
+from lerobot.configs.recipe import TrainingRecipe
+from lerobot.processor import (
+    AbsoluteActionsProcessorStep,
+    ActionTokenizerProcessorStep,
+    AddBatchDimensionProcessorStep,
+    DeviceProcessorStep,
+    NormalizerProcessorStep,
+    PolicyAction,
+    PolicyProcessorPipeline,
+    RelativeActionsProcessorStep,
+    RenameObservationsProcessorStep,
+    UnnormalizerProcessorStep,
+    policy_action_to_transition,
+    transition_to_policy_action,
+)
+
+# Import directly to keep optional language dependencies out of ``lerobot.processor``.
+from lerobot.processor.render_messages_processor import RenderMessagesStep
+from lerobot.utils.constants import POLICY_POSTPROCESSOR_DEFAULT_NAME, POLICY_PREPROCESSOR_DEFAULT_NAME
+
+from ..pi05.processor_pi05 import make_pi05_pre_post_processors
+from .configuration_pi052 import PI052Config
+from .text_processor_pi052 import PI052TextTokenizerStep
+
+
+def make_pi052_pre_post_processors(
+    config: PI052Config,
+    dataset_stats: dict[str, dict[str, torch.Tensor]] | None = None,
+    dataset_repo_id: str | None = None,
+    dataset_root: str | None = None,
+    dataset_revision: str | None = None,
+    episodes: list[int] | None = None,
+    exclude_episodes: list[int] | None = None,
+) -> tuple[
+    PolicyProcessorPipeline[dict[str, Any], dict[str, Any]],
+    PolicyProcessorPipeline[PolicyAction, PolicyAction],
+]:
+    """Build PI0.5-v2's pre/post-processor pipelines.
+
+    Falls through to π0.5's stock pipeline when ``recipe_path`` is unset.
+    """
+    if not config.recipe_path:
+        if getattr(config, "enable_fast_action_loss", False):
+            raise ValueError("PI052 FAST action loss requires recipe_path to build action supervision.")
+        return make_pi05_pre_post_processors(config, dataset_stats=dataset_stats)
+
+    recipe = _load_recipe(config.recipe_path)
+
+    relative_step = RelativeActionsProcessorStep(
+        enabled=config.use_relative_actions,
+        exclude_joints=getattr(config, "relative_exclude_joints", []),
+        action_names=getattr(config, "action_feature_names", None),
+    )
+
+    input_steps = [
+        RenameObservationsProcessorStep(rename_map={}),
+        AddBatchDimensionProcessorStep(),
+        relative_step,
+        NormalizerProcessorStep(
+            features={**config.input_features, **config.output_features},
+            norm_map=config.normalization_mapping,
+            stats=dataset_stats,
+        ),
+        RenderMessagesStep(recipe=recipe),
+        PI052TextTokenizerStep(
+            tokenizer_name="google/paligemma-3b-pt-224",
+            max_length=config.tokenizer_max_length,
+            plan_dropout_prob=getattr(config, "plan_dropout_prob", 0.0),
+            memory_dropout_prob=getattr(config, "memory_dropout_prob", 0.0),
+            subtask_dropout_prob=getattr(config, "subtask_dropout_prob", 0.0),
+        ),
+    ]
+
+    # Add FAST action-token supervision only when explicitly enabled.
+    if getattr(config, "enable_fast_action_loss", False):
+        from .fit_fast_tokenizer import resolve_fast_tokenizer  # noqa: PLC0415
+
+        input_steps.append(
+            ActionTokenizerProcessorStep(
+                action_tokenizer_name=resolve_fast_tokenizer(
+                    config,
+                    dataset_repo_id,
+                    dataset_root,
+                    dataset_stats,
+                    dataset_revision,
+                    episodes,
+                    exclude_episodes,
+                ),
+                max_action_tokens=config.max_action_tokens,
+                fast_skip_tokens=config.fast_skip_tokens,
+                paligemma_tokenizer_name="google/paligemma-3b-pt-224",
+                allow_truncation=False,
+            )
+        )
+
+    input_steps.append(DeviceProcessorStep(device=config.device))
+
+    output_steps = [
+        UnnormalizerProcessorStep(
+            features=config.output_features,
+            norm_map=config.normalization_mapping,
+            stats=dataset_stats,
+        ),
+        AbsoluteActionsProcessorStep(
+            enabled=config.use_relative_actions,
+            relative_step=relative_step,
+        ),
+        DeviceProcessorStep(device="cpu"),
+    ]
+    return (
+        PolicyProcessorPipeline[dict[str, Any], dict[str, Any]](
+            steps=input_steps,
+            name=POLICY_PREPROCESSOR_DEFAULT_NAME,
+        ),
+        PolicyProcessorPipeline[PolicyAction, PolicyAction](
+            steps=output_steps,
+            name=POLICY_POSTPROCESSOR_DEFAULT_NAME,
+            to_transition=policy_action_to_transition,
+            to_output=transition_to_policy_action,
+        ),
+    )
+
+
+def _load_recipe(path_str: str) -> TrainingRecipe:
+    """Resolve ``path_str`` to a ``TrainingRecipe``.
+
+    Accepts an absolute path or a path relative to
+    ``src/lerobot/configs/``.
+    """
+    p = Path(path_str)
+    if not p.is_absolute() and not p.exists():
+        from lerobot.configs import recipe as _recipe_module  # noqa: PLC0415
+
+        configs_dir = Path(_recipe_module.__file__).resolve().parent
+        candidate = configs_dir / path_str
+        if candidate.exists():
+            p = candidate
+    return TrainingRecipe.from_yaml(p)

src/lerobot/policies/pi052/text_processor_pi052.py

@@ -0,0 +1,521 @@
+# 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.
+
+"""Tokenize PI052 messages and build text/action supervision masks."""
+
+from __future__ import annotations
+
+import json
+import logging
+from dataclasses import dataclass
+from typing import Any
+
+import numpy as np
+import torch
+from torch import Tensor
+
+from lerobot.configs import PipelineFeatureType, PolicyFeature
+from lerobot.processor.pipeline import ProcessorStep, ProcessorStepRegistry
+from lerobot.types import EnvTransition, TransitionKey
+from lerobot.utils.constants import OBS_LANGUAGE_ATTENTION_MASK, OBS_LANGUAGE_TOKENS, OBS_STATE
+
+logger = logging.getLogger(__name__)
+
+
+def discretize_state_str(state_row: Any) -> str:
+    """Format one normalized state row with PI0.5's 256-bin convention."""
+    arr = state_row.detach().cpu().numpy() if hasattr(state_row, "detach") else np.asarray(state_row)
+    disc = np.digitize(arr, bins=np.linspace(-1, 1, 256 + 1)[:-1]) - 1
+    return " ".join(str(int(x)) for x in disc.reshape(-1).tolist())
+
+
+def _state_row_at(state_all: Any, pos: int) -> Any:
+    """Select the per-sample state row from a (possibly batched) state tensor."""
+    if state_all is None:
+        return None
+    if hasattr(state_all, "ndim") and state_all.ndim >= 2:
+        return state_all[pos]
+    return state_all
+
+
+def _content_to_text(content: Any) -> str:
+    """Collapse a message's ``content`` (string or multimodal blocks) to text."""
+    if isinstance(content, str):
+        return content
+    if isinstance(content, list):
+        parts = [
+            b["text"]
+            for b in content
+            if isinstance(b, dict) and b.get("type") == "text" and isinstance(b.get("text"), str)
+        ]
+        return "\n".join(parts)
+    return ""
+
+
+def _flatten_say_tool_calls(message: dict[str, Any]) -> dict[str, Any]:
+    """Move ``say`` tool calls into text markers that PaliGemma can learn."""
+    tool_calls = message.get("tool_calls")
+    if not tool_calls:
+        return message
+    say_texts: list[str] = []
+    for call in tool_calls:
+        if not isinstance(call, dict):
+            continue
+        fn = call.get("function") or {}
+        if fn.get("name") != "say":
+            continue
+        args = fn.get("arguments")
+        if isinstance(args, str):
+            try:
+                import json  # noqa: PLC0415
+
+                args = json.loads(args)
+            except (ValueError, TypeError):
+                args = {}
+        text = args.get("text", "") if isinstance(args, dict) else ""
+        if text:
+            say_texts.append(str(text))
+    new = dict(message)
+    new.pop("tool_calls", None)
+    if not say_texts:
+        return new
+    base = _content_to_text(new.get("content")).strip()
+    marker = "".join(f"<say>{t}</say>" for t in say_texts)
+    new["content"] = f"{base}\n{marker}" if base else marker
+    return new
+
+
+def _strip_blocks(message: dict[str, Any]) -> dict[str, Any]:
+    """Flatten text blocks and drop image blocks handled by observation inputs."""
+    new = dict(message)
+    new.pop("stream", None)
+    new.pop("target", None)
+    content = new.get("content")
+    if content is None:
+        new["content"] = ""
+    elif isinstance(content, str):
+        pass
+    elif isinstance(content, list):
+        parts: list[str] = []
+        for block in content:
+            if not isinstance(block, dict):
+                continue
+            if block.get("type") == "text":
+                t = block.get("text", "")
+                if isinstance(t, str):
+                    parts.append(t)
+        new["content"] = "\n".join(parts)
+    else:
+        new["content"] = str(content)
+    return new
+
+
+def _is_batched_messages(messages: Any) -> bool:
+    return isinstance(messages, list) and bool(messages) and isinstance(messages[0], list)
+
+
+def _sample_indices(value: Any, batch_size: int) -> list[int | None]:
+    if value is None:
+        return [None] * batch_size
+    if isinstance(value, torch.Tensor):
+        if value.numel() == 1:
+            return [int(value.item())] * batch_size
+        values = value.reshape(-1).tolist()
+        return [int(v) for v in values[:batch_size]]
+    if isinstance(value, (list, tuple)):
+        if len(value) == 1:
+            return _sample_indices(value[0], batch_size)
+        return [int(v.item() if hasattr(v, "item") else v) for v in value[:batch_size]]
+    return [int(value)] * batch_size
+
+
+_VQA_COORD_SCALE = 1000.0
+
+
+def register_paligemma_loc_tokens(tokenizer: Any) -> Any:
+    """Register PaliGemma's reserved ``<locDDDD>`` strings as single tokens.
+
+    Without registration, the stock tokenizer splits each location into generic text pieces.
+    """
+    if "<loc0000>" in getattr(tokenizer, "added_tokens_encoder", {}):
+        return tokenizer
+    tokenizer.add_tokens([f"<loc{i:04d}>" for i in range(1024)])
+    return tokenizer
+
+
+def _loc_token(coord: float, scale: float = _VQA_COORD_SCALE) -> str:
+    """PaliGemma ``<locNNNN>`` for a coord on a ``[0, scale]`` axis."""
+    idx = round(float(coord) / scale * 1023) if scale > 0 else 0
+    return f"<loc{max(0, min(1023, idx)):04d}>"
+
+
+def _vqa_answer_to_loc(answer: dict[str, Any]) -> str | None:
+    """Convert normalized bbox/keypoint answers to label-first PaliGemma locations.
+
+    Label-first targets prevent location tokens from dominating every assistant turn; non-spatial answers return ``None``.
+    """
+    point = answer.get("point")
+    if isinstance(point, list | tuple) and len(point) == 2 and "point_format" in answer:
+        try:
+            x, y = float(point[0]), float(point[1])
+        except (TypeError, ValueError):
+            return None
+        label = str(answer.get("label", "")).strip()
+        if not label:
+            return None
+        return f"{label} {_loc_token(y)}{_loc_token(x)}"
+
+    detections = answer.get("detections")
+    if isinstance(detections, list) and detections:
+        parts: list[str] = []
+        for det in detections:
+            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
+            label = str(det.get("label", "")).strip()
+            if not label:
+                continue
+            toks = f"{_loc_token(y1)}{_loc_token(x1)}{_loc_token(y2)}{_loc_token(x2)}"
+            parts.append(f"{label} {toks}")
+        return " ; ".join(parts) if parts else None
+    return None
+
+
+def _messages_vqa_to_loc(
+    messages: list[dict[str, Any]],
+    target_indices: list[int],
+) -> list[dict[str, Any]]:
+    """Rewrite spatial VQA target JSON as camera-independent ``<loc>`` text."""
+    if not target_indices:
+        return messages
+    out = list(messages)
+    for idx in target_indices:
+        if not (0 <= idx < len(out)):
+            continue
+        content = out[idx].get("content")
+        if not isinstance(content, str) or not content.strip():
+            continue
+        try:
+            answer = json.loads(content)
+        except (ValueError, TypeError):
+            continue
+        if not isinstance(answer, dict):
+            continue
+        loc_text = _vqa_answer_to_loc(answer)
+        if loc_text is not None:
+            out[idx] = {**out[idx], "content": loc_text}
+    return out
+
+
+def _format_messages(
+    messages: list[dict[str, Any]],
+    target_indices: list[int] | None = None,
+    eos_token: str | None = None,
+) -> tuple[str, list[tuple[int, int]]]:
+    """Build the flat PI0.5 prompt and each message's payload span.
+
+    Supervised targets include EOS so generation learns when to stop.
+    """
+    targets = set(target_indices or [])
+    parts: list[str] = []
+    spans: list[tuple[int, int]] = []
+    cursor = 0
+    for i, m in enumerate(messages):
+        role = m.get("role", "user")
+        content = m.get("content", "") or ""
+        header = f"{role.capitalize()}: "
+        body = content + eos_token if (eos_token and i in targets) else content
+        full = header + body + "\n"
+        start = cursor + len(header)
+        end = start + len(body)
+        parts.append(full)
+        spans.append((start, end))
+        cursor += len(full)
+    return "".join(parts), spans
+
+
+def encode_prompt_with_targets(
+    tokenizer: Any, messages: list[dict[str, Any]], target_indices: list[int]
+) -> tuple[Tensor, Tensor, Tensor]:
+    """Tokenize a flat prompt and mark the token positions of target spans.
+
+    Inference-side twin of ``PI052TextTokenizerStep._encode_messages``: same
+    serialization (role headers, target EOS) and the same offset-overlap span
+    arithmetic, but unpadded and returning a boolean target mask instead of
+    labels. Used to rebuild joint-sequence prompts whose target spans must be
+    attended causally, matching ``_mark_target_span_causal`` at train time.
+
+    Returns ``(input_ids, attention_mask, target_marks)``, each ``(1, L)``.
+    """
+    prompt, spans = _format_messages(messages, target_indices, getattr(tokenizer, "eos_token", None))
+    encoded = tokenizer(prompt, return_tensors="pt", return_offsets_mapping=True)
+    input_ids = encoded["input_ids"][0]
+    attention_mask = encoded.get("attention_mask")
+    if attention_mask is None:
+        attention_mask = torch.ones_like(input_ids, dtype=torch.bool)
+    else:
+        attention_mask = attention_mask[0].bool()
+    offsets = encoded["offset_mapping"][0]
+
+    marks = torch.zeros_like(input_ids, dtype=torch.bool)
+    for idx in target_indices:
+        if idx >= len(spans):
+            continue
+        char_start, char_end = spans[idx]
+        for token_pos in range(input_ids.shape[0]):
+            if not attention_mask[token_pos]:
+                continue
+            tok_start, tok_end = int(offsets[token_pos, 0]), int(offsets[token_pos, 1])
+            if tok_end <= char_start or tok_start >= char_end:
+                continue
+            marks[token_pos] = True
+    return input_ids.unsqueeze(0), attention_mask.unsqueeze(0), marks.unsqueeze(0)
+
+
+@dataclass
+@ProcessorStepRegistry.register(name="pi052_text_tokenizer")
+class PI052TextTokenizerStep(ProcessorStep):
+    """Convert flat role-delimited messages into tokens and supervision masks."""
+
+    tokenizer_name: str = "google/paligemma-3b-pt-224"
+    max_length: int = 200
+    padding: str = "max_length"
+    padding_side: str = "right"
+    plan_dropout_prob: float = 0.0
+    memory_dropout_prob: float = 0.0
+    subtask_dropout_prob: float = 0.0
+    interjection_dropout_prob: float = 0.0
+    dropout_seed: int | None = None
+
+    def __post_init__(self) -> None:
+        self._tokenizer: Any = None
+
+    def get_config(self) -> dict[str, Any]:
+        return {
+            "tokenizer_name": self.tokenizer_name,
+            "max_length": self.max_length,
+            "padding": self.padding,
+            "padding_side": self.padding_side,
+            "plan_dropout_prob": self.plan_dropout_prob,
+            "memory_dropout_prob": self.memory_dropout_prob,
+            "subtask_dropout_prob": self.subtask_dropout_prob,
+            "interjection_dropout_prob": self.interjection_dropout_prob,
+            "dropout_seed": self.dropout_seed,
+        }
+
+    def _ensure_tokenizer(self) -> Any:
+        if self._tokenizer is not None:
+            return self._tokenizer
+        from transformers import AutoTokenizer  # noqa: PLC0415
+
+        self._tokenizer = register_paligemma_loc_tokens(AutoTokenizer.from_pretrained(self.tokenizer_name))
+        return self._tokenizer
+
+    def __call__(self, transition: EnvTransition) -> EnvTransition | None:
+        transition = transition.copy()
+        complementary = transition.get(TransitionKey.COMPLEMENTARY_DATA, {}) or {}
+        messages = complementary.get("messages") or []
+
+        if not messages:
+            return transition
+
+        tokenizer = self._ensure_tokenizer()
+        state_all = (transition.get(TransitionKey.OBSERVATION) or {}).get(OBS_STATE)
+        if _is_batched_messages(messages):
+            indices_iter = _sample_indices(complementary.get("index"), len(messages))
+            encoded = [
+                self._encode_messages(
+                    tokenizer,
+                    msg,
+                    list(streams),
+                    list(tgt_indices),
+                    complementary,
+                    sample_idx=int(s_idx) if s_idx is not None else None,
+                    state_row=_state_row_at(state_all, pos),
+                )
+                for pos, (msg, streams, tgt_indices, s_idx) in enumerate(
+                    zip(
+                        messages,
+                        complementary.get("message_streams") or [[] for _ in messages],
+                        complementary.get("target_message_indices") or [[] for _ in messages],
+                        indices_iter,
+                        strict=False,
+                    )
+                )
+            ]
+        else:
+            sample_idx = _sample_indices(complementary.get("index"), 1)[0]
+            encoded = [
+                self._encode_messages(
+                    tokenizer,
+                    messages,
+                    list(complementary.get("message_streams") or []),
+                    list(complementary.get("target_message_indices") or []),
+                    complementary,
+                    sample_idx=sample_idx,
+                    state_row=_state_row_at(state_all, 0),
+                )
+            ]
+
+        obs = dict(transition.get(TransitionKey.OBSERVATION) or {})
+        obs[OBS_LANGUAGE_TOKENS] = torch.stack([ids for ids, _, _, _, _ in encoded])
+        obs[OBS_LANGUAGE_ATTENTION_MASK] = torch.stack([attn for _, attn, _, _, _ in encoded])
+        transition[TransitionKey.OBSERVATION] = obs
+
+        transition[TransitionKey.COMPLEMENTARY_DATA] = {
+            **complementary,
+            "text_labels": torch.stack([labels for _, _, labels, _, _ in encoded]),
+            "predict_actions": torch.stack([pred for _, _, _, pred, _ in encoded]),
+        }
+        return transition
+
+    def _encode_messages(
+        self,
+        tokenizer: Any,
+        messages: list[dict[str, Any]],
+        message_streams: list[str | None],
+        target_indices: list[int],
+        complementary: dict[str, Any],
+        sample_idx: int | None = None,
+        state_row: Any = None,
+    ) -> tuple[Tensor, Tensor, Tensor, Tensor, str]:
+        if (
+            self.plan_dropout_prob
+            or self.memory_dropout_prob
+            or self.subtask_dropout_prob
+            or self.interjection_dropout_prob
+        ):
+            messages, target_indices = self._apply_prompt_dropout(
+                messages,
+                target_indices,
+                complementary,
+                sample_idx=sample_idx,
+            )
+
+        messages = _messages_vqa_to_loc(messages, target_indices)
+
+        messages = [_strip_blocks(_flatten_say_tool_calls(m)) for m in messages]
+        # Only low-level prompts carry PI0.5-style proprioception.
+        if state_row is not None and any(s == "low_level" for s in message_streams):
+            state_str = discretize_state_str(state_row)
+            for m in reversed(messages):
+                if m.get("role") == "user":
+                    base = _content_to_text(m.get("content", ""))
+                    m["content"] = f"{base}, State: {state_str};"
+                    break
+        prompt, spans = _format_messages(messages, target_indices, getattr(tokenizer, "eos_token", None))
+
+        encoded = tokenizer(
+            prompt,
+            max_length=self.max_length,
+            padding=self.padding,
+            truncation=True,
+            return_tensors="pt",
+            return_offsets_mapping=True,
+            padding_side=self.padding_side,
+        )
+
+        input_ids = encoded["input_ids"][0]
+        attention_mask = encoded["attention_mask"][0].bool()
+        offsets = encoded["offset_mapping"][0]
+
+        labels = torch.full_like(input_ids, fill_value=-100)
+        for idx in target_indices:
+            if idx >= len(spans):
+                continue
+            char_start, char_end = spans[idx]
+            for token_pos in range(input_ids.shape[0]):
+                if not attention_mask[token_pos]:
+                    continue
+                tok_start, tok_end = int(offsets[token_pos, 0]), int(offsets[token_pos, 1])
+                if tok_end <= char_start or tok_start >= char_end:
+                    continue
+                labels[token_pos] = input_ids[token_pos]
+
+        predict_actions = torch.tensor(
+            bool(any(s == "low_level" for s in message_streams)),
+            dtype=torch.bool,
+        )
+        return input_ids, attention_mask, labels, predict_actions, prompt
+
+    def _apply_prompt_dropout(
+        self,
+        messages: list[dict[str, Any]],
+        target_indices: list[int],
+        complementary: dict[str, Any],
+        sample_idx: int | None = None,
+    ) -> tuple[list[dict[str, Any]], list[int]]:
+        """Drop sampled context messages and remap the retained target positions."""
+        import random  # noqa: PLC0415
+
+        seed = self.dropout_seed
+        if seed is None:
+            seed_src = sample_idx if sample_idx is not None else complementary.get("index", 0)
+            try:
+                if hasattr(seed_src, "item"):
+                    seed_src = seed_src.item()
+                seed = int(seed_src)
+            except (TypeError, ValueError):
+                seed = 0
+        rng = random.Random(seed)
+
+        keep_indices: list[int] = []
+        for idx, msg in enumerate(messages):
+            if idx in target_indices:
+                keep_indices.append(idx)
+                continue
+            kind = _classify_for_dropout(msg)
+            prob = {
+                "plan": self.plan_dropout_prob,
+                "memory": self.memory_dropout_prob,
+                "subtask": self.subtask_dropout_prob,
+                "interjection": self.interjection_dropout_prob,
+            }.get(kind, 0.0)
+            if prob > 0.0 and rng.random() < prob:
+                continue
+            keep_indices.append(idx)
+
+        new_messages = [messages[i] for i in keep_indices]
+        old_to_new = {old: new for new, old in enumerate(keep_indices)}
+        new_targets = [old_to_new[t] for t in target_indices if t in old_to_new]
+        return new_messages, new_targets
+
+    def transform_features(
+        self, features: dict[PipelineFeatureType, dict[str, PolicyFeature]]
+    ) -> dict[PipelineFeatureType, dict[str, PolicyFeature]]:
+        return features
+
+
+def _classify_for_dropout(message: dict[str, Any]) -> str | None:
+    """Classify context from its rendered text prefix."""
+    content = message.get("content")
+    if isinstance(content, list):
+        text_parts = [b.get("text", "") for b in content if isinstance(b, dict) and b.get("type") == "text"]
+        content = " ".join(text_parts)
+    elif content is None or not isinstance(content, str):
+        return None
+    s = content.strip()
+    if s.startswith("Plan:") or s.startswith("Previous plan"):
+        return "plan"
+    if s.startswith("Memory:") or s.startswith("Previous memory"):
+        return "memory"
+    if s.startswith("Current subtask") or s.startswith("Completed subtask"):
+        return "subtask"
+    return None

src/lerobot/policies/pi0_fast/configuration_pi0_fast.py

@@ -61,21 +61,21 @@ class PI0FastConfig(PreTrainedConfig):
     tokenizer_max_length: int = 200  # see openpi `__post_init__`
     text_tokenizer_name: str = "google/paligemma-3b-pt-224"
     action_tokenizer_name: str = "lerobot/fast-action-tokenizer"
+    auto_fit_fast_tokenizer: bool = False
+    fast_tokenizer_cache_dir: str = "~/.cache/lerobot/fast_tokenizers"
+    fast_tokenizer_fit_samples: int = 1024
     temperature: float = 0.0
     max_decoding_steps: int = 256
     fast_skip_tokens: int = 128
 
-    # Whether to validate that decoded action tokens start with "Action: " prefix
-    validate_action_token_prefix: bool = True
-
     # Whether to use KV cache for faster autoregressive decoding
     use_kv_cache: bool = True
 
     normalization_mapping: dict[str, NormalizationMode] = field(
         default_factory=lambda: {
             "VISUAL": NormalizationMode.IDENTITY,
-            "STATE": NormalizationMode.MEAN_STD,  # Pi0Fast uses quantiles for state
+            "STATE": NormalizationMode.QUANTILES,
-            "ACTION": NormalizationMode.MEAN_STD,  # Pi0Fast uses quantiles for action
+            "ACTION": NormalizationMode.QUANTILES,
         }
     )
 

src/lerobot/policies/pi0_fast/modeling_pi0_fast.py

@@ -24,13 +24,7 @@ import numpy as np
 import torch
 from torch import Tensor, nn
 
-from lerobot.utils.import_utils import _scipy_available, _transformers_available, require_package
+from lerobot.utils.import_utils import _transformers_available, require_package
-
-# Conditional import for type checking and lazy loading
-if TYPE_CHECKING or _scipy_available:
-    from scipy.fftpack import idct
-else:
-    idct = None
 
 if TYPE_CHECKING or _transformers_available:
     from transformers import AutoProcessor, AutoTokenizer
@@ -66,6 +60,32 @@ class ActionSelectKwargs(TypedDict, total=False):
     temperature: float | None
 
 
+def _gather_last_valid_language_hidden(
+    hidden_states: Tensor,
+    language_masks: Tensor,
+    image_token_count: int,
+) -> Tensor:
+    """Gather each sample's last non-padding language hidden state."""
+    last_language_indices = image_token_count + language_masks.long().sum(dim=1) - 1
+    if torch.any(last_language_indices < image_token_count):
+        raise ValueError("PI0-FAST requires at least one valid language token per sample")
+    batch_indices = torch.arange(hidden_states.shape[0], device=hidden_states.device)
+    return hidden_states[batch_indices, last_language_indices]
+
+
+def _reduce_fast_token_loss(token_loss: Tensor, token_mask: Tensor) -> Tensor:
+    """Give every sample equal weight regardless of its FAST token count."""
+    sample_loss = (token_loss * token_mask).sum(dim=1) / token_mask.sum(dim=1).clamp(min=1)
+    return sample_loss.mean()
+
+
+def _sample_next_token(logits: Tensor, temperature: float) -> Tensor:
+    if temperature > 0:
+        probabilities = torch.softmax(logits / temperature, dim=-1)
+        return torch.multinomial(probabilities, num_samples=1)
+    return torch.argmax(logits, dim=-1, keepdim=True)
+
+
 class GemmaConfig:  # see openpi `gemma.py: Config`
     """Configuration for Gemma model variants."""
 
@@ -240,7 +260,6 @@ class PI0FastPytorch(nn.Module):  # see openpi `PI0Pytorch`
         # Compile model if requested
         if config.compile_model:
             torch.set_float32_matmul_precision("high")
-            self.sample_actions_fast = torch.compile(self.sample_actions_fast, mode=config.compile_mode)
             self.forward = torch.compile(self.forward, mode=config.compile_mode)
 
     def gradient_checkpointing_enable(self):
@@ -467,18 +486,12 @@ class PI0FastPytorch(nn.Module):  # see openpi `PI0Pytorch`
         # only compute logits for the positions that predict FAST tokens
         lm_head = self.paligemma_with_expert.paligemma.lm_head
 
-        # Targets are the FAST action tokens
+        # The last valid prompt token predicts "Action:", then each FAST token predicts the next one.
-        fast_targets = fast_action_tokens  # (B, num_fast_embs)
+        fast_hidden = prefix_out[:, -num_fast_embs:, :]
-
+        last_language_hidden = _gather_last_valid_language_hidden(prefix_out, masks, total_t_images)
-        # extract logits for FAST token prediction
+        prediction_hidden = torch.cat([last_language_hidden[:, None], fast_hidden[:, :-1]], dim=1)
-        fast_hidden = prefix_out[:, -fast_targets.shape[1] :, :]
+        fast_logits_for_pred = lm_head(prediction_hidden)
-        fast_logits_for_pred = lm_head(fast_hidden)  # (B, num_fast_embs, gemma_vocab_size)
+        fast_targets = fast_action_tokens
-
-        # Shift left for next-step prediction and shift target
-        # logits[:, i] predicts targets[:, i+1]
-        fast_logits_for_pred = fast_logits_for_pred[:, :-1, :]  # shift logits left
-        fast_targets = fast_targets[:, 1:]  # shift targets right
-        fast_action_masks = fast_action_masks[:, 1:]  # shift masks to match targets
 
         # compute cross-entropy loss
         loss_fct = torch.nn.CrossEntropyLoss(reduction="none")
@@ -488,9 +501,7 @@ class PI0FastPytorch(nn.Module):  # see openpi `PI0Pytorch`
         fast_loss_per_token = loss_fct(fast_logits_flat, fast_targets_flat)
         fast_loss_per_token = fast_loss_per_token.reshape(fast_targets.shape)
 
-        # apply mask and compute mean loss
+        fast_loss = _reduce_fast_token_loss(fast_loss_per_token, fast_action_masks.float())
-        masked_fast_loss = fast_loss_per_token * fast_action_masks.float()
-        fast_loss = masked_fast_loss.sum() / fast_action_masks.sum().clamp(min=1)
 
         return {
             "ce_loss": fast_loss,
@@ -519,15 +530,7 @@ class PI0FastPytorch(nn.Module):  # see openpi `PI0Pytorch`
         device = tokens.device
         lm_head = self.paligemma_with_expert.paligemma.lm_head
 
-        # add bos token after tokens
+        # 1. Initial embedding: the prompt's existing BOS is the only BOS in the sequence.
-        bos_token = torch.full(
-            (bsize, 1), self._paligemma_tokenizer.bos_token_id, dtype=torch.long, device=device
-        )
-        tokens = torch.cat([tokens, bos_token], dim=1)
-        masks = torch.cat([masks, torch.ones((bsize, 1), dtype=torch.bool, device=device)], dim=1)
-
-        # 1. Initial Embedding (matches training prefix)
-        # prefix_embs will include [Images, Language Prompt, BOS]
         prefix_embs, prefix_pad_masks, prefix_att_masks, total_t_images, _ = self.embed_prefix_fast(
             images, img_masks, tokens, masks, fast_action_tokens=None, fast_action_masks=None
         )
@@ -539,6 +542,8 @@ class PI0FastPytorch(nn.Module):  # see openpi `PI0Pytorch`
             prefix_embs = prefix_embs.to(dtype=torch.bfloat16)
 
         generated_action_tokens = torch.zeros((bsize, max_decoding_steps), dtype=torch.long, device=device)
+        eos_token_id = self._paligemma_tokenizer.eos_token_id
+        finished = torch.zeros(bsize, dtype=torch.bool, device=device)
 
         # 2. Decoding Loop (each step re-computes full sequence)
         for t in range(max_decoding_steps):
@@ -556,16 +561,24 @@ class PI0FastPytorch(nn.Module):  # see openpi `PI0Pytorch`
                 adarms_cond=[None, None],
             )
 
-            # predict next token from the very last sequence position
+            if t == 0:
-            last_logits = lm_head(prefix_out[:, -1:, :])  # (B, 1, vocab_size)
+                prediction_hidden = _gather_last_valid_language_hidden(prefix_out, masks, total_t_images)
-
-            if temperature > 0:
-                probs = torch.softmax(last_logits[:, -1] / temperature, dim=-1)
-                next_token = torch.multinomial(probs, num_samples=1)
             else:
-                next_token = torch.argmax(last_logits[:, -1], dim=-1, keepdim=True)
+                prediction_hidden = prefix_out[:, -1]
+            next_token = _sample_next_token(lm_head(prediction_hidden), temperature)
 
-            generated_action_tokens[:, t] = next_token.squeeze(-1)
+            active = ~finished
+            generated_action_tokens[:, t] = torch.where(
+                active, next_token.squeeze(-1), torch.zeros_like(next_token.squeeze(-1))
+            )
+            finished |= active & next_token.squeeze(-1).eq(eos_token_id)
+            if finished.all():
+                break
+            next_token = torch.where(
+                finished[:, None],
+                torch.full_like(next_token, eos_token_id),
+                next_token,
+            )
 
             # 3. Update sequence for next iteration (unless it's the last step)
             if t < max_decoding_steps - 1:
@@ -612,20 +625,14 @@ class PI0FastPytorch(nn.Module):  # see openpi `PI0Pytorch`
         device = tokens.device
         lm_head = self.paligemma_with_expert.paligemma.lm_head
 
+        generated_action_tokens = torch.zeros((bsize, max_decoding_steps), dtype=torch.long, device=device)
+        if max_decoding_steps == 0:
+            return generated_action_tokens
+
         # --- 1. PREFILL PHASE ---
-        # Process Images + Text Prompt + BOS token once to populate the KV cache.
-
-        # Add BOS token to the prompt
-        bos_token = torch.full(
-            (bsize, 1), self._paligemma_tokenizer.bos_token_id, dtype=torch.long, device=device
-        )
-        tokens_in = torch.cat([tokens, bos_token], dim=1)
-        masks_in = torch.cat([masks, torch.ones((bsize, 1), dtype=torch.bool, device=device)], dim=1)
-
-        # Embed prefix [Images, Language, BOS]
         # fast_action_tokens=None means we are just embedding the condition (images+text)
         prefix_embs, prefix_pad_masks, prefix_att_masks, total_t_images, _ = self.embed_prefix_fast(
-            images, img_masks, tokens_in, masks_in, fast_action_tokens=None, fast_action_masks=None
+            images, img_masks, tokens, masks, fast_action_tokens=None, fast_action_masks=None
         )
 
         # Ensure correct precision (bfloat16/float32)
@@ -652,17 +659,18 @@ class PI0FastPytorch(nn.Module):  # see openpi `PI0Pytorch`
             adarms_cond=[None, None],
         )
 
-        # Sample the first action token from the last logit of the prefix
+        prediction_hidden = _gather_last_valid_language_hidden(prefix_out, masks, total_t_images)
-        last_logits = lm_head(prefix_out[:, -1:, :])  # (B, 1, V)
+        next_token = _sample_next_token(lm_head(prediction_hidden), temperature)
-        if temperature > 0:
-            probs = torch.softmax(last_logits[:, -1] / temperature, dim=-1)
-            next_token = torch.multinomial(probs, num_samples=1)
-        else:
-            next_token = torch.argmax(last_logits[:, -1], dim=-1, keepdim=True)
-
-        # Initialize storage for generated tokens
-        generated_action_tokens = torch.zeros((bsize, max_decoding_steps), dtype=torch.long, device=device)
         generated_action_tokens[:, 0] = next_token.squeeze(-1)
+        eos_token_id = self._paligemma_tokenizer.eos_token_id
+        finished = next_token.squeeze(-1).eq(eos_token_id)
+        if finished.all():
+            return generated_action_tokens
+        next_token = torch.where(
+            finished[:, None],
+            torch.full_like(next_token, eos_token_id),
+            next_token,
+        )
 
         # Track valid tokens mask (0 for pad, 1 for valid)
         # We need this to tell the new token what it can attend to (images + text + past actions)
@@ -703,15 +711,19 @@ class PI0FastPytorch(nn.Module):  # see openpi `PI0Pytorch`
                 adarms_cond=[None, None],
             )
 
-            # Sample next token
+            next_token = _sample_next_token(lm_head(step_out[:, -1]), temperature)
-            last_logits = lm_head(step_out[:, -1:, :])
+            active = ~finished
-            if temperature > 0:
+            generated_action_tokens[:, t] = torch.where(
-                probs = torch.softmax(last_logits[:, -1] / temperature, dim=-1)
+                active, next_token.squeeze(-1), torch.zeros_like(next_token.squeeze(-1))
-                next_token = torch.multinomial(probs, num_samples=1)
+            )
-            else:
+            finished |= active & next_token.squeeze(-1).eq(eos_token_id)
-                next_token = torch.argmax(last_logits[:, -1], dim=-1, keepdim=True)
+            if finished.all():
-
+                break
-            generated_action_tokens[:, t] = next_token.squeeze(-1)
+            next_token = torch.where(
+                finished[:, None],
+                torch.full_like(next_token, eos_token_id),
+                next_token,
+            )
 
         return generated_action_tokens
 
@@ -1024,7 +1036,7 @@ class PI0FastPolicy(PreTrainedPolicy):
         return self._paligemma_tokenizer.vocab_size - 1 - self.config.fast_skip_tokens - tokens
 
     def decode_actions_with_fast(
-        self, token_ids: list[int], time_horizon: int, action_dim: int, relaxed_decoding: bool = True
+        self, token_ids: list[Tensor], time_horizon: int, action_dim: int
     ) -> np.ndarray:
         """
         Decodes action token IDs back to continuous action values using the FAST tokenizer.
@@ -1033,8 +1045,6 @@ class PI0FastPolicy(PreTrainedPolicy):
             token_ids: List of token IDs to decode.
             time_horizon: The number of timesteps for actions.
             action_dim: The dimensionality of each action.
-            relaxed_decoding: Whether to use relaxed decoding (allows partial sequences).
-
         Returns:
             A numpy array representing the decoded actions.
         """
@@ -1042,40 +1052,23 @@ class PI0FastPolicy(PreTrainedPolicy):
 
         for token in token_ids:
             try:
-                decoded_tokens = self.action_tokenizer.bpe_tokenizer.decode(token)
+                expected_shape = (time_horizon, action_dim)
-                decoded_dct_coeff = np.array(list(map(ord, decoded_tokens))) + self.action_tokenizer.min_token
+                decoded_action = np.asarray(
-
+                    self.action_tokenizer.decode(
-                if relaxed_decoding:
+                        [token.tolist()], time_horizon=time_horizon, action_dim=action_dim
-                    # expected sequence length
+                    )[0],
-                    expected_seq_len = time_horizon * action_dim
+                    dtype=np.float32,
-                    diff = expected_seq_len - decoded_dct_coeff.shape[0]
-
-                    # apply truncation if too long
-                    if diff < 0:
-                        decoded_dct_coeff = decoded_dct_coeff[:expected_seq_len]  # truncate on the right
-
-                    # apply padding if too short
-                    elif diff > 0:
-                        decoded_dct_coeff = np.pad(
-                            decoded_dct_coeff, (0, diff), mode="constant", constant_values=0
-                        )
-
-                decoded_dct_coeff = decoded_dct_coeff.reshape(-1, action_dim)
-                assert decoded_dct_coeff.shape == (
-                    time_horizon,
-                    action_dim,
-                ), (
-                    f"Decoded DCT coefficients have shape {decoded_dct_coeff.shape}, expected ({time_horizon}, {action_dim})"
                 )
+                if decoded_action.shape != expected_shape:
+                    raise ValueError(
+                        f"decoded action shape {decoded_action.shape} does not match {expected_shape}"
+                    )
 
             except Exception as e:
-                logging.warning(f"Error decoding tokens: {e}")
+                logging.warning("Invalid FAST action sequence; returning a zero action chunk: %s", e)
-                logging.warning(f"Tokens: {token}")
+                decoded_action = np.zeros((time_horizon, action_dim))
-                decoded_dct_coeff = np.zeros((time_horizon, action_dim))
 
-            decoded_actions.append(
+            decoded_actions.append(decoded_action)
-                idct(decoded_dct_coeff / self.action_tokenizer.scale, axis=0, norm="ortho")
-            )
 
         return np.stack(decoded_actions)
 
@@ -1105,53 +1098,28 @@ class PI0FastPolicy(PreTrainedPolicy):
         if single_sample:
             tokens = tokens.unsqueeze(0)
 
-        # Convert token IDs to token strings
+        action_tokens = []
-        decoded_tokens = [self._paligemma_tokenizer.convert_ids_to_tokens(seq.tolist()) for seq in tokens]
+        for token_sequence in tokens:
-        # Get the token sequence for "Action: " to remove it
+            try:
-        action_prefix_ids = self._paligemma_tokenizer.encode("Action: ", add_special_tokens=False)
+                token_ids = token_sequence.tolist()
-        action_prefix_tokens = self._paligemma_tokenizer.convert_ids_to_tokens(action_prefix_ids)
+                eos_token_id = self._paligemma_tokenizer.eos_token_id
-        action_prefix_len = len(action_prefix_tokens)
+                if eos_token_id in token_ids:
-
+                    token_ids = token_ids[: token_ids.index(eos_token_id) + 1]
-        # Clean tokens by removing everything after the first "|" (end-of-action marker)
+                decoded_text = self._paligemma_tokenizer.decode(token_ids)
-        # and removing all occurrences of "Action: " token sequence
+                if not decoded_text.startswith("Action: ") or "|" not in decoded_text:
-        # assert that beginning contain "Action: "
+                    raise ValueError(f"expected 'Action: <codes>|', got {decoded_text!r}")
-        if self.config.validate_action_token_prefix:
+                action_text = decoded_text.removeprefix("Action: ").split("|", maxsplit=1)[0]
-            for token_seq in decoded_tokens:
+                raw_action_tokens = torch.tensor(
-                assert len(token_seq) >= 2 and token_seq[0] == "Action" and token_seq[1] == ":", (
+                    self._paligemma_tokenizer.encode(action_text, add_special_tokens=False),
-                    f"Token sequence does not start with ['Action', ':']: {token_seq}"
+                    dtype=torch.long,
+                    device=tokens.device,
                 )
-
+                if raw_action_tokens.numel() == 0:
-        cleaned_tokens = []
+                    raise ValueError("empty FAST action payload")
-        for token_seq in decoded_tokens:
+                action_tokens.append(self._paligemma_tokens_to_act_tokens(raw_action_tokens))
-            # Remove everything after "|"
+            except Exception as e:
-            if "|" in token_seq:
+                logging.warning("Invalid generated PI0-FAST text; returning zeros for this sample: %s", e)
-                token_seq = token_seq[: token_seq.index("|")]
+                action_tokens.append(torch.empty(0, dtype=torch.long, device=tokens.device))
-
-            # Remove all occurrences of "Action: " token sequence
-            i = 0
-            while i <= len(token_seq) - action_prefix_len:
-                if token_seq[i : i + action_prefix_len] == action_prefix_tokens:
-                    # Found a match, remove it
-                    token_seq = token_seq[:i] + token_seq[i + action_prefix_len :]
-                else:
-                    i += 1
-
-            cleaned_tokens.append(token_seq)
-
-        # Convert token strings back to IDs
-        raw_action_tokens = [
-            torch.tensor(
-                self._paligemma_tokenizer.convert_tokens_to_ids(token_seq),
-                dtype=torch.long,
-                device=tokens.device,
-            )
-            for token_seq in cleaned_tokens
-        ]
-
-        # Convert PaliGemma tokens to action tokens
-        action_tokens = [
-            self._paligemma_tokens_to_act_tokens(raw_action_token) for raw_action_token in raw_action_tokens
-        ]
 
         # Decode action tokens to continuous actions
         actions = self.decode_actions_with_fast(
@@ -1220,7 +1188,7 @@ class PI0FastPolicy(PreTrainedPolicy):
             )
 
         # Detokenize action tokens to continuous actions
-        action_horizon = self.config.n_action_steps
+        action_horizon = self.config.chunk_size
         action_dim = self.config.output_features[ACTION].shape[0]
 
         continuous_actions = self.detokenize_actions(

src/lerobot/policies/pi0_fast/processor_pi0_fast.py

@@ -70,7 +70,7 @@ class Pi0FastPrepareStateAndLanguageTokenizerProcessorStep(ProcessorStep):
 
         full_prompts = []
         for i, task in enumerate(tasks):
-            cleaned_text = task.strip().replace("_", " ").replace("\n", " ")
+            cleaned_text = task.strip().replace("_", " ").replace("\n", " ").lower()
             state_str = " ".join(map(str, discretized_states[i]))
             full_prompt = f"Task: {cleaned_text}, State: {state_str};\n"
             full_prompts.append(full_prompt)
@@ -92,6 +92,11 @@ class Pi0FastPrepareStateAndLanguageTokenizerProcessorStep(ProcessorStep):
 def make_pi0_fast_pre_post_processors(
     config: PI0FastConfig,
     dataset_stats: dict[str, dict[str, torch.Tensor]] | None = None,
+    dataset_repo_id: str | None = None,
+    dataset_root: str | None = None,
+    dataset_revision: str | None = None,
+    episodes: list[int] | None = None,
+    exclude_episodes: list[int] | None = None,
 ) -> tuple[
     PolicyProcessorPipeline[dict[str, Any], dict[str, Any]],
     PolicyProcessorPipeline[PolicyAction, PolicyAction],
@@ -136,6 +141,18 @@ def make_pi0_fast_pre_post_processors(
     # state from the observation but does not change it. NormalizerProcessorStep still runs
     # before Pi0FastPrepareStateAndLanguageTokenizerProcessorStep, so the state tokenizer
     # continues to receive normalized state in [-1, 1] as expected.
+    from ..pi052.fit_fast_tokenizer import resolve_fast_tokenizer  # noqa: PLC0415
+
+    action_tokenizer_path = resolve_fast_tokenizer(
+        config,
+        dataset_repo_id,
+        dataset_root,
+        dataset_stats,
+        dataset_revision,
+        episodes,
+        exclude_episodes,
+    )
+
     input_steps: list[ProcessorStep] = [
         steps.rename_observations,  # To mimic the same processor as pretrained one
         steps.add_batch_dim,
@@ -149,10 +166,11 @@ def make_pi0_fast_pre_post_processors(
             padding="max_length",
         ),
         ActionTokenizerProcessorStep(
-            action_tokenizer_name=config.action_tokenizer_name,
+            action_tokenizer_name=action_tokenizer_path,
             max_action_tokens=config.max_action_tokens,
             fast_skip_tokens=config.fast_skip_tokens,
             paligemma_tokenizer_name=config.text_tokenizer_name,
+            prepend_bos=False,
         ),
         steps.to_device,
     ]

src/lerobot/policies/pi_gemma.py

@@ -18,6 +18,7 @@ from typing import TYPE_CHECKING
 
 import torch
 from torch import nn
+from torch.nn import functional as F  # noqa: N812
 
 from lerobot.utils.import_utils import _transformers_available
 
@@ -121,7 +122,10 @@ class PiGemmaRMSNorm(nn.Module):
         if cond.shape[-1] != self.cond_dim:
             raise ValueError(f"Expected cond dim {self.cond_dim}, got {cond.shape[-1]}")
         modulation = self.dense(cond)
-        if len(x.shape) == 3:
+        # Per-sample cond (B, cond_dim) → broadcast over the sequence. A
+        # per-token cond (B, T, cond_dim) is already aligned with x and must
+        # not be unsqueezed (used by pi052's amortized K_repeat path).
+        if len(x.shape) == 3 and modulation.dim() == 2:
             modulation = modulation.unsqueeze(1)
         scale, shift, gate = modulation.chunk(3, dim=-1)
         normed = normed * (1 + scale.float()) + shift.float()
@@ -275,6 +279,8 @@ class PiGemmaModel(GemmaModel):  # type: ignore[misc]
         # Convert to bfloat16 if the first layer uses bfloat16
         if len(self.layers) > 0 and self.layers[0].self_attn.q_proj.weight.dtype == torch.bfloat16:
             hidden_states = hidden_states.to(torch.bfloat16)
+        if causal_mask is not None and torch.is_floating_point(causal_mask):
+            causal_mask = causal_mask.to(dtype=hidden_states.dtype)
 
         # create position embeddings to be shared across the decoder layers
         position_embeddings = self.rotary_emb(hidden_states, position_ids)
@@ -367,3 +373,45 @@ __all__ = [
     "PaliGemmaModelWithPiGemma",
     "PaliGemmaForConditionalGenerationWithPiGemma",
 ]
+
+
+# PI0.5 / PI052 dual-expert backbone: generic PaliGemma + Gemma action-expert
+# transformer machinery used by the pi052 policy. GemmaVariantConfig is openpi's
+# width/depth variant config (renamed from GemmaConfig to avoid clashing with
+# transformers' GemmaConfig).
+
+
+def sdpa_attention_forward(
+    module,
+    query: torch.Tensor,
+    key: torch.Tensor,
+    value: torch.Tensor,
+    attention_mask: torch.Tensor | None,
+    scaling: float,
+    dropout: float = 0.0,
+):
+    """Drop-in for ``modeling_gemma.eager_attention_forward`` using
+    ``torch.nn.functional.scaled_dot_product_attention``.
+
+    PyTorch SDPA picks the memory-efficient kernel for arbitrary additive
+    bias masks (the FA backend only accepts causal/sliding-window). On
+    H100 that is ~1.3-1.7x faster and uses ~30-40% less attention memory
+    than the eager softmax(QK^T)+matmul path. Mirrors eager's signature
+    and output shape (``(B, Lq, H, D)``) so call sites are unchanged.
+    """
+    n_rep = module.num_key_value_groups
+    if n_rep > 1:
+        key = key.repeat_interleave(n_rep, dim=1)
+        value = value.repeat_interleave(n_rep, dim=1)
+    if attention_mask is not None and attention_mask.dtype != query.dtype:
+        attention_mask = attention_mask.to(dtype=query.dtype)
+    attn_output = F.scaled_dot_product_attention(
+        query,
+        key,
+        value,
+        attn_mask=attention_mask,
+        dropout_p=dropout if module.training else 0.0,
+        is_causal=False,
+        scale=scaling,
+    )
+    return attn_output.transpose(1, 2).contiguous(), None

src/lerobot/policies/pretrained.py

@@ -34,22 +34,14 @@ from lerobot.configs import PreTrainedConfig
 from lerobot.configs.train import TrainPipelineConfig
 from lerobot.utils.device_utils import resolve_safetensors_device
 from lerobot.utils.hub import HubMixin
-from lerobot.utils.import_utils import _peft_available, require_package
 
 from .utils import log_model_loading_keys
 
-if TYPE_CHECKING or _peft_available:
+T = TypeVar("T", bound="PreTrainedPolicy")
-    from peft import PEFT_TYPE_TO_CONFIG_MAPPING, PeftType, get_peft_model
-else:
-    PEFT_TYPE_TO_CONFIG_MAPPING = None
-    PeftType = None
-    get_peft_model = None
 
 if TYPE_CHECKING:
     from lerobot.datasets.dataset_metadata import LeRobotDatasetMetadata
 
-T = TypeVar("T", bound="PreTrainedPolicy")
-
 
 def _build_card_context(
     cfg: TrainPipelineConfig | None,
@@ -346,6 +338,7 @@ class PreTrainedPolicy(nn.Module, HubMixin, abc.ABC):
             "smolvla": "lerobot/smolvla_base",
             "pi0": "lerobot/pi0_base",
             "pi05": "lerobot/pi05_base",
+            "pi052": "lerobot/pi052_base",
             "pi0_fast": "lerobot/pi0fast-base",
             "xvla": "lerobot/xvla-base",
         }
@@ -392,7 +385,7 @@ class PreTrainedPolicy(nn.Module, HubMixin, abc.ABC):
             peft_cli_overrides: Optional dict of CLI overrides (method_type, target_modules, r, etc.)
                 These are merged with policy defaults to build the final config.
         """
-        require_package("peft", extra="peft")
+        from peft import get_peft_model
 
         # If user provided a complete config, use it directly (with overrides)
         if peft_config is not None:
@@ -463,7 +456,7 @@ class PreTrainedPolicy(nn.Module, HubMixin, abc.ABC):
         Returns:
             Preprocessed dict with renamed keys and init_type mapped to method-specific key.
         """
-        require_package("peft", extra="peft")
+        from peft import PeftType
 
         cli_overrides = cli_overrides.copy()
 
@@ -488,7 +481,7 @@ class PreTrainedPolicy(nn.Module, HubMixin, abc.ABC):
 
     def _build_peft_config(self, cli_overrides: dict):
         """Build a PEFT config from policy defaults and CLI overrides."""
-        require_package("peft", extra="peft")
+        from peft import PEFT_TYPE_TO_CONFIG_MAPPING, PeftType
 
         # Determine PEFT method type (default to LORA)
         method_type_str = cli_overrides.get("method_type") or "lora"
@@ -515,7 +508,7 @@ class PreTrainedPolicy(nn.Module, HubMixin, abc.ABC):
 
     def _apply_peft_cli_overrides(self, peft_config, cli_overrides: dict):
         """Apply CLI overrides to an existing PEFT config."""
-        require_package("peft", extra="peft")
+        from peft import PEFT_TYPE_TO_CONFIG_MAPPING, PeftType
 
         # Get method type from existing config or CLI override
         method_type_str = cli_overrides.get("method_type")

src/lerobot/processor/batch_processor.py

@@ -175,9 +175,6 @@ class AddBatchDimensionComplementaryDataStep(ComplementaryDataProcessorStep):
             if isinstance(task_index_value, Tensor) and task_index_value.dim() == 0:
                 complementary_data["task_index"] = task_index_value.unsqueeze(0)
 
-        complementary_data.pop("language_persistent", None)
-        complementary_data.pop("language_events", None)
-
         if "messages" in complementary_data:
             messages = complementary_data["messages"]
             if isinstance(messages, list) and (not messages or isinstance(messages[0], dict)):

src/lerobot/processor/delta_action_processor.py

@@ -132,20 +132,10 @@ class MapDeltaActionToRobotActionStep(RobotActionProcessorStep):
     def transform_features(
         self, features: dict[PipelineFeatureType, dict[str, PolicyFeature]]
     ) -> dict[PipelineFeatureType, dict[str, PolicyFeature]]:
-        for axis in ["x", "y", "z"]:
+        for axis in ["x", "y", "z", "gripper"]:
             features[PipelineFeatureType.ACTION].pop(f"delta_{axis}", None)
-        features[PipelineFeatureType.ACTION].pop("gripper", None)
 
-        for feat in [
+        for feat in ["enabled", "target_x", "target_y", "target_z", "target_wx", "target_wy", "target_wz"]:
-            "enabled",
-            "target_x",
-            "target_y",
-            "target_z",
-            "target_wx",
-            "target_wy",
-            "target_wz",
-            "gripper_vel",
-        ]:
             features[PipelineFeatureType.ACTION][f"{feat}"] = PolicyFeature(
                 type=FeatureType.ACTION, shape=(1,)
             )

src/lerobot/processor/pipeline.py

@@ -41,7 +41,7 @@ from pathlib import Path
 from typing import Any, TypedDict, TypeVar, cast
 
 import torch
-from huggingface_hub import hf_hub_download
+from huggingface_hub import hf_hub_download, snapshot_download
 from safetensors.torch import load_file, save_file
 
 from lerobot.configs import PipelineFeatureType, PolicyFeature
@@ -205,6 +205,10 @@ class ProcessorStep(ABC):
         """
         return None
 
+    def save_artifacts(self, save_directory: Path) -> dict[str, str]:
+        """Save non-tensor assets and map constructor arguments to relative paths."""
+        return {}
+
     def reset(self) -> None:
         """Resets the internal state of the processor step, if any."""
         return None
@@ -549,6 +553,22 @@ class DataProcessorPipeline[TInput, TOutput](HubMixin):
         pipeline_config = self.get_config()
         pipeline_state_dict = self.state_dict()
 
+        for processor_step, step_entry in zip(self.steps, pipeline_config["steps"], strict=True):
+            artifacts = processor_step.save_artifacts(save_directory)
+            if artifacts:
+                for config_key, relative_path in artifacts.items():
+                    artifact_path = Path(relative_path)
+                    if artifact_path.is_absolute() or ".." in artifact_path.parts:
+                        raise ValueError(
+                            f"Processor artifact path must be relative to the checkpoint: {relative_path!r}"
+                        )
+                    if not (save_directory / artifact_path).exists():
+                        raise FileNotFoundError(
+                            f"Processor step did not save declared artifact '{relative_path}'"
+                        )
+                    step_entry["config"][config_key] = artifact_path.as_posix()
+                step_entry["artifacts"] = artifacts
+
         for state_key, step_state_dict in pipeline_state_dict.items():
             state_filename = f"{state_key}.safetensors"
             save_file(step_state_dict, save_directory / state_filename)
@@ -733,7 +753,13 @@ class DataProcessorPipeline[TInput, TOutput](HubMixin):
 
         # 3. Build steps with overrides
         steps, validated_overrides = cls._build_steps_with_overrides(
-            loaded_config, overrides or {}, model_id, base_path, hub_download_kwargs, is_local_source
+            loaded_config,
+            overrides or {},
+            model_id,
+            base_path,
+            config_filename,
+            hub_download_kwargs,
+            is_local_source,
         )
 
         # 4. Validate that all overrides were used
@@ -922,6 +948,7 @@ class DataProcessorPipeline[TInput, TOutput](HubMixin):
         overrides: dict[str, Any],
         model_id: str,
         base_path: Path | None,
+        config_filename: str,
         hub_download_kwargs: dict[str, Any],
         is_local_source: bool = False,
     ) -> tuple[list[ProcessorStep], set[str]]:
@@ -976,15 +1003,68 @@ class DataProcessorPipeline[TInput, TOutput](HubMixin):
             ImportError: If a step class cannot be imported or found in registry
             ValueError: If a step cannot be instantiated with its configuration
         """
+        loaded_config = deepcopy(loaded_config)
+        cls._resolve_artifact_paths(
+            loaded_config,
+            model_id,
+            base_path,
+            config_filename,
+            hub_download_kwargs,
+        )
         steps, remaining_override_keys = cls._build_steps_from_config(loaded_config, overrides)
 
         for step_instance, step_entry in zip(steps, loaded_config["steps"], strict=True):
             cls._load_step_state(
-                step_instance, step_entry, model_id, base_path, hub_download_kwargs, is_local_source
+                step_instance,
+                step_entry,
+                model_id,
+                base_path,
+                config_filename,
+                hub_download_kwargs,
+                is_local_source,
             )
 
         return steps, remaining_override_keys
 
+    @classmethod
+    def _resolve_artifact_paths(
+        cls,
+        loaded_config: dict[str, Any],
+        model_id: str,
+        base_path: Path | None,
+        config_filename: str,
+        hub_download_kwargs: dict[str, Any],
+    ) -> None:
+        """Resolve declared relative processor artifacts before step construction."""
+        is_local = Path(model_id).is_dir() or Path(model_id).is_file()
+
+        for step_entry in loaded_config["steps"]:
+            artifacts = step_entry.get("artifacts", {})
+            for config_key, relative_path in artifacts.items():
+                artifact_path = Path(relative_path)
+                if artifact_path.is_absolute() or ".." in artifact_path.parts:
+                    raise ValueError(
+                        f"Processor artifact path must be relative to the checkpoint: {relative_path!r}"
+                    )
+
+                resolved_path = base_path / artifact_path if base_path is not None else artifact_path
+                if not resolved_path.exists() and not is_local:
+                    repository_path = Path(config_filename).parent / artifact_path
+                    snapshot_download(
+                        repo_id=model_id,
+                        repo_type="model",
+                        allow_patterns=f"{repository_path.as_posix()}/**",
+                        **hub_download_kwargs,
+                    )
+
+                if not resolved_path.exists():
+                    step_name = step_entry.get("registry_name", step_entry.get("class", "unknown"))
+                    raise FileNotFoundError(
+                        f"Missing processor artifact '{relative_path}' for step '{step_name}' "
+                        f"next to '{config_filename}'. Checkpoint artifacts are incomplete."
+                    )
+                step_entry["config"][config_key] = str(resolved_path)
+
     @classmethod
     def _build_steps_from_config(
         cls,
@@ -1144,6 +1224,7 @@ class DataProcessorPipeline[TInput, TOutput](HubMixin):
         step_entry: dict[str, Any],
         model_id: str,
         base_path: Path | None,
+        config_filename: str,
         hub_download_kwargs: dict[str, Any],
         is_local_source: bool = False,
     ) -> None:
@@ -1209,7 +1290,7 @@ class DataProcessorPipeline[TInput, TOutput](HubMixin):
             # Download from Hub
             state_path = hf_hub_download(
                 repo_id=model_id,
-                filename=state_filename,
+                filename=(Path(config_filename).parent / state_filename).as_posix(),
                 repo_type="model",
                 **hub_download_kwargs,
             )

src/lerobot/processor/render_messages_processor.py

@@ -16,7 +16,7 @@
 
 from __future__ import annotations
 
-from dataclasses import dataclass
+from dataclasses import asdict, dataclass
 from typing import Any
 
 from lerobot.configs import PipelineFeatureType, PolicyFeature
@@ -32,17 +32,18 @@ from .pipeline import ProcessorStep, ProcessorStepRegistry
 @dataclass
 @ProcessorStepRegistry.register(name="render_messages_processor")
 class RenderMessagesStep(ProcessorStep):
-    """Processor step that turns raw language columns into rendered chat messages.
+    """Render language columns into recipe-defined messages and supervision metadata."""
-
-    Reads ``language_persistent`` and ``language_events`` from the transition's
-    complementary data, renders them through ``recipe`` at the sample timestamp,
-    and replaces the raw columns with the resulting ``messages`` /
-    ``message_streams`` / ``target_message_indices`` keys.
-    """
 
     recipe: TrainingRecipe
     dataset_ctx: Any | None = None
 
+    def __post_init__(self) -> None:
+        if isinstance(self.recipe, dict):
+            self.recipe = TrainingRecipe.from_dict(self.recipe)
+
+    def get_config(self) -> dict[str, Any]:
+        return {"recipe": asdict(self.recipe)}
+
     def __call__(self, transition: EnvTransition) -> EnvTransition | None:
         """Render messages for a single transition; return ``None`` to drop it."""
         complementary_data = transition.get(TransitionKey.COMPLEMENTARY_DATA) or {}
@@ -50,7 +51,17 @@ class RenderMessagesStep(ProcessorStep):
         events = complementary_data.get(LANGUAGE_EVENTS) or []
 
         if not persistent and not events:
-            return transition
+            rendered = _fallback_low_level_render(complementary_data.get("task"))
+            if rendered is None:
+                return transition
+            new_transition = transition.copy()
+            new_complementary_data = dict(new_transition.get(TransitionKey.COMPLEMENTARY_DATA) or {})
+            new_complementary_data.update(rendered)
+            new_transition[TransitionKey.COMPLEMENTARY_DATA] = new_complementary_data
+            return new_transition
+
+        if _is_batched_language(persistent) or _is_batched_language(events):
+            return self._call_batch(transition, complementary_data, persistent, events)
 
         timestamp = complementary_data.get("timestamp")
         if timestamp is None:
@@ -67,18 +78,147 @@ class RenderMessagesStep(ProcessorStep):
             dataset_ctx=self.dataset_ctx,
         )
         if rendered is None:
-            return None
+            rendered = _fallback_low_level_render(complementary_data.get("task"))
+            if rendered is None:
+                return None
 
         new_transition = transition.copy()
-        new_complementary_data = dict(complementary_data)
+        new_complementary_data = dict(new_transition.get(TransitionKey.COMPLEMENTARY_DATA) or {})
         new_complementary_data.pop(LANGUAGE_PERSISTENT, None)
         new_complementary_data.pop(LANGUAGE_EVENTS, None)
         new_complementary_data.update(rendered)
         new_transition[TransitionKey.COMPLEMENTARY_DATA] = new_complementary_data
         return new_transition
 
+    def _call_batch(
+        self,
+        transition: EnvTransition,
+        complementary_data: dict[str, Any],
+        persistent_batch: list,
+        events_batch: list,
+    ) -> EnvTransition | None:
+        timestamp = complementary_data.get("timestamp")
+        if timestamp is None:
+            raise KeyError("RenderMessagesStep requires sample timestamp in complementary data.")
+
+        batch_size = max(len(persistent_batch), len(events_batch))
+        messages: list[list[dict[str, Any]]] = []
+        message_streams: list[list[str | None]] = []
+        target_message_indices: list[list[int]] = []
+        keep_indices: list[int] = []
+
+        for i in range(batch_size):
+            rendered = render_sample(
+                recipe=self.recipe,
+                persistent=persistent_batch[i] if i < len(persistent_batch) else [],
+                events=events_batch[i] if i < len(events_batch) else [],
+                t=_batch_value(timestamp, i),
+                sample_idx=int(_batch_value(complementary_data.get("index", 0), i)),
+                task=_batch_value(complementary_data.get("task"), i),
+                dataset_ctx=self.dataset_ctx,
+            )
+            if rendered is None:
+                rendered = _fallback_low_level_render(_batch_value(complementary_data.get("task"), i))
+                if rendered is None:
+                    continue
+            keep_indices.append(i)
+            messages.append(rendered["messages"])
+            message_streams.append(rendered["message_streams"])
+            target_message_indices.append(rendered["target_message_indices"])
+
+        if not messages:
+            return None
+
+        new_transition = (
+            _select_batch_indices(transition, keep_indices)
+            if len(keep_indices) != batch_size
+            else transition.copy()
+        )
+        new_complementary_data = dict(new_transition.get(TransitionKey.COMPLEMENTARY_DATA) or {})
+        new_complementary_data.pop(LANGUAGE_PERSISTENT, None)
+        new_complementary_data.pop(LANGUAGE_EVENTS, None)
+        new_complementary_data["messages"] = messages
+        new_complementary_data["message_streams"] = message_streams
+        new_complementary_data["target_message_indices"] = target_message_indices
+        new_transition[TransitionKey.COMPLEMENTARY_DATA] = new_complementary_data
+        return new_transition
+
     def transform_features(
         self, features: dict[PipelineFeatureType, dict[str, PolicyFeature]]
     ) -> dict[PipelineFeatureType, dict[str, PolicyFeature]]:
         """Pass features through unchanged; rendering only touches complementary data."""
         return features
+
+
+def _scalar(value: Any) -> float | int:
+    """Unwrap a tensor/array/single-element list into a Python scalar."""
+    if hasattr(value, "item"):
+        return value.item()
+    if isinstance(value, list):
+        if len(value) != 1:
+            raise ValueError(f"Expected a scalar, got list of length {len(value)}: {value!r}")
+        return _scalar(value[0])
+    return value
+
+
+def _is_batched_language(value: Any) -> bool:
+    return isinstance(value, list) and bool(value) and isinstance(value[0], list)
+
+
+def _batch_value(value: Any, index: int) -> Any:
+    if value is None:
+        return None
+    if isinstance(value, list):
+        return value[index]
+    if hasattr(value, "ndim") and value.ndim > 0:
+        return _scalar(value[index])
+    return _scalar(value)
+
+
+def _select_batch_indices(transition: EnvTransition, indices: list[int]) -> EnvTransition:
+    selected = transition.copy()
+    for key in (TransitionKey.OBSERVATION, TransitionKey.COMPLEMENTARY_DATA):
+        data = selected.get(key)
+        if isinstance(data, dict):
+            selected[key] = {k: _select_value(v, indices) for k, v in data.items()}
+    action = selected.get(TransitionKey.ACTION)
+    if action is not None:
+        selected[TransitionKey.ACTION] = _select_value(action, indices)
+    return selected
+
+
+def _select_value(value: Any, indices: list[int]) -> Any:
+    if isinstance(value, list) and len(value) >= len(indices):
+        return [value[i] for i in indices]
+    if hasattr(value, "index_select") and hasattr(value, "new_tensor") and getattr(value, "ndim", 0) > 0:
+        return value.index_select(0, value.new_tensor(indices).long())
+    return value
+
+
+def _fallback_low_level_render(task: Any) -> dict[str, Any] | None:
+    """Keep action-only samples trainable when no recipe branch matches."""
+    if hasattr(task, "item"):
+        task = task.item()
+    if isinstance(task, list):
+        messages = []
+        message_streams = []
+        target_message_indices = []
+        for t in task:
+            rendered = _fallback_low_level_render(t)
+            if rendered is None:
+                return None
+            messages.append(rendered["messages"])
+            message_streams.append(rendered["message_streams"])
+            target_message_indices.append(rendered["target_message_indices"])
+        return {
+            "messages": messages,
+            "message_streams": message_streams,
+            "target_message_indices": target_message_indices,
+        }
+    if not isinstance(task, str) or not task:
+        return None
+    return {
+        "messages": [{"role": "user", "content": task}],
+        "message_streams": ["low_level"],
+        "target_message_indices": [],
+    }

src/lerobot/processor/tokenizer_processor.py

@@ -25,6 +25,7 @@ from __future__ import annotations
 
 import logging
 from dataclasses import dataclass, field
+from pathlib import Path
 from typing import TYPE_CHECKING, Any
 
 import torch
@@ -32,6 +33,7 @@ import torch
 from lerobot.configs import FeatureType, PipelineFeatureType, PolicyFeature
 from lerobot.types import EnvTransition, RobotObservation, TransitionKey
 from lerobot.utils.constants import (
+    ACTION_CODE_TOKEN_MASK,
     ACTION_TOKEN_MASK,
     ACTION_TOKENS,
     OBS_LANGUAGE_ATTENTION_MASK,
@@ -136,7 +138,7 @@ class TokenizerProcessorStep(ObservationProcessorStep):
         # Standardize to a list of strings for the tokenizer
         if isinstance(task, str):
             return [task]
-        elif isinstance(task, (list, tuple)) and all(isinstance(t, str) for t in task):
+        elif isinstance(task, list | tuple) and all(isinstance(t, str) for t in task):
             return list(task)
 
         return None
@@ -349,6 +351,8 @@ class ActionTokenizerProcessorStep(ActionProcessorStep):
     max_action_tokens: int = 256
     fast_skip_tokens: int = 128
     paligemma_tokenizer_name: str = "google/paligemma-3b-pt-224"
+    allow_truncation: bool = True
+    prepend_bos: bool = True
     # Internal tokenizer instance (not part of the config)
     action_tokenizer: Any = field(default=None, init=False, repr=False)
     _paligemma_tokenizer: Any = field(default=None, init=False, repr=False)
@@ -412,14 +416,15 @@ class ActionTokenizerProcessorStep(ActionProcessorStep):
             # During inference, no action is available, skip tokenization
             return new_transition
 
-        # Tokenize and get both tokens and mask
+        # Tokenize and get masks for the full formatted sequence and the discrete action codes.
-        tokens, mask = self._tokenize_action(action)
+        tokens, mask, code_mask = self._tokenize_action(action)
 
         # Store mask in complementary data
         complementary_data = new_transition.get(TransitionKey.COMPLEMENTARY_DATA, {})
         if complementary_data is None:
             complementary_data = {}
         complementary_data[ACTION_TOKEN_MASK] = mask
+        complementary_data[ACTION_CODE_TOKEN_MASK] = code_mask
         complementary_data[ACTION_TOKENS] = tokens
         new_transition[TransitionKey.COMPLEMENTARY_DATA] = complementary_data
         return new_transition
@@ -430,7 +435,7 @@ class ActionTokenizerProcessorStep(ActionProcessorStep):
         """
         return self._paligemma_tokenizer.vocab_size - 1 - self.fast_skip_tokens - tokens
 
-    def _tokenize_action(self, action: torch.Tensor) -> tuple[torch.Tensor, torch.Tensor]:
+    def _tokenize_action(self, action: torch.Tensor) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
         """
         Tokenizes the action tensor and creates a mask.
 
@@ -459,6 +464,7 @@ class ActionTokenizerProcessorStep(ActionProcessorStep):
         # The fast tokenizer expects action data and returns token IDs
         tokens_list = []
         masks_list = []
+        code_masks_list = []
 
         for i in range(batch_size):
             # Tokenize single action (move to CPU first as tokenizer uses scipy which requires numpy)
@@ -476,65 +482,79 @@ class ActionTokenizerProcessorStep(ActionProcessorStep):
             if tokens.dim() > 1:
                 tokens = tokens.flatten()
 
-            bos_id = self._paligemma_tokenizer.bos_token_id
+            action_code_tokens = self._act_tokens_to_paligemma_tokens(tokens)
-            # add bos
+            prompt_tokens = torch.tensor(
-            tokens = torch.cat(
+                self._paligemma_tokenizer.encode("Action: ", add_special_tokens=False),
-                [
+                device=action.device,
-                    torch.tensor([bos_id], device=action.device),
-                    torch.tensor(
-                        self._paligemma_tokenizer.encode("Action: ", add_special_tokens=False),
-                        device=action.device,
-                    ),
-                    self._act_tokens_to_paligemma_tokens(tokens),
-                    torch.tensor(self._paligemma_tokenizer.encode("|"), device=action.device),
-                ]
             )
+            end_tokens = torch.tensor(self._paligemma_tokenizer.encode("|"), device=action.device)
+
+            token_parts = []
+            if self.prepend_bos:
+                token_parts.append(
+                    torch.tensor([self._paligemma_tokenizer.bos_token_id], device=action.device)
+                )
+            code_start = sum(len(part) for part in token_parts) + len(prompt_tokens)
+            code_end = code_start + len(action_code_tokens)
+            tokens = torch.cat([*token_parts, prompt_tokens, action_code_tokens, end_tokens])
+            code_mask = torch.zeros(len(tokens), dtype=torch.bool, device=action.device)
+            code_mask[code_start:code_end] = True
 
             # Truncate or pad to max_action_tokens
             if len(tokens) > self.max_action_tokens:
+                if not self.allow_truncation:
+                    raise ValueError(
+                        f"FAST action sequence has {len(tokens)} tokens, exceeding "
+                        f"max_action_tokens={self.max_action_tokens}."
+                    )
                 logging.warning(
                     f"Token length ({len(tokens)}) exceeds max length ({self.max_action_tokens}), truncating. "
                     "Consider increasing the `max_action_tokens` in your model config if this happens frequently."
                 )
                 tokens = tokens[: self.max_action_tokens]
+                code_mask = code_mask[: self.max_action_tokens]
                 mask = torch.ones(self.max_action_tokens, dtype=torch.bool, device=action.device)
             else:
+                pad_len = self.max_action_tokens - len(tokens)
                 mask = torch.cat(
                     [
                         torch.ones(len(tokens), dtype=torch.bool, device=action.device),
-                        torch.zeros(
+                        torch.zeros(pad_len, dtype=torch.bool, device=action.device),
-                            self.max_action_tokens - len(tokens), dtype=torch.bool, device=action.device
-                        ),
                     ]
                 )
+                code_mask = torch.nn.functional.pad(code_mask, (0, pad_len), value=False)
                 # Pad tokens with zeros
-                tokens = torch.nn.functional.pad(tokens, (0, self.max_action_tokens - len(tokens)), value=0)
+                tokens = torch.nn.functional.pad(tokens, (0, pad_len), value=0)
 
             tokens_list.append(tokens)
             masks_list.append(mask)
+            code_masks_list.append(code_mask)
 
         # Stack into batched tensors
         tokens_batch = torch.stack(tokens_list, dim=0)  # (B, max_action_tokens)
         masks_batch = torch.stack(masks_list, dim=0)  # (B, max_action_tokens)
+        code_masks_batch = torch.stack(code_masks_list, dim=0)  # (B, max_action_tokens)
 
         # Remove batch dimension if input was single sample
         if single_sample:
             tokens_batch = tokens_batch.squeeze(0)
             masks_batch = masks_batch.squeeze(0)
+            code_masks_batch = code_masks_batch.squeeze(0)
 
         # Move to the same device as the input
         if device is not None:
             tokens_batch = tokens_batch.to(device)
             masks_batch = masks_batch.to(device)
+            code_masks_batch = code_masks_batch.to(device)
 
-        return tokens_batch, masks_batch
+        return tokens_batch, masks_batch, code_masks_batch
 
     def action(self, action: torch.Tensor) -> torch.Tensor:
         """
         This method is not used since we override __call__.
         Required by ActionProcessorStep ABC.
         """
-        tokens, _ = self._tokenize_action(action)
+        tokens, _, _ = self._tokenize_action(action)
         return tokens
 
     def get_config(self) -> dict[str, Any]:
@@ -550,6 +570,10 @@ class ActionTokenizerProcessorStep(ActionProcessorStep):
         config = {
             "trust_remote_code": self.trust_remote_code,
             "max_action_tokens": self.max_action_tokens,
+            "fast_skip_tokens": self.fast_skip_tokens,
+            "paligemma_tokenizer_name": self.paligemma_tokenizer_name,
+            "allow_truncation": self.allow_truncation,
+            "prepend_bos": self.prepend_bos,
         }
 
         # Only save tokenizer_name if it was used to create the tokenizer
@@ -558,6 +582,14 @@ class ActionTokenizerProcessorStep(ActionProcessorStep):
 
         return config
 
+    def save_artifacts(self, save_directory: Path) -> dict[str, str]:
+        artifact_path = Path("action_tokenizer")
+        save_pretrained = getattr(self.action_tokenizer, "save_pretrained", None)
+        if save_pretrained is None:
+            raise TypeError("Action tokenizer must implement save_pretrained() to save a portable pipeline.")
+        save_pretrained(save_directory / artifact_path)
+        return {"action_tokenizer_name": artifact_path.as_posix()}
+
     def transform_features(
         self, features: dict[PipelineFeatureType, dict[str, PolicyFeature]]
     ) -> dict[PipelineFeatureType, dict[str, PolicyFeature]]:

src/lerobot/rl/actor.py

@@ -91,7 +91,7 @@ from lerobot.robots import so_follower  # noqa: F401
 from lerobot.teleoperators import gamepad, so_leader  # noqa: F401
 from lerobot.teleoperators.utils import TeleopEvents
 from lerobot.utils.device_utils import get_safe_torch_device
-from lerobot.utils.process import ProcessSignalHandler, ensure_multiprocessing_start_method
+from lerobot.utils.process import ProcessSignalHandler
 from lerobot.utils.random_utils import set_seed
 from lerobot.utils.robot_utils import precise_sleep
 from lerobot.utils.transition import (
@@ -124,7 +124,9 @@ def actor_cli(cfg: TrainRLServerPipelineConfig):
     cfg.validate()
     display_pid = False
     if not use_threads(cfg):
-        ensure_multiprocessing_start_method(cfg.policy.concurrency.multiprocessing_context)
+        import torch.multiprocessing as mp
+
+        mp.set_start_method("spawn")
         display_pid = True
 
     # Create logs directory to ensure it exists

src/lerobot/rl/buffer.py

@@ -18,7 +18,7 @@ import functools
 import threading
 from collections.abc import Callable, Sequence
 from contextlib import suppress
-from typing import NotRequired, TypedDict
+from typing import TypedDict
 
 import torch
 import torch.nn.functional as F  # noqa: N812
@@ -36,7 +36,7 @@ class BatchTransition(TypedDict):
     next_state: dict[str, torch.Tensor]
     done: torch.Tensor
     truncated: torch.Tensor
-    complementary_info: NotRequired[dict[str, torch.Tensor | float | int] | None]
+    complementary_info: dict[str, torch.Tensor | float | int] | None = None
 
 
 def random_crop_vectorized(images: torch.Tensor, output_size: tuple) -> torch.Tensor:

src/lerobot/rl/learner.py

@@ -102,7 +102,7 @@ from lerobot.utils.constants import (
 )
 from lerobot.utils.device_utils import get_safe_torch_device
 from lerobot.utils.io_utils import load_json, write_json
-from lerobot.utils.process import ProcessSignalHandler, ensure_multiprocessing_start_method
+from lerobot.utils.process import ProcessSignalHandler
 from lerobot.utils.random_utils import set_seed
 from lerobot.utils.utils import (
     format_big_number,
@@ -123,7 +123,9 @@ def train_cli(cfg: TrainRLServerPipelineConfig):
     # Fail fast with a friendly error if the optional ``hilserl`` extra is missing.
     require_package("grpcio", extra="hilserl", import_name="grpc")
     if not use_threads(cfg):
-        ensure_multiprocessing_start_method(cfg.policy.concurrency.multiprocessing_context)
+        import torch.multiprocessing as mp
+
+        mp.set_start_method("spawn")
 
     # Use the job_name from the config
     train(

src/lerobot/robots/so_follower/config_so_follower.py

@@ -46,12 +46,6 @@ class SOFollowerConfig:
     position_i_coefficient: int = 0
     position_d_coefficient: int = 32
 
-    # Number of extra attempts when a `sync_read` of the motors fails. Feetech buses can occasionally
-    # return a corrupted status packet ("Incorrect status packet!"), especially when several joints move
-    # at once, which otherwise aborts the control loop. Retries are immediate (no sleep) and only happen on
-    # failure, so the steady-state read cost is unchanged.
-    num_read_retries: int = 2
-
 
 @RobotConfig.register_subclass("so101_follower")
 @RobotConfig.register_subclass("so100_follower")

src/lerobot/robots/so_follower/robot_kinematic_processor.py

@@ -510,10 +510,10 @@ class ForwardKinematicsJointsToEEAction(RobotActionProcessorStep):
         # We only use the ee pose in the dataset, so we don't need the joint positions
         for n in self.motor_names:
             features[PipelineFeatureType.ACTION].pop(f"{n}.pos", None)
-        # Store end-effector features as actions in the dataset schema
+        # We specify the dataset features of this step that we want to be stored in the dataset
         for k in ["x", "y", "z", "wx", "wy", "wz", "gripper_pos"]:
             features[PipelineFeatureType.ACTION][f"ee.{k}"] = PolicyFeature(
-                type=FeatureType.ACTION, shape=(1,)
+                type=FeatureType.STATE, shape=(1,)
             )
         return features
 

src/lerobot/robots/so_follower/so_follower.py

@@ -180,7 +180,7 @@ class SOFollower(Robot):
     def get_observation(self) -> RobotObservation:
         # Read arm position
         start = time.perf_counter()
-        obs_dict = self.bus.sync_read("Present_Position", num_retry=self.config.num_read_retries)
+        obs_dict = self.bus.sync_read("Present_Position")
         obs_dict = {f"{motor}.pos": val for motor, val in obs_dict.items()}
         dt_ms = (time.perf_counter() - start) * 1e3
         logger.debug(f"{self} read state: {dt_ms:.1f}ms")
@@ -221,7 +221,7 @@ class SOFollower(Robot):
         # Cap goal position when too far away from present position.
         # /!\ Slower fps expected due to reading from the follower.
         if self.config.max_relative_target is not None:
-            present_pos = self.bus.sync_read("Present_Position", num_retry=self.config.num_read_retries)
+            present_pos = self.bus.sync_read("Present_Position")
             goal_present_pos = {key: (g_pos, present_pos[key]) for key, g_pos in goal_pos.items()}
             goal_pos = ensure_safe_goal_position(goal_present_pos, self.config.max_relative_target)
 

src/lerobot/robots/unitree_g1/config_unitree_g1.py

@@ -62,34 +62,12 @@ class UnitreeG1Config(RobotConfig):
     # Socket config for ZMQ bridge
     robot_ip: str = "192.168.123.164"  # default G1 IP
 
-    # Run the locomotion / whole-body controller ONBOARD the robot (policy on the G1
-    # itself, against local DDS at full rate) instead of on the laptop over the ZMQ
-    # socket bridge. In this mode the robot object uses the real Unitree SDK channels
-    # and expects high-level actions (arm targets + joystick axes, or 64-D SONIC
-    # tokens) fed via send_action -- e.g. by run_g1_server's serve_onboard_controller,
-    # which receives them from the laptop over ZMQ. Mutually exclusive with is_simulation.
-    onboard: bool = False
-    # DDS network interface for onboard mode (None = SDK default, matching
-    # run_g1_server.py's ChannelFactoryInitialize(0)).
-    dds_interface: str | None = None
-    # Onboard sub-flags. On a real G1 both are True: the built-in motion services
-    # must be released before we can write lowcmd, and locomotion axes are read from
-    # the physical wireless remote. Against a DDS sim neither applies (no
-    # MotionSwitcher, no physical remote), so set both False so the controller takes
-    # its locomotion axes purely from send_action (ZMQ) input.
-    release_motion_control: bool = True
-    physical_remote: bool = True
-
     # Cameras (ZMQ-based remote cameras)
     cameras: dict[str, CameraConfig] = field(default_factory=dict)
 
     # Compensates for gravity on the unitree's arms using the arm ik solver
     gravity_compensation: bool = False
 
-    # Locomotion controller class name, e.g. "GrootLocomotionController",
+    # Lower-body controller class name, e.g. "GrootLocomotionController" or
-    # "HolosomaLocomotionController", or "SonicWholeBodyController". None disables it.
+    # "HolosomaLocomotionController". None disables it.
-    # Selecting "SonicWholeBodyController" implicitly switches the robot to the 64-D
-    # latent-token action/observation interface (``motion_token.{i}.pos`` action and a
-    # ``motion_token_state.{i}.pos`` state echo) so ``lerobot-rollout`` can drive a
-    # policy trained on SONIC motion tokens (e.g. nepyope/sonic_walk).
     controller: str | None = None

src/lerobot/robots/unitree_g1/controllers/sonic_whole_body.py

@@ -1,401 +0,0 @@
-#!/usr/bin/env python
-
-# Copyright 2025 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.
-
-"""SONIC decoder whole-body controller for the Unitree G1 (token-only).
-
-Pure-Python/ONNX re-implementation of the *decode* half of NVIDIA's SONIC deploy stack.
-The encoder is intentionally absent: a token-output VLA (e.g. ``nepyope/sonic_walk``)
-supplies the 64-D latent ``motion_token`` directly each tick, and the SONIC **decoder**
-maps ``token + recent proprioception history`` to a residual action that is scaled and
-added onto ``DEFAULT_ANGLES`` to produce 50 Hz joint-position targets for the robot's PD
-controller.
-
-Index spaces: joints exist in two orderings — **IsaacLab** (policy/training order) and
-**MuJoCo** (deploy order). ``ISAACLAB_TO_MUJOCO`` / ``MUJOCO_TO_ISAACLAB`` (in g1_utils)
-convert between them. Quaternions are scalar-first ``(w, x, y, z)``.
-"""
-
-from __future__ import annotations
-
-import logging
-
-import numpy as np
-import onnxruntime as ort
-from huggingface_hub import hf_hub_download
-
-from ..g1_utils import (
-    ISAACLAB_TO_MUJOCO,
-    MOTOR_ARMATURE,
-    MUJOCO_TO_ISAACLAB,
-    NATURAL_FREQ,
-    G1_29_JointIndex,
-    compute_pd_gains,
-    get_gravity_orientation,
-    lowstate_to_obs,
-    make_ort_session_options,
-)
-
-logger = logging.getLogger(__name__)
-
-# ── Constants (hardware-validated; see the NVIDIA SONIC deploy reference) ──────
-CONTROL_DT = 0.02  # 50 Hz control period (s)
-TOKEN_DIM = 64  # decoder latent size
-
-# Nominal standing pose (rad), 29 joints in IsaacLab order. Decoder actions are residuals
-# added on top of this.
-DEFAULT_ANGLES = np.array(
-    [
-        -0.312, 0.0, 0.0, 0.669, -0.363, 0.0,
-        -0.312, 0.0, 0.0, 0.669, -0.363, 0.0,
-        0.0, 0.0, 0.0,
-        0.2, 0.2, 0.0, 0.6, 0.0, 0.0, 0.0,
-        0.2, -0.2, 0.0, 0.6, 0.0, 0.0, 0.0,
-    ],
-    dtype=np.float32,
-)
-
-# Per-motor torque limits (N·m), used only for SONIC's residual-action scaling. The
-# armature / bandwidth constants and the PD-gain formula are shared (see g1_utils).
-EFFORT = {"5020": 25.0, "7520_14": 88.0, "7520_22": 139.0, "4010": 5.0}
-
-
-def _action_scale(k):
-    """Per-motor residual-action scale (maps policy output to joint-angle delta)."""
-    return 0.25 * EFFORT[k] / (MOTOR_ARMATURE[k] * NATURAL_FREQ**2)
-
-
-# Per-joint motor model (IsaacLab order): legs, waist, then arms. Single source of truth
-# for both ACTION_SCALE and compute_kp_kd().
-MOTOR_MODELS = (
-    ["7520_22", "7520_22", "7520_14", "7520_22", "5020", "5020"] * 2
-    + ["7520_14", "5020", "5020"]
-    + ["5020", "5020", "5020", "5020", "5020", "4010", "4010"] * 2
-)
-ACTION_SCALE = np.array([_action_scale(k) for k in MOTOR_MODELS], dtype=np.float32)  # (29,) IsaacLab
-
-
-def _to_mujoco(a):
-    """Apply the ``MUJOCO_TO_ISAACLAB`` gather to a 29-vector (deploy-order reorder).
-
-    NOTE: this returns ``a[MUJOCO_TO_ISAACLAB]``. The ``_mj`` suffixes and the exact
-    permutation direction are a fixed convention validated against the deployed SONIC ONNX
-    policy (the decoder consumes vectors in this order). Do not "correct" the table or
-    rename toward the opposite direction without re-validating on hardware.
-    """
-    return a[MUJOCO_TO_ISAACLAB]
-
-
-DEFAULT_ANGLES_MUJOCO = _to_mujoco(DEFAULT_ANGLES)
-
-
-# Ankle + waist joint indices (IsaacLab order) that get a x2 stiffness/damping factor.
-_SONIC_DOUBLE = {4, 5, 10, 11, 13, 14}
-
-
-def compute_kp_kd():
-    """SONIC per-joint PD gains (kp, kd), (29,) float32 in IsaacLab joint order."""
-    return compute_pd_gains(MOTOR_MODELS, _SONIC_DOUBLE)
-
-
-# Action-feature prefix for the latent-token interface (see _extract_token_from_action).
-TOKEN_ACTION_PREFIX = "motion_token"
-# Proprio-state prefix for the token interface: the robot echoes the last commanded token
-# here so ``lerobot-rollout`` aggregates it into a 64-D ``observation.state``.
-TOKEN_STATE_PREFIX = "motion_token_state"
-
-
-def token_action_key(i: int) -> str:
-    """Action-dict key for the i-th component of the 64-D SONIC latent token.
-
-    The ``.pos`` suffix is required so the value flows through ``lerobot-rollout``, which
-    only routes ``.pos`` scalar features onto the policy action vector.
-    """
-    return f"{TOKEN_ACTION_PREFIX}.{i}.pos"
-
-
-def token_state_key(i: int) -> str:
-    """Observation key for the i-th component of the 64-D SONIC latent token state."""
-    return f"{TOKEN_STATE_PREFIX}.{i}.pos"
-
-
-# Startup blend duration: over the first control ticks, linearly interpolate every joint
-# from the robot's initial measured pose into the policy's commanded target, so control
-# eases in without a snap on the first command.
-INIT_RAMP_S = 3.0
-
-# Neutral ("zero pose") SONIC token, held by token_mode until the first real token arrives.
-# Captured from the encoder's own output while the robot stood idle in sim: the encoder is
-# an FSQ bottleneck (~5 bit/dim, Div(16)), so its tokens live on the 1/16 grid. We store the
-# integer FSQ codes and rescale by 1/16, giving an exact on-grid token -- unlike the literal
-# all-zero token, which is off the learned manifold and decodes to a slightly goofy stance.
-# This one decodes to a stable, natural standing pose.
-_NEUTRAL_TOKEN_CODES = np.array(
-    [-1, 3, 1, -1, 1, -3, 6, 1, 1, 1, -2, -4, -2, 0, -3, -1,
-     2, -1, -3, -5, 3, 1, 1, -4, -1, -1, 1, -7, 0, 1, 2, -2,
-     5, -2, -2, -4, 0, -1, 3, -1, 0, -5, -1, 0, -4, 0, 0, -1,
-     -1, 2, -2, 1, 3, 3, 1, 0, 0, 6, 0, -7, 3, 0, 2, -2],
-    dtype=np.float32,
-)
-NEUTRAL_TOKEN = _NEUTRAL_TOKEN_CODES / 16.0  # FSQ Div(16): integer codes -> on-grid token
-
-
-def _extract_token_from_action(action: dict | None) -> np.ndarray | None:
-    """Reassemble a dense (64,) latent token from ``motion_token.{i}`` keys, or None.
-
-    The token-only interface: the caller supplies the 64-D encoder latent directly (e.g. a
-    token-output VLA's action), which the decoder consumes with the encoder bypassed.
-    Requires the full dense token; a partial one is ignored (returns None).
-    """
-    if not action:
-        return None
-    keys = [token_action_key(i) for i in range(TOKEN_DIM)]
-    if any(key not in action for key in keys):
-        return None
-    return np.fromiter((float(action[key]) for key in keys), dtype=np.float32, count=TOKEN_DIM)
-
-
-class SonicDecoder:
-    """Runs the SONIC decoder ONNX model and owns the proprioception history.
-
-    Each tick it appends the latest robot state to 10-frame history buffers, then maps the
-    supplied 64-D ``token`` + that history to a residual action added onto
-    ``DEFAULT_ANGLES``. The encoder is bypassed entirely (token supplied by the policy).
-    """
-
-    def __init__(self, decoder):
-        self.decoder = decoder
-        self.decoder_input = decoder.get_inputs()[0].name
-        dec_dim = int(decoder.get_inputs()[0].shape[1])
-        if dec_dim != 994:
-            raise RuntimeError(f"Unexpected decoder input dim {dec_dim} (expected 994)")
-        self.token = np.zeros(TOKEN_DIM, np.float32)
-        self.last_action_mj = np.zeros(29, np.float32)
-        self.h_q_mj = [np.zeros(29, np.float32)] * 10
-        self.h_dq_mj = [np.zeros(29, np.float32)] * 10
-        self.h_ang = [np.zeros(3, np.float32)] * 10
-        self.h_act_mj = [np.zeros(29, np.float32)] * 10
-        self.h_quat = [np.array([1, 0, 0, 0], np.float32)] * 10
-
-    def reset(self):
-        """Clear the token and 10-frame proprioception history.
-
-        ``UnitreeG1.reset()`` relies on this so the first decoder outputs of a new episode
-        are not contaminated by the previous episode's state.
-        """
-        self.token = np.zeros(TOKEN_DIM, np.float32)
-        self.last_action_mj = np.zeros(29, np.float32)
-        self.h_q_mj = [np.zeros(29, np.float32)] * 10
-        self.h_dq_mj = [np.zeros(29, np.float32)] * 10
-        self.h_ang = [np.zeros(3, np.float32)] * 10
-        self.h_act_mj = [np.zeros(29, np.float32)] * 10
-        self.h_quat = [np.array([1, 0, 0, 0], np.float32)] * 10
-
-    def update_history(self, q, dq, ang, quat):
-        """Push the latest proprioception (pos/vel/gyro/orientation) into the 10-frame buffers."""
-        quat = quat / (np.linalg.norm(quat) + 1e-8)
-        q_mj = _to_mujoco(q)
-        dq_mj = _to_mujoco(dq)
-        self.h_q_mj = [q_mj - DEFAULT_ANGLES_MUJOCO] + self.h_q_mj[:-1]
-        self.h_dq_mj = [dq_mj] + self.h_dq_mj[:-1]
-        self.h_ang = [ang.copy()] + self.h_ang[:-1]
-        self.h_act_mj = [self.last_action_mj.copy()] + self.h_act_mj[:-1]
-        self.h_quat = [quat.copy()] + self.h_quat[:-1]
-
-    def build_decoder_obs(self):
-        """Assemble the 994-D decoder input: token + 10-frame proprioception history + gravity."""
-        obs = np.zeros(994, np.float32)
-        off = 0
-        obs[off : off + 64] = self.token
-        off += 64
-        for h, sz in [
-            (list(reversed(self.h_ang)), 3),
-            (list(reversed(self.h_q_mj)), 29),
-            (list(reversed(self.h_dq_mj)), 29),
-            (list(reversed(self.h_act_mj)), 29),
-        ]:
-            for f in range(10):
-                obs[off : off + sz] = h[f]
-                off += sz
-        for q in reversed(self.h_quat):
-            obs[off : off + 3] = get_gravity_orientation(q)
-            off += 3
-        assert off == 994, f"Decoder obs mismatch: {off}"
-        return obs
-
-    def step(self, robot_obs, token, debug=False):
-        """One control tick: read robot obs, decode the supplied token -> joint targets.
-
-        Args:
-            robot_obs: dict with ``<joint>.q``/``.dq`` and ``imu.*`` fields.
-            token: 64-D latent supplied by the policy (encoder bypassed).
-            debug: log action/delta norms.
-
-        Returns:
-            dict of ``<joint>.q`` target positions (rad) in IsaacLab joint order.
-        """
-        self.token = np.asarray(token, np.float32)
-        jnames = [m.name for m in G1_29_JointIndex]
-        q = np.array(
-            [
-                robot_obs.get(f"{n}.q", DEFAULT_ANGLES[m.value])
-                for m, n in zip(G1_29_JointIndex, jnames, strict=False)
-            ],
-            np.float32,
-        )
-        dq = np.array([robot_obs.get(f"{n}.dq", 0.0) for n in jnames], np.float32)
-        quat = np.array(
-            [
-                robot_obs.get("imu.quat.w", 1),
-                robot_obs.get("imu.quat.x", 0),
-                robot_obs.get("imu.quat.y", 0),
-                robot_obs.get("imu.quat.z", 0),
-            ],
-            np.float32,
-        )
-        ang = np.array([robot_obs.get(f"imu.gyro.{a}", 0) for a in "xyz"], np.float32)
-        self.update_history(q, dq, ang, quat)
-        action_mj = (
-            self.decoder.run(None, {self.decoder_input: self.build_decoder_obs().reshape(1, -1)})[0]
-            .squeeze()
-            .astype(np.float32)
-        )
-        self.last_action_mj = action_mj.copy()
-        target = DEFAULT_ANGLES + action_mj[ISAACLAB_TO_MUJOCO] * ACTION_SCALE
-        if debug:
-            delta = target - q
-            logger.debug(
-                "token_norm=%.4f action_norm=%.4f delta_max=%.4f delta_rms=%.4f",
-                np.linalg.norm(self.token),
-                np.linalg.norm(action_mj),
-                np.max(np.abs(delta)),
-                np.sqrt(np.mean(delta**2)),
-            )
-        return {f"{m.name}.q": float(target[m.value]) for m in G1_29_JointIndex}
-
-
-class SonicRuntime:
-    """Loads the SONIC decoder ONNX model and owns the decode controller.
-
-    Token-only deploy: the encoder is bypassed; each tick the decoder consumes a 64-D
-    latent token supplied directly by the policy.
-    """
-
-    def __init__(self):
-        decoder_path = hf_hub_download(repo_id="nvidia/GEAR-SONIC", filename="model_decoder.onnx")
-
-        so = make_ort_session_options()
-        decoder_sess = ort.InferenceSession(decoder_path, sess_options=so)
-
-        self.kp, self.kd = compute_kp_kd()
-        self.controller = SonicDecoder(decoder_sess)
-
-    @property
-    def pipeline(self):
-        return self.controller
-
-    def reset(self):
-        self.controller.reset()
-
-    def shutdown(self):
-        pass
-
-
-class SonicWholeBodyController:
-    """Full-body SONIC controller for UnitreeG1's background controller thread."""
-
-    control_dt = CONTROL_DT
-    full_body = True
-
-    def __init__(self):
-        logger.info("Loading SONIC whole-body controller...")
-        self._runtime = SonicRuntime()
-        self.kp = self._runtime.kp
-        self.kd = self._runtime.kd
-        self.controller = self._runtime.controller
-
-        # Startup blend: ease from the robot's initial pose into the first commanded policy
-        # targets over INIT_RAMP_S (captured on the first control tick).
-        self._init_ramp_steps = max(1, round(INIT_RAMP_S / CONTROL_DT))
-        self._init_step = 0
-        self._start_pose: dict[str, float] = {}
-
-        # Token-interface state. ``token_mode`` is set True by the robot whenever a SONIC
-        # whole-body controller is selected (token-driven deploy): the controller then holds a
-        # stable *neutral* token until the first real token arrives, and afterwards holds the
-        # *last* token received between ticks (the async controller runs ~50 Hz while a token
-        # VLA streams ~30 Hz). This lives here (not in the entry-point script) so it applies
-        # uniformly to run_g1_server, lerobot-rollout and the sim replays.
-        self.token_mode = False
-        self._last_token: np.ndarray | None = None
-
-        logger.info("SONIC ready (decoder, 64-D token command path)")
-
-    def _startup_blend(self, obs: dict, out: dict) -> dict:
-        """Ease into policy control at startup: for the first ``INIT_RAMP_S`` seconds,
-        interpolate between the robot's pose captured on the first tick and the policy's
-        live commanded target, so the handoff has no snap.
-
-        ``out`` is the policy's ``<joint>.q`` target dict for this tick; the blend ratio
-        climbs 0->1 over the ramp, after which the raw policy target passes through.
-        """
-        if self._init_step >= self._init_ramp_steps or not out:
-            return out
-        if self._init_step == 0:
-            # Capture the robot's actual pose as the interpolation start point.
-            self._start_pose = {
-                f"{m.name}.q": float(obs.get(f"{m.name}.q", DEFAULT_ANGLES[m.value]))
-                for m in G1_29_JointIndex
-            }
-        self._init_step += 1
-        ratio = min(1.0, self._init_step / self._init_ramp_steps)
-        blended = {
-            k: self._start_pose.get(k, float(tgt)) * (1.0 - ratio) + float(tgt) * ratio
-            for k, tgt in out.items()
-        }
-        if self._init_step >= self._init_ramp_steps:
-            logger.info("SONIC startup blend complete -> full policy control")
-        return blended
-
-    def run_step(self, action: dict, lowstate) -> dict:
-        if lowstate is None:
-            return {}
-        obs = lowstate_to_obs(lowstate)
-
-        # Token-only interface (token-output VLA): a dense 64-D ``motion_token.{i}`` command
-        # is decoded directly, encoder bypassed.
-        token = _extract_token_from_action(action)
-        if token is not None:
-            self._last_token = token
-        elif self._last_token is None and self.token_mode:
-            # Token-driven deploy, but no token has arrived yet: hold the captured neutral
-            # token (NEUTRAL_TOKEN), which the decoder maps to a stable, natural standing pose.
-            self._last_token = NEUTRAL_TOKEN.copy()
-        if self._last_token is None:
-            # No token yet and not in token_mode: hold (keep last target).
-            return {}
-        # Either a fresh token this tick or the last one received (held between the ~30 Hz
-        # token stream and the ~50 Hz control loop).
-        return self._startup_blend(obs, self.controller.step(obs, self._last_token))
-
-    def reset(self):
-        self._runtime.reset()
-        self._init_step = 0  # re-run the startup blend after a reset
-        self._start_pose = {}
-        # Drop the held token so token_mode re-seeds the neutral token after a reset.
-        self._last_token = None
-
-    def shutdown(self):
-        self._runtime.shutdown()

src/lerobot/robots/unitree_g1/g1_utils.py

@@ -23,82 +23,11 @@ import numpy as np
 
 NUM_MOTORS = 29
 
-# Joint-order permutations between the two 29-DoF layouts used across the G1 stack:
-# IsaacLab (policy/training order) and MuJoCo (deploy order). ``a[ISAACLAB_TO_MUJOCO]``
-# reorders an IsaacLab-ordered vector into MuJoCo order, and vice-versa.
-ISAACLAB_TO_MUJOCO = np.array(
-    [
-        0,
-        3,
-        6,
-        9,
-        13,
-        17,
-        1,
-        4,
-        7,
-        10,
-        14,
-        18,
-        2,
-        5,
-        8,
-        11,
-        15,
-        19,
-        21,
-        23,
-        25,
-        27,
-        12,
-        16,
-        20,
-        22,
-        24,
-        26,
-        28,
-    ],
-    dtype=np.int32,
-)
-MUJOCO_TO_ISAACLAB = np.array(
-    [
-        0,
-        6,
-        12,
-        1,
-        7,
-        13,
-        2,
-        8,
-        14,
-        3,
-        9,
-        15,
-        22,
-        4,
-        10,
-        16,
-        23,
-        5,
-        11,
-        17,
-        24,
-        18,
-        25,
-        19,
-        26,
-        20,
-        27,
-        21,
-        28,
-    ],
-    dtype=np.int32,
-)
-
 REMOTE_AXES = ("remote.lx", "remote.ly", "remote.rx", "remote.ry")
 REMOTE_BUTTONS = tuple(f"remote.button.{i}" for i in range(16))
 REMOTE_KEYS = REMOTE_AXES + REMOTE_BUTTONS
 
+
 def default_remote_input() -> dict[str, float]:
     """Return a zeroed-out remote input dict (axes + buttons)."""
     return dict.fromkeys(REMOTE_KEYS, 0.0)
@@ -114,53 +43,6 @@ def get_gravity_orientation(quaternion: list[float] | np.ndarray) -> np.ndarray:
     return gravity_orientation
 
 
-# Unitree motor-model parameters shared by the controllers that derive their PD gains
-# from motor physics rather than hand-tuning (SONIC decoder, Holosoma). NATURAL_FREQ is
-# the target closed-loop stiffness bandwidth (rad/s); MOTOR_ARMATURE is per-model rotor
-# inertia (keys are Unitree motor model names). From these: kp = armature * w**2 and
-# kd = 4 * armature * w, with an optional x2 factor on stiff joints (ankles/waist).
-NATURAL_FREQ = 10.0 * 2.0 * np.pi
-MOTOR_ARMATURE = {"5020": 0.003609725, "7520_14": 0.010177520, "7520_22": 0.025101925, "4010": 0.00425}
-
-
-def compute_pd_gains(motor_models, double_indices=()) -> tuple[np.ndarray, np.ndarray]:
-    """Derive per-joint PD gains (kp, kd) from motor armature and target bandwidth.
-
-    ``motor_models`` is a per-joint sequence of Unitree motor model names (in the
-    controller's own joint order); joints whose index is in ``double_indices`` get a
-    x2 stiffness/damping factor. Returns two (N,) float32 arrays in that same order.
-    """
-    double = set(double_indices)
-
-    def s(k):
-        return MOTOR_ARMATURE[k] * NATURAL_FREQ**2
-
-    def d(k):
-        return 4.0 * MOTOR_ARMATURE[k] * NATURAL_FREQ
-
-    kp = np.array([2 * s(k) if i in double else s(k) for i, k in enumerate(motor_models)], dtype=np.float32)
-    kd = np.array([2 * d(k) if i in double else d(k) for i, k in enumerate(motor_models)], dtype=np.float32)
-    return kp, kd
-
-
-def make_ort_session_options(intra_op_num_threads: int | None = None, inter_op_num_threads: int | None = None):
-    """Build quiet ONNX Runtime SessionOptions, optionally capping the CPU thread pool.
-
-    These tiny MLP policies are latency-bound, not throughput-bound, so letting ORT grab
-    every core starves the real-time control loop / torch policy and causes stutter. Pass
-    1 intra + 1 inter thread for lowest-latency per-step inference.
-    """
-    import onnxruntime as ort
-
-    so = ort.SessionOptions()
-    so.log_severity_level = 3
-    if intra_op_num_threads is not None:
-        so.intra_op_num_threads = intra_op_num_threads
-    if inter_op_num_threads is not None:
-        so.inter_op_num_threads = inter_op_num_threads
-    return so
-
-
 class G1_29_JointArmIndex(IntEnum):
     # Left arm
     kLeftShoulderPitch = 15
@@ -181,55 +63,13 @@ class G1_29_JointArmIndex(IntEnum):
     kRightWristYaw = 28
 
 
-def lowstate_to_obs(lowstate) -> dict:
-    """Build a robot observation dict from a Unitree lowstate.
-
-    Shared by ``UnitreeG1.get_observation`` and the SONIC pipeline so the
-    lowstate -> obs mapping lives in exactly one place. Keys match the
-    ``<joint>.q``/``imu.*`` schema consumed across the controllers.
-    """
-    obs: dict = {}
-
-    for motor in G1_29_JointIndex:
-        idx = motor.value
-        obs[f"{motor.name}.q"] = lowstate.motor_state[idx].q
-        obs[f"{motor.name}.dq"] = lowstate.motor_state[idx].dq
-        obs[f"{motor.name}.tau"] = lowstate.motor_state[idx].tau_est
-
-    imu = lowstate.imu_state
-    if imu.gyroscope:
-        obs["imu.gyro.x"] = imu.gyroscope[0]
-        obs["imu.gyro.y"] = imu.gyroscope[1]
-        obs["imu.gyro.z"] = imu.gyroscope[2]
-    if imu.accelerometer:
-        obs["imu.accel.x"] = imu.accelerometer[0]
-        obs["imu.accel.y"] = imu.accelerometer[1]
-        obs["imu.accel.z"] = imu.accelerometer[2]
-    if imu.quaternion:
-        obs["imu.quat.w"] = imu.quaternion[0]
-        obs["imu.quat.x"] = imu.quaternion[1]
-        obs["imu.quat.y"] = imu.quaternion[2]
-        obs["imu.quat.z"] = imu.quaternion[3]
-    if imu.rpy:
-        obs["imu.rpy.roll"] = imu.rpy[0]
-        obs["imu.rpy.pitch"] = imu.rpy[1]
-        obs["imu.rpy.yaw"] = imu.rpy[2]
-
-    wr = getattr(lowstate, "wireless_remote", None)
-    if wr:
-        obs["wireless_remote"] = bytes(wr) if not isinstance(wr, (bytes, bytearray)) else wr
-
-    return obs
-
-
 def make_locomotion_controller(name: str | None):
     """Instantiate a locomotion controller by class name. Returns None if name is None."""
     if name is None:
         return None
     controllers = {
-        "GrootLocomotionController": "lerobot.robots.unitree_g1.controllers.gr00t_locomotion",
+        "GrootLocomotionController": "lerobot.robots.unitree_g1.gr00t_locomotion",
-        "HolosomaLocomotionController": "lerobot.robots.unitree_g1.controllers.holosoma_locomotion",
+        "HolosomaLocomotionController": "lerobot.robots.unitree_g1.holosoma_locomotion",
-        "SonicWholeBodyController": "lerobot.robots.unitree_g1.controllers.sonic_whole_body",
     }
     module_path = controllers.get(name)
     if module_path is None:

src/lerobot/robots/unitree_g1/gr00t_locomotion.py

@@ -14,8 +14,6 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-from __future__ import annotations
-
 import logging
 from collections import deque
 
@@ -23,7 +21,7 @@ import numpy as np
 import onnxruntime as ort
 from huggingface_hub import hf_hub_download
 
-from ..g1_utils import (
+from .g1_utils import (
     REMOTE_AXES,
     REMOTE_BUTTONS,
     G1_29_JointIndex,
@@ -70,15 +68,9 @@ def load_groot_policies(
         filename="GR00T-WholeBodyControl-Walk.onnx",
     )
 
-    # Load ONNX policies with a capped thread pool. GR00T runs at 50 Hz in a
+    # Load ONNX policies
-    # background thread alongside the (torch) upper-body policy, IK and sim; letting
+    policy_balance = ort.InferenceSession(balance_path)
-    # ORT grab every core starves those and makes the whole rollout stutter. These
+    policy_walk = ort.InferenceSession(walk_path)
-    # are small MLPs, so 1 thread is both enough and lowest-latency.
-    from ..g1_utils import make_ort_session_options
-
-    so = make_ort_session_options(intra_op_num_threads=1, inter_op_num_threads=1)
-    policy_balance = ort.InferenceSession(balance_path, sess_options=so)
-    policy_walk = ort.InferenceSession(walk_path, sess_options=so)
 
     logger.info("GR00T policies loaded successfully")
 
@@ -204,16 +196,6 @@ class GrootLocomotionController:
         # Transform action back to target joint positions
         target_dof_pos_15 = GROOT_DEFAULT_ANGLES[:15] + self.groot_action * ACTION_SCALE
 
-        # Waist override: an external upper-body IK can command the 3 waist joints
-        # (indices 12/13/14) via ``kWaist{Yaw,Roll,Pitch}.q`` in the action dict. When
-        # present, we substitute the balance policy's waist target so the torso tracks
-        # the IK while the policy keeps only the legs balanced. Single-publisher stays
-        # intact (this thread still owns joints 0-14).
-        for idx in (G1_29_JointIndex.kWaistYaw, G1_29_JointIndex.kWaistRoll, G1_29_JointIndex.kWaistPitch):
-            key = f"{idx.name}.q"
-            if key in action and action[key] is not None:
-                target_dof_pos_15[idx.value] = float(action[key])
-
         # Build action dict
         action_dict = {}
         for i in range(15):

src/lerobot/robots/unitree_g1/holosoma_locomotion.py

@@ -14,19 +14,18 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-from __future__ import annotations
+import json
-
 import logging
 
 import numpy as np
+import onnx
 import onnxruntime as ort
 from huggingface_hub import hf_hub_download
 
-from ..g1_utils import (
+from .g1_utils import (
     REMOTE_AXES,
     G1_29_JointArmIndex,
     G1_29_JointIndex,
-    compute_pd_gains,
     get_gravity_orientation,
 )
 
@@ -58,23 +57,12 @@ POLICY_FILES = {
     "ppo": "ppo_g1_29dof.onnx",
 }
 
-# Per-joint motor model in Holosoma's joint order, plus the joints that get a x2
-# stiffness/damping factor. These reproduce the kp/kd that used to be read from the
-# policy's ONNX metadata exactly (both fastsac and ppo), so gains are now derived from
-# the shared motor model (see g1_utils.compute_pd_gains) instead.
-HOLOSOMA_MOTOR_MODELS = (
-    ["7520_14", "7520_22", "7520_14", "7520_22", "5020", "5020"] * 2
-    + ["7520_14", "5020", "5020"]
-    + ["5020", "5020", "5020", "5020", "5020", "4010", "4010"] * 2
-)
-HOLOSOMA_DOUBLE = {4, 5, 10, 11, 13, 14}
-
 
 def load_policy(
     repo_id: str = DEFAULT_HOLOSOMA_REPO_ID,
     policy_type: str = "fastsac",
 ) -> tuple[ort.InferenceSession, np.ndarray, np.ndarray]:
-    """Load the Holosoma locomotion policy and its motor-model-derived PD gains.
+    """Load Holosoma locomotion policy and extract KP/KD from metadata.
 
     Args:
         repo_id: Hugging Face Hub repo ID
@@ -93,7 +81,16 @@ def load_policy(
     policy = ort.InferenceSession(policy_path)
     logger.info(f"Policy loaded: {policy.get_inputs()[0].shape} → {policy.get_outputs()[0].shape}")
 
-    kp, kd = compute_pd_gains(HOLOSOMA_MOTOR_MODELS, HOLOSOMA_DOUBLE)
+    # Extract KP/KD from ONNX metadata
+    model = onnx.load(policy_path, load_external_data=False)
+    metadata = {prop.key: prop.value for prop in model.metadata_props}
+
+    if "kp" not in metadata or "kd" not in metadata:
+        raise ValueError("ONNX model must contain 'kp' and 'kd' in metadata")
+
+    kp = np.array(json.loads(metadata["kp"]), dtype=np.float32)
+    kd = np.array(json.loads(metadata["kd"]), dtype=np.float32)
+    logger.info(f"Loaded KP/KD from ONNX ({len(kp)} joints)")
 
     return policy, kp, kd
 

src/lerobot/robots/unitree_g1/run_g1_server.py

@@ -22,33 +22,16 @@ This server runs on the robot and forwards:
 - Robot commands (LowCmd) from ZMQ to DDS (from remote clients)
 
 Uses JSON for secure serialization instead of pickle.
-
-Controller-negotiation handshake
---------------------------------
-The first message from a client agrees on which controller the server will run onboard
-(``serve_onboard_controller``); the controller NEVER runs on the laptop client.
-Test the handshake in isolation (no DDS, runs on a laptop) in two terminals::
-
-    # terminal A: handshake-only server
-    python -m lerobot.robots.unitree_g1.run_g1_server --handshake-only
-
-    # terminal B: client proposes a controller
-    python -m lerobot.robots.unitree_g1.run_g1_server \\
-        --handshake-client SonicWholeBodyController --sonic-token-action --server-ip 127.0.0.1
-
-On the real robot, add ``--handshake`` to the normal bridge to require agreement first.
 """
 
 import argparse
 import base64
 import contextlib
 import json
-import signal
 import threading
 import time
 from typing import Any
 
-import numpy as np
 import zmq
 from unitree_sdk2py.comm.motion_switcher.motion_switcher_client import MotionSwitcherClient
 from unitree_sdk2py.core.channel import ChannelFactoryInitialize, ChannelPublisher, ChannelSubscriber
@@ -67,253 +50,6 @@ LOWCMD_PORT = 6000
 LOWSTATE_PORT = 6001
 NUM_MOTORS = 35
 
-# Onboard high-level channels (serve_onboard_controller): compact actions in, state out.
-ACTION_PORT = 6004
-STATE_PORT = 6005
-
-# Controller-negotiation handshake (REQ/REP). The client's first message agrees on
-# which controller the server will run before any control data flows.
-HANDSHAKE_PORT = 6002
-PROTOCOL_VERSION = 1
-
-# Controllers that can run ONBOARD (must match g1_utils.make_locomotion_controller).
-# ``None`` (a.k.a. "bridge") means no onboard controller: the laptop owns control and
-# streams raw lowcmd over the ZMQ DDS bridge (the legacy run_g1_server behavior).
-VALID_CONTROLLERS = (
-    "GrootLocomotionController",
-    "HolosomaLocomotionController",
-    "SonicWholeBodyController",
-)
-# SONIC latent-token dimensionality (mirrors sonic_whole_body.TOKEN_DIM; kept local so
-# the handshake can run without importing the heavy controller / onnxruntime).
-TOKEN_DIM = 64
-_BRIDGE_ALIASES = {"", "none", "null", "bridge", "raw"}
-
-
-def _normalize_controller(name: str | None) -> str | None:
-    """Map a requested controller name to a canonical value (or None for raw bridge)."""
-    if name is None:
-        return None
-    low = str(name).strip().lower()
-    if low in _BRIDGE_ALIASES:
-        return None
-    for c in VALID_CONTROLLERS:
-        if c.lower() == low:
-            return c
-    raise ValueError(f"Unknown controller {name!r}. Available: {list(VALID_CONTROLLERS)} or 'bridge'")
-
-
-def _capabilities(controller: str | None, sonic_token_action: bool) -> dict[str, Any]:
-    """The interface the server advertises for an agreed controller."""
-    caps: dict[str, Any] = {
-        "controller": controller,
-        "sonic_token_action": bool(sonic_token_action),
-        "protocol": PROTOCOL_VERSION,
-    }
-    if controller is None:
-        # Raw DDS bridge: the laptop runs the controller and streams lowcmd.
-        caps["mode"] = "bridge"
-        caps["lowcmd_port"] = LOWCMD_PORT
-        caps["lowstate_port"] = LOWSTATE_PORT
-    else:
-        # Onboard: the controller runs here; the laptop ships compact high-level actions.
-        caps["mode"] = "onboard"
-        caps["action_port"] = ACTION_PORT
-        caps["state_port"] = STATE_PORT
-        if sonic_token_action:
-            caps["action_space"] = "motion_token"
-            caps["action_dim"] = TOKEN_DIM
-    return caps
-
-
-def negotiate_controller(sock: zmq.Socket, shutdown_event: threading.Event) -> dict[str, Any]:
-    """Server side of the handshake: block on one REP socket until a client sends a
-    valid ``hello``, then reply with the negotiated capabilities and return them.
-
-    Rejects malformed / unknown-controller requests with an error reply and keeps
-    waiting (a rejected client can retry). Honors ``shutdown_event`` so Ctrl-C works.
-    """
-    poller = zmq.Poller()
-    poller.register(sock, zmq.POLLIN)
-    while not shutdown_event.is_set():
-        if not dict(poller.poll(timeout=200)):
-            continue
-        raw = sock.recv()
-        try:
-            hello = json.loads(raw.decode("utf-8"))
-        except (ValueError, UnicodeDecodeError) as e:
-            sock.send_json({"type": "error", "ok": False, "error": f"bad hello: {e}"})
-            continue
-        try:
-            controller = _normalize_controller(hello.get("controller"))
-        except ValueError as e:
-            sock.send_json(
-                {"type": "error", "ok": False, "error": str(e), "available": list(VALID_CONTROLLERS)}
-            )
-            continue
-        reply = {"type": "welcome", "ok": True, **_capabilities(controller, hello.get("sonic_token_action", False))}
-        sock.send_json(reply)
-        return reply
-    raise KeyboardInterrupt
-
-
-def request_controller(
-    server_ip: str,
-    controller: str | None,
-    *,
-    sonic_token_action: bool = False,
-    port: int = HANDSHAKE_PORT,
-    timeout_s: float = 5.0,
-) -> dict[str, Any]:
-    """Client side of the handshake: propose a controller, return the server's agreed
-    capabilities (or raise on rejection / timeout)."""
-    ctx = zmq.Context.instance()
-    sock = ctx.socket(zmq.REQ)
-    sock.setsockopt(zmq.LINGER, 0)
-    sock.setsockopt(zmq.RCVTIMEO, int(timeout_s * 1000))
-    sock.setsockopt(zmq.SNDTIMEO, int(timeout_s * 1000))
-    sock.connect(f"tcp://{server_ip}:{port}")
-    hello = {
-        "type": "hello",
-        "controller": controller,
-        "sonic_token_action": bool(sonic_token_action),
-        "protocol": PROTOCOL_VERSION,
-    }
-    try:
-        sock.send_json(hello)
-        reply = sock.recv_json()
-    except zmq.Again as e:
-        raise TimeoutError(f"no handshake reply from {server_ip}:{port} within {timeout_s}s") from e
-    finally:
-        sock.close(linger=0)
-    if not reply.get("ok"):
-        raise RuntimeError(f"handshake rejected: {reply.get('error')} (available: {reply.get('available')})")
-    return reply
-
-
-def serve_onboard_controller(
-    *,
-    controller: str,
-    sonic_token_action: bool,
-    dds_interface: str | None = None,
-    sim: bool = False,
-    cameras: dict | None = None,
-    camera_fps: int = 30,
-    camera_port: int = 5555,
-    action_port: int = ACTION_PORT,
-    state_port: int = STATE_PORT,
-    state_fps: float = 30.0,
-    stop: threading.Event | None = None,
-) -> None:
-    """Run the negotiated controller ONBOARD -- the single control path on the robot.
-
-    Builds ``UnitreeG1(onboard=True, controller=...)`` so the controller/balance loop runs
-    locally against DDS at full rate (the 50 Hz ``_controller_loop`` thread lives in
-    UnitreeG1), then receives compact high-level actions from the laptop over ZMQ
-    (:action_port), decodes them via the controller, publishes ``observation.state``
-    (:state_port), and optionally serves the ego camera. The controller NEVER runs on the
-    laptop; the laptop (lerobot-rollout thin-client) only ships tokens/axes and reads back
-    state + camera frames.
-    """
-    # Imported lazily: UnitreeG1 imports request_controller from this module, so a
-    # top-level import here would be circular.
-    from lerobot.robots.unitree_g1.config_unitree_g1 import UnitreeG1Config
-    from lerobot.robots.unitree_g1.unitree_g1 import UnitreeG1
-
-    if stop is None:
-        stop = threading.Event()
-        signal.signal(signal.SIGINT, lambda *_: stop.set())
-        signal.signal(signal.SIGTERM, lambda *_: stop.set())
-
-    cfg = UnitreeG1Config(
-        is_simulation=False,
-        onboard=True,
-        controller=controller,
-        dds_interface=dds_interface,
-        release_motion_control=not sim,
-        physical_remote=not sim,
-        cameras={},
-    )
-
-    # Optional camera server (background daemon thread; independent of DDS).
-    if cameras:
-        camera_server = ImageServer({"fps": camera_fps, "cameras": cameras}, port=camera_port)
-        threading.Thread(target=camera_server.run, daemon=True).start()
-        cam_summary = ", ".join(f"{name}(dev {c['device_id']})" for name, c in cameras.items())
-        print(f"Camera server started on :{camera_port}: {cam_summary}")
-
-    robot = UnitreeG1(cfg)
-    print(f"Connecting onboard robot (controller={controller}, token={sonic_token_action})...")
-    robot.connect()
-
-    ctx = zmq.Context.instance()
-    sock = ctx.socket(zmq.PULL)
-    sock.setsockopt(zmq.CONFLATE, 1)  # only ever act on the freshest command
-    sock.setsockopt(zmq.RCVTIMEO, 200)  # keeps the loop responsive to the stop event
-    sock.bind(f"tcp://0.0.0.0:{action_port}")
-    print(f"Onboard controller live. Waiting for laptop actions on :{action_port} ...")
-    print("Ctrl-C for graceful shutdown.")
-
-    state_sock = None
-    if state_fps > 0:
-        state_sock = ctx.socket(zmq.PUB)
-        state_sock.setsockopt(zmq.SNDHWM, 2)
-        state_sock.setsockopt(zmq.LINGER, 0)
-        state_sock.bind(f"tcp://0.0.0.0:{state_port}")
-        print(f"Publishing observation.state on :{state_port} at {state_fps:.0f} Hz")
-
-        def publish_state() -> None:
-            period = 1.0 / state_fps
-            while not stop.is_set():
-                t0 = time.time()
-                obs = robot.get_observation()
-                if obs:
-                    # Forward every scalar proprio key the robot exposes (29 joint .q, IMU,
-                    # and the SONIC token echo: 64-D motion_token_state.*). Camera arrays are
-                    # streamed separately by the ImageServer, so drop ndarrays here. This
-                    # makes the laptop thin-client a pure relay.
-                    state = {
-                        k: float(v)
-                        for k, v in obs.items()
-                        if isinstance(v, (bool, int, float, np.floating, np.integer))
-                    }
-                    with contextlib.suppress(zmq.Again):
-                        state_sock.send_json(state, zmq.NOBLOCK)
-                time.sleep(max(0.0, period - (time.time() - t0)))
-
-        threading.Thread(target=publish_state, daemon=True).start()
-    else:
-        print("observation.state PUB disabled (state_fps<=0)")
-
-    n = 0
-    try:
-        while not stop.is_set():
-            try:
-                payload = sock.recv()
-            except zmq.Again:
-                continue
-            except zmq.ContextTerminated:
-                break
-
-            try:
-                action = json.loads(payload.decode("utf-8"))
-            except (ValueError, UnicodeDecodeError) as e:
-                print(f"Dropping malformed action: {e}")
-                continue
-
-            robot.send_action(action)
-
-            n += 1
-            if n % 60 == 0:
-                print(f"Applied {n} actions")
-    finally:
-        print("Shutting down onboard controller...")
-        stop.set()
-        if state_sock is not None:
-            with contextlib.suppress(Exception):
-                state_sock.close(linger=0)
-        robot.disconnect()
-
 
 def lowstate_to_dict(msg: hg_LowState) -> dict[str, Any]:
     """Convert LowState SDK message to a JSON-serializable dictionary."""
@@ -424,86 +160,8 @@ def main() -> None:
     parser.add_argument("--camera-width", type=int, default=640, help="Camera width (default: 640)")
     parser.add_argument("--camera-height", type=int, default=480, help="Camera height (default: 480)")
     parser.add_argument("--camera-port", type=int, default=5555, help="Camera ZMQ port (default: 5555)")
-    # Controller-negotiation handshake (first message agrees on the controller).
-    parser.add_argument("--handshake", action="store_true",
-                        help="Wait for a client to negotiate the controller before bridging")
-    parser.add_argument("--handshake-port", type=int, default=HANDSHAKE_PORT,
-                        help=f"Handshake REQ/REP port (default: {HANDSHAKE_PORT})")
-    parser.add_argument("--handshake-only", action="store_true",
-                        help="Run ONLY the handshake server (no DDS/cameras) to test negotiation")
-    parser.add_argument("--handshake-client", default=None, metavar="CONTROLLER",
-                        help="Act as a client: propose CONTROLLER (or 'bridge') to --server-ip and print the reply")
-    parser.add_argument("--server-ip", default="127.0.0.1", help="[--handshake-client] server IP")
-    parser.add_argument("--sonic-token-action", action="store_true",
-                        help="[handshake] negotiate the 64-D SONIC token action interface")
     args = parser.parse_args()
 
-    # --- Isolated handshake test paths (no DDS, safe to run on a laptop) ---
-    if args.handshake_client is not None:
-        controller = None if args.handshake_client.strip().lower() in _BRIDGE_ALIASES else args.handshake_client
-        reply = request_controller(
-            args.server_ip, controller,
-            sonic_token_action=args.sonic_token_action, port=args.handshake_port,
-        )
-        print(json.dumps(reply, indent=2))
-        return
-
-    if args.handshake_only:
-        ctx = zmq.Context.instance()
-        rep = ctx.socket(zmq.REP)
-        rep.bind(f"tcp://0.0.0.0:{args.handshake_port}")
-        print(f"[handshake] server listening on :{args.handshake_port} (no DDS). Ctrl-C to stop.")
-        shutdown = threading.Event()
-        try:
-            while True:
-                reply = negotiate_controller(rep, shutdown)
-                print(f"[handshake] agreed: controller={reply['controller']} mode={reply['mode']} "
-                      f"sonic_token_action={reply['sonic_token_action']}")
-        except KeyboardInterrupt:
-            print("\n[handshake] stopping")
-        finally:
-            rep.close(linger=0)
-            ctx.term()
-        return
-
-    # Controller-negotiation handshake: the client's first message agrees on the
-    # controller, which we then run ONBOARD (the controller NEVER runs on the laptop).
-    # Bridge/None falls through to the legacy raw DDS forward (deprecated laptop control).
-    if args.handshake:
-        ctx = zmq.Context.instance()
-        hs = ctx.socket(zmq.REP)
-        hs.bind(f"tcp://0.0.0.0:{args.handshake_port}")
-        print(f"[handshake] waiting for client controller agreement on :{args.handshake_port} ...")
-        shutdown = threading.Event()
-        try:
-            agreed = negotiate_controller(hs, shutdown)
-        except KeyboardInterrupt:
-            print("[handshake] interrupted before agreement; exiting")
-            hs.close(linger=0)
-            ctx.term()
-            return
-        hs.close(linger=0)
-        if agreed["controller"] is not None:
-            print(f"[handshake] running controller ONBOARD: {agreed['controller']} "
-                  f"(sonic_token_action={agreed['sonic_token_action']})")
-            cameras = None
-            if args.camera:
-                cameras = {
-                    "head_camera": {
-                        "device_id": args.camera_device,
-                        "shape": [args.camera_height, args.camera_width],
-                    }
-                }
-            serve_onboard_controller(
-                controller=agreed["controller"],
-                sonic_token_action=bool(agreed["sonic_token_action"]),
-                cameras=cameras,
-                camera_fps=args.camera_fps,
-                camera_port=args.camera_port,
-            )
-            return
-        print("[handshake] client selected raw DDS bridge (laptop owns control) -> legacy forward.")
-
     # Optionally start camera server in background thread
     camera_thread = None
     if args.camera:
@@ -547,7 +205,6 @@ def main() -> None:
 
     # initialize ZMQ
     ctx = zmq.Context.instance()
-    shutdown_event = threading.Event()
 
     # receive commands from remote client
     lowcmd_sock = ctx.socket(zmq.PULL)
@@ -558,6 +215,7 @@ def main() -> None:
     lowstate_sock.bind(f"tcp://0.0.0.0:{LOWSTATE_PORT}")
 
     state_period = 0.002  # ~500 hz
+    shutdown_event = threading.Event()
 
     # start observation forwarding in background thread
     t_state = threading.Thread(

src/lerobot/robots/unitree_g1/unitree_g1.py

@@ -16,8 +16,6 @@
 
 from __future__ import annotations
 
-import contextlib
-import json
 import logging
 import threading
 import time
@@ -28,7 +26,6 @@ from typing import TYPE_CHECKING, Protocol, runtime_checkable
 import numpy as np
 
 from lerobot.cameras import make_cameras_from_configs
-from lerobot.utils.errors import DeviceNotConnectedError
 from lerobot.types import RobotAction, RobotObservation
 from lerobot.utils.import_utils import _unitree_sdk_available, require_package
 
@@ -37,10 +34,10 @@ from .config_unitree_g1 import UnitreeG1Config
 from .g1_kinematics import G1_29_ArmIK
 from .g1_utils import (
     REMOTE_AXES,
+    REMOTE_KEYS,
     G1_29_JointArmIndex,
     G1_29_JointIndex,
     default_remote_input,
-    lowstate_to_obs,
     make_locomotion_controller,
 )
 
@@ -50,9 +47,7 @@ if TYPE_CHECKING or _unitree_sdk_available:
         ChannelPublisher as _SDKChannelPublisher,
         ChannelSubscriber as _SDKChannelSubscriber,
     )
-    from unitree_sdk2py.idl.default import (
+    from unitree_sdk2py.idl.default import unitree_hg_msg_dds__LowCmd_
-        unitree_hg_msg_dds__LowCmd_,
-    )
     from unitree_sdk2py.idl.unitree_hg.msg.dds_ import (
         LowCmd_ as hg_LowCmd,
         LowState_ as hg_LowState,
@@ -84,14 +79,6 @@ class LocomotionController(Protocol):
 kTopicLowCommand_Debug = "rt/lowcmd"
 kTopicLowState = "rt/lowstate"
 
-# Wireless-remote button byte layout, mapped to the positional button indices the
-# locomotion controllers expect. Used in onboard mode to read the physical Unitree
-# remote from lowstate (mirrors the exo teleoperator's RemoteController).
-_REMOTE_BUTTON_MAP: list[str] = [
-    "RB", "LB", "start", "back", "RT", "LT", "", "",
-    "A", "B", "X", "Y", "up", "right", "down", "left",
-]
-
 
 @dataclass
 class MotorState:
@@ -132,37 +119,24 @@ class UnitreeG1(Robot):
         self.config = config
         self.control_dt = config.control_dt
 
-        # Three mutually-exclusive roles:
-        #   * simulation : local DDS + controller run in-process against a MuJoCo world.
-        #   * onboard    : local DDS + controller run in-process on the robot NX.
-        #   * client     : thin laptop client. No DDS, no controller. It negotiates a
-        #                  controller with ``run_g1_server`` (which runs it onboard),
-        #                  PUSHes high-level actions and reads back state + cameras over
-        #                  ZMQ. The controller *always* runs on the robot, never here.
-        self._client = not config.is_simulation and not config.onboard
-
         # Initialize cameras config (ZMQ-based) - actual connection in connect()
         self._cameras = make_cameras_from_configs(config.cameras)
 
-        # DDS channel classes are only needed by the in-process control roles. The thin
+        # Import channel classes based on mode
-        # client never touches DDS, so we don't import the socket shim at all.
+        if config.is_simulation:
-        if config.is_simulation or config.onboard:
             self._ChannelFactoryInitialize = _SDKChannelFactoryInitialize
             self._ChannelPublisher = _SDKChannelPublisher
             self._ChannelSubscriber = _SDKChannelSubscriber
         else:
-            self._ChannelFactoryInitialize = None
+            from .unitree_sdk2_socket import (
-            self._ChannelPublisher = None
+                ChannelFactoryInitialize,
-            self._ChannelSubscriber = None
+                ChannelPublisher,
+                ChannelSubscriber,
+            )
 
-        # Client-side ZMQ handles / negotiated capabilities (populated in connect()).
+            self._ChannelFactoryInitialize = ChannelFactoryInitialize
-        self._client_action_sock = None
+            self._ChannelPublisher = ChannelPublisher
-        self._client_state_sock = None
+            self._ChannelSubscriber = ChannelSubscriber
-        self._client_state_latest: dict[str, float] = {}
-        self._client_caps: dict | None = None
-
-        # Optional arm gravity compensation (feed-forward torque via the arm IK solver).
-        self.arm_ik = G1_29_ArmIK() if config.gravity_compensation else None
 
         # Initialize state variables
         self.sim_env = None
@@ -172,69 +146,24 @@ class UnitreeG1(Robot):
         self._shutdown_event = threading.Event()
         self.subscribe_thread = None
 
-        # Lower-body controller loaded dynamically. GUARDRAIL: the controller must never
+        self.arm_ik = G1_29_ArmIK() if config.gravity_compensation else None
-        # be built or run on the laptop client -- it always runs onboard (or in sim).
-        if self._client:
-            self.controller: LocomotionController | None = None
-        else:
-            self.controller = make_locomotion_controller(config.controller)
 
-            # Token-driven deploy: a SONIC whole-body controller always runs in token
+        # Lower-body controller loaded dynamically
-            # mode -- it holds a neutral token until the first real one arrives, then
+        self.controller: LocomotionController | None = make_locomotion_controller(config.controller)
-            # holds the last token between control ticks.
-            if hasattr(self.controller, "token_mode"):
-                self.controller.token_mode = True
 
         # Controller thread state
         self._controller_thread = None
-        # When set, the controller loop stops publishing low commands so reset() can
-        # drive the joints directly without two publishers fighting (single-publisher).
-        self._controller_paused = threading.Event()
         self._controller_action_lock = threading.Lock()
         self.controller_input = default_remote_input()
         self.controller_output = {}
 
-        # Onboard-only: parser for the physical Unitree wireless remote (read straight
-        # from local lowstate so joystick locomotion works without a laptop round-trip).
-        self._joystick = None
-
-        # Token-mode state: last 64-D SONIC latent token commanded by the policy,
-        # echoed back as ``observation.state`` so a token-output VLA closes the loop
-        # on its own previous token. Implicit whenever the SONIC whole-body controller
-        # is active. Seeded to zeros; the controller's startup blend eases joints in.
-        self._last_token: np.ndarray | None = None
-        if self._sonic_token:
-            from .controllers.sonic_whole_body import TOKEN_DIM
-
-            self._last_token = np.zeros(TOKEN_DIM, dtype=np.float32)
-
-    @property
-    def _sonic_token(self) -> bool:
-        """Whether the SONIC whole-body decoder is active.
-
-        A SONIC controller consumes a 64-D latent motion token as its action and echoes
-        the last commanded token as ``observation.state``. Keyed purely off the selected
-        controller so the token interface is implicit -- no separate config flag.
-        """
-        return self.config.controller == "SonicWholeBodyController"
-
     def _subscribe_lowstate(self):  # polls robot state @ 250Hz
         while not self._shutdown_event.is_set():
             start_time = time.time()
 
             # Step simulation if in simulation mode
             if self.config.is_simulation and self.sim_env is not None:
-                try:
+                self.sim_env.step()
-                    self.sim_env.step()
-                except ValueError as e:
-                    # Startup race: the sim thread can step once before reset() has
-                    # written a valid base pose, giving a zero-norm pelvis quaternion
-                    # (scipy>=1.11 raises instead of normalizing). Skip and retry so
-                    # the thread survives instead of dying and freezing the sim.
-                    if "zero norm" not in str(e).lower():
-                        raise
-                    time.sleep(self.control_dt)
-                    continue
 
             msg = self.lowstate_subscriber.Read()
             if msg is not None:
@@ -302,46 +231,15 @@ class UnitreeG1(Robot):
                 features[f"{cam}_depth"] = (cfg.height, cfg.width, 1)
         return features
 
-    @property
-    def _token_state_ft(self) -> dict[str, type]:
-        """64-D SONIC latent-token proprio state (``motion_token_state.{i}.pos``).
-
-        Exposed only when a SONIC whole-body controller is active; aggregated by the
-        rollout into a 64-D ``observation.state`` (the last token the policy commanded).
-        """
-        if not self._sonic_token:
-            return {}
-        from .controllers.sonic_whole_body import TOKEN_DIM, token_state_key
-
-        return {token_state_key(i): float for i in range(TOKEN_DIM)}
-
     @cached_property
     def observation_features(self) -> dict[str, type | tuple]:
-        return {
+        return {**self._motors_ft, **self._cameras_ft}
-            **self._motors_ft,
-            **self._token_state_ft,
-            **self._cameras_ft,
-        }
 
     @cached_property
     def action_features(self) -> dict[str, type]:
-        # Role-agnostic: the schema is a pure function of the controller name. The thin
+        if self.controller is None:
-        # client advertises the same schema as the onboard robot so the exact same
-        # policy output routes straight through.
-
-        # No controller configured at all: raw 29-DoF joint teleop.
-        if self.config.controller is None:
             return {f"{G1_29_JointIndex(motor).name}.q": float for motor in G1_29_JointIndex}
 
-        # Token-output VLA (SONIC decoder): advertise a 64-D latent-token action space
-        # (``motion_token.{i}.pos``) so ``lerobot-rollout`` maps a 64-D policy output
-        # straight onto the decoder, bypassing the encoder.
-        if self._sonic_token:
-            from .controllers.sonic_whole_body import TOKEN_DIM, token_action_key
-
-            return {token_action_key(i): float for i in range(TOKEN_DIM)}
-
-        # Locomotion controllers (GR00T / Holosoma): arm joint targets + joystick axes.
         arm_features = {f"{G1_29_JointArmIndex(motor).name}.q": float for motor in G1_29_JointArmIndex}
         remote_features = dict.fromkeys(REMOTE_AXES, float)
         return {**arm_features, **remote_features}
@@ -357,11 +255,6 @@ class UnitreeG1(Robot):
         while not self._shutdown_event.is_set():
             start_time = time.time()
 
-            # Paused during reset() so the reset routine is the sole low-cmd publisher.
-            if self._controller_paused.is_set():
-                time.sleep(control_dt)
-                continue
-
             with self._lowstate_lock:
                 lowstate = self._lowstate
 
@@ -378,13 +271,6 @@ class UnitreeG1(Robot):
                 with self._controller_action_lock:
                     controller_input = dict(self.controller_input)
 
-                # Onboard: the physical Unitree remote (in local lowstate) takes
-                # priority for locomotion when active; otherwise laptop/ZMQ axes stand.
-                if self.config.onboard:
-                    wl = self._wireless_remote_input(lowstate)
-                    if wl is not None:
-                        controller_input.update(wl)
-
                 # Run controller step
                 controller_action = self.controller.run_step(controller_input, lowstate)
 
@@ -407,163 +293,7 @@ class UnitreeG1(Robot):
     def configure(self) -> None:
         pass
 
-    def _wireless_remote_input(self, lowstate) -> dict | None:
-        """Parse the physical Unitree remote from lowstate into controller inputs.
-
-        Onboard only. Returns None when the remote is idle so the laptop-provided
-        (ZMQ) axes keep control; otherwise the physical remote takes priority.
-        """
-        js = self._joystick
-        if js is None:
-            return None
-        wr = getattr(lowstate, "wireless_remote", None)
-        if not wr or len(wr) < 24:
-            return None
-        try:
-            js.extract(wr)
-        except Exception:  # noqa: BLE001
-            return None
-
-        axes = {
-            "remote.lx": float(js.lx.data),
-            "remote.ly": float(js.ly.data),
-            "remote.rx": float(js.rx.data),
-            "remote.ry": float(js.ry.data),
-        }
-        active = any(abs(v) > 1e-2 for v in axes.values())
-        out = dict(axes)
-        for i, name in enumerate(_REMOTE_BUTTON_MAP):
-            if name:
-                val = float(getattr(js, name).data)
-                out[f"remote.button.{i}"] = val
-                if val:
-                    active = True
-        return out if active else None
-
-    def _release_motion_control(self) -> None:
-        """Release the robot's built-in motion services so we can send raw lowcmd.
-
-        Onboard-only. Mirrors run_g1_server.py: on the real robot the factory
-        locomotion/hand services must relinquish control before our controller can
-        write to ``rt/lowcmd``, otherwise commands are ignored or fought.
-        """
-        from unitree_sdk2py.comm.motion_switcher.motion_switcher_client import MotionSwitcherClient
-
-        msc = MotionSwitcherClient()
-        msc.SetTimeout(5.0)
-        msc.Init()
-        _, result = msc.CheckMode()
-        while result is not None and "name" in result and result["name"]:
-            logger.info("[UnitreeG1] Releasing built-in mode '%s'...", result["name"])
-            msc.ReleaseMode()
-            _, result = msc.CheckMode()
-            time.sleep(1.0)
-
-    # ------------------------------------------------------------------ #
-    # Thin-client role (laptop): no DDS, no controller. Talks to run_g1_server
-    # over ZMQ. The controller ALWAYS runs onboard; we only relay high-level
-    # actions and read back the state echo + camera frames.
-    # ------------------------------------------------------------------ #
-    def _connect_client(self) -> None:
-        import zmq
-
-        from .run_g1_server import ACTION_PORT, HANDSHAKE_PORT, STATE_PORT, request_controller
-
-        server_ip = self.config.robot_ip
-        if not server_ip:
-            raise ValueError("client mode requires config.robot_ip (the G1 running run_g1_server)")
-
-        # 1) Handshake: agree with the server on which controller it will run onboard.
-        logger.info(
-            "[client] handshaking with %s:%d (controller=%s, token=%s)...",
-            server_ip, HANDSHAKE_PORT, self.config.controller, self._sonic_token,
-        )
-        self._client_caps = request_controller(
-            server_ip,
-            self.config.controller,
-            sonic_token_action=self._sonic_token,
-            port=HANDSHAKE_PORT,
-        )
-        logger.info("[client] server agreed: %s", self._client_caps)
-
-        ctx = zmq.Context.instance()
-
-        # 2) Action PUSH: ship compact high-level actions to the onboard controller.
-        self._client_action_sock = ctx.socket(zmq.PUSH)
-        self._client_action_sock.setsockopt(zmq.SNDHWM, 2)
-        self._client_action_sock.setsockopt(zmq.LINGER, 0)
-        self._client_action_sock.connect(f"tcp://{server_ip}:{ACTION_PORT}")
-
-        # 3) State SUB: read the onboard observation.state echo (last token / joints).
-        self._client_state_sock = ctx.socket(zmq.SUB)
-        self._client_state_sock.setsockopt(zmq.CONFLATE, 1)
-        self._client_state_sock.setsockopt_string(zmq.SUBSCRIBE, "")
-        self._client_state_sock.connect(f"tcp://{server_ip}:{STATE_PORT}")
-
-        # 4) Cameras (ZMQ ImageServer served by run_g1_server) - same as any client.
-        for cam in self._cameras.values():
-            if not cam.is_connected:
-                cam.connect()
-        logger.info("[client] connected: actions ->:%d, state <-:%d, %d camera(s).",
-                    ACTION_PORT, STATE_PORT, len(self._cameras))
-
-    def _recv_client_state(self) -> None:
-        """Drain the state SUB (CONFLATE keeps only the freshest) into the latest cache."""
-        import zmq
-
-        if self._client_state_sock is None:
-            return
-        while True:
-            try:
-                state = self._client_state_sock.recv_json(flags=zmq.NOBLOCK)
-            except zmq.Again:
-                break
-            except (ValueError, zmq.ZMQError):
-                break
-            if isinstance(state, dict):
-                self._client_state_latest = {k: float(v) for k, v in state.items()}
-
-    def _get_observation_client(self) -> RobotObservation:
-        self._recv_client_state()
-        obs: dict = dict(self._client_state_latest)
-        for cam_name, cam in self._cameras.items():
-            if getattr(cam, "use_rgb", True):
-                obs[cam_name] = cam.read_latest()
-            if getattr(cam, "use_depth", False):
-                obs[f"{cam_name}_depth"] = cam.read_latest_depth()
-        return obs
-
-    def _send_action_client(self, action: RobotAction) -> RobotAction:
-        """Relay the raw action straight to the onboard controller. NO processing here:
-        the controller negotiated in the handshake interprets it (token / wb / arm)."""
-        import zmq
-
-        if self._client_action_sock is None:
-            raise DeviceNotConnectedError("UnitreeG1 client is not connected")
-        payload = json.dumps({k: float(v) for k, v in action.items()}).encode("utf-8")
-        with contextlib.suppress(zmq.Again):
-            self._client_action_sock.send(payload, zmq.NOBLOCK)
-        return action
-
-    def _disconnect_client(self) -> None:
-        for sock in (self._client_action_sock, self._client_state_sock):
-            if sock is not None:
-                with contextlib.suppress(Exception):
-                    sock.close(linger=0)
-        self._client_action_sock = None
-        self._client_state_sock = None
-        for cam in self._cameras.values():
-            with contextlib.suppress(Exception):
-                cam.disconnect()
-
     def connect(self, calibrate: bool = True) -> None:  # connect to DDS
-        # Thin-client role: no DDS, no controller. Negotiate the controller with
-        # run_g1_server (which runs it onboard), then open the high-level ZMQ links:
-        # PUSH actions on :ACTION_PORT, SUB state echo on :STATE_PORT, cameras via ZMQ.
-        if self._client:
-            self._connect_client()
-            return
-
         # Initialize DDS channel and simulation environment
         if self.config.is_simulation:
             from lerobot.envs import make_env
@@ -572,28 +302,6 @@ class UnitreeG1(Robot):
             self._env_wrapper = make_env("lerobot/unitree-g1-mujoco", trust_remote_code=True)
             # Extract the actual gym env from the dict structure
             self.sim_env = self._env_wrapper["hub_env"][0].envs[0]
-        elif self.config.onboard:
-            # Real robot, controller running onboard against local DDS. Initialize the
-            # real SDK channel factory on the robot's DDS interface and take low-level
-            # control from the built-in services before we start writing lowcmd.
-            if self.config.dds_interface:
-                self._ChannelFactoryInitialize(0, self.config.dds_interface)
-            else:
-                self._ChannelFactoryInitialize(0)
-            # Real robot: hand low-level control over from the built-in services.
-            # A DDS sim has no MotionSwitcher, so this is skipped there.
-            if self.config.release_motion_control:
-                self._release_motion_control()
-            # Real robot: read the physical wireless remote from lowstate for
-            # locomotion. A sim has no physical remote, so leave _joystick=None and
-            # let send_action (ZMQ) drive the locomotion axes instead.
-            if self.config.physical_remote:
-                from unitree_sdk2py.utils.joystick import Joystick
-
-                self._joystick = Joystick()
-                for axis in (self._joystick.lx, self._joystick.ly, self._joystick.rx, self._joystick.ry):
-                    axis.smooth = 1.0
-                    axis.deadzone = 0.0
         else:
             self._ChannelFactoryInitialize(0, config=self.config)
 
@@ -635,9 +343,6 @@ class UnitreeG1(Robot):
 
         self.kp = np.array(self.config.kp, dtype=np.float32)
         self.kd = np.array(self.config.kd, dtype=np.float32)
-        if self.controller is not None and hasattr(self.controller, "kp"):
-            self.kp = np.array(self.controller.kp, dtype=np.float32)
-            self.kd = np.array(self.controller.kd, dtype=np.float32)
 
         for joint in G1_29_JointIndex:
             self.msg.motor_cmd[joint].mode = 1
@@ -645,8 +350,7 @@ class UnitreeG1(Robot):
             self.msg.motor_cmd[joint].kd = self.kd[joint.value]
             self.msg.motor_cmd[joint].q = lowstate.motor_state[joint.value].q
 
-        # Start the 50 Hz controller thread (runs the locomotion/whole-body policy and
+        # Start controller thread if enabled
-        # publishes low commands to DDS).
         if self.controller is not None:
             self._controller_thread = threading.Thread(target=self._controller_loop, daemon=True)
             self._controller_thread.start()
@@ -668,34 +372,12 @@ class UnitreeG1(Robot):
             logger.warning(f"Failed to send zero-torque on disconnect: {e}")
 
     def disconnect(self):
-        if self._client:
+        # Put robot in passive mode before stopping threads
-            self._disconnect_client()
+        if not self.config.is_simulation:
-            return
-
-        # Stop the controller loop first so it isn't fighting the shutdown ramp.
-        self._shutdown_event.set()
-        controller_stopped = True
-        if self._controller_thread is not None:
-            # Wait long enough for any in-flight inference tick to finish and the loop
-            # to observe the shutdown flag, so no stray low command is published while
-            # the ramp runs (the shutdown routine must be the single publisher).
-            self._controller_thread.join(timeout=5.0)
-            if self._controller_thread.is_alive():
-                controller_stopped = False
-                logger.error(
-                    "Controller thread did not stop; skipping graceful ramp to avoid "
-                    "concurrent low commands (fail-safe: joints keep last command until exit)"
-                )
-
-        # Put the robot in passive mode (zero-torque) before stopping the rest (real
-        # robot only; the subscribe thread is still alive here to supply the current
-        # pose). Only publish once the controller thread has definitely exited so the
-        # two aren't publishing at once.
-        if not self.config.is_simulation and controller_stopped:
             self._send_zero_torque()
 
-        if self.controller is not None and hasattr(self.controller, "shutdown"):
+        # Signal thread to stop and unblock any waits
-            self.controller.shutdown()
+        self._shutdown_event.set()
 
         # Wait for subscribe thread to finish
         if self.subscribe_thread is not None:
@@ -703,6 +385,12 @@ class UnitreeG1(Robot):
             if self.subscribe_thread.is_alive():
                 logger.warning("Subscribe thread did not stop cleanly")
 
+        # Wait for controller thread to finish
+        if self._controller_thread is not None:
+            self._controller_thread.join(timeout=2.0)
+            if self._controller_thread.is_alive():
+                logger.warning("Controller thread did not stop cleanly")
+
         # Close simulation environment
         if self.config.is_simulation and self.sim_env is not None:
             try:
@@ -729,25 +417,49 @@ class UnitreeG1(Robot):
             cam.disconnect()
 
     def get_observation(self) -> RobotObservation:
-        if self._client:
-            return self._get_observation_client()
-
         with self._lowstate_lock:
             lowstate = self._lowstate
         if lowstate is None:
             return {}
 
-        # Motors + IMU + wireless remote (shared lowstate -> obs mapping)
+        obs = {}
-        obs = lowstate_to_obs(lowstate)
 
-        # Token mode: echo the last commanded latent token as observation.state so a
+        # Motors - q, dq, tau for all joints
-        # token-output VLA closes the loop on its own previous token.
+        for motor in G1_29_JointIndex:
-        if self._sonic_token:
+            name = motor.name
-            from .controllers.sonic_whole_body import token_state_key
+            idx = motor.value
+            obs[f"{name}.q"] = lowstate.motor_state[idx].q
+            obs[f"{name}.dq"] = lowstate.motor_state[idx].dq
+            obs[f"{name}.tau"] = lowstate.motor_state[idx].tau_est
 
-            token = self._last_token if self._last_token is not None else []
+        # IMU - gyroscope
-            for i, v in enumerate(token):
+        if lowstate.imu_state.gyroscope:
-                obs[token_state_key(i)] = float(v)
+            obs["imu.gyro.x"] = lowstate.imu_state.gyroscope[0]
+            obs["imu.gyro.y"] = lowstate.imu_state.gyroscope[1]
+            obs["imu.gyro.z"] = lowstate.imu_state.gyroscope[2]
+
+        # IMU - accelerometer
+        if lowstate.imu_state.accelerometer:
+            obs["imu.accel.x"] = lowstate.imu_state.accelerometer[0]
+            obs["imu.accel.y"] = lowstate.imu_state.accelerometer[1]
+            obs["imu.accel.z"] = lowstate.imu_state.accelerometer[2]
+
+        # IMU - quaternion
+        if lowstate.imu_state.quaternion:
+            obs["imu.quat.w"] = lowstate.imu_state.quaternion[0]
+            obs["imu.quat.x"] = lowstate.imu_state.quaternion[1]
+            obs["imu.quat.y"] = lowstate.imu_state.quaternion[2]
+            obs["imu.quat.z"] = lowstate.imu_state.quaternion[3]
+
+        # IMU - rpy
+        if lowstate.imu_state.rpy:
+            obs["imu.rpy.roll"] = lowstate.imu_state.rpy[0]
+            obs["imu.rpy.pitch"] = lowstate.imu_state.rpy[1]
+            obs["imu.rpy.yaw"] = lowstate.imu_state.rpy[2]
+
+        # Wireless remote (raw bytes for teleoperator)
+        if lowstate.wireless_remote:
+            obs["wireless_remote"] = lowstate.wireless_remote
 
         # Cameras - read images from ZMQ cameras
         for cam_name, cam in self._cameras.items():
@@ -759,22 +471,11 @@ class UnitreeG1(Robot):
         return obs
 
     def send_action(self, action: RobotAction) -> RobotAction:
-        if self._client:
-            return self._send_action_client(action)
-
         action_to_publish = action
         if self.controller is not None:
-            if self._sonic_token:
-                from .controllers.sonic_whole_body import _extract_token_from_action
-
-                token = _extract_token_from_action(action)
-                if token is not None:
-                    self._last_token = token
-            self._update_controller_action(action)
-            if getattr(self.controller, "full_body", False):
-                return action
             # Controller thread owns legs/waist. Here we only update joystick inputs
             # and publish arm targets from the teleoperator.
+            self._update_controller_action(action)
             arm_prefixes = tuple(j.name for j in G1_29_JointArmIndex)
             action_to_publish = {
                 key: value
@@ -802,17 +503,11 @@ class UnitreeG1(Robot):
         return action
 
     def _update_controller_action(self, action: RobotAction) -> None:
-        """Update controller input state from an incoming teleop action.
+        """Update controller input state from incoming teleop action."""
-
-        Controller-agnostic: every value-carrying key (e.g. locomotion ``remote.*``
-        axes/buttons) is forwarded verbatim into ``controller_input`` and each
-        controller extracts only the keys it understands. The robot deliberately does
-        not enumerate any controller's key schema here.
-        """
         with self._controller_action_lock:
-            for key, value in action.items():
+            for key in REMOTE_KEYS:
-                if isinstance(key, str) and value is not None:
+                if key in action:
-                    self.controller_input[key] = value
+                    self.controller_input[key] = action[key]
 
     @property
     def is_calibrated(self) -> bool:
@@ -820,8 +515,6 @@ class UnitreeG1(Robot):
 
     @property
     def is_connected(self) -> bool:
-        if self._client:
-            return self._client_action_sock is not None
         with self._lowstate_lock:
             return self._lowstate is not None
 
@@ -844,64 +537,43 @@ class UnitreeG1(Robot):
         if default_positions is None:
             default_positions = np.array(self.config.default_positions, dtype=np.float32)
 
-        # Full-body controllers (SONIC / OpenHLM) own the whole 29-DoF command and
+        if self.config.is_simulation and self.sim_env is not None:
-        # ignore ``<joint>.q`` in send_action(), so reset() must publish the default
+            self.sim_env.reset()
-        # pose directly. Pause the background controller first so the two aren't both
+            self.publish_lowcmd(
-        # writing low commands while the robot moves to the default pose.
+                {f"{motor.name}.q": float(default_positions[motor.value]) for motor in G1_29_JointIndex}
-        full_body = getattr(self.controller, "full_body", False)
+            )
-        paused = False
+        else:
-        if full_body and self._controller_thread is not None:
+            total_time = 3.0
-            self._controller_paused.set()
+            num_steps = int(total_time / control_dt)
-            paused = True
-            time.sleep(control_dt)  # let any in-flight controller tick settle
 
-        try:
+            # get current state
-            if self.config.is_simulation and self.sim_env is not None:
+            obs = self.get_observation()
-                self.sim_env.reset()
-                self.publish_lowcmd(
-                    {f"{motor.name}.q": float(default_positions[motor.value]) for motor in G1_29_JointIndex}
-                )
-            else:
-                total_time = 3.0
-                num_steps = int(total_time / control_dt)
 
-                # get current state
+            # record current positions
-                obs = self.get_observation()
+            init_dof_pos = np.zeros(29, dtype=np.float32)
+            for motor in G1_29_JointIndex:
+                init_dof_pos[motor.value] = obs[f"{motor.name}.q"]
 
-                # record current positions
+            # Interpolate to default position
-                init_dof_pos = np.zeros(29, dtype=np.float32)
+            for step in range(num_steps):
+                start_time = time.time()
+
+                alpha = step / num_steps
+                action_dict = {}
                 for motor in G1_29_JointIndex:
-                    init_dof_pos[motor.value] = obs[f"{motor.name}.q"]
+                    target_pos = default_positions[motor.value]
+                    interp_pos = init_dof_pos[motor.value] * (1 - alpha) + target_pos * alpha
+                    action_dict[f"{motor.name}.q"] = float(interp_pos)
 
-                # Interpolate to default position
+                self.send_action(action_dict)
-                for step in range(num_steps):
-                    start_time = time.time()
 
-                    alpha = step / num_steps
+                # Maintain constant control rate
-                    action_dict = {}
+                elapsed = time.time() - start_time
-                    for motor in G1_29_JointIndex:
+                sleep_time = max(0, control_dt - elapsed)
-                        target_pos = default_positions[motor.value]
+                time.sleep(sleep_time)
-                        interp_pos = init_dof_pos[motor.value] * (1 - alpha) + target_pos * alpha
-                        action_dict[f"{motor.name}.q"] = float(interp_pos)
 
-                    # Full-body controllers no-op in send_action(); publish the pose
+        # Reset controller internal state (gait phase, obs history, etc.)
-                    # directly (arm-only controllers keep the send_action() path).
+        if self.controller is not None and hasattr(self.controller, "reset"):
-                    if full_body:
+            self.controller.reset()
-                        self.publish_lowcmd(action_dict)
-                    else:
-                        self.send_action(action_dict)
-
-                    # Maintain constant control rate
-                    elapsed = time.time() - start_time
-                    sleep_time = max(0, control_dt - elapsed)
-                    time.sleep(sleep_time)
-
-            # Reset controller internal state (gait phase, obs history, etc.) before
-            # resuming so its buffers reflect the post-reset pose.
-            if self.controller is not None and hasattr(self.controller, "reset"):
-                self.controller.reset()
-        finally:
-            if paused:
-                self._controller_paused.clear()
 
         logger.info("Reached default position")

src/lerobot/robots/utils.py

@@ -21,6 +21,8 @@ from lerobot.utils.import_utils import make_device_from_device_class
 from .config import RobotConfig
 from .robot import Robot
 
+logger = logging.getLogger(__name__)
+
 
 def make_robot_from_config(config: RobotConfig) -> Robot:
     # TODO(Steven): Consider just using the make_device_from_device_class for all types
@@ -118,7 +120,7 @@ def ensure_safe_goal_position(
             }
 
     if warnings_dict:
-        logging.warning(
+        logger.warning(
             "Relative goal position magnitude had to be clamped to be safe.\n"
             f"{pformat(warnings_dict, indent=4)}"
         )

src/lerobot/rollout/context.py

@@ -24,7 +24,6 @@ from __future__ import annotations
 import logging
 from dataclasses import dataclass, field
 from threading import Event
-from typing import TYPE_CHECKING
 
 import torch
 
@@ -48,7 +47,6 @@ from lerobot.processor.relative_action_processor import RelativeActionsProcessor
 from lerobot.robots import make_robot_from_config
 from lerobot.teleoperators import Teleoperator, make_teleoperator_from_config
 from lerobot.utils.feature_utils import combine_feature_dicts, hw_to_dataset_features
-from lerobot.utils.import_utils import _peft_available, require_package
 
 from .configs import BaseStrategyConfig, DAggerStrategyConfig, RolloutConfig
 from .inference import (
@@ -59,12 +57,6 @@ from .inference import (
 )
 from .robot_wrapper import ThreadSafeRobot
 
-if TYPE_CHECKING or _peft_available:
-    from peft import PeftConfig, PeftModel
-else:
-    PeftConfig = None
-    PeftModel = None
-
 logger = logging.getLogger(__name__)
 
 
@@ -179,7 +171,7 @@ def _load_pretrained_policy(policy_config: PreTrainedConfig) -> PreTrainedPolicy
             revision=pretrained_revision,
         )
 
-    require_package("peft", extra="peft")
+    from peft import PeftConfig, PeftModel
 
     peft_path = policy_config.pretrained_path
     peft_config = PeftConfig.from_pretrained(peft_path, revision=pretrained_revision)
@@ -302,22 +294,12 @@ def build_rollout_context(
     # ``observation_features`` values are either a tuple (camera shape) or the
     # ``float`` type itself used as a sentinel for scalar motor features —
     # see ``dict[str, type | tuple]`` annotation on ``Robot.observation_features``.
-    # Keep cameras (tuple) plus both joint-position (.pos) and base-velocity (.vel)
-    # scalar state features. LeKiwi's observation.state is 9-dim (6 arm .pos +
-    # x/y/theta.vel) and the policy was trained/normalized on all 9; the old .pos-only
-    # filter fed a 6-dim state into a 9-dim normalizer → RuntimeError (size 6 vs 9).
-    # Pure-arm robots have no .vel state keys, so this is a no-op for them.
     observation_features_hw = {
         k: v
         for k, v in all_obs_features.items()
-        if isinstance(v, tuple) or (v is float and k.endswith((".pos", ".vel")))
+        if isinstance(v, tuple) or (v is float and k.endswith(".pos"))
     }
-    # Keep both joint-position (.pos) and base-velocity (.vel) action features so
+    action_features_hw = {k: v for k, v in robot.action_features.items() if k.endswith(".pos")}
-    # mobile manipulators command the base too (e.g. LeKiwi: 6 arm .pos +
-    # x/y/theta.vel = 9-dim action). Pure-arm robots have no .vel keys, so this is
-    # a no-op for them. Without the .vel keys the base velocities are silently
-    # dropped from dataset_features[ACTION]/ordered_action_keys and the base never moves.
-    action_features_hw = {k: v for k, v in robot.action_features.items() if k.endswith((".pos", ".vel"))}
 
     # The action side is always needed: sync inference reads action names from
     # ``dataset_features[ACTION]`` to map policy tensors back to robot actions.

src/lerobot/runtime/__init__.py

@@ -0,0 +1,38 @@
+# 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.
+
+"""Policy-agnostic runtime for language-conditioned policies.
+
+Adapters registered in :mod:`lerobot.runtime.registry` are served by ``lerobot-rollout --language``.
+"""
+
+from .adapter import BaseLanguageAdapter, GenerationConfig, LanguageDiagnostics
+from .language_runtime import (
+    LanguageConditionedPolicyAdapter,
+    LanguageConditionedRuntime,
+    RuntimeState,
+    Tick,
+    TickClock,
+)
+
+__all__ = [
+    "BaseLanguageAdapter",
+    "GenerationConfig",
+    "LanguageConditionedPolicyAdapter",
+    "LanguageConditionedRuntime",
+    "LanguageDiagnostics",
+    "RuntimeState",
+    "Tick",
+    "TickClock",
+]

src/lerobot/runtime/adapter.py

@@ -0,0 +1,165 @@
+# 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.
+
+"""Policy adapters for the language runtime.
+
+The base adapter owns generation control and diagnostics while subclasses provide policy-specific actions and text.
+"""
+
+from __future__ import annotations
+
+import re
+from abc import ABC, abstractmethod
+from dataclasses import dataclass, field
+from typing import Any
+
+from .language_runtime import RuntimeState
+
+_SAY_RE = re.compile(r"<\s*say\s*>(.*?)<\s*/\s*say\s*>", re.IGNORECASE | re.DOTALL)
+
+
+@dataclass
+class GenerationConfig:
+    """Text-generation settings fixed for the adapter's lifetime."""
+
+    min_new_tokens: int = 0
+    temperature: float = 0.0
+    top_p: float = 1.0
+    chunks_per_regen: int = 1  # regenerate the language context every N action chunks
+    enable_memory: bool = True  # generate a running memory note on subtask change
+    enable_subtask: bool = True  # generate the low-level subtask (off => use the given text directly)
+
+
+@dataclass
+class LanguageDiagnostics:
+    """Runtime-panel generation counters keyed by text kind."""
+
+    last_raw: dict[str, str] = field(default_factory=dict)
+    empty: dict[str, int] = field(default_factory=dict)
+    repeat: int = 0
+
+    def _bump(self, table: dict[str, int], kind: str) -> int:
+        table[kind] = table.get(kind, 0) + 1
+        return table[kind]
+
+
+class BaseLanguageAdapter(ABC):
+    """Batteries-included adapter: generic high-level control, policy primitives abstract."""
+
+    def __init__(self, policy: Any, gen: GenerationConfig | None = None) -> None:
+        self.policy = policy
+        self.gen = gen or GenerationConfig()
+        self.diag = LanguageDiagnostics()
+        self._chunks_until_regen = 0
+
+    @abstractmethod
+    def select_action(self, observation: dict[str, Any], state: RuntimeState) -> Any:
+        """Produce an action chunk from the observation + current language context."""
+
+    @abstractmethod
+    def generate_text(
+        self,
+        kind: str,
+        observation: dict[str, Any] | None,
+        state: RuntimeState,
+        user_text: str | None = None,
+    ) -> str:
+        """Generate one text stream (``kind``) and return the decoded string."""
+
+    def update_language_state(self, observation: dict[str, Any] | None, state: RuntimeState) -> None:
+        """Throttled regeneration of the language context (subtask / memory / ...)."""
+        if self._chunks_until_regen > 0:
+            self._chunks_until_regen -= 1
+            return
+        self._chunks_until_regen = max(1, self.gen.chunks_per_regen) - 1
+        self._regenerate_context(observation, state)
+
+    def handle_interjection(
+        self, user_text: str, observation: dict[str, Any] | None, state: RuntimeState
+    ) -> None:
+        """React to a mid-run user message by regenerating the plan."""
+        out = self.generate_text("interjection", observation, state, user_text=user_text)
+        plan = self.plan_from_text(out)
+        if plan:
+            state.set_context("plan", plan, label="plan")
+
+    def plan_from_text(self, text: str) -> str:
+        """Strip ``<say>`` speech markers from a generated plan."""
+        plan, _speech = split_plan_and_say(text)
+        return plan
+
+    def _regenerate_context(self, observation: dict[str, Any] | None, state: RuntimeState) -> None:
+        """Default hierarchy: regenerate the subtask, then memory when it changes.
+
+        Override for a policy with a different language hierarchy.
+        """
+        if not self.gen.enable_subtask:
+            # Preserve operator-provided subtasks in direct mode.
+            return
+        subtask = self._generate_filtered("subtask", observation, state)
+        if subtask is None:
+            return
+        previous = state.language_context.get("subtask")
+        if not state.set_context("subtask", subtask, label="subtask"):
+            self.diag.repeat += 1
+            return
+        self.diag.repeat = 0
+        if previous:
+            state.extra["prior_subtask"] = previous
+        if not self.gen.enable_memory:
+            return
+        memory = self._generate_filtered("memory", observation, state)
+        if memory is not None:
+            state.set_context("memory", memory, label="memory")
+
+    def _generate_filtered(
+        self, kind: str, observation: dict[str, Any] | None, state: RuntimeState
+    ) -> str | None:
+        """Generate one ``kind``, record diagnostics, and drop empty output."""
+        text = self.generate_text(kind, observation, state)
+        self.diag.last_raw[kind] = text or ""
+        if not text:
+            count = self.diag._bump(self.diag.empty, kind)
+            if count == 1 or count % 5 == 0:
+                state.log(f"  [info] {kind} gen returned empty (x{count})")
+            return None
+        return text
+
+
+class DirectTaskPolicyAdapter(BaseLanguageAdapter):
+    """Adapter for flat policies whose preprocessors condition actions on the operator's task."""
+
+    def select_action(self, observation: dict[str, Any], state: RuntimeState) -> Any:
+        return self.policy.predict_action_chunk(observation)
+
+    def generate_text(
+        self,
+        kind: str,
+        observation: dict[str, Any] | None,
+        state: RuntimeState,
+        user_text: str | None = None,
+    ) -> str:
+        return ""
+
+
+def split_plan_and_say(text: str) -> tuple[str, str]:
+    """Split ``plan <say>speech</say>`` into ``(plan, 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/runtime/language_runtime.py

@@ -0,0 +1,349 @@
+# 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.
+
+"""Small reusable runtime for language-conditioned robot policies."""
+
+from __future__ import annotations
+
+import logging
+import threading
+import time
+from collections import deque
+from collections.abc import Callable
+from dataclasses import dataclass, field
+from typing import Any, Protocol
+
+logger = logging.getLogger(__name__)
+
+
+@dataclass
+class RuntimeState:
+    """Explicit state shared by the runtime and policy adapter."""
+
+    task: str = ""
+    language_context: dict[str, str] = field(default_factory=dict)
+    action_queue: deque[Any] = field(default_factory=deque)
+    events: set[str] = field(default_factory=set)
+    log_lines: list[str] = field(default_factory=list)
+    mode: str = "action"
+    stop: bool = False
+    tick: Tick | None = None
+    actions_dispatched: int = 0
+    action_deadline: float | None = None
+    extra: dict[str, Any] = field(default_factory=dict)
+    revision: int = 0
+    lock: Any = field(default_factory=threading.RLock, repr=False)
+
+    def emit(self, event_name: str) -> None:
+        self.events.add(event_name)
+
+    def take_event(self, event_name: str) -> bool:
+        if event_name not in self.events:
+            return False
+        self.events.remove(event_name)
+        return True
+
+    def log(self, line: str) -> None:
+        self.log_lines.append(line)
+
+    def set_context(self, key: str, value: str | None, *, label: str | None = None) -> bool:
+        with self.lock:
+            previous = self.language_context.get(key)
+            if previous == value:
+                return False
+            if value is None:
+                self.language_context.pop(key, None)
+            else:
+                self.language_context[key] = value
+            self.revision += 1
+            if label is not None and value:
+                self.log(f"  {label}: {value}")
+            return True
+
+    def get(self, key: str, default: Any = None) -> Any:
+        try:
+            return self[key]
+        except KeyError:
+            return default
+
+    def setdefault(self, key: str, default: Any = None) -> Any:
+        current = self.get(key, None)
+        if current is not None:
+            return current
+        self[key] = default
+        return default
+
+    def __getitem__(self, key: str) -> Any:
+        if hasattr(self, key):
+            return getattr(self, key)
+        if key in self.extra:
+            return self.extra[key]
+        raise KeyError(key)
+
+    def __setitem__(self, key: str, value: Any) -> None:
+        with self.lock:
+            if hasattr(self, key):
+                if key == "mode" and self.mode != value:
+                    self.revision += 1
+                setattr(self, key, value)
+            else:
+                self.extra[key] = value
+
+
+class LanguageConditionedPolicyAdapter(Protocol):
+    """Runtime policy contract, implemented directly or through ``BaseLanguageAdapter``."""
+
+    def select_action(self, observation: dict[str, Any], state: RuntimeState) -> Any: ...
+
+    def update_language_state(self, observation: dict[str, Any] | None, state: RuntimeState) -> None: ...
+
+    def handle_interjection(
+        self, user_text: str, observation: dict[str, Any] | None, state: RuntimeState
+    ) -> None: ...
+
+
+@dataclass
+class Tick:
+    index: int
+    monotonic_seconds: float
+
+
+@dataclass
+class TickClock:
+    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)
+
+
+@dataclass
+class _RateGate:
+    hz: float
+    _last_seconds: float | None = None
+
+    def due(self, tick: Tick, *, force: bool = False) -> bool:
+        if force:
+            self._last_seconds = tick.monotonic_seconds
+            return True
+        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:
+        self._last_seconds = None
+
+
+@dataclass
+class LanguageConditionedRuntime:
+    """Generic tick loop for language-conditioned robot policies."""
+
+    policy_adapter: LanguageConditionedPolicyAdapter
+    observation_provider: Callable[[], dict[str, Any] | None] | None = None
+    action_executor: Callable[[Any], None] | None = None
+    event_collector: Callable[[RuntimeState], None] | None = None
+    chunk_hz: float = 4.0
+    ctrl_hz: float = 50.0
+    high_level_hz: float = 1.0
+    max_rate_hz: float = 50.0
+
+    state: RuntimeState = field(default_factory=RuntimeState)
+    _chunk_gate: _RateGate = field(init=False)
+    _ctrl_gate: _RateGate = field(init=False)
+    _language_gate: _RateGate = field(init=False)
+    _stop: bool = field(default=False, init=False)
+    _last_dispatch_seconds: float | None = field(default=None, init=False)
+
+    def __post_init__(self) -> None:
+        self._chunk_gate = _RateGate(self.chunk_hz)
+        self._ctrl_gate = _RateGate(self.ctrl_hz)
+        self._language_gate = _RateGate(self.high_level_hz)
+
+    @property
+    def policy(self) -> Any:
+        return getattr(self.policy_adapter, "policy", self.policy_adapter)
+
+    def set_task(self, task: str) -> None:
+        with self.state.lock:
+            if self.state.task != task:
+                self.state.revision += 1
+            self.state.task = task
+            self.state.log(f"Task: {task}")
+
+    def stop(self) -> None:
+        self._stop = True
+        self.state.stop = True
+
+    def run(self, *, max_ticks: int | None = None) -> None:
+        clock = TickClock(max_rate_hz=self.max_rate_hz)
+        while not self._stop:
+            tick = clock.advance()
+            self._run_tick(tick)
+            self._flush_logs()
+            if self.state.stop:
+                self._stop = True
+            if max_ticks is not None and tick.index >= max_ticks:
+                break
+        self._on_shutdown()
+
+    def step_once(self) -> list[str]:
+        previous = self.state.tick.index if self.state.tick is not None else 0
+        tick = Tick(index=previous + 1, monotonic_seconds=time.monotonic())
+        self._run_tick(tick, force_rates=True)
+        return list(self.state.log_lines)
+
+    def _run_tick(self, tick: Tick, *, force_rates: bool = False) -> None:
+        self.state.tick = tick
+        self.state.log_lines = []
+        if self.event_collector is not None:
+            self.event_collector(self.state)
+        self._handle_action_deadline()
+        if self.state.stop:
+            return
+        self.maybe_update_language_state(force=force_rates)
+        self.maybe_handle_user_events()
+        self.maybe_enqueue_action_chunk(force=force_rates)
+        self.dispatch_action(force=force_rates)
+        self.state.events.clear()
+
+    def _current_observation(self) -> dict[str, Any] | None:
+        if self.observation_provider is None:
+            return None
+        try:
+            return self.observation_provider()
+        except Exception as exc:  # noqa: BLE001
+            logger.debug("observation_provider failed: %s", exc)
+            return None
+
+    def maybe_update_language_state(self, *, force: bool = False) -> None:
+        if self.state.mode != "action" or not self.state.task:
+            return
+        if self.state.action_queue:
+            self._language_gate.rearm()
+            return
+        if self.state.tick is None or not self._language_gate.due(self.state.tick, force=force):
+            return
+        observation = self._current_observation()
+        try:
+            self.policy_adapter.update_language_state(observation, self.state)
+        except Exception as exc:  # noqa: BLE001
+            logger.warning("language update failed: %s", exc, exc_info=logger.isEnabledFor(logging.DEBUG))
+            self.state.log(f"  [warn] language update failed: {type(exc).__name__}: {exc}")
+
+    def maybe_handle_user_events(self) -> None:
+        if self.state.take_event("user_interjection"):
+            self._handle_user_interjection()
+
+    def _handle_user_interjection(self) -> None:
+        text = str(self.state.extra.get("recent_interjection") or "")
+        if not text:
+            return
+        observation = self._current_observation()
+        self.policy_adapter.handle_interjection(text, observation, self.state)
+        self.state.extra["recent_interjection"] = None
+
+    def maybe_enqueue_action_chunk(self, *, force: bool = False) -> None:
+        with self.state.lock:
+            if self.state.mode != "action" or not self.state.task:
+                return
+            if self.state.action_queue:
+                return
+            if self.state.tick is None or not self._chunk_gate.due(self.state.tick, force=force):
+                return
+            revision = self.state.revision
+        observation = self._current_observation()
+        if observation is None:
+            return
+        try:
+            chunk = self.policy_adapter.select_action(observation, self.state)
+        except Exception as exc:  # noqa: BLE001
+            logger.warning("select_action failed: %s", exc, exc_info=logger.isEnabledFor(logging.DEBUG))
+            self.state.log(f"  [warn] select_action failed: {type(exc).__name__}: {exc}")
+            return
+        with self.state.lock:
+            if (
+                self.state.revision != revision
+                or self.state.mode != "action"
+                or self.state.stop
+                or self._stop
+            ):
+                logger.info("Discarded an action chunk invalidated during inference.")
+                return
+            self._enqueue_chunk(chunk)
+
+    def _enqueue_chunk(self, chunk: Any) -> None:
+        if chunk is None:
+            return
+        chunk_iter = chunk[0] if getattr(chunk, "ndim", None) == 3 else chunk
+        if getattr(chunk_iter, "ndim", None) == 1:
+            chunk_iter = chunk_iter.unsqueeze(0)
+        for step in chunk_iter:
+            self.state.action_queue.append(step.unsqueeze(0) if hasattr(step, "unsqueeze") else step)
+        try:
+            self.state.extra["last_chunk_size"] = int(chunk_iter.shape[0])
+        except Exception:  # noqa: BLE001
+            self.state.extra["last_chunk_size"] = len(self.state.action_queue)
+
+    def dispatch_action(self, *, force: bool = False) -> None:
+        if self.state.mode != "action":
+            self._last_dispatch_seconds = None
+            return
+        if self.state.tick is None or not self._ctrl_gate.due(self.state.tick, force=force):
+            return
+        queue = self.state.action_queue
+        if not queue:
+            self._last_dispatch_seconds = None
+            return
+        now = time.monotonic()
+        if self._last_dispatch_seconds is None or self.ctrl_hz <= 0:
+            n_to_pop = 1
+        else:
+            n_to_pop = max(1, min(len(queue), int(round((now - self._last_dispatch_seconds) * self.ctrl_hz))))
+        self._last_dispatch_seconds = now
+        latest = None
+        for _ in range(n_to_pop):
+            if not queue:
+                break
+            latest = queue.popleft()
+            self.state.actions_dispatched += 1
+        if latest is not None and self.action_executor is not None:
+            self.action_executor(latest)
+
+    def _handle_action_deadline(self) -> None:
+        deadline = self.state.action_deadline
+        if self.state.mode == "action" and deadline is not None and time.monotonic() >= deadline:
+            self.state.mode = "paused"
+            self.state.action_deadline = None
+            self.state.action_queue.clear()
+            self.state.log("timed action elapsed — paused")
+
+    def _flush_logs(self) -> None:
+        for line in self.state.log_lines:
+            print(f"[runtime] {line}", flush=True)
+
+    def _on_shutdown(self) -> None:
+        self.state.action_queue.clear()
+        print("[runtime] stopped", flush=True)

src/lerobot/runtime/registry.py

@@ -0,0 +1,39 @@
+# 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.
+
+"""Lazy mapping from policy types to language-runtime adapters."""
+
+from __future__ import annotations
+
+import importlib
+from collections.abc import Callable
+from typing import Any
+
+_ADAPTERS: dict[str, str] = {
+    "pi052": "lerobot.policies.pi052.inference.pi052_adapter:PI052PolicyAdapter",
+    "pi05": "lerobot.runtime.adapter:DirectTaskPolicyAdapter",
+    "molmoact2": "lerobot.runtime.adapter:DirectTaskPolicyAdapter",
+}
+
+
+def get_language_adapter_factory(policy_type: str) -> Callable[..., Any]:
+    """Return the adapter class registered for ``policy_type``."""
+    spec = _ADAPTERS.get(policy_type)
+    if spec is None:
+        raise ValueError(
+            f"No language-runtime adapter registered for policy type {policy_type!r}. "
+            f"Registered: {sorted(_ADAPTERS)}. Add an entry to lerobot.runtime.registry."
+        )
+    module_path, class_name = spec.split(":")
+    return getattr(importlib.import_module(module_path), class_name)

src/lerobot/runtime/sim_robocasa.py

@@ -0,0 +1,406 @@
+# 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.
+
+"""RoboCasa backend for interactive language-conditioned rollouts.
+
+It reuses the eval observation/action pipeline while prompts control a persistent selected scene.
+"""
+
+from __future__ import annotations
+
+import logging
+from collections.abc import Callable
+from datetime import datetime
+from pathlib import Path
+from typing import Any
+
+import numpy as np
+import torch
+
+from lerobot.utils.io_utils import StreamingVideoWriter
+from lerobot.utils.video_annotation import annotate_frame
+
+logger = logging.getLogger(__name__)
+
+
+def _short_cam_name(cam: str) -> str:
+    """Human-friendly view label for a RoboCasa camera name."""
+    c = cam.replace("robot0_", "")
+    return {
+        "agentview_left": "left",
+        "agentview_right": "right",
+        "eye_in_hand": "wrist",
+    }.get(c, c)
+
+
+def _label_panel(img: np.ndarray, label: str) -> np.ndarray:
+    """Draw a small camera-view label in the bottom-left corner of a panel."""
+    try:
+        import cv2  # noqa: PLC0415
+    except ImportError:
+        return img
+    y = img.shape[0] - 6
+    cv2.putText(img, label, (5, y), cv2.FONT_HERSHEY_SIMPLEX, 0.45, (0, 0, 0), 3, cv2.LINE_AA)
+    cv2.putText(img, label, (5, y), cv2.FONT_HERSHEY_SIMPLEX, 0.45, (255, 255, 0), 1, cv2.LINE_AA)
+    return img
+
+
+# Two workers avoid broken single-worker EGL rendering; only env 0 is displayed.
+_SIM_N_ENVS = 2
+
+
+def create_sim_env(
+    *,
+    task: str,
+    split: str | None,
+    obj_registries: list[str],
+    seed: int | None,
+    render_size: int = 384,
+) -> tuple[Any, dict]:
+    """Create and reset the vectorized RoboCasa environment before CUDA initializes.
+
+    Two workers keep EGL stable, while only env 0 is driven and displayed.
+    """
+    from lerobot.envs.configs import RoboCasaEnv as RoboCasaEnvConfig  # noqa: PLC0415
+
+    # The policy resizes inputs, so render_size only affects display quality and cost.
+    env_cfg = RoboCasaEnvConfig(
+        task=task,
+        split=split,
+        obj_registries=list(obj_registries),
+        observation_height=render_size,
+        observation_width=render_size,
+    )
+    # Keep one kitchen alive across sequential prompts.
+    envs = env_cfg.create_envs(
+        n_envs=_SIM_N_ENVS,
+        use_async_envs=True,
+        terminate_on_success=False,
+        horizon=100_000,
+    )
+    env = envs[next(iter(envs))][0]
+    logger.info("[sim] resetting RoboCasa scene task=%r split=%r (n_envs=%d)", task, split, _SIM_N_ENVS)
+    seeds = None if seed is None else [seed + i for i in range(_SIM_N_ENVS)]
+    obs, _ = env.reset(seed=seeds)
+    return env, obs
+
+
+def start_mjpeg_server(port: int, get_frame: Callable[[], np.ndarray | None]) -> Any:
+    """Start an MJPEG server that shows a placeholder until ``get_frame`` returns frames."""
+    import io  # noqa: PLC0415
+    import threading  # noqa: PLC0415
+    import time  # noqa: PLC0415
+    from http.server import BaseHTTPRequestHandler, ThreadingHTTPServer  # noqa: PLC0415
+
+    from PIL import Image  # noqa: PLC0415
+
+    _placeholder = Image.new("RGB", (256, 256), (17, 17, 17))
+
+    class _Handler(BaseHTTPRequestHandler):
+        def log_message(self, *args):  # silence per-request logging
+            pass
+
+        def do_GET(self):  # noqa: N802
+            if self.path in ("/", "/index.html"):
+                self.send_response(200)
+                self.send_header("Content-Type", "text/html")
+                self.end_headers()
+                self.wfile.write(
+                    b"<html><body style='margin:0;background:#111;text-align:center'>"
+                    b"<img src='/stream' style='max-width:100vw;max-height:100vh;"
+                    b"image-rendering:pixelated'></body></html>"
+                )
+                return
+            if self.path != "/stream":
+                self.send_response(404)
+                self.end_headers()
+                return
+            self.send_response(200)
+            self.send_header("Content-Type", "multipart/x-mixed-replace; boundary=frame")
+            self.end_headers()
+            try:
+                while True:
+                    frame = get_frame()
+                    buf = io.BytesIO()
+                    img = Image.fromarray(frame) if frame is not None else _placeholder
+                    img.save(buf, format="JPEG", quality=80)
+                    data = buf.getvalue()
+                    self.wfile.write(
+                        b"--frame\r\nContent-Type: image/jpeg\r\nContent-Length: "
+                        + str(len(data)).encode()
+                        + b"\r\n\r\n"
+                        + data
+                        + b"\r\n"
+                    )
+                    time.sleep(0.05)
+            except (BrokenPipeError, ConnectionResetError):
+                pass
+
+    try:
+        # Bind all interfaces intentionally so the viewer remains reachable
+        # through the documented SSH port-forwarding workflow.
+        server = ThreadingHTTPServer(("0.0.0.0", port), _Handler)  # nosec B104
+    except OSError as exc:
+        logger.warning("[sim] could not start live stream on port %d: %s", port, exc)
+        print(f"[runtime] WARNING: live stream port {port} unavailable ({exc})", flush=True)
+        return None
+    threading.Thread(target=server.serve_forever, daemon=True, name="sim-mjpeg").start()
+    print(
+        f"[runtime] live view: http://localhost:{port}  "
+        f"(over SSH: ssh -L {port}:localhost:{port} <host>) — loading until scene is ready",
+        flush=True,
+    )
+    return server
+
+
+class RoboCasaSimBackend:
+    """Expose a RoboCasa environment through the runtime observation/action contract.
+
+    The environment must be created before the policy initializes CUDA.
+    """
+
+    def __init__(
+        self,
+        *,
+        env: Any,
+        last_obs: dict,
+        task: str,
+        seed: int | None,
+        device: str,
+        preprocessor: Any,
+        postprocessor: Any,
+        record: bool = True,
+        output_dir: str | None = None,
+        view_cams: list[str] | None = None,
+    ) -> None:
+        self.env = env
+        self._last_obs = last_obs
+        self._scene_task = task
+        self._view_cams = view_cams or [
+            "robot0_agentview_left",
+            "robot0_eye_in_hand",
+            "robot0_agentview_right",
+        ]
+        self.device = torch.device(device) if isinstance(device, str) else device
+        self.preprocessor = preprocessor
+        self.postprocessor = postprocessor
+        self.seed = seed
+        self.record = record
+        self.output_dir = Path(output_dir) if output_dir else Path("outputs/runtime_sim")
+
+        self._video_writer: StreamingVideoWriter | None = None
+        self._video_path: Path | None = None
+        self._live_counter = 0
+        self._latest_frame: np.ndarray | None = None
+        self._stream_server: Any = None
+        self._reset_count = 0
+        # Bind these after runtime construction for live annotations.
+        self._task_getter: Callable[[], str | None] | None = None
+        self._subtask_getter: Callable[[], str | None] | None = None
+        self._memory_getter: Callable[[], str | None] | None = None
+        logger.info("[sim] scene ready — task_description=%r", self._scene_description())
+
+    def bind_runtime(self, runtime: Any) -> None:
+        """Wire live task/subtask/memory getters from the runtime state."""
+        self._task_getter = lambda: runtime.state.get("task")
+        self._subtask_getter = lambda: runtime.state.language_context.get("subtask")
+        self._memory_getter = lambda: (runtime.state.get("language_context") or {}).get("memory")
+
+    def _scene_description(self) -> str:
+        try:
+            return str(self.env.get_attr("task_description")[0]) or self._scene_task
+        except Exception:  # noqa: BLE001
+            return self._scene_task
+
+    def _current_task(self) -> str:
+        task = self._task_getter() if self._task_getter else None
+        return task or self._scene_description() or self._scene_task
+
+    def reset_scene(self) -> None:
+        """Re-roll the kitchen: reset the env to a fresh scene (new layout/style).
+
+        Uses a new seed each call so ``/reset`` explores different kitchens.
+        """
+        self._reset_count += 1
+        n = self.env.num_envs
+        if self.seed is None:
+            seeds = None
+        else:
+            base = self.seed + self._reset_count * 1000
+            seeds = [base + i for i in range(n)]
+        obs, _ = self.env.reset(seed=seeds)
+        self._last_obs = obs
+        logger.info("[sim] scene reset (#%d)", self._reset_count)
+
+    def _env0_obs(self) -> dict:
+        """Slice env 0 out of the batched vec-env observation (batch of 1)."""
+        raw = self._last_obs or {}
+        pixels = raw.get("pixels")
+        out: dict[str, Any] = {}
+        if isinstance(pixels, dict):
+            out["pixels"] = {k: np.asarray(v)[0:1] for k, v in pixels.items()}
+        agent_pos = raw.get("agent_pos")
+        if agent_pos is not None:
+            out["agent_pos"] = np.asarray(agent_pos)[0:1]
+        return out
+
+    def observation_provider(self) -> dict | None:
+        from lerobot.envs.utils import preprocess_observation  # noqa: PLC0415
+
+        try:
+            obs = preprocess_observation(self._env0_obs())
+        except Exception as exc:  # noqa: BLE001
+            logger.warning("[sim] preprocess_observation failed: %s", exc)
+            return None
+        # The adapter later replaces this recipe input with its generated subtask.
+        obs["task"] = [self._current_task()]
+        if self.preprocessor is not None:
+            try:
+                obs = self.preprocessor(obs)
+            except Exception as exc:  # noqa: BLE001
+                logger.warning("[sim] preprocessor failed: %s", exc)
+                return None
+        return {
+            k: (v.to(self.device) if isinstance(v, torch.Tensor) else v)
+            for k, v in obs.items()
+            if isinstance(k, str) and k.startswith("observation.")
+        }
+
+    def action_executor(self, action: Any) -> None:
+        try:
+            if self.postprocessor is not None:
+                action = self.postprocessor(action)
+            if isinstance(action, torch.Tensor):
+                if action.ndim > 1 and action.shape[0] == 1:
+                    action = action.squeeze(0)
+                action = action.detach().to("cpu").numpy()
+            # Tile env 0's action because the extra workers exist only for EGL stability.
+            action_row = np.asarray(action, dtype=np.float32).reshape(-1)
+            action_np = np.tile(action_row, (self.env.num_envs, 1))
+            obs, _reward, terminated, truncated, _info = self.env.step(action_np)
+            self._last_obs = obs
+            self._capture_frame()
+            # AsyncVectorEnv resets terminated sub-environments automatically.
+            if bool(np.any(terminated)) or bool(np.any(truncated)):
+                logger.info("[sim] episode ended — scene auto-reset")
+        except Exception as exc:  # noqa: BLE001
+            logger.error("[sim] env.step failed: %s", exc, exc_info=True)
+
+    def _multiview_frame(self) -> np.ndarray | None:
+        """Label and compose env 0's existing observation views without extra rendering."""
+        pixels = (self._last_obs or {}).get("pixels")
+        if not isinstance(pixels, dict) or not pixels:
+            return None
+        panels: list[np.ndarray] = []
+        for cam in self._view_cams:
+            v = pixels.get(cam)
+            if v is None:
+                continue
+            img = np.asarray(v)
+            if img.ndim == 4:  # (n_envs, H, W, C) -> env 0
+                img = img[0]
+            if img.ndim != 3 or img.shape[-1] != 3:
+                continue
+            panels.append(_label_panel(np.ascontiguousarray(img.astype(np.uint8)), _short_cam_name(cam)))
+        if not panels:
+            return None
+        h = min(p.shape[0] for p in panels)
+        panels = [p[:h] for p in panels]
+        return np.concatenate(panels, axis=1)
+
+    def _capture_frame(self) -> None:
+        frame = self._multiview_frame()
+        if frame is None:  # fallback to single env.render()
+            try:
+                rendered = self.env.call("render")[0]
+                if isinstance(rendered, np.ndarray) and rendered.ndim == 3:
+                    frame = rendered
+            except Exception as exc:  # noqa: BLE001
+                logger.debug("[sim] render failed: %s", exc)
+        if frame is None:
+            return
+        subtask = self._subtask_getter() if self._subtask_getter else None
+        memory = self._memory_getter() if self._memory_getter else None
+        annotated = annotate_frame(
+            frame,
+            (("Task", self._current_task()), ("Subtask", subtask), ("Memory", memory)),
+        )
+        self._latest_frame = annotated  # served by the live MJPEG stream
+        self._write_live_frame(annotated)
+        if self.record:
+            self._write_recording_frame(annotated)
+
+    def _write_recording_frame(self, frame: np.ndarray) -> None:
+        try:
+            if self._video_writer is None:
+                self.output_dir.mkdir(parents=True, exist_ok=True)
+                stamp = datetime.now().strftime("%Y%m%d_%H%M%S")
+                self._video_path = self.output_dir / f"sim_{stamp}.mp4"
+                fps = int((getattr(self.env, "metadata", None) or {}).get("render_fps", 20))
+                self._video_writer = StreamingVideoWriter(self._video_path, fps)
+            self._video_writer.add_frame(frame)
+        except Exception as exc:  # noqa: BLE001
+            logger.warning("[sim] video encoding failed: %s", exc)
+            self.record = False
+
+    def _write_live_frame(self, frame: np.ndarray) -> None:
+        """Write a rolling latest.png every few frames for live viewing over SSH.
+
+        Open ``{output_dir}/latest.png`` in an editor/viewer and refresh to watch
+        the rollout in near-real-time without a GUI window. Written atomically
+        (temp + replace) so a reader never sees a half-written file.
+        """
+        self._live_counter += 1
+        if self._live_counter % 3 != 0:
+            return
+        try:
+            import os  # noqa: PLC0415
+
+            from PIL import Image  # noqa: PLC0415
+
+            self.output_dir.mkdir(parents=True, exist_ok=True)
+            tmp = self.output_dir / ".latest.tmp.png"
+            Image.fromarray(frame).save(tmp)
+            os.replace(tmp, self.output_dir / "latest.png")
+        except Exception as exc:  # noqa: BLE001
+            logger.debug("[sim] live frame write failed: %s", exc)
+
+    def _flush_video(self) -> None:
+        if self._video_writer is None:
+            return
+        writer = self._video_writer
+        self._video_writer = None
+        try:
+            writer.close()
+            logger.info("[sim] wrote video (%d frames) to %s", writer.frames_written, self._video_path)
+            print(f"[runtime] sim video saved to {self._video_path}", flush=True)
+        except Exception as exc:  # noqa: BLE001
+            logger.warning("[sim] video close failed: %s", exc)
+
+    def attach_stream_server(self, server: Any) -> None:
+        """Attach an already-running MJPEG server so disconnect() can stop it."""
+        self._stream_server = server
+
+    def disconnect(self) -> None:
+        """Match the robot backend's cleanup contract."""
+        if self._stream_server is not None:
+            try:
+                self._stream_server.shutdown()
+            except Exception as exc:  # noqa: BLE001
+                logger.debug("[sim] stream server shutdown raised %s", exc)
+        self._flush_video()
+        try:
+            self.env.close()
+        except Exception as exc:  # noqa: BLE001
+            logger.debug("[sim] env.close raised %s", exc)

src/lerobot/scripts/augment_dataset_quantile_stats.py

@@ -36,7 +36,6 @@ python src/lerobot/scripts/augment_dataset_quantile_stats.py \
 import argparse
 import concurrent.futures
 import logging
-import os
 from pathlib import Path
 
 import numpy as np
@@ -53,7 +52,6 @@ from lerobot.datasets import (
     get_feature_stats,
     write_stats,
 )
-from lerobot.datasets.compute_stats import sample_indices
 from lerobot.utils.utils import init_logging
 
 
@@ -79,14 +77,12 @@ def has_quantile_stats(stats: dict[str, dict] | None, quantile_list_keys: list[s
     return False
 
 
-def process_single_episode(dataset: LeRobotDataset, episode_idx: int, use_sampling: bool = True) -> dict:
+def process_single_episode(dataset: LeRobotDataset, episode_idx: int) -> dict:
     """Process a single episode and return its statistics.
 
     Args:
         dataset: The LeRobot dataset
         episode_idx: Index of the episode to process
-        use_sampling: If True, sub-sample image/video frames per episode to bound
-            memory. If False, use every frame (exact, higher memory).
 
     Returns:
         Dictionary containing episode statistics
@@ -96,31 +92,16 @@ def process_single_episode(dataset: LeRobotDataset, episode_idx: int, use_sampli
     start_idx = dataset.meta.episodes[episode_idx]["dataset_from_index"]
     end_idx = dataset.meta.episodes[episode_idx]["dataset_to_index"]
 
-    episode_len = end_idx - start_idx
-
-    # Images/video are the memory hog, so sub-sample those frames per episode;
-    # numeric columns are cheap, so read them in full (exact).
-    image_keys = [k for k in dataset.features if dataset.features[k]["dtype"] in ("image", "video")]
-    numeric_keys = [
-        k for k in dataset.features if dataset.features[k]["dtype"] not in ("image", "video", "string")
-    ]
-
     collected_data: dict[str, list] = {}
+    for idx in range(start_idx, end_idx):
+        item = dataset[idx]
+        for key, value in item.items():
+            if key not in dataset.features:
+                continue
 
-    # Numeric features: every frame, read directly from the underlying table.
+            if key not in collected_data:
-    if numeric_keys:
+                collected_data[key] = []
-        numeric_cols = dataset.hf_dataset.select_columns(numeric_keys)[start_idx:end_idx]
+            collected_data[key].append(value)
-        for key in numeric_keys:
-            collected_data[key] = [torch.as_tensor(v) for v in numeric_cols[key]]
-
-    # Image/video features: decode only a sampled subset of frames.
-    if image_keys:
-        sampled_offsets = sample_indices(episode_len) if use_sampling else list(range(episode_len))
-        for offset in sampled_offsets:
-            item = dataset[start_idx + offset]
-            for key in image_keys:
-                if key in item:
-                    collected_data.setdefault(key, []).append(item[key])
 
     ep_stats = {}
     for key, data_list in collected_data.items():
@@ -150,13 +131,11 @@ def process_single_episode(dataset: LeRobotDataset, episode_idx: int, use_sampli
     return ep_stats
 
 
-def compute_quantile_stats_for_dataset(dataset: LeRobotDataset, use_sampling: bool = True) -> dict[str, dict]:
+def compute_quantile_stats_for_dataset(dataset: LeRobotDataset) -> dict[str, dict]:
     """Compute quantile statistics for all episodes in the dataset.
 
     Args:
         dataset: The LeRobot dataset to compute statistics for
-        use_sampling: If True, sub-sample image/video frames per episode to bound
-            memory. If False, use every frame (exact, higher memory).
 
     Returns:
         Dictionary containing aggregated statistics with quantiles
@@ -174,15 +153,15 @@ def compute_quantile_stats_for_dataset(dataset: LeRobotDataset, use_sampling: bo
     if has_videos:
         logging.info("Dataset contains video keys - using sequential processing for thread safety")
         for episode_idx in tqdm(range(dataset.num_episodes), desc="Processing episodes"):
-            ep_stats = process_single_episode(dataset, episode_idx, use_sampling)
+            ep_stats = process_single_episode(dataset, episode_idx)
             episode_stats_list.append(ep_stats)
     else:
         logging.info("Dataset has no video keys - using parallel processing for better performance")
-        max_workers = min(dataset.num_episodes, int(os.environ.get("LEROBOT_STATS_MAX_WORKERS", 16)))
+        max_workers = min(dataset.num_episodes, 16)
 
         with concurrent.futures.ThreadPoolExecutor(max_workers=max_workers) as executor:
             future_to_episode = {
-                executor.submit(process_single_episode, dataset, episode_idx, use_sampling): episode_idx
+                executor.submit(process_single_episode, dataset, episode_idx): episode_idx
                 for episode_idx in range(dataset.num_episodes)
             }
 
@@ -209,7 +188,6 @@ def augment_dataset_with_quantile_stats(
     repo_id: str,
     root: str | Path | None = None,
     overwrite: bool = False,
-    use_sampling: bool = True,
 ) -> None:
     """Augment a dataset with quantile statistics if they are missing.
 
@@ -217,8 +195,6 @@ def augment_dataset_with_quantile_stats(
         repo_id: Repository ID of the dataset
         root: Local root directory for the dataset
         overwrite: Overwrite existing quantile statistics if they already exist
-        use_sampling: If True, sub-sample image/video frames per episode to bound
-            memory. If False, use every frame (exact, higher memory).
     """
     logging.info(f"Loading dataset: {repo_id}")
     dataset = LeRobotDataset(
@@ -232,7 +208,7 @@ def augment_dataset_with_quantile_stats(
 
     logging.info("Dataset does not contain quantile statistics. Computing them now...")
 
-    new_stats = compute_quantile_stats_for_dataset(dataset, use_sampling=use_sampling)
+    new_stats = compute_quantile_stats_for_dataset(dataset)
 
     logging.info("Updating dataset metadata with new quantile statistics")
     dataset.meta.stats = new_stats
@@ -272,14 +248,6 @@ def main():
         action="store_true",
         help="Overwrite existing quantile statistics if they already exist",
     )
-    parser.add_argument(
-        "--no-sampling",
-        action="store_true",
-        help=(
-            "Compute stats over every frame (exact, higher memory). By default, "
-            "image/video frames are sub-sampled per episode to bound memory."
-        ),
-    )
 
     args = parser.parse_args()
     root = Path(args.root) if args.root else None
@@ -290,7 +258,6 @@ def main():
         repo_id=args.repo_id,
         root=root,
         overwrite=args.overwrite,
-        use_sampling=not args.no_sampling,
     )
 
 

src/lerobot/scripts/convert_dataset_v21_to_v30.py

@@ -94,8 +94,6 @@ from lerobot.datasets.video_utils import concatenate_video_files, get_video_dura
 from lerobot.utils.constants import HF_LEROBOT_HOME
 from lerobot.utils.utils import flatten_dict, init_logging
 
-logger = logging.getLogger(__name__)
-
 V21 = "v2.1"
 V30 = "v3.0"
 
@@ -478,11 +476,11 @@ def convert_dataset(
     # First check if the dataset already has a v3.0 version
     if root is None and not force_conversion:
         try:
-            logger.info("Trying to download v3.0 version of the dataset from the hub...")
+            print("Trying to download v3.0 version of the dataset from the hub...")
             snapshot_download(repo_id, repo_type="dataset", revision=V30, local_dir=HF_LEROBOT_HOME / repo_id)
             return
         except Exception:
-            logger.info("Dataset does not have an uploaded v3.0 version. Continuing with conversion.")
+            print("Dataset does not have an uploaded v3.0 version. Continuing with conversion.")
 
     # Set root based on whether local dataset path is provided
     use_local_dataset = False
@@ -490,7 +488,7 @@ def convert_dataset(
     if root.exists():
         validate_local_dataset_version(root)
         use_local_dataset = True
-        logger.info(f"Using local dataset at {root}")
+        print(f"Using local dataset at {root}")
 
     old_root = root.parent / f"{root.name}_old"
     new_root = root.parent / f"{root.name}_v30"
@@ -525,7 +523,7 @@ def convert_dataset(
         try:
             hub_api.delete_tag(repo_id, tag=CODEBASE_VERSION, repo_type="dataset")
         except (HTTPError, RevisionNotFoundError) as e:
-            logger.warning(f"tag={CODEBASE_VERSION} probably doesn't exist. Skipping exception ({e})")
+            print(f"tag={CODEBASE_VERSION} probably doesn't exist. Skipping exception ({e})")
             pass
         hub_api.delete_files(
             delete_patterns=["data/chunk*/episode_*", "meta/*.jsonl", "videos/chunk*"],

src/lerobot/scripts/lerobot_annotate.py

@@ -154,14 +154,14 @@ def _push_to_hub(root: Path, cfg: AnnotationPipelineConfig) -> None:
     repo_id = cfg.new_repo_id or cfg.repo_id
     commit_message = cfg.push_commit_message or "Add steerable annotations (lerobot-annotate)"
     api = HfApi()
-    logger.info(f"[lerobot-annotate] creating/locating dataset repo {repo_id}...")
+    print(f"[lerobot-annotate] creating/locating dataset repo {repo_id}...", flush=True)
     api.create_repo(
         repo_id=repo_id,
         repo_type="dataset",
         private=cfg.push_private,
         exist_ok=True,
     )
-    logger.info(f"[lerobot-annotate] uploading {root} -> {repo_id}...")
+    print(f"[lerobot-annotate] uploading {root} -> {repo_id}...", flush=True)
     commit_info = api.upload_folder(
         folder_path=str(root),
         repo_id=repo_id,
@@ -172,7 +172,7 @@ def _push_to_hub(root: Path, cfg: AnnotationPipelineConfig) -> None:
         # at the source dataset; a fresh card is generated below instead.
         ignore_patterns=[".annotate_staging/**", "**/.DS_Store", "README.md"],
     )
-    logger.info(f"[lerobot-annotate] uploaded to https://huggingface.co/datasets/{repo_id}")
+    print(f"[lerobot-annotate] uploaded to https://huggingface.co/datasets/{repo_id}", flush=True)
 
     dataset_info = load_info(root)
     card = create_lerobot_dataset_card(dataset_info=dataset_info, license="apache-2.0", repo_id=repo_id)
@@ -200,13 +200,14 @@ def _push_to_hub(root: Path, cfg: AnnotationPipelineConfig) -> None:
         with suppress(RevisionNotFoundError):
             api.delete_tag(repo_id, tag=version_tag, repo_type="dataset")
         api.create_tag(**tag_kwargs)
-        logger.info(f"[lerobot-annotate] tagged {repo_id} as {version_tag}")
+        print(f"[lerobot-annotate] tagged {repo_id} as {version_tag}", flush=True)
     except Exception as exc:  # noqa: BLE001
-        logger.warning(
+        print(
             f"[lerobot-annotate] WARNING: could not create tag {version_tag!r} on {repo_id}: {exc}. "
             "Dataset is uploaded but ``LeRobotDataset`` won't be able to load it until it's tagged. "
             "Run: from huggingface_hub import HfApi; "
-            f"HfApi().create_tag({repo_id!r}, tag={version_tag!r}, repo_type='dataset', exist_ok=True)"
+            f"HfApi().create_tag({repo_id!r}, tag={version_tag!r}, repo_type='dataset', exist_ok=True)",
+            flush=True,
         )
 
 

src/lerobot/scripts/lerobot_dataset_viz.py

@@ -89,8 +89,6 @@ from lerobot.datasets import LeRobotDataset
 from lerobot.utils.constants import ACTION, DONE, OBS_STATE, REWARD, SUCCESS
 from lerobot.utils.utils import init_logging
 
-logger = logging.getLogger(__name__)
-
 DEFAULT_FOXGLOVE_PORT = 8765
 DEFAULT_RERUN_PORT = 9090
 
@@ -301,7 +299,7 @@ def visualize_dataset(
             while True:
                 time.sleep(1)
         except KeyboardInterrupt:
-            logger.info("Ctrl-C received. Exiting.")
+            print("Ctrl-C received. Exiting.")
 
 
 def main():

src/lerobot/scripts/lerobot_eval.py

@@ -62,7 +62,7 @@ from dataclasses import asdict
 from functools import partial
 from pathlib import Path
 from pprint import pformat
-from typing import TYPE_CHECKING, Any, TypedDict
+from typing import Any, TypedDict
 
 import einops
 import gymnasium as gym
@@ -87,21 +87,26 @@ from lerobot.processor import PolicyProcessorPipeline
 from lerobot.types import PolicyAction
 from lerobot.utils.constants import ACTION, DONE, OBS_IMAGE, OBS_IMAGES, OBS_STR, REWARD
 from lerobot.utils.device_utils import get_safe_torch_device
-from lerobot.utils.import_utils import _peft_available, register_third_party_plugins, require_package
+from lerobot.utils.import_utils import register_third_party_plugins
 from lerobot.utils.io_utils import write_video
 from lerobot.utils.random_utils import set_seed
 from lerobot.utils.utils import (
     init_logging,
     inside_slurm,
 )
-
+from lerobot.utils.video_annotation import annotate_frame
-if TYPE_CHECKING or _peft_available:
-    from peft import PeftModel
-else:
-    PeftModel = None
 
 
-logger = logging.getLogger(__name__)
+def _annotate_eval_frames(frames: np.ndarray, task: str | None, subtask: str | None) -> np.ndarray:
+    """Overlay the high-level task and predicted subtask onto rendered frames.
+
+    ``frames`` is ``(n_envs, H, W, C)`` uint8. Best-effort: if OpenCV isn't
+    available the frames are returned unchanged so eval never fails over a
+    visualization concern.
+    """
+    if frames.ndim != 4 or frames.shape[-1] != 3:
+        return frames
+    return np.stack([annotate_frame(frame, (("Task", task), ("Subtask", subtask))) for frame in frames])
 
 
 def _env_features_to_dataset_features(env_features: dict) -> dict:
@@ -452,11 +457,13 @@ def eval_policy(
         exc = ValueError(
             f"Policy of type 'PreTrainedPolicy' is expected, but type '{type(policy)}' was provided."
         )
-        if not _peft_available:
+        try:
-            raise exc
+            from peft import PeftModel
-        require_package("peft", extra="peft")
+
-        if not isinstance(policy, PeftModel):
+            if not isinstance(policy, PeftModel):
-            raise exc
+                raise exc
+        except ImportError:
+            raise exc from None
 
     start = time.time()
     # Preserve the mode for direct callers. eval_policy_all scopes the mode
@@ -483,11 +490,36 @@ def eval_policy(
             return
         n_to_render_now = min(max_episodes_rendered - n_episodes_rendered, env.num_envs)
         if isinstance(env, gym.vector.SyncVectorEnv):
-            ep_frames.append(np.stack([env.envs[i].render() for i in range(n_to_render_now)]))  # noqa: B023
+            frames = np.stack([env.envs[i].render() for i in range(n_to_render_now)])  # noqa: B023
         elif hasattr(env, "call"):
             # Here we must render all frames and discard any we don't need.
             # Covers AsyncVectorEnv and _LazyAsyncVectorEnv (which wraps one).
-            ep_frames.append(np.stack(env.call("render")[:n_to_render_now]))
+            frames = np.stack(env.call("render")[:n_to_render_now])
+        else:
+            return
+
+        # Overlay the high-level task and (for hierarchical policies like
+        # pi052) the predicted low-level subtask onto each frame. Both are
+        # best-effort: missing values just skip that line.
+        try:
+            tasks = list(env.call("task_description"))
+        except (AttributeError, NotImplementedError):
+            try:
+                tasks = list(env.call("task"))
+            except (AttributeError, NotImplementedError):
+                tasks = None
+        subtasks = getattr(policy, "last_subtasks", None)
+        annotated = []
+        for i in range(frames.shape[0]):
+            subtask_i = subtasks[i] if subtasks is not None and i < len(subtasks) else None
+            annotated.append(
+                _annotate_eval_frames(
+                    frames[i : i + 1],
+                    tasks[i] if tasks is not None and i < len(tasks) else None,
+                    subtask_i,
+                )[0]
+            )
+        ep_frames.append(np.stack(annotated))
 
     if max_episodes_rendered > 0:
         video_paths: list[str] = []
@@ -564,7 +596,7 @@ def eval_policy(
         if seeds:
             all_seeds.extend(seeds)
         else:
-            all_seeds.extend([None] * env.num_envs)
+            all_seeds.append(None)
 
         # FIXME: episode_data is either None or it doesn't exist
         if return_episode_data:
@@ -802,13 +834,13 @@ def eval_main(cfg: EvalPipelineConfig):
             recording_repo_id=cfg.eval.recording_repo_id,
             recording_private=cfg.eval.recording_private,
         )
-        logger.info("Overall Aggregated Metrics:")
+        print("Overall Aggregated Metrics:")
-        logger.info(info["overall"])
+        print(info["overall"])
 
         # Print per-suite stats
         for task_group, task_group_info in info.items():
-            logger.info(f"\nAggregated Metrics for {task_group}:")
+            print(f"\nAggregated Metrics for {task_group}:")
-            logger.info(task_group_info)
+            print(task_group_info)
     # Close all vec envs
     close_envs(envs)
 

src/lerobot/scripts/lerobot_find_cameras.py

@@ -28,6 +28,7 @@ lerobot-find-cameras
 # NOTE(Steven): macOS cameras sometimes report different FPS at init time, not an issue here as we don't specify FPS when opening the cameras, but the information displayed might not be truthful.
 
 import argparse
+import concurrent.futures
 import logging
 import time
 from pathlib import Path
@@ -39,7 +40,6 @@ from PIL import Image
 from lerobot.cameras import ColorMode
 from lerobot.cameras.opencv import OpenCVCamera, OpenCVCameraConfig
 from lerobot.cameras.realsense import RealSenseCamera, RealSenseCameraConfig
-from lerobot.utils.utils import init_logging
 
 logger = logging.getLogger(__name__)
 
@@ -132,7 +132,7 @@ def save_image(
     camera_identifier: str | int,
     images_dir: Path,
     camera_type: str,
-) -> None:
+):
     """
     Saves a single image to disk using Pillow. Handles color conversion if necessary.
     """
@@ -151,7 +151,7 @@ def save_image(
         logger.error(f"Failed to save image for camera {camera_identifier} (type {camera_type}): {e}")
 
 
-def create_camera_instance(cam_meta: dict[str, Any], *, warmup_s: int = 1) -> dict[str, Any] | None:
+def create_camera_instance(cam_meta: dict[str, Any]) -> dict[str, Any] | None:
     """Create and connect to a camera instance based on metadata."""
     cam_type = cam_meta.get("type")
     cam_id = cam_meta.get("id")
@@ -164,14 +164,12 @@ def create_camera_instance(cam_meta: dict[str, Any], *, warmup_s: int = 1) -> di
             cv_config = OpenCVCameraConfig(
                 index_or_path=cam_id,
                 color_mode=ColorMode.RGB,
-                warmup_s=warmup_s,
             )
             instance = OpenCVCamera(cv_config)
         elif cam_type == "RealSense":
             rs_config = RealSenseCameraConfig(
                 serial_number_or_name=cam_id,
                 color_mode=ColorMode.RGB,
-                warmup_s=warmup_s,
             )
             instance = RealSenseCamera(rs_config)
         else:
@@ -189,7 +187,9 @@ def create_camera_instance(cam_meta: dict[str, Any], *, warmup_s: int = 1) -> di
         return None
 
 
-def process_camera_image(cam_dict: dict[str, Any], output_dir: Path, current_time: float) -> None:
+def process_camera_image(
+    cam_dict: dict[str, Any], output_dir: Path, current_time: float
+) -> concurrent.futures.Future | None:
     """Capture and process an image from a single camera."""
     cam = cam_dict["instance"]
     meta = cam_dict["meta"]
@@ -199,7 +199,7 @@ def process_camera_image(cam_dict: dict[str, Any], output_dir: Path, current_tim
     try:
         image_data = cam.read()
 
-        save_image(
+        return save_image(
             image_data,
             cam_id_str,
             output_dir,
@@ -214,21 +214,21 @@ def process_camera_image(cam_dict: dict[str, Any], output_dir: Path, current_tim
     return None
 
 
-def cleanup_camera(cam_dict: dict[str, Any]) -> None:
+def cleanup_cameras(cameras_to_use: list[dict[str, Any]]):
     """Disconnect all cameras."""
-    logger.info(f"Disconnecting camera with ID {cam_dict['meta'].get('id')}...")
+    logger.info(f"Disconnecting {len(cameras_to_use)} cameras...")
-    try:
+    for cam_dict in cameras_to_use:
-        if cam_dict["instance"] and cam_dict["instance"].is_connected:
+        try:
-            cam_dict["instance"].disconnect()
+            if cam_dict["instance"] and cam_dict["instance"].is_connected:
-    except Exception as e:
+                cam_dict["instance"].disconnect()
-        logger.error(f"Error disconnecting camera {cam_dict['meta'].get('id')}: {e}")
+        except Exception as e:
+            logger.error(f"Error disconnecting camera {cam_dict['meta'].get('id')}: {e}")
 
 
 def save_images_from_all_cameras(
     output_dir: Path,
     record_time_s: float = 2.0,
     camera_type: str | None = None,
-    warmup_s: int = 1,
 ):
     """
     Connects to detected cameras (optionally filtered by type) and saves images from each.
@@ -239,7 +239,6 @@ def save_images_from_all_cameras(
         record_time_s: Duration in seconds to record images.
         camera_type: Optional string to filter cameras ("realsense" or "opencv").
                             If None, uses all detected cameras.
-        warmup_s: Duration in seconds to warmup camera before recording images.
     """
     output_dir.mkdir(parents=True, exist_ok=True)
     logger.info(f"Saving images to {output_dir}")
@@ -249,32 +248,47 @@ def save_images_from_all_cameras(
         logger.warning("No cameras detected matching the criteria. Cannot save images.")
         return
 
-    logger.info(
+    cameras_to_use = []
-        f"Starting image capture for {record_time_s} seconds from {len(all_camera_metadata)} cameras."
+    for cam_meta in all_camera_metadata:
-    )
+        camera_instance = create_camera_instance(cam_meta)
+        if camera_instance:
+            cameras_to_use.append(camera_instance)
 
-    try:
+    if not cameras_to_use:
-        for cam_meta in all_camera_metadata:
+        logger.warning("No cameras could be connected. Aborting image save.")
-            cam_dict = create_camera_instance(cam_meta, warmup_s=warmup_s)
+        return
-            if cam_dict is None:
+
-                continue
+    logger.info(f"Starting image capture for {record_time_s} seconds from {len(cameras_to_use)} cameras.")
-            start_time = time.perf_counter()
+    start_time = time.perf_counter()
+
+    with concurrent.futures.ThreadPoolExecutor(max_workers=len(cameras_to_use) * 2) as executor:
+        try:
             while time.perf_counter() - start_time < record_time_s:
+                futures = []
                 current_capture_time = time.perf_counter()
-                process_camera_image(cam_dict, output_dir, current_capture_time)
+
-            cleanup_camera(cam_dict)
+                for cam_dict in cameras_to_use:
-    except KeyboardInterrupt:
+                    future = process_camera_image(cam_dict, output_dir, current_capture_time)
-        logger.info("Capture interrupted by user.")
+                    if future:
-    finally:
+                        futures.append(future)
-        print(f"Image capture finished. Images saved to {output_dir}")
+
+                if futures:
+                    concurrent.futures.wait(futures)
+
+        except KeyboardInterrupt:
+            logger.info("Capture interrupted by user.")
+        finally:
+            print("\nFinalizing image saving...")
+            executor.shutdown(wait=True)
+            cleanup_cameras(cameras_to_use)
+            print(f"Image capture finished. Images saved to {output_dir}")
 
 
 def main():
-    init_logging()
-
     parser = argparse.ArgumentParser(
         description="Unified camera utility script for listing cameras and capturing images."
     )
+
     parser.add_argument(
         "camera_type",
         type=str,
@@ -292,14 +306,8 @@ def main():
     parser.add_argument(
         "--record-time-s",
         type=float,
-        default=2.0,
+        default=6.0,
-        help="Time duration to attempt capturing frames. Default: 2 seconds.",
+        help="Time duration to attempt capturing frames. Default: 6 seconds.",
-    )
-    parser.add_argument(
-        "--warmup-s",
-        type=int,
-        default=1,
-        help="Time duration to warmup camera before attempting to capture frames. Default: 1 second.",
     )
     args = parser.parse_args()
     save_images_from_all_cameras(**vars(args))

src/lerobot/scripts/lerobot_rollout.py

@@ -151,6 +151,7 @@ Usage examples
 """
 
 import logging
+import sys
 
 from lerobot.cameras.opencv import OpenCVCameraConfig  # noqa: F401
 from lerobot.cameras.realsense import RealSenseCameraConfig  # noqa: F401
@@ -165,7 +166,6 @@ from lerobot.robots import (  # noqa: F401
     earthrover_mini_plus,
     hope_jr,
     koch_follower,
-    lekiwi,
     omx_follower,
     openarm_follower,
     reachy2,
@@ -242,10 +242,69 @@ def rollout(cfg: RolloutConfig):
     logger.info("Rollout finished")
 
 
-def main():
+_LANGUAGE_RUNTIME_FLAGS = {
-    """CLI entry point for ``lerobot-rollout``."""
+    "--language",
+    "--no_robot",
+    "--sim",
+    "--direct_subtask",
+    "--sim.direct_subtask",
+    "--disable_memory",
+    "--fp8",
+}
+_LANGUAGE_RUNTIME_PREFIXES = (
+    "--sim.",
+    "--chunk_hz",
+    "--ctrl_hz",
+    "--high_level_hz",
+    "--subtask_chunks_per_gen",
+    "--text_min_new_tokens",
+    "--text_temperature",
+    "--text_top_p",
+)
+
+
+def _uses_language_runtime(argv: list[str]) -> bool:
+    """Return whether *argv* selects the interactive language runtime.
+
+    ``--language`` is the explicit selector for real-robot runs whose other
+    options overlap with the standard rollout CLI. Language-only options also
+    select it automatically, which keeps the former language-runtime examples
+    working after replacing their command name with ``lerobot-rollout``.
+    """
+    return any(
+        arg.split("=", 1)[0] in _LANGUAGE_RUNTIME_FLAGS or arg.startswith(_LANGUAGE_RUNTIME_PREFIXES)
+        for arg in argv
+    )
+
+
+def main(argv: list[str] | None = None):
+    """CLI entry point for ``lerobot-rollout``.
+
+    Standard policy deployment continues through :class:`RolloutConfig`.
+    Interactive language-conditioned and RoboCasa runs share this entry point
+    and are selected with ``--language`` or any language-runtime-only option.
+    """
     register_third_party_plugins()
-    rollout()
+    cli_args = list(sys.argv[1:] if argv is None else argv)
+    if _uses_language_runtime(cli_args):
+        from lerobot.runtime.cli import run as run_language_runtime
+
+        # ``--language`` is a dispatcher flag, not part of the runtime's own
+        # argparse surface. All other arguments pass through unchanged.
+        runtime_args = [arg for arg in cli_args if arg != "--language"]
+        return run_language_runtime(runtime_args, prog="lerobot-rollout")
+
+    if argv is None:
+        return rollout()
+
+    # draccus reads sys.argv. Supporting an explicit argv keeps this entry
+    # point easy to smoke-test and mirrors the language-runtime branch above.
+    previous_argv = sys.argv
+    try:
+        sys.argv = [previous_argv[0], *cli_args]
+        return rollout()
+    finally:
+        sys.argv = previous_argv
 
 
 if __name__ == "__main__":

src/lerobot/scripts/lerobot_train.py

@@ -20,10 +20,11 @@ Requires: pip install 'lerobot[training]'  (includes dataset + accelerate + wand
 
 import dataclasses
 import logging
+import os
 import sys
 import time
-from collections.abc import Iterator
+from contextlib import nullcontext
-from contextlib import contextmanager, nullcontext
+from datetime import timedelta
 from pprint import pformat
 from typing import TYPE_CHECKING, Any
 
@@ -58,7 +59,7 @@ from lerobot.optim.factory import make_optimizer_and_scheduler
 from lerobot.policies import PreTrainedPolicy, make_policy, make_pre_post_processors
 from lerobot.rewards import make_reward_pre_post_processors
 from lerobot.utils.collate import lerobot_collate_fn
-from lerobot.utils.import_utils import _peft_available, register_third_party_plugins, require_package
+from lerobot.utils.import_utils import register_third_party_plugins
 from lerobot.utils.logging_utils import AverageMeter, MetricsTracker
 from lerobot.utils.random_utils import set_seed
 from lerobot.utils.utils import (
@@ -69,28 +70,9 @@ from lerobot.utils.utils import (
     inside_slurm,
 )
 
-if TYPE_CHECKING or _peft_available:
-    from peft import PeftModel
-else:
-    PeftModel = None
-
 from .lerobot_eval import eval_policy_all
 
 
-@contextmanager
-def _make_eval_envs(cfg: TrainPipelineConfig) -> Iterator[dict[str, dict[int, Any]]]:
-    """Create evaluation environments for one run and always dispose of them."""
-    envs = make_env(
-        cfg.env,
-        n_envs=cfg.eval.batch_size,
-        use_async_envs=cfg.eval.use_async_envs,
-    )
-    try:
-        yield envs
-    finally:
-        close_envs(envs)
-
-
 def _dataloader_worker_kwargs(cfg: TrainPipelineConfig) -> dict[str, Any]:
     """Return worker-only DataLoader options, disabling them for single-process loading."""
     workers_enabled = cfg.num_workers > 0
@@ -111,6 +93,7 @@ def update_policy(
     lr_scheduler=None,
     lock=None,
     sample_weighter=None,
+    log_metrics: bool = True,
 ) -> tuple[MetricsTracker, dict | None]:
     """
     Performs a single training step to update the policy's weights.
@@ -128,6 +111,7 @@ def update_policy(
         lr_scheduler: An optional learning rate scheduler.
         lock: An optional lock for thread-safe optimizer updates.
         sample_weighter: Optional SampleWeighter instance for per-sample loss weighting.
+        log_metrics: Whether to synchronize and record GPU metrics this step.
 
     Returns:
         A tuple containing:
@@ -195,12 +179,20 @@ def update_policy(
     if has_method(accelerator.unwrap_model(policy, keep_fp32_wrapper=True), "update"):
         accelerator.unwrap_model(policy, keep_fp32_wrapper=True).update()
 
-    train_metrics.loss = loss.item()
-    train_metrics.grad_norm = grad_norm.item()
     train_metrics.lr = optimizer.param_groups[0]["lr"]
-    train_metrics.update_s = time.perf_counter() - start_time
     if torch.cuda.is_available():
         train_metrics.gpu_mem_gb = torch.cuda.max_memory_allocated() / (1024**3)
+    train_metrics.accumulate_tensor("loss", loss)
+    train_metrics.accumulate_tensor("grad_norm", grad_norm)
+    train_metrics.update_s = time.perf_counter() - start_time
+    # Synchronize accumulated GPU metrics only when logging.
+    if log_metrics:
+        train_metrics.materialize_tensors()
+        # Materialize detached loss components during the same logging synchronization.
+        if output_dict:
+            output_dict = {
+                k: (v.item() if isinstance(v, torch.Tensor) else v) for k, v in output_dict.items()
+            }
     # Aggregate the policy's scalar outputs for logging and rank-reduction across the log window.
     if output_dict:
         train_metrics.update_metrics(output_dict)
@@ -227,9 +219,11 @@ def train(cfg: TrainPipelineConfig, accelerator: "Accelerator | None" = None):
     if cfg.job.is_remote:
         return submit_to_hf(cfg)
 
+    from lerobot.utils.import_utils import require_package
+
     require_package("accelerate", extra="training")
     from accelerate import Accelerator
-    from accelerate.utils import DistributedDataParallelKwargs, DistributedType
+    from accelerate.utils import DistributedDataParallelKwargs, DistributedType, InitProcessGroupKwargs
 
     cfg.validate()
 
@@ -238,7 +232,16 @@ def train(cfg: TrainPipelineConfig, accelerator: "Accelerator | None" = None):
     # We set step_scheduler_with_optimizer=False to prevent accelerate from adjusting the lr_scheduler steps based on the num_processes
     # We set find_unused_parameters=True to handle models with conditional computation
     if accelerator is None:
-        ddp_kwargs = DistributedDataParallelKwargs(find_unused_parameters=True)
+        # Static graphs restore DDP overlap when conditional parameter usage is stable.
+        # Environment flags retain the existing defaults.
+        ddp_find_unused = os.environ.get("LEROBOT_DDP_FIND_UNUSED", "1") == "1"
+        ddp_static_graph = os.environ.get("LEROBOT_DDP_STATIC_GRAPH", "0") == "1"
+        ddp_kwargs = DistributedDataParallelKwargs(
+            find_unused_parameters=ddp_find_unused and not ddp_static_graph,
+            static_graph=ddp_static_graph,
+        )
+        # Allow rank 0 enough time to index large datasets before other ranks leave the barrier.
+        ipg_kwargs = InitProcessGroupKwargs(timeout=timedelta(hours=2))
         # Accelerate auto-detects the device based on the available hardware and ignores the policy.device setting.
         # Force the device to be CPU when the active config's device is set to CPU (works for both policy and reward model training).
         force_cpu = cfg.trainable_config.device == "cpu"
@@ -248,7 +251,7 @@ def train(cfg: TrainPipelineConfig, accelerator: "Accelerator | None" = None):
         accelerator = Accelerator(
             step_scheduler_with_optimizer=False,
             mixed_precision=mixed_precision,
-            kwargs_handlers=[ddp_kwargs],
+            kwargs_handlers=[ddp_kwargs, ipg_kwargs],
             cpu=force_cpu,
         )
 
@@ -295,6 +298,14 @@ def train(cfg: TrainPipelineConfig, accelerator: "Accelerator | None" = None):
     if not is_main_process:
         dataset, eval_dataset = make_train_eval_datasets(cfg)
 
+    # Create environment used for evaluating checkpoints during training on simulation data.
+    # On real-world data, no need to create an environment as evaluations are done outside train.py,
+    # using the eval.py instead, with gym_dora environment and dora-rs.
+    eval_env = None
+    if cfg.env_eval_freq > 0 and cfg.env is not None and is_main_process:
+        logging.info("Creating env")
+        eval_env = make_env(cfg.env, n_envs=cfg.eval.batch_size, use_async_envs=cfg.eval.use_async_envs)
+
     if cfg.is_reward_model_training:
         if is_main_process:
             logging.info("Creating reward model")
@@ -322,7 +333,7 @@ def train(cfg: TrainPipelineConfig, accelerator: "Accelerator | None" = None):
     if cfg.peft is not None:
         if cfg.is_reward_model_training:
             raise ValueError("PEFT is only supported for policy training. ")
-        require_package("peft", extra="peft")
+        from peft import PeftModel
 
         if isinstance(policy, PeftModel):
             logging.info("PEFT adapter already loaded from checkpoint, skipping wrap_with_peft.")
@@ -336,6 +347,14 @@ def train(cfg: TrainPipelineConfig, accelerator: "Accelerator | None" = None):
 
     active_cfg = cfg.trainable_config
     processor_pretrained_path = active_cfg.pretrained_path
+    # A weight checkpoint may contain PI05 or differently configured PI052 processors.
+    if cfg.policy.type == "pi052" and processor_pretrained_path is not None and not cfg.resume:
+        logging.warning(
+            "pi052 is loading pretrained weights from %s, but building processors from the current "
+            "pi052 config so recipe text labels and FAST action labels are generated.",
+            processor_pretrained_path,
+        )
+        processor_pretrained_path = None
 
     processor_kwargs = {}
     if (processor_pretrained_path and not cfg.resume) or not processor_pretrained_path:
@@ -344,6 +363,13 @@ def train(cfg: TrainPipelineConfig, accelerator: "Accelerator | None" = None):
     if cfg.is_reward_model_training:
         processor_kwargs["dataset_meta"] = dataset.meta
 
+    if cfg.policy.type in {"pi0_fast", "pi052"}:
+        processor_kwargs["dataset_repo_id"] = cfg.dataset.repo_id
+        processor_kwargs["dataset_revision"] = cfg.dataset.revision
+        processor_kwargs["dataset_episodes"] = cfg.dataset.episodes
+        processor_kwargs["dataset_exclude_episodes"] = cfg.dataset.exclude_episodes
+        processor_kwargs["dataset_root"] = cfg.dataset.root
+
     if not cfg.is_reward_model_training and processor_pretrained_path is not None:
         preprocessor_overrides = {
             "device_processor": {"device": device.type},
@@ -440,13 +466,17 @@ def train(cfg: TrainPipelineConfig, accelerator: "Accelerator | None" = None):
         # same permutation. accelerate then shards it disjointly across ranks via BatchSamplerShard
         # without needing a `generator` attribute to synchronize an RNG, and resume is sample-exact.
         shuffle = False
+        from_indices = dataset.meta.episodes["dataset_from_index"]
+        to_indices = dataset.meta.episodes["dataset_to_index"]
+        seed = cfg.seed if cfg.seed is not None else 0
+
         sampler = EpisodeAwareSampler(
-            dataset.meta.episodes["dataset_from_index"],
+            from_indices,
-            dataset.meta.episodes["dataset_to_index"],
+            to_indices,
             episode_indices_to_use=dataset.episodes,
             drop_n_last_frames=getattr(active_cfg, "drop_n_last_frames", 0),
             shuffle=True,
-            seed=cfg.seed if cfg.seed is not None else 0,
+            seed=seed,
             absolute_to_relative_idx=dataset.absolute_to_relative_idx,
         )
         if cfg.resume and step > 0:
@@ -593,7 +623,10 @@ def train(cfg: TrainPipelineConfig, accelerator: "Accelerator | None" = None):
         batch = preprocessor(batch)
         train_tracker.dataloading_s = time.perf_counter() - start_time
 
-        train_tracker, _ = update_policy(
+        # Synchronize GPU metrics only for updates that will be logged.
+        log_metrics = cfg.log_freq > 0 and (step + 1) % cfg.log_freq == 0
+
+        train_tracker, output_dict = update_policy(
             train_tracker,
             policy,
             batch,
@@ -602,6 +635,7 @@ def train(cfg: TrainPipelineConfig, accelerator: "Accelerator | None" = None):
             accelerator=accelerator,
             lr_scheduler=lr_scheduler,
             sample_weighter=sample_weighter,
+            log_metrics=log_metrics,
         )
 
         # Note: eval and checkpoint happens *after* the `step`th training update has completed, so we
@@ -702,10 +736,11 @@ def train(cfg: TrainPipelineConfig, accelerator: "Accelerator | None" = None):
             if is_main_process:
                 step_id = get_step_identifier(step, cfg.steps)
                 logging.info(f"Eval policy at step {step}")
-                with _make_eval_envs(cfg) as eval_env, torch.no_grad(), accelerator.autocast():
+                eval_target_policy = accelerator.unwrap_model(policy)
+                with torch.no_grad(), accelerator.autocast():
                     eval_info = eval_policy_all(
                         envs=eval_env,  # dict[suite][task_id] -> vec_env
-                        policy=accelerator.unwrap_model(policy),
+                        policy=eval_target_policy,
                         env_preprocessor=env_preprocessor,
                         env_postprocessor=env_postprocessor,
                         preprocessor=preprocessor,
@@ -750,6 +785,9 @@ def train(cfg: TrainPipelineConfig, accelerator: "Accelerator | None" = None):
     if is_main_process:
         progbar.close()
 
+    if eval_env:
+        close_envs(eval_env)
+
     is_fsdp = accelerator.distributed_type == DistributedType.FSDP
     model_state_dict = accelerator.get_state_dict(policy) if is_fsdp else None
     if is_main_process:

src/lerobot/scripts/lerobot_train_tokenizer.py

@@ -45,7 +45,6 @@ lerobot-train-tokenizer \
 """
 
 import json
-import logging
 from dataclasses import dataclass
 from pathlib import Path
 from typing import TYPE_CHECKING
@@ -64,9 +63,6 @@ else:
 from lerobot.configs import NormalizationMode, parser
 from lerobot.datasets import LeRobotDataset
 from lerobot.utils.constants import ACTION, OBS_STATE
-from lerobot.utils.utils import init_logging
-
-logger = logging.getLogger(__name__)
 
 
 @dataclass
@@ -278,8 +274,11 @@ def process_episode(args):
 
         return action_chunks
 
-    except Exception:
+    except Exception as e:
-        logger.exception("Error processing episode %s", ep_idx)
+        print(f"Error processing episode {ep_idx}: {e}")
+        import traceback
+
+        traceback.print_exc()
         return None
 
 
@@ -301,10 +300,10 @@ def train_fast_tokenizer(
     Returns:
         Trained FAST tokenizer
     """
-    logger.info(f"Training FAST tokenizer on {len(action_chunks)} action chunks...")
+    print(f"Training FAST tokenizer on {len(action_chunks)} action chunks...")
-    logger.info(f"Action chunk shape: {action_chunks.shape}")
+    print(f"Action chunk shape: {action_chunks.shape}")
-    logger.info(f"Vocab size: {vocab_size}")
+    print(f"Vocab size: {vocab_size}")
-    logger.info(f"DCT scale: {scale}")
+    print(f"DCT scale: {scale}")
 
     # download the tokenizer source code (not pretrained weights)
     # we'll train a new tokenizer on our own data
@@ -315,7 +314,7 @@ def train_fast_tokenizer(
 
     # train the new tokenizer on our action data using .fit()
     # this trains the BPE tokenizer on DCT coefficients
-    logger.info("Training new tokenizer (this may take a few minutes)...")
+    print("Training new tokenizer (this may take a few minutes)...")
     tokenizer = base_tokenizer.fit(
         action_data_list,
         scale=scale,
@@ -323,21 +322,21 @@ def train_fast_tokenizer(
         time_horizon=action_chunks.shape[1],  # action_horizon
         action_dim=action_chunks.shape[2],  # encoded dimensions
     )
-    logger.info("✓ Tokenizer training complete!")
+    print("✓ Tokenizer training complete!")
 
     # validate it works
     sample_chunk = action_chunks[0]
     encoded = tokenizer(sample_chunk[None])[0]
     if isinstance(encoded, list):
         encoded = np.array(encoded)
-    logger.info(f"Sample encoding: {len(encoded)} tokens for chunk shape {sample_chunk.shape}")
+    print(f"Sample encoding: {len(encoded)} tokens for chunk shape {sample_chunk.shape}")
 
     return tokenizer
 
 
 def compute_compression_stats(tokenizer, action_chunks: np.ndarray):
     """Compute compression statistics."""
-    logger.info("\nComputing compression statistics...")
+    print("\nComputing compression statistics...")
 
     # sample for stats (use max 1000 chunks for speed)
     sample_size = min(1000, len(action_chunks))
@@ -367,12 +366,12 @@ def compute_compression_stats(tokenizer, action_chunks: np.ndarray):
         "max_token_length": float(np.max(token_lengths)),
     }
 
-    logger.info("Compression Statistics:")
+    print("Compression Statistics:")
-    logger.info(f"  Average compression ratio: {stats['compression_ratio']:.2f}x")
+    print(f"  Average compression ratio: {stats['compression_ratio']:.2f}x")
-    logger.info(f"  Mean token length: {stats['mean_token_length']:.1f}")
+    print(f"  Mean token length: {stats['mean_token_length']:.1f}")
-    logger.info(f"  P99 token length: {stats['p99_token_length']:.0f}")
+    print(f"  P99 token length: {stats['p99_token_length']:.0f}")
-    logger.info(f"  Min token length: {stats['min_token_length']:.0f}")
+    print(f"  Min token length: {stats['min_token_length']:.0f}")
-    logger.info(f"  Max token length: {stats['max_token_length']:.0f}")
+    print(f"  Max token length: {stats['max_token_length']:.0f}")
 
     return stats
 
@@ -386,9 +385,9 @@ def train_tokenizer(cfg: TokenizerTrainingConfig):
         cfg: TokenizerTrainingConfig dataclass with all configuration parameters
     """
     # load dataset
-    logger.info(f"Loading dataset: {cfg.repo_id}")
+    print(f"Loading dataset: {cfg.repo_id}")
     dataset = LeRobotDataset(repo_id=cfg.repo_id, root=cfg.root)
-    logger.info(f"Dataset loaded: {dataset.num_episodes} episodes, {dataset.num_frames} frames")
+    print(f"Dataset loaded: {dataset.num_episodes} episodes, {dataset.num_frames} frames")
 
     # parse normalization mode
     try:
@@ -398,7 +397,7 @@ def train_tokenizer(cfg: TokenizerTrainingConfig):
             f"Invalid normalization_mode: {cfg.normalization_mode}. "
             f"Must be one of: {', '.join([m.value for m in NormalizationMode])}"
         ) from err
-    logger.info(f"Normalization mode: {norm_mode.value}")
+    print(f"Normalization mode: {norm_mode.value}")
 
     # parse encoded dimensions
     encoded_dim_ranges = []
@@ -407,38 +406,38 @@ def train_tokenizer(cfg: TokenizerTrainingConfig):
         encoded_dim_ranges.append((start, end))
 
     total_encoded_dims = sum(end - start for start, end in encoded_dim_ranges)
-    logger.info(f"Encoding {total_encoded_dims} dimensions: {cfg.encoded_dims}")
+    print(f"Encoding {total_encoded_dims} dimensions: {cfg.encoded_dims}")
 
     # parse relative dimensions
     relative_dim_list = None
     if cfg.relative_dims is not None and cfg.relative_dims.strip():
         relative_dim_list = [int(d.strip()) for d in cfg.relative_dims.split(",")]
-        logger.info(f"Relative dimensions: {relative_dim_list}")
+        print(f"Relative dimensions: {relative_dim_list}")
     else:
-        logger.info("No relative dimensions specified")
+        print("No relative dimensions specified")
 
-    logger.info(f"Use relative transform: {cfg.use_relative_transform}")
+    print(f"Use relative transform: {cfg.use_relative_transform}")
     if cfg.use_relative_transform and (relative_dim_list is None or len(relative_dim_list) == 0):
-        logger.warning(
+        print(
             "Warning: use_relative_transform=True but no relative_dims specified. "
             "No relative transform will be applied."
         )
 
-    logger.info(f"Action horizon: {cfg.action_horizon}")
+    print(f"Action horizon: {cfg.action_horizon}")
-    logger.info(f"State key: {cfg.state_key}")
+    print(f"State key: {cfg.state_key}")
 
     # determine episodes to process
     num_episodes = dataset.num_episodes
     if cfg.max_episodes is not None:
         num_episodes = min(cfg.max_episodes, num_episodes)
 
-    logger.info(f"Processing {num_episodes} episodes...")
+    print(f"Processing {num_episodes} episodes...")
 
     # process episodes sequentially (to avoid pickling issues with dataset)
     all_chunks = []
     for ep_idx in range(num_episodes):
         if ep_idx % 10 == 0:
-            logger.info(f"  Processing episode {ep_idx}/{num_episodes}...")
+            print(f"  Processing episode {ep_idx}/{num_episodes}...")
 
         chunks = process_episode(
             (
@@ -456,19 +455,19 @@ def train_tokenizer(cfg: TokenizerTrainingConfig):
 
     # concatenate all chunks
     all_chunks = np.concatenate(all_chunks, axis=0)
-    logger.info(f"Collected {len(all_chunks)} action chunks")
+    print(f"Collected {len(all_chunks)} action chunks")
 
     # extract only encoded dimensions FIRST (before normalization)
     encoded_chunks = []
     for start, end in encoded_dim_ranges:
         encoded_chunks.append(all_chunks[:, :, start:end])
     encoded_chunks = np.concatenate(encoded_chunks, axis=-1)  # [N, H, D_encoded]
-    logger.info(f"Extracted {encoded_chunks.shape[-1]} encoded dimensions")
+    print(f"Extracted {encoded_chunks.shape[-1]} encoded dimensions")
 
     # apply normalization to encoded dimensions
-    logger.info("\nBefore normalization - overall stats:")
+    print("\nBefore normalization - overall stats:")
-    logger.info(f"  Min: {np.min(encoded_chunks):.4f}, Max: {np.max(encoded_chunks):.4f}")
+    print(f"  Min: {np.min(encoded_chunks):.4f}, Max: {np.max(encoded_chunks):.4f}")
-    logger.info(f"  Mean: {np.mean(encoded_chunks):.4f}, Std: {np.std(encoded_chunks):.4f}")
+    print(f"  Mean: {np.mean(encoded_chunks):.4f}, Std: {np.std(encoded_chunks):.4f}")
 
     # get normalization stats from dataset
     norm_stats = dataset.meta.stats
@@ -490,9 +489,9 @@ def train_tokenizer(cfg: TokenizerTrainingConfig):
                     encoded_stats[stat_name] = stat_array[encoded_dim_indices]
 
         if encoded_stats:
-            logger.info(f"\nNormalization stats for encoded dimensions (mode: {norm_mode.value}):")
+            print(f"\nNormalization stats for encoded dimensions (mode: {norm_mode.value}):")
             for stat_name, stat_values in encoded_stats.items():
-                logger.info(
+                print(
                     f"  {stat_name}: shape={stat_values.shape}, "
                     f"range=[{np.min(stat_values):.4f}, {np.max(stat_values):.4f}]"
                 )
@@ -500,27 +499,27 @@ def train_tokenizer(cfg: TokenizerTrainingConfig):
             # apply normalization based on mode
             try:
                 encoded_chunks = apply_normalization(encoded_chunks, encoded_stats, norm_mode, eps=1e-8)
-                logger.info(f"\nApplied {norm_mode.value} normalization")
+                print(f"\nApplied {norm_mode.value} normalization")
             except ValueError as e:
-                logger.warning(f"Warning: {e}. Using raw actions without normalization.")
+                print(f"Warning: {e}. Using raw actions without normalization.")
 
-            logger.info("\nAfter normalization - overall stats:")
+            print("\nAfter normalization - overall stats:")
-            logger.info(f"  Min: {np.min(encoded_chunks):.4f}, Max: {np.max(encoded_chunks):.4f}")
+            print(f"  Min: {np.min(encoded_chunks):.4f}, Max: {np.max(encoded_chunks):.4f}")
-            logger.info(f"  Mean: {np.mean(encoded_chunks):.4f}, Std: {np.std(encoded_chunks):.4f}")
+            print(f"  Mean: {np.mean(encoded_chunks):.4f}, Std: {np.std(encoded_chunks):.4f}")
 
-            logger.info("\nPer-dimension stats (after normalization):")
+            print("\nPer-dimension stats (after normalization):")
             for d in range(encoded_chunks.shape[-1]):
                 dim_data = encoded_chunks[:, :, d]
-                logger.info(
+                print(
                     f"  Dim {d}: min={np.min(dim_data):7.4f}, max={np.max(dim_data):7.4f}, "
                     f"mean={np.mean(dim_data):7.4f}, std={np.std(dim_data):7.4f}"
                 )
         else:
-            logger.warning("Warning: Could not extract stats for encoded dimensions, using raw actions")
+            print("Warning: Could not extract stats for encoded dimensions, using raw actions")
     else:
-        logger.warning("Warning: No normalization stats found in dataset, using raw actions")
+        print("Warning: No normalization stats found in dataset, using raw actions")
 
-    logger.info(f"Encoded chunks shape: {encoded_chunks.shape}")
+    print(f"Encoded chunks shape: {encoded_chunks.shape}")
 
     # train FAST tokenizer
     tokenizer = train_fast_tokenizer(
@@ -562,8 +561,8 @@ def train_tokenizer(cfg: TokenizerTrainingConfig):
     with open(output_path / "metadata.json", "w") as f:
         json.dump(metadata, f, indent=2)
 
-    logger.info(f"\nSaved FAST tokenizer to {output_path}")
+    print(f"\nSaved FAST tokenizer to {output_path}")
-    logger.info(f"Metadata: {json.dumps(metadata, indent=2)}")
+    print(f"Metadata: {json.dumps(metadata, indent=2)}")
 
     # push to Hugging Face Hub if requested
     if cfg.push_to_hub:
@@ -571,10 +570,10 @@ def train_tokenizer(cfg: TokenizerTrainingConfig):
         hub_repo_id = cfg.hub_repo_id
         if hub_repo_id is None:
             hub_repo_id = output_path.name
-            logger.info(f"\nNo hub_repo_id provided, using: {hub_repo_id}")
+            print(f"\nNo hub_repo_id provided, using: {hub_repo_id}")
 
-        logger.info(f"\nPushing tokenizer to Hugging Face Hub: {hub_repo_id}")
+        print(f"\nPushing tokenizer to Hugging Face Hub: {hub_repo_id}")
-        logger.info(f"   Private: {cfg.hub_private}")
+        print(f"   Private: {cfg.hub_private}")
 
         try:
             # use the tokenizer's push_to_hub method
@@ -594,15 +593,14 @@ def train_tokenizer(cfg: TokenizerTrainingConfig):
                 commit_message="Upload tokenizer metadata",
             )
 
-            logger.info(f"Successfully pushed tokenizer to: https://huggingface.co/{hub_repo_id}")
+            print(f"Successfully pushed tokenizer to: https://huggingface.co/{hub_repo_id}")
         except Exception as e:
-            logger.error(f"Error pushing to hub: {e}")
+            print(f"Error pushing to hub: {e}")
-            logger.error("   Make sure you're logged in with `huggingface-cli login`")
+            print("   Make sure you're logged in with `huggingface-cli login`")
 
 
 def main():
     """CLI entry point that parses arguments and runs the tokenizer training."""
-    init_logging()
     train_tokenizer()