Merge branch 'main' into mrussi/glove_visualizer

* chore(docs): update readme + gr00t libero results

* chore(docs): update template and in-tree policy steps

.dockerignore

@@ -22,6 +22,10 @@ outputs
 rl
 media
 
+# Local virtualenvs (the image provides its own)
+.venv
+venv
+
 
 # Logging
 logs

.github/workflows/benchmark_tests.yml

@@ -167,9 +167,9 @@ jobs:
 
       # ── LIBERO TRAIN+EVAL SMOKE ──────────────────────────────────────────────
       # Train SmolVLA for 1 step (batch_size=1, dataset episode 0 only) then
-      # immediately runs eval inside the training loop (eval_freq=1, 1 episode).
+      # immediately runs eval inside the training loop (env_eval_freq=1, 1 episode).
       # Tests the full train→eval-within-training pipeline end-to-end.
-      - name: Run Libero train+eval smoke (1 step, eval_freq=1)
+      - name: Run Libero train+eval smoke (1 step, env_eval_freq=1)
         if: env.HF_USER_TOKEN != ''
         run: |
           docker run --name libero-train-smoke --gpus all \
@@ -196,7 +196,7 @@ jobs:
                 --output_dir=/tmp/train-smoke \
                 --steps=1 \
                 --batch_size=1 \
-                --eval_freq=1 \
+                --env_eval_freq=1 \
                 --eval.n_episodes=1 \
                 --eval.batch_size=1 \
                 --eval.use_async_envs=false \
@@ -382,6 +382,7 @@ jobs:
                 --policy.path=\"\$ROBOTWIN_POLICY\" \
                 --env.type=robotwin \
                 --env.task=\"\$ROBOTWIN_TASKS\" \
+                --env.max_parallel_tasks=5 \
                 --eval.batch_size=1 \
                 --eval.n_episodes=1 \
                 --eval.use_async_envs=false \
@@ -482,6 +483,7 @@ jobs:
                 --policy.path=lerobot/smolvla_robocasa \
                 --env.type=robocasa \
                 --env.task=CloseFridge,OpenCabinet,OpenDrawer,TurnOnMicrowave,TurnOffStove,CloseToasterOvenDoor,SlideDishwasherRack,TurnOnSinkFaucet,NavigateKitchen,TurnOnElectricKettle \
+                --env.max_parallel_tasks=5 \
                 --eval.batch_size=1 \
                 --eval.n_episodes=1 \
                 --eval.use_async_envs=false \
@@ -693,6 +695,7 @@ jobs:
                 --env.task=\"\$ROBOMME_TASKS\" \
                 --env.dataset_split=test \
                 --env.task_ids=[0] \
+                --env.max_parallel_tasks=5 \
                 --eval.batch_size=1 \
                 --eval.n_episodes=1 \
                 --eval.use_async_envs=false \
@@ -800,6 +803,7 @@ jobs:
                 --env.type=libero_plus \
                 --env.task=\"\$LIBERO_PLUS_SUITE\" \
                 --env.task_ids=\"\$LIBERO_PLUS_TASK_IDS\" \
+                --env.max_parallel_tasks=5 \
                 --eval.batch_size=1 \
                 --eval.n_episodes=1 \
                 --eval.use_async_envs=false \
@@ -900,6 +904,8 @@ jobs:
                 --policy.path=lerobot/smolvla_vlabench \
                 --env.type=vlabench \
                 --env.task=select_fruit,select_toy,select_book,select_painting,select_drink,select_ingredient,select_billiards,select_poker,add_condiment,insert_flower \
+                --env.episode_length=50 \
+                --env.max_parallel_tasks=5 \
                 --eval.batch_size=1 \
                 --eval.n_episodes=1 \
                 --eval.use_async_envs=false \

.github/workflows/release.yml

@@ -152,13 +152,14 @@ jobs:
             BASE_VERSION="${VERSION%%-*}"
             echo "Installing pre-release version $BASE_VERSION from TestPyPI..."
             uv pip install \
+              --torch-backend cpu \
               --index-url https://test.pypi.org/simple/ \
               --extra-index-url https://pypi.org/simple \
               --index-strategy unsafe-best-match \
                "lerobot[all]==$BASE_VERSION"
           else
             echo "Installing release version $VERSION from PyPI..."
-            uv pip install "lerobot[all]==$VERSION"
+            uv pip install --torch-backend cpu "lerobot[all]==$VERSION"
           fi
       - name: Check lerobot version
         run: uv run python -c "import lerobot; print(lerobot.__version__)"

.github/workflows/stale.yml

@@ -19,8 +19,8 @@ on:
   workflow_dispatch:
 
   # Runs at 02:00
-  schedule:
-    - cron: "0 2 * * *"
+  # schedule:
+  #   - cron: "0 2 * * *"
 
 env:
   CLOSE_ISSUE_MESSAGE: >

.pre-commit-config.yaml

@@ -65,6 +65,9 @@ repos:
         name: Format Markdown with Prettier
         types_or: [markdown, mdx]
         args: [--prose-wrap=preserve]
+        # Jinja2 model-card templates use a .md extension but contain {% ... %} /
+        # {{ ... }} tags that prettier's Markdown formatter mangles (e.g. table loops).
+        exclude: ^src/lerobot/templates/.*\.md$
 
   ##### Security #####
   - repo: https://github.com/gitleaks/gitleaks

AGENT_GUIDE.md

@@ -138,7 +138,7 @@ lerobot-replay --robot.type=so101_follower --robot.port=<FOLLOWER_PORT> --robot.
   --dataset.repo_id=${HF_USER}/my_task --dataset.episode=0
 ```
 
-**4.9 Train** (default: ACT — fastest, lowest memory). Apple silicon: `--policy.device=mps`. See §6/§7 for policy and duration.
+**4.9 Train** (default: ACT — fastest, lowest memory). Apple silicon: `--policy.device=mps`. No local GPU? Add `--job.target=<flavor>` (e.g. `a10g-small`, list them with `hf jobs hardware`) to run on Hugging Face Jobs instead. See §6/§7 for policy and duration.
 
 ```bash
 lerobot-train \
@@ -232,6 +232,8 @@ Match the policy to the user's **GPU memory** and **time budget**. Numbers below
 
 All policies typically train for **5–10 epochs** (see §7).
 
+> **Human-facing version:** the [Compute Hardware Guide](./docs/source/hardware_guide.mdx) reuses the table below and adds a cloud-GPU tier guide and a Hugging Face Jobs pointer.
+
 | Policy      | Batch | Update (ms) | Peak GPU mem (GB) | Best for                                                                                         |
 | ----------- | ----: | ----------: | ----------------: | ------------------------------------------------------------------------------------------------ |
 | `act`       |     4 |    **83.9** |          **0.94** | First-time users, laptops, single-task. Fast and reliable.                                       |

Makefile

@@ -58,7 +58,7 @@ test-act-ete-train:
 		--dataset.episodes="[0]" \
 		--batch_size=2 \
 		--steps=4 \
-		--eval_freq=2 \
+		--env_eval_freq=2 \
 		--eval.n_episodes=1 \
 		--eval.batch_size=1 \
 		--save_freq=2 \
@@ -96,7 +96,7 @@ test-diffusion-ete-train:
 		--dataset.episodes="[0]" \
 		--batch_size=2 \
 		--steps=2 \
-		--eval_freq=2 \
+		--env_eval_freq=2 \
 		--eval.n_episodes=1 \
 		--eval.batch_size=1 \
 		--save_checkpoint=true \
@@ -126,7 +126,7 @@ test-tdmpc-ete-train:
 		--dataset.episodes="[0]" \
 		--batch_size=2 \
 		--steps=2 \
-		--eval_freq=2 \
+		--env_eval_freq=2 \
 		--eval.n_episodes=1 \
 		--eval.batch_size=1 \
 		--save_checkpoint=true \
@@ -161,7 +161,7 @@ test-smolvla-ete-train:
 		--dataset.episodes="[0]" \
 		--batch_size=2 \
 		--steps=4 \
-		--eval_freq=2 \
+		--env_eval_freq=2 \
 		--eval.n_episodes=1 \
 		--eval.batch_size=1 \
 		--save_freq=2 \
@@ -178,3 +178,9 @@ test-smolvla-ete-eval:
 		--env.episode_length=5 \
 		--eval.n_episodes=1 \
 		--eval.batch_size=1
+
+# E2E annotation pipeline smoke test against a tiny in-memory fixture
+# dataset. Opt-in (not part of `make test-end-to-end`) and uses a stub VLM
+# backend, so it does not require a real model checkpoint or GPU.
+annotation-e2e:
+	uv run python -m tests.annotations.run_e2e_smoke

README.md

@@ -58,7 +58,7 @@ action = model.select_action(obs)
 robot.send_action(action)
 ```
 
-**Supported Hardware:** SO100, LeKiwi, Koch, HopeJR, OMX, EarthRover, Reachy2, Gamepads, Keyboards, Phones, OpenARM, Unitree G1.
+**Supported Hardware:** SO100, LeKiwi, Koch, HopeJR, OMX, EarthRover, Reachy2, Gamepads, Keyboards, Phones, OpenARM, Unitree G1, reBot B601.
 
 While these devices are natively integrated into the LeRobot codebase, the library is designed to be extensible. You can easily implement the Robot interface to utilize LeRobot's data collection, training, and visualization tools for your own custom robot.
 
@@ -87,7 +87,7 @@ Learn more about it in the [LeRobotDataset Documentation](https://huggingface.co
 
 ## SoTA Models
 
-LeRobot implements state-of-the-art policies in pure PyTorch, covering Imitation Learning, Reinforcement Learning, and Vision-Language-Action (VLA) models, with more coming soon. It also provides you with the tools to instrument and inspect your training process.
+LeRobot implements state-of-the-art policies in pure PyTorch, covering Imitation Learning, Reinforcement Learning, Vision-Language-Action (VLA) models, World Models, and Reward Models, with more coming soon. It also provides you with the tools to instrument and inspect your training process.
 
 <p align="center">
   <img alt="Gr00t Architecture" src="./media/readme/VLA_architecture.jpg" width="640px">
@@ -97,19 +97,21 @@ Training a policy is as simple as running a script configuration:
 
 ```bash
 lerobot-train \
-  --policy=act \
+  --policy.type=act \
   --dataset.repo_id=lerobot/aloha_mobile_cabinet
 ```
 
-| 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) |
-| **Reinforcement Learning** | [HIL-SERL](./docs/source/hilserl.mdx), [TDMPC](./docs/source/policy_tdmpc_README.md) & QC-FQL (coming soon)                                                                                                             |
-| **VLAs Models**            | [Pi0Fast](./docs/source/pi0fast.mdx), [Pi0.5](./docs/source/pi05.mdx), [GR00T N1.5](./docs/source/policy_groot_README.md), [SmolVLA](./docs/source/policy_smolvla_README.md), [XVLA](./docs/source/xvla.mdx)            |
+| 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)                                                                                                                                                                    |
+| **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) |
+| **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)                                                                                                                                                                                                                                                                         |
 
 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
 
-For detailed policy setup guides, see the [Policy Documentation](https://huggingface.co/docs/lerobot/bring_your_own_policies).
+For detailed policy setup guides, see the [Policy Documentation](https://huggingface.co/docs/lerobot/bring_your_own_policies). For GPU/RAM requirements and expected training time per policy, see the [Compute Hardware Guide](https://huggingface.co/docs/lerobot/hardware_guide).
 
 ## Inference & Evaluation
 
@@ -133,6 +135,8 @@ Learn how to implement your own simulation environment or benchmark and distribu
 - **[Discord](https://discord.gg/q8Dzzpym3f):** Join the `LeRobot` server to discuss with the community.
 - **[X](https://x.com/LeRobotHF):** Follow us on X to stay up-to-date with the latest developments.
 - **[Robot Learning Tutorial](https://huggingface.co/spaces/lerobot/robot-learning-tutorial):** A free, hands-on course to learn robot learning using LeRobot.
+- **[T-Shirt Folding Experiment](https://huggingface.co/spaces/lerobot/robot-folding):** An end-to-end demonstration of folding t-shirts with LeRobot.
+- **[LeLab](https://github.com/huggingface/leLab):** A web interface for LeRobot — teleoperate, calibrate, record datasets, replay, and train your SO arm from the browser, no CLI required.
 
 ## Citation
 
@@ -140,7 +144,7 @@ If you use LeRobot in your project, please cite the GitHub repository to acknowl
 
 ```bibtex
 @misc{cadene2024lerobot,
-    author = {Cadene, Remi and Alibert, Simon and Soare, Alexander and Gallouedec, Quentin and Zouitine, Adil and Palma, Steven and Kooijmans, Pepijn and Aractingi, Michel and Shukor, Mustafa and Aubakirova, Dana and Russi, Martino and Capuano, Francesco and Pascal, Caroline and Choghari, Jade and Moss, Jess and Wolf, Thomas},
+    author = {Cadene, Remi and Alibert, Simon and Soare, Alexander and Gallouedec, Quentin and Zouitine, Adil and Palma, Steven and Kooijmans, Pepijn and Aractingi, Michel and Shukor, Mustafa and Aubakirova, Dana and Russi, Martino and Capuano, Francesco and Pascal, Caroline and Choghari, Jade and Meftah, Khalil and Ellerbach, Maxime and Moss, Jess and Wolf, Thomas},
     title = {LeRobot: State-of-the-art Machine Learning for Real-World Robotics in Pytorch},
     howpublished = "\url{https://github.com/huggingface/lerobot}",
     year = {2024}

benchmarks/video/README.md

@@ -1,288 +0,0 @@
-# Video benchmark
-
-## Questions
-
-What is the optimal trade-off between:
-
-- maximizing loading time with random access,
-- minimizing memory space on disk,
-- maximizing success rate of policies,
-- compatibility across devices/platforms for decoding videos (e.g. video players, web browsers).
-
-How to encode videos?
-
-- Which video codec (`-vcodec`) to use? h264, h265, AV1?
-- What pixel format to use (`-pix_fmt`)? `yuv444p` or `yuv420p`?
-- How much compression (`-crf`)? No compression with `0`, intermediate compression with `25` or extreme with `50+`?
-- Which frequency to chose for key frames (`-g`)? A key frame every `10` frames?
-
-How to decode videos?
-
-- Which `decoder`? `torchvision`, `torchaudio`, `ffmpegio`, `decord`, or `nvc`?
-- What scenarios to use for the requesting timestamps during benchmark? (`timestamps_mode`)
-
-## Variables
-
-**Image content & size**
-We don't expect the same optimal settings for a dataset of images from a simulation, or from real-world in an apartment, or in a factory, or outdoor, or with lots of moving objects in the scene, etc. Similarly, loading times might not vary linearly with the image size (resolution).
-For these reasons, we run this benchmark on four representative datasets:
-
-- `lerobot/pusht_image`: (96 x 96 pixels) simulation with simple geometric shapes, fixed camera.
-- `lerobot/aloha_mobile_shrimp_image`: (480 x 640 pixels) real-world indoor, moving camera.
-- `lerobot/paris_street`: (720 x 1280 pixels) real-world outdoor, moving camera.
-- `lerobot/kitchen`: (1080 x 1920 pixels) real-world indoor, fixed camera.
-
-Note: The datasets used for this benchmark need to be image datasets, not video datasets.
-
-**Data augmentations**
-We might revisit this benchmark and find better settings if we train our policies with various data augmentations to make them more robust (e.g. robust to color changes, compression, etc.).
-
-### Encoding parameters
-
-| parameter   | values                                                       |
-| ----------- | ------------------------------------------------------------ |
-| **vcodec**  | `libx264`, `libx265`, `libsvtav1`                            |
-| **pix_fmt** | `yuv444p`, `yuv420p`                                         |
-| **g**       | `1`, `2`, `3`, `4`, `5`, `6`, `10`, `15`, `20`, `40`, `None` |
-| **crf**     | `0`, `5`, `10`, `15`, `20`, `25`, `30`, `40`, `50`, `None`   |
-
-Note that `crf` value might be interpreted differently by various video codecs. In other words, the same value used with one codec doesn't necessarily translate into the same compression level with another codec. In fact, the default value (`None`) isn't the same amongst the different video codecs. Importantly, it is also the case for many other ffmpeg arguments like `g` which specifies the frequency of the key frames.
-
-For a comprehensive list and documentation of these parameters, see the ffmpeg documentation depending on the video codec used:
-
-- h264: https://trac.ffmpeg.org/wiki/Encode/H.264
-- h265: https://trac.ffmpeg.org/wiki/Encode/H.265
-- AV1: https://trac.ffmpeg.org/wiki/Encode/AV1
-
-### Decoding parameters
-
-**Decoder**
-We tested two video decoding backends from torchvision:
-
-- `pyav`
-- `video_reader` (requires to build torchvision from source)
-
-**Requested timestamps**
-Given the way video decoding works, once a keyframe has been loaded, the decoding of subsequent frames is fast.
-This of course is affected by the `-g` parameter during encoding, which specifies the frequency of the keyframes. Given our typical use cases in robotics policies which might request a few timestamps in different random places, we want to replicate these use cases with the following scenarios:
-
-- `1_frame`: 1 frame,
-- `2_frames`: 2 consecutive frames (e.g. `[t, t + 1 / fps]`),
-- `6_frames`: 6 consecutive frames (e.g. `[t + i / fps for i in range(6)]`)
-
-Note that this differs significantly from a typical use case like watching a movie, in which every frame is loaded sequentially from the beginning to the end and it's acceptable to have big values for `-g`.
-
-Additionally, because some policies might request single timestamps that are a few frames apart, we also have the following scenario:
-
-- `2_frames_4_space`: 2 frames with 4 consecutive frames of spacing in between (e.g `[t, t + 5 / fps]`),
-
-However, due to how video decoding is implemented with `pyav`, we don't have access to an accurate seek so in practice this scenario is essentially the same as `6_frames` since all 6 frames between `t` and `t + 5 / fps` will be decoded.
-
-## Metrics
-
-**Data compression ratio (lower is better)**
-`video_images_size_ratio` is the ratio of the memory space on disk taken by the encoded video over the memory space taken by the original images. For instance, `video_images_size_ratio=25%` means that the video takes 4 times less memory space on disk compared to the original images.
-
-**Loading time ratio (lower is better)**
-`video_images_load_time_ratio` is the ratio of the time it takes to decode frames from the video at a given timestamps over the time it takes to load the exact same original images. Lower is better. For instance, `video_images_load_time_ratio=200%` means that decoding from video is 2 times slower than loading the original images.
-
-**Average Mean Square Error (lower is better)**
-`avg_mse` is the average mean square error between each decoded frame and its corresponding original image over all requested timestamps, and also divided by the number of pixels in the image to be comparable when switching to different image sizes.
-
-**Average Peak Signal to Noise Ratio (higher is better)**
-`avg_psnr` measures the ratio between the maximum possible power of a signal and the power of corrupting noise that affects the fidelity of its representation. Higher PSNR indicates better quality.
-
-**Average Structural Similarity Index Measure (higher is better)**
-`avg_ssim` evaluates the perceived quality of images by comparing luminance, contrast, and structure. SSIM values range from -1 to 1, where 1 indicates perfect similarity.
-
-One aspect that can't be measured here with those metrics is the compatibility of the encoding across platforms, in particular on web browser, for visualization purposes.
-h264, h265 and AV1 are all commonly used codecs and should not pose an issue. However, the chroma subsampling (`pix_fmt`) format might affect compatibility:
-
-- `yuv420p` is more widely supported across various platforms, including web browsers.
-- `yuv444p` offers higher color fidelity but might not be supported as broadly.
-
-<!-- **Loss of a pretrained policy (higher is better)** (not available)
-`loss_pretrained` is the result of evaluating with the selected encoding/decoding settings a policy pretrained on original images. It is easier to understand than `avg_l2_error`.
-
-**Success rate after retraining (higher is better)** (not available)
-`success_rate` is the result of training and evaluating a policy with the selected encoding/decoding settings. It is the most difficult metric to get but also the very best. -->
-
-## How the benchmark works
-
-The benchmark evaluates both encoding and decoding of video frames on the first episode of each dataset.
-
-**Encoding:** for each `vcodec` and `pix_fmt` pair, we use a default value for `g` and `crf` upon which we change a single value (either `g` or `crf`) to one of the specified values (we don't test every combination of those as this would be computationally too heavy).
-This gives a unique set of encoding parameters which is used to encode the episode.
-
-**Decoding:** Then, for each of those unique encodings, we iterate through every combination of the decoding parameters `backend` and `timestamps_mode`. For each of them, we record the metrics of a number of samples (given by `--num-samples`). This is parallelized for efficiency and the number of processes can be controlled with `--num-workers`. Ideally, it's best to have a `--num-samples` that is divisible by `--num-workers`.
-
-Intermediate results saved for each `vcodec` and `pix_fmt` combination in csv tables.
-These are then all concatenated to a single table ready for analysis.
-
-## Caveats
-
-We tried to measure the most impactful parameters for both encoding and decoding. However, for computational reasons we can't test out every combination.
-
-Additional encoding parameters exist that are not included in this benchmark. In particular:
-
-- `-preset` which allows for selecting encoding presets. This represents a collection of options that will provide a certain encoding speed to compression ratio. By leaving this parameter unspecified, it is considered to be `medium` for libx264 and libx265 and `8` for libsvtav1.
-- `-tune` which allows to optimize the encoding for certain aspects (e.g. film quality, fast decoding, etc.).
-
-See the documentation mentioned above for more detailed info on these settings and for a more comprehensive list of other parameters.
-
-Similarly on the decoding side, other decoders exist but are not implemented in our current benchmark. To name a few:
-
-- `torchaudio`
-- `ffmpegio`
-- `decord`
-- `nvc`
-
-Note as well that since we are mostly interested in the performance at decoding time (also because encoding is done only once before uploading a dataset), we did not measure encoding times nor have any metrics regarding encoding.
-However, besides the necessity to build ffmpeg from source, encoding did not pose any issue and it didn't take a significant amount of time during this benchmark.
-
-## Install
-
-Building ffmpeg from source is required to include libx265 and libaom/libsvtav1 (av1) video codecs ([compilation guide](https://trac.ffmpeg.org/wiki/CompilationGuide/Ubuntu)).
-
-**Note:** While you still need to build torchvision with a conda-installed `ffmpeg<4.3` to use the `video_reader` decoder (as described in [#220](https://github.com/huggingface/lerobot/pull/220)), you also need another version which is custom-built with all the video codecs for encoding. For the script to then use that version, you can prepend the command above with `PATH="$HOME/bin:$PATH"`, which is where ffmpeg should be built.
-
-## Adding a video decoder
-
-Right now, we're only benchmarking the two video decoder available with torchvision: `pyav` and `video_reader`.
-You can easily add a new decoder to benchmark by adding it to this function in the script:
-
-```diff
-def decode_video_frames(
-    video_path: str,
-    timestamps: list[float],
-    tolerance_s: float,
-    backend: str,
-) -> torch.Tensor:
-    if backend in ["pyav", "video_reader"]:
-        return decode_video_frames_torchvision(
-            video_path, timestamps, tolerance_s, backend
-        )
-+    elif backend == ["your_decoder"]:
-+        return your_decoder_function(
-+            video_path, timestamps, tolerance_s, backend
-+        )
-    else:
-        raise NotImplementedError(backend)
-```
-
-## Example
-
-For a quick run, you can try these parameters:
-
-```bash
-python benchmark/video/run_video_benchmark.py \
-    --output-dir outputs/video_benchmark \
-    --repo-ids \
-        lerobot/pusht_image \
-        lerobot/aloha_mobile_shrimp_image \
-    --vcodec libx264 libx265 \
-    --pix-fmt yuv444p yuv420p \
-    --g 2 20 None \
-    --crf 10 40 None \
-    --timestamps-modes 1_frame 2_frames \
-    --backends pyav video_reader \
-    --num-samples 5 \
-    --num-workers 5 \
-    --save-frames 0
-```
-
-## Results
-
-### Reproduce
-
-We ran the benchmark with the following parameters:
-
-```bash
-# h264 and h265 encodings
-python benchmark/video/run_video_benchmark.py \
-    --output-dir outputs/video_benchmark \
-    --repo-ids \
-        lerobot/pusht_image \
-        lerobot/aloha_mobile_shrimp_image \
-        lerobot/paris_street \
-        lerobot/kitchen \
-    --vcodec libx264 libx265 \
-    --pix-fmt yuv444p yuv420p \
-    --g 1 2 3 4 5 6 10 15 20 40 None \
-    --crf 0 5 10 15 20 25 30 40 50 None \
-    --timestamps-modes 1_frame 2_frames 6_frames \
-    --backends pyav video_reader \
-    --num-samples 50 \
-    --num-workers 5 \
-    --save-frames 1
-
-# av1 encoding (only compatible with yuv420p and pyav decoder)
-python benchmark/video/run_video_benchmark.py \
-    --output-dir outputs/video_benchmark \
-    --repo-ids \
-        lerobot/pusht_image \
-        lerobot/aloha_mobile_shrimp_image \
-        lerobot/paris_street \
-        lerobot/kitchen \
-    --vcodec libsvtav1 \
-    --pix-fmt yuv420p \
-    --g 1 2 3 4 5 6 10 15 20 40 None \
-    --crf 0 5 10 15 20 25 30 40 50 None \
-    --timestamps-modes 1_frame 2_frames 6_frames \
-    --backends pyav \
-    --num-samples 50 \
-    --num-workers 5 \
-    --save-frames 1
-```
-
-The full results are available [here](https://docs.google.com/spreadsheets/d/1OYJB43Qu8fC26k_OyoMFgGBBKfQRCi4BIuYitQnq3sw/edit?usp=sharing)
-
-### Parameters selected for LeRobotDataset
-
-Considering these results, we chose what we think is the best set of encoding parameter:
-
-- vcodec: `libsvtav1`
-- pix-fmt: `yuv420p`
-- g: `2`
-- crf: `30`
-
-Since we're using av1 encoding, we're choosing the `pyav` decoder as `video_reader` does not support it (and `pyav` doesn't require a custom build of `torchvision`).
-
-### Summary
-
-These tables show the results for `g=2` and `crf=30`, using `timestamps-modes=6_frames` and `backend=pyav`
-
-| video_images_size_ratio           | vcodec     | pix_fmt |           |           |           |
-| --------------------------------- | ---------- | ------- | --------- | --------- | --------- |
-|                                   | libx264    |         | libx265   |           | libsvtav1 |
-| repo_id                           | yuv420p    | yuv444p | yuv420p   | yuv444p   | yuv420p   |
-| lerobot/pusht_image               | **16.97%** | 17.58%  | 18.57%    | 18.86%    | 22.06%    |
-| lerobot/aloha_mobile_shrimp_image | 2.14%      | 2.11%   | 1.38%     | **1.37%** | 5.59%     |
-| lerobot/paris_street              | 2.12%      | 2.13%   | **1.54%** | **1.54%** | 4.43%     |
-| lerobot/kitchen                   | 1.40%      | 1.39%   | **1.00%** | **1.00%** | 2.52%     |
-
-| video_images_load_time_ratio      | vcodec  | pix_fmt |          |         |           |
-| --------------------------------- | ------- | ------- | -------- | ------- | --------- |
-|                                   | libx264 |         | libx265  |         | libsvtav1 |
-| repo_id                           | yuv420p | yuv444p | yuv420p  | yuv444p | yuv420p   |
-| lerobot/pusht_image               | 6.45    | 5.19    | **1.90** | 2.12    | 2.47      |
-| lerobot/aloha_mobile_shrimp_image | 11.80   | 7.92    | 0.71     | 0.85    | **0.48**  |
-| lerobot/paris_street              | 2.21    | 2.05    | 0.36     | 0.49    | **0.30**  |
-| lerobot/kitchen                   | 1.46    | 1.46    | 0.28     | 0.51    | **0.26**  |
-
-|                                   |          | vcodec   | pix_fmt      |          |           |              |
-| --------------------------------- | -------- | -------- | ------------ | -------- | --------- | ------------ |
-|                                   |          | libx264  |              | libx265  |           | libsvtav1    |
-| repo_id                           | metric   | yuv420p  | yuv444p      | yuv420p  | yuv444p   | yuv420p      |
-| lerobot/pusht_image               | avg_mse  | 2.90E-04 | **2.03E-04** | 3.13E-04 | 2.29E-04  | 2.19E-04     |
-|                                   | avg_psnr | 35.44    | 37.07        | 35.49    | **37.30** | 37.20        |
-|                                   | avg_ssim | 98.28%   | **98.85%**   | 98.31%   | 98.84%    | 98.72%       |
-| lerobot/aloha_mobile_shrimp_image | avg_mse  | 2.76E-04 | 2.59E-04     | 3.17E-04 | 3.06E-04  | **1.30E-04** |
-|                                   | avg_psnr | 35.91    | 36.21        | 35.88    | 36.09     | **40.17**    |
-|                                   | avg_ssim | 95.19%   | 95.18%       | 95.00%   | 95.05%    | **97.73%**   |
-| lerobot/paris_street              | avg_mse  | 6.89E-04 | 6.70E-04     | 4.03E-03 | 4.02E-03  | **3.09E-04** |
-|                                   | avg_psnr | 33.48    | 33.68        | 32.05    | 32.15     | **35.40**    |
-|                                   | avg_ssim | 93.76%   | 93.75%       | 89.46%   | 89.46%    | **95.46%**   |
-| lerobot/kitchen                   | avg_mse  | 2.50E-04 | 2.24E-04     | 4.28E-04 | 4.18E-04  | **1.53E-04** |
-|                                   | avg_psnr | 36.73    | 37.33        | 36.56    | 36.75     | **39.12**    |
-|                                   | avg_ssim | 95.47%   | 95.58%       | 95.52%   | 95.53%    | **96.82%**   |

benchmarks/video/run_video_benchmark.py

@@ -1,488 +0,0 @@
-#!/usr/bin/env python
-
-# Copyright 2024 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.
-"""Assess the performance of video decoding in various configurations.
-
-This script will benchmark different video encoding and decoding parameters.
-See the provided README.md or run `python benchmark/video/run_video_benchmark.py --help` for usage info.
-"""
-
-import argparse
-import datetime as dt
-import itertools
-import random
-import shutil
-from collections import OrderedDict
-from concurrent.futures import ThreadPoolExecutor, as_completed
-from pathlib import Path
-from threading import Lock
-
-import einops
-import numpy as np
-import pandas as pd
-import PIL
-import torch
-from skimage.metrics import mean_squared_error, peak_signal_noise_ratio, structural_similarity
-from tqdm import tqdm
-
-from lerobot.datasets.lerobot_dataset import LeRobotDataset
-from lerobot.datasets.video_utils import (
-    decode_video_frames,
-    encode_video_frames,
-)
-from lerobot.utils.constants import OBS_IMAGE
-from lerobot.utils.utils import TimerManager
-
-BASE_ENCODING = OrderedDict(
-    [
-        ("vcodec", "libx264"),
-        ("pix_fmt", "yuv444p"),
-        ("g", 2),
-        ("crf", None),
-        # TODO(aliberts): Add fastdecode
-        # ("fastdecode", 0),
-    ]
-)
-
-
-# TODO(rcadene, aliberts): move to `utils.py` folder when we want to refactor
-def parse_int_or_none(value) -> int | None:
-    if value.lower() == "none":
-        return None
-    try:
-        return int(value)
-    except ValueError as e:
-        raise argparse.ArgumentTypeError(f"Invalid int or None: {value}") from e
-
-
-def check_datasets_formats(repo_ids: list) -> None:
-    for repo_id in repo_ids:
-        dataset = LeRobotDataset(repo_id)
-        if len(dataset.meta.video_keys) > 0:
-            raise ValueError(
-                f"Use only image dataset for running this benchmark. Video dataset provided: {repo_id}"
-            )
-
-
-def get_directory_size(directory: Path) -> int:
-    total_size = 0
-    for item in directory.rglob("*"):
-        if item.is_file():
-            total_size += item.stat().st_size
-    return total_size
-
-
-def load_original_frames(imgs_dir: Path, timestamps: list[float], fps: int) -> torch.Tensor:
-    frames = []
-    for ts in timestamps:
-        idx = int(ts * fps)
-        frame = PIL.Image.open(imgs_dir / f"frame-{idx:06d}.png")
-        frame = torch.from_numpy(np.array(frame))
-        frame = frame.type(torch.float32) / 255
-        frame = einops.rearrange(frame, "h w c -> c h w")
-        frames.append(frame)
-    return torch.stack(frames)
-
-
-def save_decoded_frames(
-    imgs_dir: Path, save_dir: Path, frames: torch.Tensor, timestamps: list[float], fps: int
-) -> None:
-    if save_dir.exists() and len(list(save_dir.glob("frame-*.png"))) == len(timestamps):
-        return
-
-    save_dir.mkdir(parents=True, exist_ok=True)
-    for i, ts in enumerate(timestamps):
-        idx = int(ts * fps)
-        frame_hwc = (frames[i].permute((1, 2, 0)) * 255).type(torch.uint8).cpu().numpy()
-        PIL.Image.fromarray(frame_hwc).save(save_dir / f"frame-{idx:06d}_decoded.png")
-        shutil.copyfile(imgs_dir / f"frame-{idx:06d}.png", save_dir / f"frame-{idx:06d}_original.png")
-
-
-def save_first_episode(imgs_dir: Path, dataset: LeRobotDataset) -> None:
-    episode_index = 0
-    ep_num_images = dataset.meta.episodes["length"][episode_index]
-    if imgs_dir.exists() and len(list(imgs_dir.glob("frame-*.png"))) == ep_num_images:
-        return
-
-    imgs_dir.mkdir(parents=True, exist_ok=True)
-    hf_dataset = dataset.hf_dataset.with_format(None)
-
-    # We only save images from the first camera
-    img_keys = [key for key in hf_dataset.features if key.startswith(OBS_IMAGE)]
-    imgs_dataset = hf_dataset.select_columns(img_keys[0])
-
-    for i, item in enumerate(
-        tqdm(imgs_dataset, desc=f"saving {dataset.repo_id} first episode images", leave=False)
-    ):
-        img = item[img_keys[0]]
-        img.save(str(imgs_dir / f"frame-{i:06d}.png"), quality=100)
-
-        if i >= ep_num_images - 1:
-            break
-
-
-def sample_timestamps(timestamps_mode: str, ep_num_images: int, fps: int) -> list[float]:
-    # Start at 5 to allow for 2_frames_4_space and 6_frames
-    idx = random.randint(5, ep_num_images - 1)
-    match timestamps_mode:
-        case "1_frame":
-            frame_indexes = [idx]
-        case "2_frames":
-            frame_indexes = [idx - 1, idx]
-        case "2_frames_4_space":
-            frame_indexes = [idx - 5, idx]
-        case "6_frames":
-            frame_indexes = [idx - i for i in range(6)][::-1]
-        case _:
-            raise ValueError(timestamps_mode)
-
-    return [idx / fps for idx in frame_indexes]
-
-
-def benchmark_decoding(
-    imgs_dir: Path,
-    video_path: Path,
-    timestamps_mode: str,
-    backend: str,
-    ep_num_images: int,
-    fps: int,
-    num_samples: int = 50,
-    num_workers: int = 4,
-    save_frames: bool = False,
-) -> dict:
-    def process_sample(sample: int, lock: Lock):
-        time_benchmark = TimerManager(log=False)
-        timestamps = sample_timestamps(timestamps_mode, ep_num_images, fps)
-        num_frames = len(timestamps)
-        result = {
-            "psnr_values": [],
-            "ssim_values": [],
-            "mse_values": [],
-        }
-
-        with time_benchmark, lock:
-            frames = decode_video_frames(video_path, timestamps=timestamps, tolerance_s=5e-1, backend=backend)
-        result["load_time_video_ms"] = (time_benchmark.last * 1000) / num_frames
-
-        with time_benchmark:
-            original_frames = load_original_frames(imgs_dir, timestamps, fps)
-        result["load_time_images_ms"] = (time_benchmark.last * 1000) / num_frames
-
-        frames_np, original_frames_np = frames.numpy(), original_frames.numpy()
-        for i in range(num_frames):
-            result["mse_values"].append(mean_squared_error(original_frames_np[i], frames_np[i]))
-            result["psnr_values"].append(
-                peak_signal_noise_ratio(original_frames_np[i], frames_np[i], data_range=1.0)
-            )
-            result["ssim_values"].append(
-                structural_similarity(original_frames_np[i], frames_np[i], data_range=1.0, channel_axis=0)
-            )
-
-        if save_frames and sample == 0:
-            save_dir = video_path.with_suffix("") / f"{timestamps_mode}_{backend}"
-            save_decoded_frames(imgs_dir, save_dir, frames, timestamps, fps)
-
-        return result
-
-    load_times_video_ms = []
-    load_times_images_ms = []
-    mse_values = []
-    psnr_values = []
-    ssim_values = []
-
-    # A sample is a single set of decoded frames specified by timestamps_mode (e.g. a single frame, 2 frames, etc.).
-    # For each sample, we record metrics (loading time and quality metrics) which are then averaged over all samples.
-    # As these samples are independent, we run them in parallel threads to speed up the benchmark.
-    # Use a single shared lock for all worker threads
-    shared_lock = Lock()
-    with ThreadPoolExecutor(max_workers=num_workers) as executor:
-        futures = [executor.submit(process_sample, i, shared_lock) for i in range(num_samples)]
-        for future in tqdm(as_completed(futures), total=num_samples, desc="samples", leave=False):
-            result = future.result()
-            load_times_video_ms.append(result["load_time_video_ms"])
-            load_times_images_ms.append(result["load_time_images_ms"])
-            psnr_values.extend(result["psnr_values"])
-            ssim_values.extend(result["ssim_values"])
-            mse_values.extend(result["mse_values"])
-
-    avg_load_time_video_ms = float(np.array(load_times_video_ms).mean())
-    avg_load_time_images_ms = float(np.array(load_times_images_ms).mean())
-    video_images_load_time_ratio = avg_load_time_video_ms / avg_load_time_images_ms
-
-    return {
-        "avg_load_time_video_ms": avg_load_time_video_ms,
-        "avg_load_time_images_ms": avg_load_time_images_ms,
-        "video_images_load_time_ratio": video_images_load_time_ratio,
-        "avg_mse": float(np.mean(mse_values)),
-        "avg_psnr": float(np.mean(psnr_values)),
-        "avg_ssim": float(np.mean(ssim_values)),
-    }
-
-
-def benchmark_encoding_decoding(
-    dataset: LeRobotDataset,
-    video_path: Path,
-    imgs_dir: Path,
-    encoding_cfg: dict,
-    decoding_cfg: dict,
-    num_samples: int,
-    num_workers: int,
-    save_frames: bool,
-    overwrite: bool = False,
-    seed: int = 1337,
-) -> list[dict]:
-    fps = dataset.fps
-
-    if overwrite or not video_path.is_file():
-        tqdm.write(f"encoding {video_path}")
-        encode_video_frames(
-            imgs_dir=imgs_dir,
-            video_path=video_path,
-            fps=fps,
-            vcodec=encoding_cfg["vcodec"],
-            pix_fmt=encoding_cfg["pix_fmt"],
-            g=encoding_cfg.get("g"),
-            crf=encoding_cfg.get("crf"),
-            # fast_decode=encoding_cfg.get("fastdecode"),
-            overwrite=True,
-        )
-
-    episode_index = 0
-    ep_num_images = dataset.meta.episodes["length"][episode_index]
-    width, height = tuple(dataset[0][dataset.meta.camera_keys[0]].shape[-2:])
-    num_pixels = width * height
-    video_size_bytes = video_path.stat().st_size
-    images_size_bytes = get_directory_size(imgs_dir)
-    video_images_size_ratio = video_size_bytes / images_size_bytes
-
-    random.seed(seed)
-    benchmark_table = []
-    for timestamps_mode in tqdm(
-        decoding_cfg["timestamps_modes"], desc="decodings (timestamps_modes)", leave=False
-    ):
-        for backend in tqdm(decoding_cfg["backends"], desc="decodings (backends)", leave=False):
-            benchmark_row = benchmark_decoding(
-                imgs_dir,
-                video_path,
-                timestamps_mode,
-                backend,
-                ep_num_images,
-                fps,
-                num_samples,
-                num_workers,
-                save_frames,
-            )
-            benchmark_row.update(
-                **{
-                    "repo_id": dataset.repo_id,
-                    "resolution": f"{width} x {height}",
-                    "num_pixels": num_pixels,
-                    "video_size_bytes": video_size_bytes,
-                    "images_size_bytes": images_size_bytes,
-                    "video_images_size_ratio": video_images_size_ratio,
-                    "timestamps_mode": timestamps_mode,
-                    "backend": backend,
-                },
-                **encoding_cfg,
-            )
-            benchmark_table.append(benchmark_row)
-
-    return benchmark_table
-
-
-def main(
-    output_dir: Path,
-    repo_ids: list[str],
-    vcodec: list[str],
-    pix_fmt: list[str],
-    g: list[int],
-    crf: list[int],
-    # fastdecode: list[int],
-    timestamps_modes: list[str],
-    backends: list[str],
-    num_samples: int,
-    num_workers: int,
-    save_frames: bool,
-):
-    check_datasets_formats(repo_ids)
-    encoding_benchmarks = {
-        "g": g,
-        "crf": crf,
-        # "fastdecode": fastdecode,
-    }
-    decoding_benchmarks = {
-        "timestamps_modes": timestamps_modes,
-        "backends": backends,
-    }
-    headers = ["repo_id", "resolution", "num_pixels"]
-    headers += list(BASE_ENCODING.keys())
-    headers += [
-        "timestamps_mode",
-        "backend",
-        "video_size_bytes",
-        "images_size_bytes",
-        "video_images_size_ratio",
-        "avg_load_time_video_ms",
-        "avg_load_time_images_ms",
-        "video_images_load_time_ratio",
-        "avg_mse",
-        "avg_psnr",
-        "avg_ssim",
-    ]
-    file_paths = []
-    for video_codec in tqdm(vcodec, desc="encodings (vcodec)"):
-        for pixel_format in tqdm(pix_fmt, desc="encodings (pix_fmt)", leave=False):
-            benchmark_table = []
-            for repo_id in tqdm(repo_ids, desc="encodings (datasets)", leave=False):
-                dataset = LeRobotDataset(repo_id)
-                imgs_dir = output_dir / "images" / dataset.repo_id.replace("/", "_")
-                # We only use the first episode
-                save_first_episode(imgs_dir, dataset)
-                for duet in [
-                    dict(zip(encoding_benchmarks.keys(), unique_combination, strict=False))
-                    for unique_combination in itertools.product(*encoding_benchmarks.values())
-                ]:
-                    encoding_cfg = BASE_ENCODING.copy()
-                    encoding_cfg["vcodec"] = video_codec
-                    encoding_cfg["pix_fmt"] = pixel_format
-                    for key, value in duet.items():
-                        encoding_cfg[key] = value
-                    args_path = Path("_".join(str(value) for value in encoding_cfg.values()))
-                    video_path = output_dir / "videos" / args_path / f"{repo_id.replace('/', '_')}.mp4"
-                    benchmark_table += benchmark_encoding_decoding(
-                        dataset,
-                        video_path,
-                        imgs_dir,
-                        encoding_cfg,
-                        decoding_benchmarks,
-                        num_samples,
-                        num_workers,
-                        save_frames,
-                    )
-
-            # Save intermediate results
-            benchmark_df = pd.DataFrame(benchmark_table, columns=headers)
-            now = dt.datetime.now()
-            csv_path = (
-                output_dir
-                / f"{now:%Y-%m-%d}_{now:%H-%M-%S}_{video_codec}_{pixel_format}_{num_samples}-samples.csv"
-            )
-            benchmark_df.to_csv(csv_path, header=True, index=False)
-            file_paths.append(csv_path)
-            del benchmark_df
-
-    # Concatenate all results
-    df_list = [pd.read_csv(csv_path) for csv_path in file_paths]
-    concatenated_df = pd.concat(df_list, ignore_index=True)
-    concatenated_path = output_dir / f"{now:%Y-%m-%d}_{now:%H-%M-%S}_all_{num_samples}-samples.csv"
-    concatenated_df.to_csv(concatenated_path, header=True, index=False)
-
-
-if __name__ == "__main__":
-    parser = argparse.ArgumentParser()
-    parser.add_argument(
-        "--output-dir",
-        type=Path,
-        default=Path("outputs/video_benchmark"),
-        help="Directory where the video benchmark outputs are written.",
-    )
-    parser.add_argument(
-        "--repo-ids",
-        type=str,
-        nargs="*",
-        default=[
-            "lerobot/pusht_image",
-            "lerobot/aloha_mobile_shrimp_image",
-            "lerobot/paris_street",
-            "lerobot/kitchen",
-        ],
-        help="Datasets repo-ids to test against. First episodes only are used. Must be images.",
-    )
-    parser.add_argument(
-        "--vcodec",
-        type=str,
-        nargs="*",
-        default=["h264", "hevc", "libsvtav1"],
-        help="Video codecs to be tested",
-    )
-    parser.add_argument(
-        "--pix-fmt",
-        type=str,
-        nargs="*",
-        default=["yuv444p", "yuv420p"],
-        help="Pixel formats (chroma subsampling) to be tested",
-    )
-    parser.add_argument(
-        "--g",
-        type=parse_int_or_none,
-        nargs="*",
-        default=[1, 2, 3, 4, 5, 6, 10, 15, 20, 40, 100, None],
-        help="Group of pictures sizes to be tested.",
-    )
-    parser.add_argument(
-        "--crf",
-        type=parse_int_or_none,
-        nargs="*",
-        default=[0, 5, 10, 15, 20, 25, 30, 40, 50, None],
-        help="Constant rate factors to be tested.",
-    )
-    # parser.add_argument(
-    #     "--fastdecode",
-    #     type=int,
-    #     nargs="*",
-    #     default=[0, 1],
-    #     help="Use the fastdecode tuning option. 0 disables it. "
-    #         "For libx264 and libx265/hevc, only 1 is possible. "
-    #         "For libsvtav1, 1, 2 or 3 are possible values with a higher number meaning a faster decoding optimization",
-    # )
-    parser.add_argument(
-        "--timestamps-modes",
-        type=str,
-        nargs="*",
-        default=[
-            "1_frame",
-            "2_frames",
-            "2_frames_4_space",
-            "6_frames",
-        ],
-        help="Timestamps scenarios to be tested.",
-    )
-    parser.add_argument(
-        "--backends",
-        type=str,
-        nargs="*",
-        default=["torchcodec", "pyav"],
-        help="Torchvision decoding backend to be tested.",
-    )
-    parser.add_argument(
-        "--num-samples",
-        type=int,
-        default=50,
-        help="Number of samples for each encoding x decoding config.",
-    )
-    parser.add_argument(
-        "--num-workers",
-        type=int,
-        default=10,
-        help="Number of processes for parallelized sample processing.",
-    )
-    parser.add_argument(
-        "--save-frames",
-        type=int,
-        default=0,
-        help="Whether to save decoded frames or not. Enter a non-zero number for true.",
-    )
-    args = parser.parse_args()
-    main(**vars(args))

docker/Dockerfile.benchmark.robotwin

@@ -35,7 +35,7 @@ USER root
 ARG ROBOTWIN_SHA=0aeea2d669c0f8516f4d5785f0aa33ba812c14b4
 RUN apt-get update \
     && apt-get install -y --no-install-recommends \
-         cuda-nvcc-12-6 cuda-cudart-dev-12-6 \
+         cuda-nvcc-12-8 cuda-cudart-dev-12-8 \
          libvulkan1 vulkan-tools \
     && mkdir -p /usr/share/vulkan/icd.d \
     && echo '{"file_format_version":"1.0.0","ICD":{"library_path":"libGLX_nvidia.so.0","api_version":"1.3.0"}}' \

docker/Dockerfile.internal

@@ -18,7 +18,7 @@
 # docker build -f docker/Dockerfile.internal -t lerobot-internal .
 
 # Configure the base image for CI with GPU access
-ARG CUDA_VERSION=12.6.3
+ARG CUDA_VERSION=12.8.1
 ARG OS_VERSION=24.04
 FROM nvidia/cuda:${CUDA_VERSION}-base-ubuntu${OS_VERSION}
 

docs/source/_toctree.yml

@@ -3,12 +3,16 @@
     title: LeRobot
   - local: installation
     title: Installation
+  - local: cheat-sheet
+    title: Cheat sheet
   title: Get started
 - sections:
   - local: il_robots
     title: Imitation Learning for Robots
+  - local: lelab
+    title: LeLab - Lerobot GUI
   - local: bring_your_own_policies
-    title: Bring Your Own Policies
+    title: Adding a Policy
   - local: integrate_hardware
     title: Bring Your Own Hardware
   - local: hilserl
@@ -24,6 +28,12 @@
   - local: rename_map
     title: Using Rename Map and Empty Cameras
   title: "Tutorials"
+- sections:
+  - local: hardware_guide
+    title: Compute Hardware Guide
+  - local: torch_accelerators
+    title: PyTorch accelerators
+  title: "Compute & Hardware"
 - sections:
   - local: lerobot-dataset-v3
     title: Using LeRobotDataset
@@ -31,8 +41,14 @@
     title: Porting Large Datasets
   - local: using_dataset_tools
     title: Using the Dataset Tools
-  - local: dataset_subtask
-    title: Using Subtasks in the Dataset
+  - local: language_and_recipes
+    title: Language Columns and Recipes
+  - local: tools
+    title: Tools
+  - local: annotation_pipeline
+    title: Annotation Pipeline
+  - local: video_encoding_parameters
+    title: Video encoding parameters
   - local: streaming_video_encoding
     title: Streaming Video Encoding
   title: "Datasets"
@@ -47,10 +63,20 @@
     title: π₀-FAST (Pi0Fast)
   - local: pi05
     title: π₀.₅ (Pi05)
+  - local: molmoact2
+    title: MolmoAct2
+  - local: vla_jepa
+    title: VLA-JEPA
   - local: eo1
     title: EO-1
+  - local: lingbot_va
+    title: LingBot-VA
+  - local: fastwam
+    title: FastWAM
+  - local: evo1
+    title: EVO1
   - local: groot
-    title: NVIDIA GR00T N1.5
+    title: NVIDIA GR00T
   - local: xvla
     title: X-VLA
   - local: multi_task_dit
@@ -61,6 +87,10 @@
 - sections:
   - local: sarm
     title: SARM
+  - local: robometer
+    title: ROBOMETER
+  - local: topreward
+    title: TOPReward
   title: "Reward Models"
 - sections:
   - local: inference
@@ -133,6 +163,8 @@
     title: OMX
   - local: openarm
     title: OpenArm
+  - local: rebot_b601
+    title: reBot B601-DM
   title: "Robots"
 - sections:
   - local: phone_teleop
@@ -142,10 +174,6 @@
   - local: cameras
     title: Cameras
   title: "Sensors"
-- sections:
-  - local: torch_accelerators
-    title: PyTorch accelerators
-  title: "Supported Hardware"
 - sections:
   - local: notebooks
     title: Notebooks

docs/source/act.mdx

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

docs/source/annotation_pipeline.mdx

@@ -0,0 +1,291 @@
+# Annotation Pipeline
+
+`lerobot-annotate` watches each episode's video with a vision-language
+model (VLM) and writes natural-language annotations back into your
+dataset. It fills the two language columns from the
+[Language Columns and Recipes](./language_and_recipes) page —
+`language_persistent` and `language_events` — straight into
+`data/chunk-*/file-*.parquet`.
+
+In short: point it at a LeRobot dataset, and it adds subtasks, plans,
+memory, interjections, speech, and visual Q&A that a policy can be
+trained on.
+
+## How it fits together
+
+```text
+  your dataset                  lerobot-annotate
+  (LeRobot v3.1)
+        │
+        ▼
+  ┌─────────────────────────────────────────────────────┐
+  │                    read episodes                     │
+  └──────────────────────────┬──────────────────────────┘
+                             │
+        ┌────────────────────┼────────────────────┐
+        ▼                    ▼                     ▼
+  ┌──────────┐      ┌───────────────┐        ┌──────────┐       one shared Qwen-VL
+  │   plan   │      │ interjections │        │   vqa    │  ◀──   server (vLLM, OpenAI
+  └────┬─────┘      └───────┬───────┘        └────┬─────┘        API) drives all three
+       └────────────────────┼─────────────────────┘
+                            │   each module stages raw JSONL
+                            ▼   into .annotate_staging/
+                  ┌─────────────────┐
+                  │    validator    │  ◀──  checks everything
+                  └────────┬────────┘
+                           ▼
+                  ┌─────────────────┐
+                  │     writer      │
+                  └────────┬────────┘
+                           ▼
+              data/chunk-*/file-*.parquet
+              (+ meta/info.json tools)
+```
+
+Three modules (`plan`, `interjections`, `vqa`) all talk to **one** shared
+VLM. Each module stages its output to disk, a validator checks it, and a
+single writer rewrites the dataset shards in place.
+
+## What the pipeline produces
+
+Each module emits a few kinds of annotation ("styles"), routed to one of
+the two language columns:
+
+| Style / atom                                | Column                | Module          |
+| ------------------------------------------- | --------------------- | --------------- |
+| `subtask` (Pi0.7-style "how, not what")     | `language_persistent` | `plan`          |
+| `plan` (initial + refresh on interjection)  | `language_persistent` | `plan`          |
+| `memory` (MEM-style compression)            | `language_persistent` | `plan`          |
+| `task_aug` (rephrasings of the task)        | `language_persistent` | `plan`          |
+| `interjection`                              | `language_events`     | `interjections` |
+| speech tool-call atom (`style=null`, `say`) | `language_events`     | `interjections` |
+| `vqa` (user / assistant pair)               | `language_events`     | `vqa`           |
+
+### How subtasks are generated
+
+The `plan` module doesn't ask the VLM for subtasks in one shot. Instead
+it uses a two-step **describe → segment** flow:
+
+1. **Describe** — the VLM narrates only what it actually sees in the
+   chosen camera (no guessing about the task).
+2. **Segment** — that description is fed back in, and the VLM splits the
+   episode into consecutive atomic subtasks.
+
+Both passes see the episode as **timestamped contact sheets** — frames
+sampled at `frames_per_second` (0.5s by default) and packed into JPEG
+grids with each frame's time burned into its corner, so the VLM cites
+exact boundary times directly. This is far cheaper in vision tokens than
+one image per frame, so the sampling can stay dense; episodes longer than
+`max_frames_per_prompt` are split into windows at the same density and
+merged. Both prompts also carry a causal **event-boundary** definition (a
+new event starts when an object becomes held / is released / reaches a new
+location / a lid changes state / contents move) to sharpen where cuts land.
+
+The resulting spans are then stitched into a gap-free, full-episode
+cover, so **every frame has exactly one active subtask**. See
+[`run_hf_job.py`](https://github.com/huggingface/lerobot/blob/main/examples/annotations/run_hf_job.py)
+for the production settings (single camera, timestamped contact sheets,
+auto-windowed subtask generation).
+
+### Tools
+
+The writer does **not** add a `tools` column to the parquet. The tool
+catalog lives in `meta/info.json["tools"]` instead (see [Tools](./tools)).
+After every run, the pipeline makes sure the canonical `say` schema is in
+that list, keeping any tools you declared beforehand.
+
+Want to add your own tool? Edit `meta/info.json["tools"]` directly — the
+pipeline preserves whatever is already there. That makes the tool visible
+to the chat template, so the model can learn to _generate_ the call. The
+runtime layer that actually _executes_ a generated call (the `Tool`
+protocol / `TOOL_REGISTRY` under `src/lerobot/tools/`) is not part of
+this PR — the [Tools](./tools) doc marks those pieces as
+not-yet-implemented.
+
+## Running on Hugging Face Jobs
+
+Annotation runs on [Hugging Face Jobs](https://huggingface.co/docs/hub/en/jobs).
+The repo ships a launcher script you copy and tweak for your dataset:
+
+```bash
+HF_TOKEN=hf_... uv run python examples/annotations/run_hf_job.py
+```
+
+[`run_hf_job.py`](https://github.com/huggingface/lerobot/blob/main/examples/annotations/run_hf_job.py)
+starts a single-GPU `h200` job (bump it to `h200x4` for big datasets)
+that:
+
+1. installs `lerobot` (from `main`) plus the annotation extras,
+2. boots one vLLM server per GPU (using the `vllm/vllm-openai` image) and
+   drives it over the OpenAI-compatible API,
+3. runs the `plan` / `interjections` / `vqa` modules across the dataset
+   with `lerobot-annotate`,
+4. with `--push_to_hub=true`, uploads the result to `--new_repo_id` (or
+   back to `--repo_id` in place if you leave that unset).
+
+To use a different dataset, model, or hub repo, edit the `CMD` block in
+the script. Every flag there maps directly to a `lerobot-annotate` flag
+(run `lerobot-annotate --help` for the full list).
+
+## Key options
+
+These are the flags you'll reach for most often. Run
+`lerobot-annotate --help` for everything else; the defaults are tuned for
+short manipulation episodes.
+
+### Dataset in / out
+
+| Flag              | Default | What it does                                                            |
+| ----------------- | ------- | ----------------------------------------------------------------------- |
+| `--repo_id`       | —       | Hub dataset to annotate (downloaded if `--root` unset).                 |
+| `--root`          | —       | Annotate a local dataset directory instead.                             |
+| `--new_repo_id`   | —       | Push the result to a new repo (leaves the source repo untouched).       |
+| `--push_to_hub`   | `false` | Upload after annotating (to `--new_repo_id`, else back to `--repo_id`). |
+| `--only_episodes` | all     | Annotate just these episode indices (handy for a test run).             |
+| `--seed`          | `1729`  | Seeds the RNGs that pick interjection timestamps + VQA question types.  |
+
+### Which modules run
+
+Every module is on by default and can be toggled independently (set to
+`false` to skip it, e.g. to iterate on one module at a time):
+
+| Flag                      | Default | Turns off                           |
+| ------------------------- | ------- | ----------------------------------- |
+| `--plan.enabled`          | `true`  | subtasks + plan + memory + task_aug |
+| `--interjections.enabled` | `true`  | interjections + speech atoms        |
+| `--vqa.enabled`           | `true`  | the VQA pairs                       |
+
+### The VLM (`--vlm.*`)
+
+| Flag                       | Default            | What it does                                                                        |
+| -------------------------- | ------------------ | ----------------------------------------------------------------------------------- |
+| `--vlm.model_id`           | `Qwen/Qwen3.6-27B` | The model to serve and prompt.                                                      |
+| `--vlm.camera_key`         | first `images.*`   | Which camera every prompt is grounded on.                                           |
+| `--vlm.serve_command`      | auto               | The exact `vllm serve …` command (set TP size, GPU memory, `--max-model-len` here). |
+| `--vlm.parallel_servers`   | `1`                | Independent servers for round-robin routing (one per GPU).                          |
+| `--vlm.num_gpus`           | `0`                | GPUs per server (`0` = one each).                                                   |
+| `--vlm.client_concurrency` | `16`               | In-flight requests across all servers.                                              |
+| `--vlm.max_new_tokens`     | `512`              | Generation cap per call.                                                            |
+| `--vlm.temperature`        | `0.2`              | Sampling temperature.                                                               |
+
+### Subtasks / plan / memory (`--plan.*`)
+
+| Flag                            | Default    | What it does                                                                                                              |
+| ------------------------------- | ---------- | ------------------------------------------------------------------------------------------------------------------------- |
+| `--plan.frames_per_second`      | `2.0`      | Frame sampling rate for the contact sheets (`2.0` = one frame every 0.5s).                                                |
+| `--plan.max_frames_per_prompt`  | `60`       | Frame budget per VLM call. Episodes whose sampling exceeds this are auto-windowed at the same density, then stitched.     |
+| `--plan.contact_sheet_columns`  | `5`        | Columns per contact-sheet grid (`contact_sheet_frames_per_sheet` tiles, time row-major).                                  |
+| `--plan.plan_max_steps`         | `8`        | Upper bound on subtasks per episode.                                                                                      |
+| `--plan.subtask_describe_first` | `true`     | Run the describe→segment grounding pass (best subtask quality; +1 call/episode).                                          |
+| `--plan.emit_plan`              | `true`     | Emit the numbered `plan` rows (`false` = subtasks + memory only).                                                         |
+| `--plan.emit_memory`            | `true`     | Emit the `memory` rows (`false` = subtasks + plan only); symmetric to `emit_plan`.                                        |
+| `--plan.n_task_rephrasings`     | `10`       | How many `task_aug` rephrasings to emit (`0` disables).                                                                   |
+| `--plan.derive_task_from_video` | `if_short` | Use the dataset task as-is (`off`), only when it's missing/short (`if_short`), or always re-derive from video (`always`). |
+
+### Interjections + VQA
+
+| Flag                                            | Default | What it does                                               |
+| ----------------------------------------------- | ------- | ---------------------------------------------------------- |
+| `--interjections.max_interjections_per_episode` | `3`     | Cap on interjection/speech pairs per episode.              |
+| `--vqa.vqa_emission_hz`                         | `1.0`   | How often VQA pairs are emitted.                           |
+| `--vqa.restrict_to_default_camera`              | `false` | Ground VQA only on `--vlm.camera_key` (else every camera). |
+| `--executor.episode_parallelism`                | `16`    | Episodes processed concurrently within each phase.         |
+
+## Contributing new modules
+
+The pipeline is built to grow, and **contributions are very welcome** —
+a brand-new module (say, trajectory traces or affordances), a new prompt
+template, a smarter grounding flow, or quality fixes to the existing
+`plan` / `interjections` / `vqa` modules.
+
+Every module lives under
+`src/lerobot/annotations/steerable_pipeline/modules/`, shares the VLM
+client and the keyframe cache, writes its raw output to the staging
+tree, and plugs into the executor as its own phase. Got an idea? Open an
+issue or PR on [the repo](https://github.com/huggingface/lerobot).
+
+## How recipes consume the output
+
+The annotations are meant to be read by recipes (see
+[Language Columns and Recipes](./language_and_recipes)). Typically:
+
+- low-level / high-level / memory-update branches read
+  `subtask` / `plan` / `memory` from `language_persistent`.
+- an interjection-response branch reads `interjection` events plus the
+  paired speech atom (merged into one assistant turn via `tool_calls_from`)
+  and the matching `plan` refresh at the same timestamp.
+- a VQA branch reads the `(vqa, user)` and `(vqa, assistant)` pairs from
+  `language_events`.
+
+## Why state and events are split
+
+Two ideas shape the design:
+
+1. **Persistent state vs. exact events.** Persistent rows (`subtask`,
+   `plan`, `memory`) apply to the whole episode and answer "what's true
+   right now?". Event rows (`interjection`, `vqa`, speech) appear only on
+   the one frame whose timestamp matches. Timestamps are copied straight
+   from the source parquet — never recomputed in floating point.
+2. **One VLM pass.** All three modules share a single VLM client (the
+   OpenAI-compatible client talking to the job's vLLM server), so you pay
+   for one model load per dataset, not three.
+
+## Re-running a single module
+
+Each module stages its raw output to
+`<root>/.annotate_staging/episode_{N:06d}/<module>.jsonl`. This makes
+prompt iteration cheap: re-running one module overwrites only its own
+JSONL, then the writer recomposes the final parquet. Disable modules you
+don't want with `--plan.enabled=false` (and likewise
+`--interjections.enabled` / `--vqa.enabled`) to test one at a time.
+
+## What the validator checks
+
+Before the writer runs, `StagingValidator` confirms:
+
+- every event row lands exactly on a real frame timestamp;
+- no speech / interjection pairs are left orphaned;
+- `plan` is refreshed at every interjection timestamp;
+- `memory` rows fall on subtask boundaries (a warning, not an error);
+- each VQA assistant `content` is valid JSON in one of the
+  bbox / keypoint / count / attribute / spatial shapes;
+- every row goes to the column chosen by `column_for_style(style)`.
+
+Any error aborts the writer. Pass `--skip_validation=true` to override
+while debugging.
+
+## Where each module's ideas come from
+
+- **`plan` — subtasks.** Hi Robot ([Shi 2025](https://arxiv.org/abs/2502.19417))
+  for atom granularity ("pick up one piece of lettuce", "place bowl to
+  box"); Pi0.7 ([Physical Intelligence 2025](https://pi.website/pi07))
+  for "how, not what" detail.
+- **`plan` — memory.** MEM ([Torne 2026](https://arxiv.org/abs/2603.03596)):
+  keep only the minimal relevant information — preserve outcomes, drop
+  specific attributes.
+- **`interjections`.** Hi Robot's scenario taxonomy: negative task,
+  situated correction, specific constraint, preference. Speech is a
+  tool-call-only atom
+  (`tool_calls=[{type:function, function:{name:"say", arguments:{text:...}}}]`).
+- **`vqa`.** ECoT ([Zawalski 2024](https://arxiv.org/abs/2407.08693)) for
+  grounded features (pixel bounding boxes `[x_min, y_min, x_max, y_max]`,
+  keypoints) and Steerable VLA Policies
+  ([Zhao 2025](https://arxiv.org/abs/2509.07626)) for multi-abstraction
+  grounding. Pi0.7 also grounds answers across abstraction levels.
+
+When improving a module, tweak its prompt template in
+`src/lerobot/annotations/steerable_pipeline/prompts/` rather than
+rewriting from scratch.
+
+## Roughly how much it costs
+
+Per episode, the pipeline makes about `max_steps` plan calls,
+`max_interjections_per_episode` interjection calls, and
+`vqa_emission_hz × episode_seconds` VQA calls. With the defaults (8
+subtasks, 1 interjection, 1 Hz × 3 pairs) on a 30-second episode, that's
+~50 VLM calls.
+
+Storage stays small: `language_persistent` is at most tens of KB per
+episode (parquet dictionary-encodes the one entry that repeats across
+frames), and `language_events` is empty on most frames — its size scales
+with the number of emissions, not `num_frames × num_emissions`.

docs/source/bring_your_own_policies.mdx

@@ -1,60 +1,37 @@
-# Bring Your Own Policies
+# Adding a Policy
 
-This tutorial explains how to integrate your own custom policy implementations into the LeRobot ecosystem, allowing you to leverage all LeRobot tools for training, evaluation, and deployment while using your own algorithms.
+This guide walks you through implementing a custom policy and getting it to work with LeRobot's training, evaluation, and deployment tools. There are two paths:
 
-## Step 1: Create a Policy Package
+- **Plugin (out-of-tree)** — ship your policy as a standalone `lerobot_policy_*` package. Faster, no PR required, easy to iterate. Right for experimentation, internal use, or when you want to publish independently.
+- **In-tree (contributed to LeRobot)** — land your policy directly in `src/lerobot/policies/`. Requires a PR, but makes your policy a first-class citizen of the library.
 
-Your custom policy should be organized as an installable Python package following LeRobot's plugin conventions.
+The plugin route is usually the right starting point — promote to in-tree once the policy has stabilized and there's clear value in shipping it with the library.
 
-### Package Structure
+Either way, the building blocks are the same: a configuration class, a policy class, and a processor factory. The first half of this guide covers those shared pieces; the second half covers the path-specific scaffolding ([Path A](#path-a-out-of-tree-plugin), [Path B](#path-b-contributing-in-tree)).
 
-Create a package with the prefix `lerobot_policy_` (IMPORTANT!) followed by your policy name:
+A note on tone: robot-learning is an actively evolving field, and "what a policy looks like" can shift with each new architecture. The conventions described here exist because they let `lerobot-train` and `lerobot-eval` work uniformly across very different models. When a new policy genuinely doesn't fit them, raise it (in your PR, or an issue) — the conventions are not sacred.
 
-```bash
-lerobot_policy_my_custom_policy/
-├── pyproject.toml
-└── src/
-    └── lerobot_policy_my_custom_policy/
-        ├── __init__.py
-        ├── configuration_my_custom_policy.py
-        ├── modeling_my_custom_policy.py
-        └── processor_my_custom_policy.py
-```
+---
 
-### Package Configuration
+## Anatomy of a policy
 
-Set up your `pyproject.toml`:
+Three building blocks make up every policy. The names below use `my_policy` as a placeholder — replace with your policy's name. That name is load-bearing: it must match the string you pass to `@PreTrainedConfig.register_subclass`, the `MyPolicy.name` class attribute, and the `make_<name>_pre_post_processors` factory function (more on each below).
 
-```toml
-[project]
-name = "lerobot_policy_my_custom_policy"
-version = "0.1.0"
-dependencies = [
-    # your policy-specific dependencies
-]
-requires-python = ">= 3.12"
+### Configuration class
 
-[build-system]
-build-backend = # your-build-backend
-requires = # your-build-system
-```
-
-## Step 2: Define the Policy Configuration
-
-Create a configuration class that inherits from [`PreTrainedConfig`](https://github.com/huggingface/lerobot/blob/main/src/lerobot/configs/policies.py) and registers your policy type:
-Here is a template to get you started, customize the parameters and methods as needed for your policy's architecture and training requirements.
+Inherit from [`PreTrainedConfig`](https://github.com/huggingface/lerobot/blob/main/src/lerobot/configs/policies.py) and register your policy type. Here is a template — customize the parameters and methods as needed for your policy's architecture and training requirements.
 
 ```python
-# configuration_my_custom_policy.py
+# configuration_my_policy.py
 from dataclasses import dataclass, field
 from lerobot.configs import PreTrainedConfig
 from lerobot.optim import AdamWConfig
 from lerobot.optim import CosineDecayWithWarmupSchedulerConfig
 
-@PreTrainedConfig.register_subclass("my_custom_policy")
+@PreTrainedConfig.register_subclass("my_policy")
 @dataclass
-class MyCustomPolicyConfig(PreTrainedConfig):
-    """Configuration class for MyCustomPolicy.
+class MyPolicyConfig(PreTrainedConfig):
+    """Configuration class for MyPolicy.
 
     Args:
         n_obs_steps: Number of observation steps to use as input
@@ -77,16 +54,20 @@ class MyCustomPolicyConfig(PreTrainedConfig):
             raise ValueError("n_action_steps cannot exceed horizon")
 
     def validate_features(self) -> None:
-        """Validate input/output feature compatibility."""
+        """Validate input/output feature compatibility.
+
+        Call this explicitly from your policy's __init__ — the base class does not.
+        """
         if not self.image_features:
-            raise ValueError("MyCustomPolicy requires at least one image feature.")
+            raise ValueError("MyPolicy requires at least one image feature.")
         if self.action_feature is None:
-            raise ValueError("MyCustomPolicy requires 'action' in output_features.")
+            raise ValueError("MyPolicy requires 'action' in output_features.")
 
     def get_optimizer_preset(self) -> AdamWConfig:
         return AdamWConfig(lr=self.optimizer_lr, weight_decay=self.optimizer_weight_decay)
 
     def get_scheduler_preset(self):
+        """Return a LRSchedulerConfig from lerobot.optim, or None."""
         return None
 
     @property
@@ -101,8 +82,7 @@ class MyCustomPolicyConfig(PreTrainedConfig):
 
     @property
     def action_delta_indices(self) -> list[int]:
-        """Relative timestep offsets for the action chunk the dataset loader returns.
-        """
+        """Relative timestep offsets for the action chunk the dataset loader returns."""
         return list(range(self.horizon))
 
     @property
@@ -110,32 +90,34 @@ class MyCustomPolicyConfig(PreTrainedConfig):
         return None
 ```
 
-## Step 3: Implement the Policy Class
+The string you pass to `@register_subclass` must match `MyPolicy.name` (next section) and is what users supply as `--policy.type` on the CLI. Default to `AdamW` from `lerobot.optim` for `get_optimizer_preset` unless you genuinely need otherwise.
 
-Create your policy implementation by inheriting from [`PreTrainedPolicy`](https://github.com/huggingface/lerobot/blob/main/src/lerobot/policies/pretrained.py):
+### Policy class
+
+Inherit from [`PreTrainedPolicy`](https://github.com/huggingface/lerobot/blob/main/src/lerobot/policies/pretrained.py) and set two class attributes — both are checked by `__init_subclass__`:
 
 ```python
-# modeling_my_custom_policy.py
+# modeling_my_policy.py
 import torch
 import torch.nn as nn
 from typing import Any
 
 from lerobot.policies import PreTrainedPolicy
 from lerobot.utils.constants import ACTION
-from .configuration_my_custom_policy import MyCustomPolicyConfig
+from .configuration_my_policy import MyPolicyConfig
 
-class MyCustomPolicy(PreTrainedPolicy):
-    config_class = MyCustomPolicyConfig  # must match the string in @register_subclass
-    name = "my_custom_policy"
+class MyPolicy(PreTrainedPolicy):
+    config_class = MyPolicyConfig  # must match the string in @register_subclass
+    name = "my_policy"
 
-    def __init__(self, config: MyCustomPolicyConfig, dataset_stats: dict[str, Any] = None):
+    def __init__(self, config: MyPolicyConfig, dataset_stats: dict[str, Any] = None):
         super().__init__(config, dataset_stats)
         config.validate_features()  # not called automatically by the base class
         self.config = config
         self.model = ...  # your nn.Module here
 
     def reset(self):
-        """Reset episode state."""
+        """Reset per-episode state. Called by lerobot-eval at the start of each episode."""
         ...
 
     def get_optim_params(self) -> dict:
@@ -147,35 +129,51 @@ class MyCustomPolicy(PreTrainedPolicy):
         ...
 
     def select_action(self, batch: dict[str, torch.Tensor], **kwargs) -> torch.Tensor:
-        """Return a single action for the current timestep (called at inference)."""
+        """Return a single action for the current timestep (called every step at inference)."""
         ...
 
-    def forward(self, batch: dict[str, torch.Tensor]) -> dict[str, torch.Tensor]:
+    def forward(self, batch: dict[str, torch.Tensor]) -> tuple[torch.Tensor, dict | None]:
         """Compute the training loss.
 
+        Returns `(loss, output_dict)`. `output_dict` may be `None`; everything in it must be
+        logging-friendly Python natives (no tensors with gradients).
+
         `batch["action_is_pad"]` is a bool mask of shape (B, horizon) that marks
-        timesteps padded because the episode ended before `horizon` steps, you
+        timesteps padded because the episode ended before `horizon` steps; you
         can exclude those from your loss.
         """
         actions = batch[ACTION]
         action_is_pad = batch.get("action_is_pad")
         ...
-        return {"loss": ...}
+        return loss, {"some_loss_component": some_loss_component.item()}
 ```
 
-## Step 4: Add Data Processors
+The methods called by the train/eval loops:
 
-Create processor functions. For a concrete reference, see [processor_act.py](https://github.com/huggingface/lerobot/blob/main/src/lerobot/policies/act/processor_act.py) or [processor_diffusion.py](https://github.com/huggingface/lerobot/blob/main/src/lerobot/policies/diffusion/processor_diffusion.py).
+| Method                                                            | Used by           | What it does                                                                                                                                                                                                                                         |
+| ----------------------------------------------------------------- | ----------------- | ---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
+| `reset() -> None`                                                 | `lerobot-eval`    | Clear per-episode state at the start of each episode.                                                                                                                                                                                                |
+| `select_action(batch, **kwargs) -> Tensor`                        | `lerobot-eval`    | Return the next action `(B, action_dim)`. Called every step.                                                                                                                                                                                         |
+| `predict_action_chunk(batch, **kwargs) -> Tensor`                 | the policy itself | Return an action chunk `(B, chunk_size, action_dim)`. Currently abstract on the base class — raise `NotImplementedError` if your policy doesn't chunk.                                                                                               |
+| `forward(batch, reduction="mean") -> tuple[Tensor, dict \| None]` | `lerobot-train`   | Return `(loss, output_dict)`. Accept `reduction="none"` if you want to support per-sample weighting.                                                                                                                                                 |
+| `get_optim_params() -> dict`                                      | the optimizer     | Return `self.parameters()` for simple policies; return a named parameter dict for [multi-optimizer policies](https://github.com/huggingface/lerobot/blob/ecd38c50d7d15b4184cf42649ff1185ee2e11eeb/src/lerobot/policies/sac/modeling_sac.py#L61-L73). |
+| `update() -> None` _(optional)_                                   | `lerobot-train`   | Called after each optimizer step _if defined_. Use for EMA, target nets, replay buffers (TDMPC uses this).                                                                                                                                           |
+
+Batches are flat dictionaries keyed by the constants in [`lerobot.utils.constants`](https://github.com/huggingface/lerobot/blob/main/src/lerobot/utils/constants.py): `OBS_STATE` (`observation.state.<motor>`), `OBS_IMAGES` (`observation.images.<camera>`), `OBS_LANGUAGE`, `ACTION`, etc. Reuse the constants — don't invent new prefixes.
+
+### Processor functions
+
+LeRobot uses `PolicyProcessorPipeline`s to normalize inputs and de-normalize outputs around your policy. For a concrete reference, see [`processor_act.py`](https://github.com/huggingface/lerobot/blob/main/src/lerobot/policies/act/processor_act.py) or [`processor_diffusion.py`](https://github.com/huggingface/lerobot/blob/main/src/lerobot/policies/diffusion/processor_diffusion.py).
 
 ```python
-# processor_my_custom_policy.py
+# processor_my_policy.py
 from typing import Any
 import torch
 
 from lerobot.processor import PolicyAction, PolicyProcessorPipeline
 
 
-def make_my_custom_policy_pre_post_processors(
+def make_my_policy_pre_post_processors(
     config,
     dataset_stats: dict[str, dict[str, torch.Tensor]] | None = None,
 ) -> tuple[
@@ -187,11 +185,48 @@ def make_my_custom_policy_pre_post_processors(
     return preprocessor, postprocessor
 ```
 
-**Important - function naming:** LeRobot discovers your processor by name. The function **must** be called `make_{policy_name}_pre_post_processors` (matching the string you passed to `@PreTrainedConfig.register_subclass`).
+**Important — function naming:** LeRobot discovers your processor by name. The function **must** be called `make_{policy_name}_pre_post_processors` (matching the string you passed to `@PreTrainedConfig.register_subclass`).
 
-## Step 5: Package Initialization
+---
 
-Expose your classes in the package's `__init__.py`:
+## 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.
+
+### Package structure
+
+Create a package with the prefix `lerobot_policy_` (IMPORTANT!) followed by your policy name:
+
+```bash
+lerobot_policy_my_policy/
+├── pyproject.toml
+└── src/
+    └── lerobot_policy_my_policy/
+        ├── __init__.py
+        ├── configuration_my_policy.py
+        ├── modeling_my_policy.py
+        └── processor_my_policy.py
+```
+
+### `pyproject.toml`
+
+```toml
+[project]
+name = "lerobot_policy_my_policy"
+version = "0.1.0"
+dependencies = [
+    # your policy-specific dependencies
+]
+requires-python = ">= 3.12"
+
+[build-system]
+build-backend = # your-build-backend
+requires = # your-build-system
+```
+
+### Package `__init__.py`
+
+Expose your classes in the package's `__init__.py` and guard against missing `lerobot`:
 
 ```python
 # __init__.py
@@ -204,44 +239,151 @@ except ImportError:
         "lerobot is not installed. Please install lerobot to use this policy package."
     )
 
-from .configuration_my_custom_policy import MyCustomPolicyConfig
-from .modeling_my_custom_policy import MyCustomPolicy
-from .processor_my_custom_policy import make_my_custom_policy_pre_post_processors
+from .configuration_my_policy import MyPolicyConfig
+from .modeling_my_policy import MyPolicy
+from .processor_my_policy import make_my_policy_pre_post_processors
 
 __all__ = [
-    "MyCustomPolicyConfig",
-    "MyCustomPolicy",
-    "make_my_custom_policy_pre_post_processors",
+    "MyPolicyConfig",
+    "MyPolicy",
+    "make_my_policy_pre_post_processors",
 ]
 ```
 
-## Step 6: Installation and Usage
-
-### Install Your Policy Package
+### Install and use
 
 ```bash
-cd lerobot_policy_my_custom_policy
+cd lerobot_policy_my_policy
 pip install -e .
 
 # Or install from PyPI if published
-pip install lerobot_policy_my_custom_policy
+pip install lerobot_policy_my_policy
 ```
 
-### Use Your Policy
-
 Once installed, your policy automatically integrates with LeRobot's training and evaluation tools:
 
 ```bash
 lerobot-train \
-    --policy.type my_custom_policy \
+    --policy.type my_policy \
     --env.type pusht \
     --steps 200000
 ```
 
-## Examples and Community Contributions
+---
+
+## Path B: Contributing in-tree
+
+When your policy has stabilized and there's clear value in shipping it with the library, you can land it directly in LeRobot. Read the general [contribution guide](./contributing) and the [PR template](https://github.com/huggingface/lerobot/blob/main/.github/PULL_REQUEST_TEMPLATE.md) first — that's where you'll find the testing/quality expectations every PR has to meet (`pre-commit run -a`, `pytest`, the community-review rule, etc.). What's below is the policy-specific layer on top of that.
+
+### In-tree layout
+
+```
+src/lerobot/policies/my_policy/
+├── __init__.py                    # re-exports config + modeling + processor factory
+├── configuration_my_policy.py     # MyPolicyConfig + @register_subclass
+├── modeling_my_policy.py          # MyPolicy(PreTrainedPolicy)
+├── processor_my_policy.py         # make_my_policy_pre_post_processors
+└── README.md                      # symlink → ../../../../docs/source/policy_my_policy_README.md
+```
+
+Two notes:
+
+- The `README.md` next to the source is a **symlink** into `docs/source/policy_<name>_README.md` — the actual file lives under `docs/`. Existing policies (act, smolvla, diffusion, …) all do this; copy one of those symlinks. The policy README is conventionally minimal: paper link + BibTeX citation.
+- The user-facing tutorial — what to install, how to train, hyperparameters, benchmark numbers — lives separately at `docs/source/<my_policy>.mdx` and is registered in `_toctree.yml` under "Policies".
+
+The file names are load-bearing: the factory does lazy imports by name, and the processor is discovered by the `make_<policy_name>_pre_post_processors` convention.
+
+### Wiring
+
+Four places need to know about your policy. All by name.
+
+1. **`policies/__init__.py`** — re-export `MyPolicyConfig` and add it to `__all__`. **Don't** re-export the modeling class; it loads lazily through the factory (so `import lerobot` stays fast).
+2. **`factory.py:get_policy_class`** — add a branch returning `MyPolicy` from a lazy import.
+3. **`factory.py:make_policy_config`** and **`factory.py:make_pre_post_processors`** — same idea, two more branches.
+4. **`templates/lerobot_modelcard_template.md` and the root `README.md`** — the template is what `push_model_to_hub` renders into the model card of every checkpoint trained with your policy: add a one-line description of your policy in the `model_name` branches, map it in `policy_docs` so cards link to your MDX guide, and optionally add an architecture image to `diagrams`. Then add your policy to the models table in the root `README.md`, under the right category, linking to your doc page.
+
+Mirror an existing policy that's structurally similar to yours; the diff is small.
+
+### Heavy / optional dependencies
+
+Most policies need a heavy backbone (transformers, diffusers, a specific VLM SDK). The convention is **two-step gating**: a `TYPE_CHECKING`-guarded import at module top, and a `require_package` runtime check in the constructor. [`modeling_diffusion.py`](https://github.com/huggingface/lerobot/blob/main/src/lerobot/policies/diffusion/modeling_diffusion.py) is the canonical reference:
+
+```python
+from typing import TYPE_CHECKING
+from lerobot.utils.import_utils import _diffusers_available, require_package
+
+if TYPE_CHECKING or _diffusers_available:
+    from diffusers.schedulers.scheduling_ddim import DDIMScheduler
+else:
+    DDIMScheduler = None  # keeps the symbol bindable at import time
+
+class DiffusionPolicy(PreTrainedPolicy):
+    def __init__(self, config):
+        require_package("diffusers", extra="diffusion")
+        super().__init__(config)
+        ...
+```
+
+This way:
+
+- `import lerobot.policies` keeps working without the extra installed (the symbol is just bound to `None`).
+- Type checkers see the real symbol.
+- Instantiating the policy without the extra raises a clear `ImportError` pointing at `pip install 'lerobot[diffusion]'`.
+
+Add a matching extra to [`pyproject.toml`](https://github.com/huggingface/lerobot/blob/main/pyproject.toml) `[project.optional-dependencies]` and include it in the `all` extra so `pip install 'lerobot[all]'` keeps installing everything.
+
+### Benchmarks and a published checkpoint
+
+A new policy is much easier to review — and far more useful — when it ships with a working checkpoint and at least one number you can reproduce.
+
+**Pick at least one in-tree benchmark.** LeRobot ships sim benchmarks with per-benchmark Docker images (LIBERO, LIBERO-plus, Meta-World, RoboTwin 2.0, RoboCasa365, RoboCerebra, RoboMME, VLABench and more). Pick the one that matches your policy's modality — VLAs usually go to LIBERO or VLABench; image-only BC to LIBERO or Meta-World. The full list lives under [Benchmarks](./libero) in the docs sidebar.
+
+**Push the checkpoint & processors** to the Hub under `lerobot/<policy>_<benchmark>` (or your namespace if you don't have write access; a maintainer can mirror it). Use `PreTrainedPolicy.push_model_to_hub` so the repo gets `config.json`, `model.safetensors`, and a model card.
+
+**Report results in your policy's MDX**, with the exact `lerobot-eval` command and hardware so anyone can re-run:
+
+```markdown
+## Results
+
+Evaluated on LIBERO with `lerobot/<policy>_libero`:
+
+| Suite          | Success rate | n_episodes |
+| -------------- | -----------: | ---------: |
+| libero_spatial |        87.5% |         50 |
+| libero_object  |        93.0% |         50 |
+| libero_goal    |        81.5% |         50 |
+| libero_10      |        62.0% |         50 |
+| **average**    |    **81.0%** |        200 |
+
+Reproduce: `lerobot-eval --policy.path=lerobot/<policy>_libero --env.type=libero --env.task=libero_spatial --eval.n_episodes=50` (1× A100 40 GB).
+```
+
+Use `n_episodes ≥ 50` per suite for stable success-rate estimates.
+
+If your policy is real-robot-only and no sim benchmark applies, swap the sim eval for: a public training dataset on the Hub, the `lerobot-train` command, the checkpoint, and a real-robot success rate over ≥10 episodes via `lerobot-rollout --policy.path=...`.
+
+### PR checklist
+
+The general expectations are in [`CONTRIBUTING.md`](https://github.com/huggingface/lerobot/blob/main/CONTRIBUTING.md) and the [PR template](https://github.com/huggingface/lerobot/blob/main/.github/PULL_REQUEST_TEMPLATE.md). On top of those, reviewers will look for:
+
+- [ ] `MyPolicy` and `MyPolicyConfig` cover the surface above; `__init_subclass__` accepts the class.
+- [ ] `factory.py` and `policies/__init__.py` are wired (lazy imports for modeling).
+- [ ] `make_my_policy_pre_post_processors` follows the naming convention.
+- [ ] Optional deps live behind a `[project.optional-dependencies]` extra and the `TYPE_CHECKING + require_package` guard.
+- [ ] `tests/policies/` updated; backward-compat artifact committed & policy-specific tests.
+- [ ] `src/lerobot/policies/<name>/README.md` symlinked into `docs/source/policy_<name>_README.md`; user-facing `docs/source/<name>.mdx` written and added to `_toctree.yml`.
+- [ ] `templates/lerobot_modelcard_template.md` has a description entry and a `policy_docs` link for your policy.
+- [ ] The models table in the root `README.md` lists your policy in the right category, linking to your doc page.
+- [ ] At least one reproducible benchmark eval in the policy MDX with a published checkpoint (sim benchmark, or real-robot dataset + checkpoint).
+
+The fastest way to get a clean PR is to copy the directory of the existing policy closest to yours, rename, and replace contents method by method. Don't wait until everything is polished — open a draft PR early and iterate with us; reviewers would much rather give feedback on a half-finished branch than a fully-merged one.
+
+---
+
+## Examples and community contributions
 
 Check out these example policy implementations:
 
-- [DiTFlow Policy](https://github.com/danielsanjosepro/lerobot_policy_ditflow) - Diffusion Transformer policy with flow-matching objective. Try it out in this example: [DiTFlow Example](https://github.com/danielsanjosepro/test_lerobot_policy_ditflow)
+- [DiTFlow Policy](https://github.com/danielsanjosepro/lerobot_policy_ditflow) — Diffusion Transformer policy with flow-matching objective. Try it out in this example: [DiTFlow Example](https://github.com/danielsanjosepro/test_lerobot_policy_ditflow)
 
-Share your policy implementations with the community! 🤗
+Thanks for taking the time to bring a new policy into LeRobot. Every architecture that lands in `main` — and every plugin published by the community — makes the library a little more useful for the next person, and a little more representative of where robot learning is going. We're looking forward to seeing what you ship. 🤗

docs/source/cameras.mdx

@@ -157,6 +157,14 @@ finally:
 </hfoption>
 </hfoptions>
 
+### Working with depth
+
+The Intel RealSense and Reachy 2 cameras can capture both color and depth in lockstep. Calling `read()` returns the **color** frame as `(H, W, 3)` `uint8`. Calling `read_depth()` returns the **depth map** as `(H, W, 1)` `uint16`, where each pixel value is the distance from the sensor expressed in **millimetres**. A pixel value of `0` typically means "no measurement available" (out-of-range, occluded, or low-confidence).
+
+During recording, the control loop peeks the freshest buffered frames non-blockingly via `read_latest()` (color) and `read_latest_depth()` (depth), adding the depth map as a sibling feature (e.g. `front_depth` next to `front`).
+
+For how depth streams are stored and encoded when recording a dataset, see the [Depth streams](./video_encoding_parameters#depth-streams) section of the video encoding guide.
+
 ## Use your phone's camera
 
 <hfoptions id="use phone">

docs/source/cheat-sheet.mdx

@@ -0,0 +1,177 @@
+# Cheat sheet
+
+All of the LeRobot commands in one place. If you forgot how to use a specific command or want to learn about a new one you can do it here.
+
+> [!WARNING]
+> For all of the commands listed below remember to change the ports/names/ids to your own values!
+
+> [!TIP]
+> Another great way to look at all the commands and get them configured for your specific setup is to use this [Jupyter Notebook](https://github.com/huggingface/lerobot/blob/main/examples/notebooks/quickstart.ipynb).
+
+### Setup and installation
+
+For installation please look at [LeRobot Installation](https://huggingface.co/docs/lerobot/main/en/installation).
+
+### Useful tools
+
+###### Find port
+
+Use this to identify which serial ports your robots are connected to. Follow the instructions in your terminal: you will be asked to unplug the USB cable and press Enter. The script will then detect and print the correct serial port for that robot.
+
+```bash
+lerobot-find-port
+```
+
+###### Find cameras
+
+Quickly find camera indices and verify their output. This command prints camera information to the terminal and saves test frames from each detected camera to `lerobot/outputs/captured_images`
+
+```bash
+lerobot-find-cameras
+```
+
+### Calibration
+
+In most cases you will need to perform calibration just once for each robot and teleoperation device. Before performing the calibration make sure that all the joints are roughly in the middle position.
+
+```bash
+lerobot-calibrate \
+    --robot.type=so101_follower \
+    --robot.port=/dev/ttyACM0 \
+    --robot.id=my_follower_arm
+```
+
+Make sure that you use the same IDs used during calibration later for the other scripts. That's how LeRobot finds the calibration files.
+
+### Teleoperation
+
+Teleoperating with two cameras and displaying the data with Rerun.
+
+```bash
+lerobot-teleoperate \
+    --robot.type=so101_follower \
+    --robot.port=/dev/ttyACM0 \
+    --robot.id=my_follower_arm \
+    --robot.cameras="{ top: {type: opencv, index_or_path: 1, width: 640, height: 480, fps: 30}, wrist: {type: opencv, index_or_path: 0, width: 640, height: 480, fps: 30} }" \
+    --teleop.type=so101_leader \
+    --teleop.port=/dev/ttyACM1 \
+    --teleop.id=my_leader_arm \
+    --display_data=true
+```
+
+### Recording a dataset
+
+The dataset is automatically uploaded to the server and saved under repo_id, make sure you are logged in to your HF account with CLI:
+`hf auth login`
+
+You can get the token from: [https://huggingface.co/settings/tokens](https://huggingface.co/settings/tokens)
+
+```bash
+lerobot-record \
+    --robot.type=so101_follower \
+    --robot.port=/dev/ttyACM0 \
+    --robot.id=my_follower_arm \
+    --robot.cameras="{ top: {type: opencv, index_or_path: 1, width: 640, height: 480, fps: 30}, wrist: {type: opencv, index_or_path: 0, width: 640, height: 480, fps: 30} }" \
+    --teleop.type=so101_leader \
+    --teleop.port=/dev/ttyACM1 \
+    --teleop.id=my_leader_arm \
+    --dataset.repo_id=${HF_USER}/so101_dataset_test \
+    --dataset.num_episodes=30 \
+    --dataset.single_task="put the red brick in a bowl" \
+    --dataset.streaming_encoding=true \
+    --display_data=true
+```
+
+While collecting the dataset you can control the process with your keyboard:
+Control the data recording flow using keyboard shortcuts:
+
+- Press **Right Arrow (`→`)**: Save episode and move to the next.
+- Press **Left Arrow (`←`)**: Delete current episode and retry.
+- Press **Escape (`ESC`)**: Stop, encode videos, and upload.
+
+### Recording depth
+
+Intel RealSense cameras (`type: intelrealsense`) record a depth stream when you set `use_depth: true`. Depth is quantized to 12-bit codes and stored as its own video.
+
+```bash
+lerobot-record \
+    ... \
+    --robot.cameras="{ head: {type: intelrealsense, serial_number_or_name: \"0123456789\", width: 640, height: 480, fps: 30, use_depth: true} }" \
+    --dataset.repo_id=${HF_USER}/so101_depth_test \
+    --dataset.single_task="put the red brick in a bowl" \
+    --dataset.depth_encoder.depth_min=0.01 \
+    --dataset.depth_encoder.depth_max=10.0 \
+    --dataset.depth_encoder.shift=0.0 \
+    --dataset.depth_encoder.use_log=true
+```
+
+### Video encoding parameters
+
+RGB and depth streams are encoded independently via the `--dataset.rgb_encoder.*` and `--dataset.depth_encoder.*` keys.
+
+```bash
+lerobot-record \
+    ... \
+    --dataset.rgb_encoder.vcodec=h264 \
+    --dataset.rgb_encoder.pix_fmt=yuv420p \
+    --dataset.rgb_encoder.crf=23 \
+    --dataset.depth_encoder.vcodec=hevc \
+    --dataset.depth_encoder.extra_options='{"x265-params": "lossless=1"}'
+```
+
+### Training
+
+Depending on your hardware training the policy might take a few hours. That's how you train simple `ACT` policy:
+
+```bash
+lerobot-train \
+    --dataset.repo_id=${HF_USER}/so101_dataset_test \
+    --policy.type=act \
+    --output_dir=outputs/train/act_so101_test \
+    --job_name=act_so101_test \
+    --policy.device=cuda \
+    --wandb.enable=true \
+    --policy.repo_id=${HF_USER}/policy_test \
+    --steps=20000
+```
+
+- Policy Types: `act`, `diffusion`, `smolvla`, `pi05`
+- Devices: `cuda` (NVIDIA), `mps` (Apple Silicon), `cpu`
+
+If you want to fine-tune a specific model you can provide the path to the model. In this case path is enough and type can be skipped.
+
+```bash
+lerobot-train \
+    --dataset.repo_id=${HF_USER}/so101_dataset_test \
+    --policy.path=username/the_policy_to_finetune \
+    --policy.device=cuda \
+    --policy.repo_id=${HF_USER}/policy_test \
+    --output_dir=outputs/train/act_so101_test \
+    --steps=20000
+```
+
+No local GPU? Add `--job.target=<flavor>` (e.g. `a10g-small`) to either command and `lerobot-train` runs it on [Hugging Face Jobs](https://huggingface.co/docs/hub/jobs) instead — it uploads a local-only dataset for you and pushes the trained model. List flavors with `hf jobs hardware`.
+
+To resume, point `--config_path` at a checkpoint and add `--resume=true`. It accepts a local path or a Hub repo id (the latest checkpoint is fetched), and works locally or on a job by adding `--job.target=<flavor>`:
+
+```bash
+lerobot-train --config_path=${HF_USER}/policy_test --resume=true --job.target=a10g-small
+```
+
+### Inference
+
+Inference means running the trained policy/model on a robot. For that we use `lerobot-rollout`. You will need to provide a path to your policy. It can be a local path or a path to Hugging Face for example "lerobot/folding_latest". Your cameras configuration needs to match what was used when collecting the dataset. Duration is in seconds if unspecified, it will run forever.
+
+> [!TIP]
+> If you are using the previous release V0.5.1 instead of `lerobot-rollout` you need to use `lerobot-record`. More information [here](https://huggingface.co/docs/lerobot/v0.5.1/en/il_robots#run-inference-and-evaluate-your-policy).
+
+```bash
+lerobot-rollout \
+    --strategy.type=base \
+    --policy.path=${HF_USER}/my_policy \
+    --robot.type=so101_follower \
+    --robot.port=/dev/ttyACM1 \
+    --robot.cameras="{ up: {type: opencv, index_or_path: /dev/video1, width: 640, height: 480, fps: 30}, side: {type: opencv, index_or_path: /dev/video5, width: 640, height: 480, fps: 30}}" \
+    --task="Put lego brick into the transparent box" \
+    --duration=60
+```

docs/source/dataset_subtask.mdx

@@ -1,277 +0,0 @@
-# Using Subtasks in LeRobot Datasets
-
-Subtask support in robotics datasets has proven effective in improving robot reasoning and understanding. Subtasks are particularly useful for:
-
-- **Hierarchical policies**: Building policies that include subtask predictions to visualize robot reasoning in real time
-- **Reward modeling**: Helping reward models understand task progression (e.g., SARM-style stage-aware reward models)
-- **Task decomposition**: Breaking down complex manipulation tasks into atomic, interpretable steps
-
-LeRobotDataset now supports subtasks as part of its dataset structure, alongside tasks.
-
-## What are Subtasks?
-
-While a **task** describes the overall goal (e.g., "Pick up the apple and place it in the basket"), **subtasks** break down the execution into finer-grained steps:
-
-1. "Approach the apple"
-2. "Grasp the apple"
-3. "Lift the apple"
-4. "Move to basket"
-5. "Release the apple"
-
-Each frame in the dataset can be annotated with its corresponding subtask, enabling models to learn and predict these intermediate stages.
-
-<img
-  src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/lerobot/subtask-asset.png"
-  alt="An overview of subtask annotation showing how frames are labeled with intermediate subtask stages"
-  width="80%"
-/>
-
-<p>
-  <em>Figure: Overview of subtask annotation.</em>
-</p>
-
-**Reference:** _Subtask-learning based for robot self-assembly in flexible collaborative assembly in manufacturing_, Original Article, Published: 19 April 2022.
-
-## Dataset Structure
-
-Subtask information is stored in the dataset metadata:
-
-```
-my-dataset/
-├── data/
-│   └── ...
-├── meta/
-│   ├── info.json
-│   ├── stats.json
-│   ├── tasks.parquet
-│   ├── subtasks.parquet      # Subtask index → subtask string mapping
-│   └── episodes/
-│       └── ...
-└── videos/
-    └── ...
-```
-
-### Subtasks Parquet File
-
-The `meta/subtasks.parquet` file maps subtask indices to their natural language descriptions:
-
-| subtask_index | subtask (index column) |
-| ------------- | ---------------------- |
-| 0             | "Approach the apple"   |
-| 1             | "Grasp the apple"      |
-| 2             | "Lift the apple"       |
-| ...           | ...                    |
-
-### Frame-Level Annotations
-
-Each frame in the dataset can include a `subtask_index` field that references the subtasks parquet file:
-
-```python
-# Example frame data in the parquet file
-{
-    "index": 42,
-    "timestamp": 1.4,
-    "episode_index": 0,
-    "task_index": 0,
-    "subtask_index": 2,  # References "Lift the apple"
-    "observation.state": [...],
-    "action": [...],
-}
-```
-
-## Annotating Datasets with Subtasks
-
-We provide a HuggingFace Space for easily annotating any LeRobotDataset with subtasks:
-
-**[https://huggingface.co/spaces/lerobot/annotate](https://huggingface.co/spaces/lerobot/annotate)**
-
-After completing your annotation:
-
-1. Click "Push to Hub" to upload your annotated dataset
-2. You can also run the annotation space locally by following the instructions at [github.com/huggingface/lerobot-annotate](https://github.com/huggingface/lerobot-annotate)
-
-## Loading Datasets with Subtasks
-
-When you load a dataset with subtask annotations, the subtask information is automatically available:
-
-```python
-from lerobot.datasets import LeRobotDataset
-
-# Load a dataset with subtask annotations
-dataset = LeRobotDataset("jadechoghari/collect-fruit-annotated")
-
-# Access a sample
-sample = dataset[100]
-
-# The sample includes both task and subtask information
-print(sample["task"])        # "Collect the fruit"
-print(sample["subtask"])     # "Grasp the apple"
-print(sample["task_index"])  # tensor(0)
-print(sample["subtask_index"])  # tensor(2)
-```
-
-### Checking for Subtask Support
-
-You can check if a dataset has subtask annotations:
-
-```python
-# Check if subtasks are available
-has_subtasks = (
-    "subtask_index" in dataset.features
-    and dataset.meta.subtasks is not None
-)
-
-if has_subtasks:
-    print(f"Dataset has {len(dataset.meta.subtasks)} unique subtasks")
-    print("Subtasks:", list(dataset.meta.subtasks.index))
-```
-
-## Using Subtasks for Training
-
-### With the Tokenizer Processor
-
-The `TokenizerProcessor` automatically handles subtask tokenization for Vision-Language Action (VLA) models:
-
-```python
-from lerobot.processor import TokenizerProcessorStep
-
-# Create a tokenizer processor step
-tokenizer_processor = TokenizerProcessorStep(
-    tokenizer_name_or_path="google/paligemma-3b-pt-224",
-    padding="max_length",
-    max_length=64,
-)
-
-# The processor will automatically tokenize subtasks if present in the batch
-# and add them to the observation under:
-# - "observation.subtask.tokens"
-# - "observation.subtask.attention_mask"
-```
-
-When subtasks are available in the batch, the tokenizer processor adds:
-
-- `observation.subtask.tokens`: Tokenized subtask text
-- `observation.subtask.attention_mask`: Attention mask for the subtask tokens
-
-### DataLoader with Subtasks
-
-```python
-import torch
-from lerobot.datasets import LeRobotDataset
-
-dataset = LeRobotDataset("jadechoghari/collect-fruit-annotated")
-
-dataloader = torch.utils.data.DataLoader(
-    dataset,
-    batch_size=16,
-    shuffle=True,
-)
-
-for batch in dataloader:
-    # Access subtask information in the batch
-    subtasks = batch["subtask"]  # List of subtask strings
-    subtask_indices = batch["subtask_index"]  # Tensor of subtask indices
-
-    # Use for training hierarchical policies or reward models
-    print(f"Batch subtasks: {set(subtasks)}")
-```
-
-## Example Datasets with Subtask Annotations
-
-Try loading a dataset with subtask annotations:
-
-```python
-from lerobot.datasets import LeRobotDataset
-
-# Example dataset with subtask annotations
-dataset = LeRobotDataset("jadechoghari/collect-fruit-annotated")
-
-# Explore the subtasks
-print("Available subtasks:")
-for subtask_name in dataset.meta.subtasks.index:
-    print(f"  - {subtask_name}")
-
-# Get subtask distribution
-subtask_counts = {}
-for i in range(len(dataset)):
-    sample = dataset[i]
-    subtask = sample["subtask"]
-    subtask_counts[subtask] = subtask_counts.get(subtask, 0) + 1
-
-print("\nSubtask distribution:")
-for subtask, count in sorted(subtask_counts.items(), key=lambda x: -x[1]):
-    print(f"  {subtask}: {count} frames")
-```
-
-## Use Cases
-
-### 1. Hierarchical Policy Training
-
-Train policies that predict both actions and current subtask:
-
-```python
-class HierarchicalPolicy(nn.Module):
-    def __init__(self, num_subtasks):
-        super().__init__()
-        self.action_head = nn.Linear(hidden_dim, action_dim)
-        self.subtask_head = nn.Linear(hidden_dim, num_subtasks)
-
-    def forward(self, observations):
-        features = self.encoder(observations)
-        actions = self.action_head(features)
-        subtask_logits = self.subtask_head(features)
-        return actions, subtask_logits
-```
-
-### 2. Stage-Aware Reward Modeling (SARM)
-
-Build reward models that understand task progression:
-
-```python
-# SARM predicts:
-# - Stage: Which subtask is being executed (discrete)
-# - Progress: How far along the subtask (continuous 0-1)
-
-class SARMRewardModel(nn.Module):
-    def forward(self, observations):
-        features = self.encoder(observations)
-        stage_logits = self.stage_classifier(features)
-        progress = self.progress_regressor(features)
-        return stage_logits, progress
-```
-
-### 3. Progress Visualization
-
-Monitor robot execution by tracking subtask progression:
-
-```python
-def visualize_execution(model, observations):
-    for t, obs in enumerate(observations):
-        action, subtask_logits = model(obs)
-        predicted_subtask = subtask_names[subtask_logits.argmax()]
-        print(f"t={t}: Executing '{predicted_subtask}'")
-```
-
-## API Reference
-
-### LeRobotDataset Properties
-
-| Property                    | Type                   | Description                                |
-| --------------------------- | ---------------------- | ------------------------------------------ |
-| `meta.subtasks`             | `pd.DataFrame \| None` | DataFrame mapping subtask names to indices |
-| `features["subtask_index"]` | `dict`                 | Feature spec for subtask_index if present  |
-
-### Sample Keys
-
-When subtasks are available, each sample includes:
-
-| Key             | Type           | Description                          |
-| --------------- | -------------- | ------------------------------------ |
-| `subtask_index` | `torch.Tensor` | Integer index of the current subtask |
-| `subtask`       | `str`          | Natural language subtask description |
-
-## Related Resources
-
-- [SARM Paper](https://arxiv.org/pdf/2509.25358) - Stage-Aware Reward Modeling for Long Horizon Robot Manipulation
-- [LeRobot Annotate Space](https://huggingface.co/spaces/lerobot/annotate) - Interactive annotation tool
-- [LeRobotDataset v3.0](./lerobot-dataset-v3) - Dataset format documentation

docs/source/earthrover_mini_plus.mdx

@@ -194,7 +194,7 @@ lerobot-record \
     --dataset.single_task="Navigate around obstacles" \
     --dataset.streaming_encoding=true \
     --dataset.encoder_threads=2 \
-    # --dataset.vcodec=auto \
+    # --dataset.rgb_encoder.vcodec=auto \
     --display_data=true
 ```
 

docs/source/envhub_isaaclab_arena.mdx

@@ -193,7 +193,7 @@ To learn more about training policies with LeRobot, please refer to the training
 
 - [SmolVLA](./smolvla)
 - [Pi0.5](./pi05)
-- [GR00T N1.5](./groot)
+- [GR00T N1.7](./groot)
 
 Sample IsaacLab Arena datasets are available on HuggingFace Hub for experimentation:
 

docs/source/evo1.mdx

@@ -0,0 +1,191 @@
+# EVO1
+
+EVO1 is a Vision-Language-Action policy for robot control built around an InternVL3 backbone and a continuous flow-matching action head. This LeRobot integration exposes EVO1 as a standard policy type so it can be trained and evaluated with the usual LeRobot dataset, checkpoint, and processor APIs.
+
+## Model Overview
+
+The policy embeds one or more camera images and the language task prompt with InternVL3, pads robot state/action vectors to fixed maximum dimensions, and predicts future action chunks with a flow-matching action head. During inference, the policy samples an action chunk and returns `n_action_steps` actions from that chunk before sampling again.
+
+### What the LeRobot Integration Covers
+
+- Standard `policy.type=evo1` configuration through LeRobot
+- InternVL3 image/text embedding with optional FlashAttention fallback
+- Stage-based finetuning controls for action-head-only and VLM finetuning runs
+- Continuous flow-matching action prediction
+- Checkpoint save/load through LeRobot policy APIs
+- Training with `lerobot-train` and evaluation with standard policy inference APIs
+
+The broader EVO1 project may include additional training scripts and dataset tooling. This page focuses on the LeRobot robot-control policy path.
+
+## Installation Requirements
+
+1. Install LeRobot by following the [Installation Guide](./installation).
+2. Install EVO1 dependencies:
+
+   ```bash
+   pip install -e ".[evo1]"
+   ```
+
+   For LIBERO evaluation, install the LIBERO extra as well:
+
+   ```bash
+   pip install -e ".[evo1,libero]"
+   ```
+
+3. Install a `flash-attn` wheel only if it is compatible with your Python, PyTorch, CUDA, and GPU stack. EVO1 falls back to standard attention when `flash_attn` is not available.
+
+EVO1 uses the native Hugging Face `transformers` InternVL implementation, so `policy.vlm_model_name` must point to a natively converted checkpoint such as `OpenGVLab/InternVL3-1B-hf` (note the `-hf` suffix). The first run may download the configured VLM checkpoint unless `policy.vlm_model_name` points to a local model directory.
+
+## Data Requirements
+
+EVO1 expects a LeRobot dataset with:
+
+- One to `policy.max_views` visual observations, for example `observation.images.image`
+- `observation.state`
+- `action`
+- A language task instruction in the dataset `task` field, or another field configured with `policy.task_field`
+
+State and action vectors are padded to `policy.max_state_dim` and `policy.max_action_dim`. Predictions are cropped back to the dataset action dimension before being returned.
+
+## Usage
+
+To use EVO1 in a LeRobot configuration, specify:
+
+```python
+policy.type=evo1
+```
+
+By default, a new EVO1 policy initializes its VLM from:
+
+```python
+policy.vlm_model_name=OpenGVLab/InternVL3-1B-hf
+```
+
+Once a LeRobot-format EVO1 checkpoint is available, load it with:
+
+```python
+policy.path=your-org/your-evo1-checkpoint
+```
+
+## Training
+
+### Stage 1
+
+Stage 1 freezes the VLM and trains the action head:
+
+```bash
+lerobot-train \
+  --dataset.repo_id=your_org/your_dataset \
+  --policy.type=evo1 \
+  --policy.training_stage=stage1 \
+  --policy.vlm_model_name=OpenGVLab/InternVL3-1B-hf \
+  --policy.device=cuda \
+  --policy.chunk_size=50 \
+  --policy.n_action_steps=50 \
+  --policy.max_state_dim=24 \
+  --policy.max_action_dim=24 \
+  --policy.optimizer_lr=1e-5 \
+  --batch_size=4 \
+  --steps=5000 \
+  --output_dir=./outputs/evo1_stage1
+```
+
+### Stage 2
+
+Stage 2 finetunes the VLM branches and action head. A common workflow starts from a Stage 1 checkpoint:
+
+```bash
+lerobot-train \
+  --dataset.repo_id=your_org/your_dataset \
+  --policy.path=./outputs/evo1_stage1/checkpoints/005000/pretrained_model \
+  --policy.training_stage=stage2 \
+  --policy.vlm_model_name=OpenGVLab/InternVL3-1B-hf \
+  --policy.device=cuda \
+  --policy.chunk_size=50 \
+  --policy.n_action_steps=50 \
+  --policy.max_state_dim=24 \
+  --policy.max_action_dim=24 \
+  --policy.optimizer_lr=1e-5 \
+  --batch_size=4 \
+  --steps=80000 \
+  --output_dir=./outputs/evo1_stage2
+```
+
+By default, `policy.training_stage` reapplies the finetuning defaults for that stage. This is important when
+starting Stage 2 from a Stage 1 checkpoint, because the Stage 1 checkpoint config stores the VLM finetuning
+flags as disabled. These stage defaults take precedence over saved or manually supplied `policy.finetune_*`
+flags unless `policy.apply_training_stage_defaults=false`, so set that flag only when manually controlling
+every finetuning flag.
+
+### Key Training Parameters
+
+| Parameter                                     | Default                     | Description                                                       |
+| --------------------------------------------- | --------------------------- | ----------------------------------------------------------------- |
+| `policy.vlm_model_name`                       | `OpenGVLab/InternVL3-1B-hf` | Natively converted InternVL3 checkpoint or local model directory  |
+| `policy.training_stage`                       | `stage1`                    | `stage1` trains the action head; `stage2` finetunes VLM branches  |
+| `policy.apply_training_stage_defaults`        | `true`                      | Reapplies stage finetuning defaults after loading a checkpoint    |
+| `policy.vlm_num_layers`                       | `14`                        | Number of InternVL3 language layers kept for the policy           |
+| `policy.vlm_dtype`                            | `bfloat16`                  | Requested VLM dtype                                               |
+| `policy.use_flash_attn`                       | `true`                      | Requests FlashAttention when installed; otherwise falls back      |
+| `policy.enable_gradient_checkpointing`        | `true`                      | Enables checkpointing on supported InternVL3 modules              |
+| `policy.gradient_checkpointing_use_reentrant` | `false`                     | Reentrant setting passed to gradient checkpointing when supported |
+| `policy.chunk_size`                           | `50`                        | Number of future actions predicted per chunk                      |
+| `policy.n_action_steps`                       | `50`                        | Number of actions consumed from a sampled chunk                   |
+| `policy.max_state_dim`                        | `24`                        | State padding dimension                                           |
+| `policy.max_action_dim`                       | `24`                        | Action padding dimension                                          |
+| `policy.postprocess_action_dim`               | `null`                      | Optional action dimension returned after EVO1 postprocessing      |
+| `policy.binarize_gripper`                     | `false`                     | Binarizes the postprocessed gripper channel for LIBERO-style eval |
+| `policy.task_field`                           | `task`                      | Batch field used as the language prompt                           |
+
+## Inference
+
+Try it out with a trained EVO1 checkpoint:
+
+```bash
+lerobot-rollout \
+  --policy.path=your-org/your-evo1-checkpoint \
+  --inference.type=rtc \ # optional
+  ...
+```
+
+## Results
+
+### LIBERO Evaluation
+
+> [!NOTE]
+> Benchmark results for a `lerobot`-hosted LIBERO checkpoint trained with this implementation
+> will be added once training completes.
+
+The official EVO1 LIBERO rollout protocol uses the raw LIBERO camera feature names
+(`observation.images.agentview_image` and `observation.images.robot0_eye_in_hand_image`), replans every
+14 actions, and binarizes the gripper command before stepping the simulator. The EVO1 policy postprocessor
+can crop the padded 24D action back to the 7D LIBERO action space and apply that gripper binarization. To
+evaluate a LIBERO checkpoint under the same one-episode-per-task setting, keep the raw camera names instead
+of the default `image`/`image2` mapping and set the LIBERO action postprocessing flags:
+
+```bash
+lerobot-eval \
+  --policy.path=your-org/your-evo1-libero-checkpoint \
+  --policy.vlm_model_name=OpenGVLab/InternVL3-1B-hf \
+  --policy.device=cuda \
+  --policy.use_flash_attn=true \
+  --policy.n_action_steps=14 \
+  --policy.postprocess_action_dim=7 \
+  --policy.binarize_gripper=true \
+  --env.type=libero \
+  --env.task=libero_object \
+  --env.camera_name_mapping="{agentview_image: agentview_image, robot0_eye_in_hand_image: robot0_eye_in_hand_image}" \
+  --env.observation_height=448 \
+  --env.observation_width=448 \
+  --eval.batch_size=1 \
+  --eval.n_episodes=1
+```
+
+## References
+
+- [EVO1 repository](https://github.com/MINT-SJTU/Evo-1)
+- [InternVL3-1B-hf](https://huggingface.co/OpenGVLab/InternVL3-1B-hf)
+
+## License
+
+This LeRobot integration follows the Apache 2.0 License used by LeRobot. Check the upstream EVO1 and InternVL3 model pages for the licenses of released checkpoints and data.

docs/source/fastwam.mdx

@@ -0,0 +1,167 @@
+# FastWAM
+
+FastWAM is a World Action Model policy for robot control. The LeRobot integration exposes FastWAM through the standard policy API so it can be configured with `policy.type=fastwam`, trained with `lerobot-train`, and loaded through the LeRobot pretrained policy interface.
+
+## Model Overview
+
+FastWAM keeps video modeling during training, but uses direct action prediction at inference time instead of iteratively generating future observations. This LeRobot policy wraps the FastWAM action model, adapts LeRobot batches to FastWAM training samples, and provides the standard processor pipeline for normalization and action postprocessing.
+
+The implementation initializes the visual world-model components from `Wan-AI/Wan2.2-TI2V-5B` by default and predicts action chunks with shape `[batch, action_horizon, action_dim]`.
+
+### What the LeRobot Integration Covers
+
+- Standard `policy.type=fastwam` configuration through LeRobot
+- Image, state, action, and language-task batch adaptation
+- Action chunk inference through `select_action` and `predict_action_chunk`
+- Checkpoint save/load through the LeRobot policy APIs
+- Configurable LIBERO gripper action postprocessing
+
+## Installation Requirements
+
+Install LeRobot from source, then install FastWAM dependencies:
+
+```bash
+pip install -e ".[fastwam]"
+```
+
+This installs the FastWAM policy extra from `pyproject.toml`: `transformers`,
+`diffusers`, `ftfy`, and `regex`, plus LeRobot's base dependencies.
+
+For LIBERO evaluation, install the benchmark dependencies too:
+
+```bash
+pip install -e ".[fastwam,libero]"
+```
+
+This installs both extras. In addition to the FastWAM dependencies above, the
+`libero` extra installs LeRobot dataset dependencies, `hf-libero` on Linux, and
+`scipy`.
+
+FastWAM uses the Wan2.2 TI2V backbone. The default model id is:
+
+```python
+policy.model_id=Wan-AI/Wan2.2-TI2V-5B
+```
+
+## Data Requirements
+
+FastWAM expects a LeRobot dataset with:
+
+- one or more visual observations whose widths concatenate to `policy.image_size[1]`
+- `observation.state` when `policy.proprio_dim` is not `None`
+- `action`
+- a language task instruction through the dataset task field, or precomputed `context` and `context_mask` tensors
+
+The default visual setup is one image feature named `observation.images.image` with shape `(3, 224, 448)`. If the dataset uses two cameras, configure `policy.input_features` so their heights match `224` and their widths sum to `448`.
+
+## Usage
+
+Create a new FastWAM policy with:
+
+```bash
+lerobot-train \
+  --dataset.repo_id=your-org/your-dataset \
+  --policy.type=fastwam \
+  --policy.action_dim=7 \
+  --policy.proprio_dim=8 \
+  --policy.action_horizon=32 \
+  --policy.n_action_steps=10 \
+  --policy.image_size='[224,448]' \
+  --output_dir=./outputs/fastwam_training \
+  --job_name=fastwam_training \
+  --steps=300000 \
+  --batch_size=8 \
+  --policy.device=cuda
+```
+
+Evaluate an existing LeRobot-format checkpoint on LIBERO-10 with:
+
+```bash
+lerobot-eval \
+  --policy.path=ZibinDong/fastwam_libero_uncond_2cam224 \
+  --policy.device=cuda \
+  --policy.torch_dtype=float32 \
+  --policy.n_action_steps=10 \
+  --env.type=libero \
+  --env.task=libero_10 \
+  --env.observation_height=224 \
+  --env.observation_width=224 \
+  --eval.batch_size=1 \
+  --eval.n_episodes=50 \
+  --seed=0 \
+  --env.episode_length=600
+```
+
+For `libero_goal`, `libero_spatial`, and `libero_object`, use
+`--env.episode_length=300`.
+
+For real-robot rollout, use the same checkpoint path:
+
+```bash
+lerobot-rollout \
+  --robot.type=so101_follower \
+  --robot.port=/dev/ttyACM0 \
+  --policy.path=your-org/fastwam-real-robot
+```
+
+## Configuration Notes
+
+### Image Features
+
+`policy.image_size` is the size of the concatenated FastWAM image tensor as `(height, width)`. Each configured image feature must have shape `(3, height, camera_width)`, and all camera widths must sum to the configured width.
+
+### Action Chunking
+
+`policy.action_horizon` controls the number of future actions supervised during training and predicted during inference. `policy.n_action_steps` controls how many actions are consumed before the policy predicts a fresh chunk. `policy.n_action_steps` must be less than or equal to `policy.action_horizon`.
+
+### Wan Components
+
+FastWAM loads the Wan VAE, video DiT, text encoder, and tokenizer from the configured Wan model directory or Hugging Face Hub model id. LeRobot-format FastWAM checkpoints saved by `save_pretrained` also copy the local Wan component files needed by `from_pretrained`.
+
+### Attention Backend
+
+FastWAM's DiT uses PyTorch's `scaled_dot_product_attention` (SDPA) for all attention. It does **not** use FlashAttention: its Mixture-of-Transformers (MoT) routing needs arbitrary boolean `[query, key]` attention masks, which the FlashAttention varlen API cannot express. Installing the `flash-attn` package therefore has no effect on the FastWAM path. (Note that SDPA itself may still select PyTorch's own flash / memory-efficient / math kernel internally — this is unrelated to the `flash-attn` package.)
+
+### LIBERO Action Toggle
+
+FastWAM LIBERO checkpoints use `policy.toggle_action_dimensions=[-1]` by
+default to match the gripper action convention used by the original FastWAM
+evaluation pipeline:
+
+```bash
+--policy.toggle_action_dimensions='[-1]'
+```
+
+## Results
+
+Evaluated on LIBERO with [`ZibinDong/fastwam_libero_uncond_2cam224`](https://huggingface.co/ZibinDong/fastwam_libero_uncond_2cam224):
+
+| Suite          | Success rate | n_episodes |
+| -------------- | -----------: | ---------: |
+| libero_spatial |        97.6% |        500 |
+| libero_object  |        99.0% |        500 |
+| libero_goal    |        95.0% |        500 |
+| libero_10      |        94.0% |        500 |
+| **average**    |    **96.4%** |       2000 |
+
+Reproduce: `lerobot-eval --policy.path=ZibinDong/fastwam_libero_uncond_2cam224 --policy.device=cuda --policy.torch_dtype=float32 --policy.n_action_steps=10 --env.type=libero --env.task=libero_spatial --env.observation_height=256 --env.observation_width=256 --eval.batch_size=1 --eval.n_episodes=50 --seed=0 --env.episode_length=300` (1x H20 140 GB).
+
+## References
+
+- [Fast-WAM paper](https://arxiv.org/abs/2603.16666)
+- [Fast-WAM project page](https://yuantianyuan01.github.io/FastWAM/)
+- [Fast-WAM code](https://github.com/yuantianyuan01/FastWAM)
+- [Released upstream checkpoints](https://huggingface.co/yuanty/fastwam)
+- [Wan2.2 TI2V 5B](https://huggingface.co/Wan-AI/Wan2.2-TI2V-5B)
+
+## Citation
+
+```bibtex
+@article{yuan2026fastwam,
+  title = {Fast-WAM: Do World Action Models Need Test-time Future Imagination?},
+  author = {Tianyuan Yuan and Zibin Dong and Yicheng Liu and Hang Zhao},
+  journal = {arXiv preprint arXiv:2603.16666},
+  year = {2026},
+  url = {https://arxiv.org/abs/2603.16666}
+}
+```

docs/source/groot.mdx

@@ -1,16 +1,19 @@
-# GR00T N1.5 Policy
+# GR00T Policy
 
-GR00T N1.5 is an open foundation model from NVIDIA designed for generalized humanoid robot reasoning and skills. It is a cross-embodiment model that accepts multimodal input, including language and images, to perform manipulation tasks in diverse environments.
+GR00T is an NVIDIA foundation model family for generalized humanoid robot reasoning and skills. It is a cross-embodiment policy that accepts multimodal input, including language, images, and proprioception, to perform manipulation tasks in diverse environments.
 
-This document outlines the specifics of its integration and usage within the LeRobot framework.
+LeRobot integrates GR00T N1.7 through the `groot` policy type.
+
+> [!WARNING]
+> **Breaking change:** GR00T N1.5 support was removed from LeRobot, and current releases support GR00T N1.7 only. N1.5 checkpoints and configs are rejected with a migration note. To keep using an N1.5 checkpoint, pin the last release that supports it: `pip install 'lerobot==0.5.1'`. To use the current release, migrate to GR00T N1.7 (base model [`nvidia/GR00T-N1.7-3B`](https://huggingface.co/nvidia/GR00T-N1.7-3B)).
 
 ## Model Overview
 
-NVIDIA Isaac GR00T N1.5 is an upgraded version of the GR00T N1 foundation model. It is built to improve generalization and language-following abilities for humanoid robots.
+GR00T N1.7 uses a Cosmos-Reason2/Qwen3-VL backbone and provides checkpoints for SimplerEnv, DROID, and LIBERO.
 
-Developers and researchers can post-train GR00T N1.5 with their own real or synthetic data to adapt it for specific humanoid robots or tasks.
+Developers and researchers can post-train GR00T with their own real or synthetic data to adapt it for specific humanoid robots or tasks.
 
-GR00T N1.5 (specifically the GR00T-N1.5-3B model) is built using pre-trained vision and language encoders. It utilizes a flow matching action transformer to model a chunk of actions, conditioned on vision, language, and proprioception.
+GR00T uses pre-trained vision and language encoders with a flow matching action transformer to model a chunk of actions conditioned on vision, language, and proprioception.
 
 <img
   src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/lerobot/lerobot-groot-paper1%20(1).png"
@@ -28,33 +31,24 @@ This approach allows the model to be highly adaptable through post-training for
 
 ## Installation Requirements
 
-As of today, GR00T N1.5 requires flash attention for it's internal working.
-
-We are working on making this optional, but in the meantime that means that we require an extra installation step and it can only be used in CUDA enabled devices.
-
-1. Following the Environment Setup of our [Installation Guide](./installation). **Attention** don't install `lerobot` in this step.
-2. Install [Flash Attention](https://github.com/Dao-AILab/flash-attention) by running:
+GR00T is intended for NVIDIA GPU-accelerated systems. Install LeRobot with the GR00T extra:
 
 ```bash
-# Check https://pytorch.org/get-started/locally/ for your system
-pip install "torch>=2.2.1,<2.8.0" "torchvision>=0.21.0,<0.23.0" # --index-url https://download.pytorch.org/whl/cu1XX
-pip install ninja "packaging>=24.2,<26.0" # flash attention dependencies
-pip install "flash-attn>=2.5.9,<3.0.0" --no-build-isolation
-python -c "import flash_attn; print(f'Flash Attention {flash_attn.__version__} imported successfully')"
+pip install "lerobot[groot]"
 ```
 
-3. Install LeRobot by running:
+For a source checkout:
 
 ```bash
-pip install lerobot[groot]
+pip install -e ".[groot]"
 ```
 
 ## Usage
 
-To use GR00T in your LeRobot configuration, specify the policy type as:
+To use GR00T N1.7:
 
-```python
-policy.type=groot
+```bash
+--policy.type=groot
 ```
 
 ## Training
@@ -63,72 +57,171 @@ policy.type=groot
 
 Here's a complete training command for finetuning the base GR00T model on your own dataset:
 
+This command is using the `new_embodiment` flag, which is used for the SO-101 robot, [read more about how GR00T handles different embodiments.](https://github.com/NVIDIA/Isaac-GR00T/blob/main/getting_started/policy.md#--embodiment-tag).
+
 ```bash
-# Using a multi-GPU setup
-accelerate launch \
-  --multi_gpu \
-  --num_processes=$NUM_GPUS \
-  $(which lerobot-train) \
-  --output_dir=$OUTPUT_DIR \
-  --save_checkpoint=true \
-  --batch_size=$BATCH_SIZE \
-  --steps=$NUM_STEPS \
-  --save_freq=$SAVE_FREQ \
-  --log_freq=$LOG_FREQ \
-  --policy.push_to_hub=true \
+# install extra deps for training
+pip install "lerobot[training]"
+
+hf auth login
+wandb login
+
+export DATASET_NAME=your_data_set
+export HF_USER=your_hf_username
+export DATASET=$HF_USER/$DATASET_NAME
+export REPO_ID="${DATASET}_GR00T17" #this is the model that will be uploaded to huggingface
+export OUTPUT_DIR=outputs/train/$REPO_ID
+
+lerobot-train \
+  --dataset.repo_id=$DATASET \
+  --dataset.image_transforms.enable=true \
   --policy.type=groot \
+  --policy.device=cuda \
+  --policy.base_model_path=nvidia/GR00T-N1.7-3B \
+  --policy.embodiment_tag=new_embodiment \
+  --policy.chunk_size=16 \
+  --policy.n_action_steps=16 \
+  --policy.use_relative_actions=true \
+  --policy.relative_exclude_joints='["gripper"]' \
+  --policy.use_bf16=true \
+  --policy.push_to_hub=true \
   --policy.repo_id=$REPO_ID \
-  --policy.tune_diffusion_model=false \
-  --dataset.repo_id=$DATASET_ID \
+  --seed=42 \
+  --batch_size=64 \
+  --steps=20000 \
+  --save_checkpoint=true \
+  --save_freq=5000 \
+  --use_policy_training_preset=true \
+  --env_eval_freq=0 \
+  --eval_steps=0 \
+  --log_freq=10 \
+  --output_dir=$OUTPUT_DIR \
+  --job_name=$DATASET \
   --wandb.enable=true \
-  --wandb.disable_artifact=true \
-  --job_name=$JOB_NAME
+  --wandb.disable_artifact=true
+
 ```
 
 ## Performance Results
 
-### Libero Benchmark Results
+### LIBERO Benchmark Results
 
 > [!NOTE]
-> Follow our instructions for Libero usage: [Libero](./libero)
+> Follow the [LIBERO](./libero) setup instructions before running `lerobot-eval`.
 
-GR00T has demonstrated strong performance on the Libero benchmark suite. To compare and test its LeRobot implementation, we finetuned the GR00T N1.5 model for 30k steps on the Libero dataset and compared the results to the GR00T reference results.
+GR00T N1.7 has demonstrated strong performance on the LIBERO benchmark suite. To reproduce LeRobot results, follow the instructions in the [LIBERO](./libero) section.
 
-| Benchmark          | LeRobot Implementation | GR00T Reference |
-| ------------------ | ---------------------- | --------------- |
-| **Libero Spatial** | 82.0%                  | 92.0%           |
-| **Libero Object**  | 99.0%                  | 92.0%           |
-| **Libero Long**    | 82.0%                  | 76.0%           |
-| **Average**        | 87.0%                  | 87.0%           |
+### Train on LIBERO
 
-These results demonstrate GR00T's strong generalization capabilities across diverse robotic manipulation tasks. To reproduce these results, you can follow the instructions in the [Libero](https://huggingface.co/docs/lerobot/libero) section.
+Example training command for a LIBERO suite (here `libero_spatial`):
+
+```bash
+IMAGE_TRANSFORMS='{
+  "brightness": {"weight": 1.0, "type": "ColorJitter", "kwargs": {"brightness": [0.7, 1.3]}},
+  "contrast":   {"weight": 1.0, "type": "ColorJitter", "kwargs": {"contrast":   [0.6, 1.4]}},
+  "saturation": {"weight": 1.0, "type": "ColorJitter", "kwargs": {"saturation": [0.5, 1.5]}},
+  "hue":        {"weight": 1.0, "type": "ColorJitter", "kwargs": {"hue":        [-0.08, 0.08]}}
+}'
+
+lerobot-train \
+  --dataset.repo_id=IPEC-COMMUNITY/libero_spatial_no_noops_1.0.0_lerobot \
+  --dataset.root=/datasets/libero_spatial \
+  --dataset.revision=main \
+  --dataset.video_backend=pyav \
+  --dataset.image_transforms.enable=true \
+  --dataset.image_transforms.max_num_transforms=4 \
+  --dataset.image_transforms.tfs="$IMAGE_TRANSFORMS" \
+  --policy.type=groot \
+  --policy.base_model_path=nvidia/GR00T-N1.7-3B \
+  --policy.embodiment_tag=libero_sim \
+  --policy.push_to_hub=false \
+  --policy.use_relative_actions=false \
+  --policy.max_steps=20000 \
+  --batch_size=320 \
+  --steps=20000 \
+  --save_freq=2000 \
+  --env_eval_freq=0 \
+  --eval_steps=0 \
+  --log_freq=10 \
+  --wandb.enable=true \
+  --wandb.project=lerobot \
+  --wandb.mode=online \
+  --wandb.disable_artifact=true \
+  --num_workers=4 \
+  --prefetch_factor=2 \
+  --persistent_workers=true \
+  --output_dir=$OUTPUT_DIR \
+  --job_name=$JOB_NAME
+```
+
+This will follow the recipe found [here](https://github.com/NVIDIA/Isaac-GR00T/blob/main/examples/LIBERO/README.md).
+
+### GR00T N1.7 LIBERO Results
+
+Preliminary LeRobot integration results (GR00T-LeRobot, `eval.n_episodes >= 50` per suite):
+
+| Suite            | Success rate | Checkpoint                                                                                                    |
+| ---------------- | -----------: | ------------------------------------------------------------------------------------------------------------- |
+| LIBERO Spatial   |          91% | [nvidia/gr00t17-lerobot-libero_spatial-640](https://huggingface.co/nvidia/gr00t17-lerobot-libero_spatial-640) |
+| LIBERO Object    |          81% | [nvidia/gr00t17-lerobot-libero_object-640](https://huggingface.co/nvidia/gr00t17-lerobot-libero_object-640)   |
+| LIBERO Goal      |          97% | [nvidia/gr00t17-lerobot-libero_goal-640](https://huggingface.co/nvidia/gr00t17-lerobot-libero_goal-640)       |
+| LIBERO 10 (Long) |          84% | [nvidia/gr00t17-lerobot-libero_10-640](https://huggingface.co/nvidia/gr00t17-lerobot-libero_10-640)           |
+| **Average**      |   **88.25%** |                                                                                                               |
+
+```bash
+export MODEL_ID=your_trained_model_on_huggingface
+
+lerobot-eval \
+  --policy.type=groot \
+  --policy.base_model_path=$MODEL_ID \
+  --policy.embodiment_tag=libero_sim \
+  --env.type=libero \
+  --env.task=libero_spatial \
+  --eval.n_episodes=50
+```
+
+Use `eval.n_episodes >= 50` per suite when reporting success rates.
 
 ### Evaluate in your hardware setup
 
-Once you have trained your model using your parameters you can run inference in your downstream task. Follow the instructions in [Imitation Learning for Robots](./il_robots). For example:
+Once you have trained your model using your parameters you can run inference in your downstream task. Follow the instructions in [Policy Deployment (lerobot-rollout)](./inference). For example:
 
 ```bash
-lerobot-record \
-  --robot.type=bi_so_follower \
-  --robot.left_arm_port=/dev/ttyACM1 \
-  --robot.right_arm_port=/dev/ttyACM0 \
-  --robot.id=bimanual_follower \
-  --robot.cameras='{ right: {"type": "opencv", "index_or_path": 0, "width": 640, "height": 480, "fps": 30},
-    left: {"type": "opencv", "index_or_path": 2, "width": 640, "height": 480, "fps": 30},
-    top: {"type": "opencv", "index_or_path": 4, "width": 640, "height": 480, "fps": 30},
-  }' \
+# install extra deps for roullout and real hardware
+pip install "lerobot[feetech,viz]"
+
+export MODEL_ID=your_trained_model_on_huggingface
+
+# make sure that camera index matches your setup!
+# find index using `uv run lerobot-find-cameras opencv`
+WRIST_CAM='wrist: {type: opencv, index_or_path: 2, width: 640, height: 480, fps: 30, fourcc: "MJPG"}'
+FRONT_CAM='front: {type: opencv, index_or_path: 0, width: 640, height: 480, fps: 30, fourcc: "MJPG"}'
+export ROBOT_CAMERAS="{ $WRIST_CAM, $FRONT_CAM }"
+export ROBOT_ID=follower_robot
+export ROBOT_PORT=/dev/ttyACM0
+
+uv run lerobot-rollout \
+  --strategy.type=base \
+  --policy.path=$MODEL_ID \
+  --policy.base_model_path=nvidia/GR00T-N1.7-3B \
+  --policy.n_action_steps=8 \
+  --robot.type=so101_follower \
+  --robot.port=$ROBOT_PORT \
+  --robot.id=$ROBOT_ID \
+  --robot.cameras="$ROBOT_CAMERAS" \
+  --task="place the vial in the rack" \
+  --duration=60 \
+  --device=cuda \
   --display_data=true \
-  --dataset.repo_id=<user>/eval_groot-bimanual  \
-  --dataset.num_episodes=10 \
-  --dataset.single_task="Grab and handover the red cube to the other arm" \
-  --dataset.streaming_encoding=true \
-  --dataset.encoder_threads=2 \
-  # --dataset.vcodec=auto \
-  --policy.path=<user>/groot-bimanual \ # your trained model
-  --dataset.episode_time_s=30 \
-  --dataset.reset_time_s=10
+  --inference.type=rtc \
+  --inference.rtc.enabled=True \ # set to False if it causes inference instability
+  --inference.rtc.execution_horizon=8 \
+  --inference.queue_threshold=0
 ```
 
+> [!NOTE]
+> Value of `inference.queue_threshold` should not exceed 5 to ensure stable inference.
+
 ## License
 
-This model follows NVIDIA's proprietary license, consistent with the original [GR00T repository](https://github.com/NVIDIA/Isaac-GR00T). Future versions (starting from N1.7) will follow **Apache 2.0 License**.
+GR00T N1.7 is released under the [NVIDIA Open Model License Agreement](https://www.nvidia.com/en-us/agreements/enterprise-software/nvidia-open-model-license/).

docs/source/hardware_guide.mdx

@@ -0,0 +1,99 @@
+# Compute HW Guide for LeRobot Training
+
+Rough sizing for training a LeRobot policy: how much VRAM each policy needs, what training time looks like, and where to run when local hardware isn't enough.
+
+The numbers below are **indicative** — order-of-magnitude figures for picking hardware, not exact predictions. Throughput depends heavily on dataset I/O, image resolution, batch size, and number of GPUs.
+
+## Memory by policy group
+
+Policies cluster by backbone size; the groupings below give a single VRAM envelope per group instead of repeating numbers per policy. Memory scales roughly linearly with batch size; AdamW (the LeRobot default) carries optimizer state that adds ~30–100% over a forward+backward pass alone.
+
+| Group      | Policies                                    | Peak VRAM (BS 8, AdamW) | Suitable starter GPUs             |
+| ---------- | ------------------------------------------- | ----------------------: | --------------------------------- |
+| Light BC   | `act`, `vqbet`, `tdmpc`                     |                  ~2–6GB | Laptop GPU (RTX 3060), L4, A10G   |
+| Diffusion  | `diffusion`, `multi_task_dit`               |                 ~8–14GB | RTX 4070+ / L4 / A10G             |
+| Small VLA  | `smolvla`                                   |                ~10–16GB | RTX 4080+ / L4 / A10G             |
+| Large VLA  | `pi0`, `pi0_fast`, `pi05`, `xvla`, `wall_x` |                ~24–40GB | A100 40 GB+ (24 GB tight at BS 1) |
+| Multimodal | `groot`, `eo1`                              |                ~24–40GB | A100 40 GB+                       |
+| RL         | `sac`                                       |             config-dep. | See [HIL-SERL guide](./hilserl)   |
+
+Memory-bound? Drop the batch size (~linear), use gradient accumulation to recover effective batch, or for SmolVLA leave `freeze_vision_encoder=True`.
+
+## Training time
+
+Robotics imitation learning typically converges in **5–10 epochs over the dataset**, not hundreds of thousands of raw steps. Once you know your epoch count, wall-clock is essentially:
+
+```text
+total_frames    = sum of frames over all episodes      # 50 ep × 30 fps × 30 s ≈ 45,000
+steps_per_epoch = ceil(total_frames / (num_gpus × batch_size))
+total_steps     = epochs × steps_per_epoch
+wall_clock      ≈ total_steps × per_step_time
+```
+
+Per-step time depends on the policy and the GPU. The numbers in the table below are anchors — pick the row closest to your setup and scale linearly with `total_steps` if you train longer or shorter.
+
+### Common scenarios
+
+Indicative wall-clock for **5 epochs on a ~50-episode dataset (~45k frames at 30 fps × 30 s)**, default optimizer (AdamW), 640×480 images:
+
+| Setup                                | Policy         | Batch | Wall-clock |
+| ------------------------------------ | -------------- | ----- | ---------: |
+| Single RTX 4090 / RTX 3090 (24 GB)   | `act`          | 8     |  ~30–60min |
+| Single RTX 4090 / RTX 3090 (24 GB)   | `diffusion`    | 8     |      ~2–4h |
+| Single L4 / A10G (24 GB)             | `act`          | 8     |      ~1–2h |
+| Single L4 / A10G (24 GB)             | `smolvla`      | 4     |      ~3–6h |
+| Single A100 40 GB                    | `smolvla`      | 16    |      ~1–2h |
+| Single A100 40 GB                    | `pi0` / `pi05` | 4     |      ~4–8h |
+| 4× H100 80 GB cluster (`accelerate`) | `diffusion`    | 32    |  ~30–60min |
+| 4× H100 80 GB cluster (`accelerate`) | `smolvla`      | 32    |      ~1–2h |
+| Apple Silicon M1/M2/M3 Max (MPS)     | `act`          | 4     |     ~6–14h |
+
+These are order-of-magnitude figures. Real runs deviate by ±50% depending on image resolution, dataset I/O, dataloader threading, and exact GPU SKU. They are useful as "is this run going to take an hour or a day?" intuition, not as SLAs.
+
+### Multi-GPU matters a lot
+
+`accelerate launch --num_processes=N` is the easiest way to cut training time. Each optimizer step processes `N × batch_size` samples in roughly the same wall-clock as a single-GPU step, so 4 GPUs ≈ 4× speedup for compute-bound runs. See the [Multi GPU training](./multi_gpu_training) guide for the full setup.
+
+Reference data points on a 4×H100 80 GB cluster (`accelerate launch --num_processes=4`), 5000 steps, batch 32, AdamW, dataset [`imstevenpmwork/super_poulain_draft`](https://huggingface.co/datasets/imstevenpmwork/super_poulain_draft) (~50 episodes, ~640×480 images):
+
+| Policy      | Wall-clock | `update_s` | `dataloading_s` | GPU util | Notable flags                                                                                                                  |
+| ----------- | ---------- | ---------: | --------------: | -------- | ------------------------------------------------------------------------------------------------------------------------------ |
+| `diffusion` | 16m 17s    |      0.167 |           0.015 | ~90%     | defaults (training from scratch)                                                                                               |
+| `smolvla`   | 27m 49s    |      0.312 |           0.011 | ~80%     | `--policy.path=lerobot/smolvla_base`, `freeze_vision_encoder=false`, `train_expert_only=false`                                 |
+| `pi05`      | 3h 41m     |      2.548 |           0.014 | ~95%     | `--policy.pretrained_path=lerobot/pi05_base`, `gradient_checkpointing=true`, `dtype=bfloat16`, vision encoder + expert trained |
+
+The `dataloading_s` vs. `update_s` ratio is the diagnostic that matters: when `dataloading_s` approaches `update_s`, more GPUs stop helping — your dataloader is the bottleneck and you should look at `--num_workers`, image resolution, and disk speed before adding compute.
+
+### Schedule and checkpoints
+
+If you shorten training (e.g. 5k–10k steps on a small dataset), also shorten the LR schedule with `--policy.scheduler_decay_steps≈--steps`. Otherwise the LR stays near its peak and never decays. Same for `--save_freq`.
+
+## Where to run
+
+VRAM is the first filter. Within a tier, pick by budget and availability — the `$`–`$$$$` columns are relative; check current pricing on the provider you actually use.
+
+| Class                      | VRAM  | Tier   | Comfortable for                                             |
+| -------------------------- | ----- | ------ | ----------------------------------------------------------- |
+| RTX 3090 / 4090 (consumer) | 24 GB | `$`    | Light BC, Diffusion, SmolVLA. Tight for VLAs at batch 1.    |
+| L4 / A10G (cloud)          | 24 GB | `$–$$` | Same envelope; common on Google Cloud, RunPod, AWS `g5/g6`. |
+| A100 40 GB                 | 40 GB | `$$$`  | Any policy at reasonable batch sizes.                       |
+| A100 80 GB / H100 80 GB    | 80 GB | `$$$$` | Multi-GPU clusters; large batches for VLAs.                 |
+| **CPU only**               | —     | —      | Don't train. Use Colab or rent a GPU.                       |
+
+### Hugging Face Jobs
+
+[Hugging Face Jobs](https://huggingface.co/docs/hub/jobs) lets you run training on managed HF infrastructure, billed by the second, without owning a GPU. `lerobot-train` submits and streams the job for you — just add `--job.target=<flavor>` to a normal training command:
+
+```bash
+lerobot-train \
+  --policy.type=act --dataset.repo_id=<USER>/<DATASET> \
+  --policy.repo_id=<USER>/act_<task> \
+  --job.target=a10g-large
+```
+
+Notes:
+
+- Run `hf auth login` once before submitting, the job runs under your token.
+- `--job.target` maps onto the table above: `t4-small`/`t4-medium` (T4, ACT only), `l4x1`/`l4x4` (L4 24 GB), `a10g-small/large/largex2/largex4` (A10G 24 GB scaled out), `a100-large` (A100). List the current catalogue with pricing via `hf jobs hardware`, or see [https://huggingface.co/docs/hub/jobs](https://huggingface.co/docs/hub/jobs).
+- The job defaults to a `2d` (48h) timeout. Override it with `--job.timeout=4h` (or any other valid duration string) to shorten or extend the timeout. The job automatically stops when the command completes.
+- For the full walkthrough — dataset upload, checkpoint streaming, resuming a run on a job — see the [imitation-learning training guide](./il_robots#train-using-hugging-face-jobs).

docs/source/hil_data_collection.mdx

@@ -57,11 +57,11 @@ The `lerobot-rollout --strategy.type=dagger` mode requires **teleoperators with
 
 **Compatible teleoperators:**
 
-- `openarm_mini` - OpenArm Mini
+- `bi_openarm_mini` - Bimanual OpenArm Mini
 - `so_leader` - SO100 / SO101 leader arm
 
 > [!IMPORTANT]
-> The provided commands default to `bi_openarm_follower` + `openarm_mini`.
+> The provided commands default to `bi_openarm_follower` + `bi_openarm_mini`.
 > `so_follower` + `so_leader` configs are also registered and can be used via CLI flags.
 
 ---
@@ -104,9 +104,9 @@ lerobot-rollout --strategy.type=dagger \
     --robot.right_arm_config.port=can0 \
     --robot.right_arm_config.side=right \
     --robot.cameras='{left_wrist: {type: opencv, index_or_path: "/dev/video0", width: 1280, height: 720, fps: 30}, right_wrist: {type: opencv, index_or_path: "/dev/video4", width: 1280, height: 720, fps: 30}, base: {type: opencv, index_or_path: "/dev/video2", width: 640, height: 480, fps: 30}}' \
-    --teleop.type=openarm_mini \
-    --teleop.port_left=/dev/ttyACM0 \
-    --teleop.port_right=/dev/ttyACM1 \
+    --teleop.type=bi_openarm_mini \
+    --teleop.left_arm_config.port=/dev/ttyACM0 \
+    --teleop.right_arm_config.port=/dev/ttyACM1 \
     --policy.path=outputs/pretrain/checkpoints/last/pretrained_model \
     --dataset.repo_id=your-username/rollout_hil_dataset \
     --dataset.single_task="Fold the T-shirt properly" \
@@ -131,9 +131,9 @@ lerobot-rollout --strategy.type=dagger \
     --robot.right_arm_config.port=can0 \
     --robot.right_arm_config.side=right \
     --robot.cameras='{left_wrist: {type: opencv, index_or_path: "/dev/video0", width: 1280, height: 720, fps: 30}, right_wrist: {type: opencv, index_or_path: "/dev/video4", width: 1280, height: 720, fps: 30}, base: {type: opencv, index_or_path: "/dev/video2", width: 640, height: 480, fps: 30}}' \
-    --teleop.type=openarm_mini \
-    --teleop.port_left=/dev/ttyACM0 \
-    --teleop.port_right=/dev/ttyACM1 \
+    --teleop.type=bi_openarm_mini \
+    --teleop.left_arm_config.port=/dev/ttyACM0 \
+    --teleop.right_arm_config.port=/dev/ttyACM1 \
     --policy.path=outputs/pretrain/checkpoints/last/pretrained_model \
     --dataset.repo_id=your-username/rollout_hil_rtc_dataset \
     --dataset.single_task="Fold the T-shirt properly" \

docs/source/hilserl.mdx

@@ -62,7 +62,7 @@ pip install -e ".[hilserl]"
 
 ### Understanding Configuration
 
-The training process begins with proper configuration for the HILSerl environment. The main configuration class is `GymManipulatorConfig` in `lerobot/rl/gym_manipulator.py`, which contains nested `HILSerlRobotEnvConfig` and `DatasetConfig`. The configuration is organized into focused, nested sub-configs:
+The training process begins with proper configuration for the HILSERl environment. The main configuration class is `GymManipulatorConfig` in `lerobot/rl/gym_manipulator.py`, which contains nested `HILSerlRobotEnvConfig` (defined in `lerobot/envs/configs.py`) and `DatasetConfig`. The configuration is organized into focused, nested sub-configs:
 
 <!-- prettier-ignore-start -->
 ```python
@@ -95,6 +95,7 @@ class HILSerlProcessorConfig:
 class ObservationConfig:
     add_joint_velocity_to_observation: bool = False    # Add joint velocities to state
     add_current_to_observation: bool = False    # Add motor currents to state
+    add_ee_pose_to_observation: bool = False    # Add end-effector pose to state
     display_cameras: bool = False    # Display camera feeds during execution
 
 class ImagePreprocessingConfig:
@@ -326,14 +327,22 @@ lerobot-find-joint-limits \
    Max joint positions [-20.0, -20.0, -20.0, -20.0, -20.0, -20.0]
    Min joint positions [50.0, 50.0, 50.0, 50.0, 50.0, 50.0]
    ```
-3. Use these values in the configuration of your teleoperation device (TeleoperatorConfig) under the `end_effector_bounds` field
+3. Use these values in your environment configuration under `env.processor.inverse_kinematics.end_effector_bounds` (see `InverseKinematicsConfig` in `lerobot/envs/configs.py`)
 
 **Example Configuration**
 
 ```json
-"end_effector_bounds": {
-    "max": [0.24, 0.20, 0.10],
-    "min": [0.16, -0.08, 0.03]
+{
+  "env": {
+    "processor": {
+      "inverse_kinematics": {
+        "end_effector_bounds": {
+          "max": [0.24, 0.2, 0.1],
+          "min": [0.16, -0.08, 0.03]
+        }
+      }
+    }
+  }
 }
 ```
 
@@ -404,30 +413,24 @@ We support using a gamepad or a keyboard or the leader arm of the robot.
 
 HIL-Serl learns actions in the end-effector space of the robot. Therefore, the teleoperation will control the end-effector's x,y,z displacements.
 
-For that we need to define a version of the robot that takes actions in the end-effector space. Check the robot class `SO100FollowerEndEffector` and its configuration `SO100FollowerEndEffectorConfig` for the default parameters related to the end-effector space.
+The end-effector transformation is applied by the processor pipeline (`InverseKinematicsRLStep`, `EEBoundsAndSafety`, `EEReferenceAndDelta`, `GripperVelocityToJoint`) configured under `env.processor.inverse_kinematics` (`InverseKinematicsConfig`) and `env.processor.gripper` / `env.processor.max_gripper_pos`. The defaults related to the end-effector space are:
 
 <!-- prettier-ignore-start -->
 ```python
-class SO100FollowerEndEffectorConfig(SO100FollowerConfig):
-    """Configuration for the SO100FollowerEndEffector robot."""
+class InverseKinematicsConfig:
+    """Configuration for inverse kinematics processing."""
 
-    # Default bounds for the end-effector position (in meters)
-    end_effector_bounds: dict[str, list[float]] = field( # bounds for the end-effector in x,y,z direction
-        default_factory=lambda: {
-            "min": [-1.0, -1.0, -1.0],  # min x, y, z
-            "max": [1.0, 1.0, 1.0],  # max x, y, z
-        }
-    )
+    urdf_path: str | None = None
+    target_frame_name: str | None = None
+    # bounds for the end-effector in x,y,z direction
+    end_effector_bounds: dict[str, list[float]] | None = None
+    # maximum step size for the end-effector in x,y,z direction
+    end_effector_step_sizes: dict[str, float] | None = None
 
-    max_gripper_pos: float = 50 # maximum gripper position that the gripper will be open at
-
-    end_effector_step_sizes: dict[str, float] = field( # maximum step size for the end-effector in x,y,z direction
-        default_factory=lambda: {
-            "x": 0.02,
-            "y": 0.02,
-            "z": 0.02,
-        }
-    )
+class HILSerlProcessorConfig:
+    ...
+    # maximum gripper position that the gripper will be open at
+    max_gripper_pos: float | None = 100.0
 ```
 <!-- prettier-ignore-end -->
 
@@ -606,11 +609,11 @@ This guide explains how to train a reward classifier for human-in-the-loop reinf
 
 **Note**: Training a reward classifier is optional. You can start the first round of RL experiments by annotating the success manually with your gamepad or keyboard device.
 
-The reward classifier implementation in `modeling_classifier.py` uses a pretrained vision model to process the images. It can output either a single value for binary rewards to predict success/fail cases or multiple values for multi-class settings.
+The reward classifier implementation in `lerobot/rewards/classifier/modeling_classifier.py` uses a pretrained vision model to process the images. It can output either a single value for binary rewards to predict success/fail cases or multiple values for multi-class settings.
 
 **Collecting a Dataset for the reward classifier**
 
-Before training, you need to collect a dataset with labeled examples. The `record_dataset` function in `gym_manipulator.py` enables the process of collecting a dataset of observations, actions, and rewards.
+Before training, you need to collect a dataset with labeled examples. Setting `mode: "record"` in your config and running `gym_manipulator.py` enables the process of collecting a dataset of observations, actions, and rewards.
 
 To collect a dataset, you need to modify some parameters in the environment configuration based on HILSerlRobotEnvConfig.
 
@@ -658,7 +661,7 @@ Example configuration section for data collection:
   },
   "dataset": {
     "repo_id": "hf_username/dataset_name",
-    "dataset_root": "data/your_dataset",
+    "root": "data/your_dataset",
     "task": "reward_classifier_task",
     "num_episodes_to_record": 20,
     "replay_episode": null,
@@ -671,7 +674,7 @@ Example configuration section for data collection:
 
 **Reward Classifier Configuration**
 
-The reward classifier is configured using `configuration_classifier.py`. Here are the key parameters:
+The reward classifier is configured using `lerobot/rewards/classifier/configuration_classifier.py`. Here are the key parameters:
 
 - **model_name**: Base model architecture (e.g., we mainly use `"helper2424/resnet10"`)
 - **model_type**: `"cnn"` or `"transformer"`
@@ -689,7 +692,7 @@ Example configuration for training the [reward classifier](https://huggingface.c
     "repo_id": "hf_username/dataset_name",
     "root": null
   },
-  "policy": {
+  "reward_model": {
     "type": "reward_classifier",
     "model_name": "helper2424/resnet10",
     "model_type": "cnn",
@@ -699,7 +702,6 @@ Example configuration for training the [reward classifier](https://huggingface.c
     "dropout_rate": 0.1,
     "learning_rate": 1e-4,
     "device": "cuda",
-    "use_amp": true,
     "input_features": {
       "observation.images.front": {
         "type": "VISUAL",
@@ -717,7 +719,7 @@ Example configuration for training the [reward classifier](https://huggingface.c
   "num_workers": 4,
   "steps": 5000,
   "log_freq": 10,
-  "eval_freq": 1000,
+  "env_eval_freq": 1000,
   "save_freq": 1000,
   "save_checkpoint": true,
   "seed": 2,
@@ -818,13 +820,14 @@ The LeRobot system uses a distributed actor-learner architecture for training. T
 
 **Configuration Setup**
 
-Create a training configuration file (example available [here](https://huggingface.co/datasets/lerobot/config_examples/resolve/main/rl/train_config.json)). The training config is based on the main `TrainRLServerPipelineConfig` class in `lerobot/configs/train.py`.
+Create a training configuration file (example available [here](https://huggingface.co/datasets/lerobot/config_examples/resolve/main/rl/train_config.json)). The training config is based on the main `TrainRLServerPipelineConfig` class in `lerobot/rl/train_rl.py`.
 
-1. Configure the policy settings (`type="sac"`, `device`, etc.)
-2. Set `dataset` to your cropped dataset
-3. Configure environment settings with crop parameters
-4. Check the other parameters related to SAC in [configuration_sac.py](https://github.com/huggingface/lerobot/blob/main/src/lerobot/policies/sac/configuration_sac.py#L79).
-5. Verify that the `policy` config is correct with the right `input_features` and `output_features` for your task.
+1. Configure the policy settings (`type="gaussian_actor"`, `device`, etc.)
+2. Configure the algorithm settings under the top-level `algorithm` block (`type="sac"`, learning rates, discount, etc., defined in `lerobot/rl/algorithms/sac/configuration_sac.py`).
+3. Set `dataset` to your cropped dataset
+4. Configure environment settings with crop parameters
+5. Check the other parameters related to the Gaussian Actor in [configuration_gaussian_actor.py](https://github.com/huggingface/lerobot/blob/main/src/lerobot/policies/gaussian_actor/configuration_gaussian_actor.py#L79).
+6. Verify that the `policy` config is correct with the right `input_features` and `output_features` for your task.
 
 **Starting the Learner**
 
@@ -926,7 +929,7 @@ The ideal behaviour is that your intervention rate should drop gradually during
 
 Some configuration values have a disproportionate impact on training stability and speed:
 
-- **`temperature_init`** (`policy.temperature_init`) – initial entropy temperature in SAC. Higher values encourage more exploration; lower values make the policy more deterministic early on. A good starting point is `1e-2`. We observed that setting it too high can make human interventions ineffective and slow down learning.
+- **`temperature_init`** (`algorithm.temperature_init`) – initial entropy temperature in SAC. Higher values encourage more exploration; lower values make the policy more deterministic early on. A good starting point is `1e-2`. We observed that setting it too high can make human interventions ineffective and slow down learning.
 - **`policy_parameters_push_frequency`** (`policy.actor_learner_config.policy_parameters_push_frequency`) – interval in _seconds_ between two weight pushes from the learner to the actor. The default is `4 s`. Decrease to **1-2 s** to provide fresher weights (at the cost of more network traffic); increase only if your connection is slow, as this will reduce sample efficiency.
 - **`storage_device`** (`policy.storage_device`) – device on which the learner keeps the policy parameters. If you have spare GPU memory, set this to `"cuda"` (instead of the default `"cpu"`). Keeping the weights on-GPU removes CPU→GPU transfer overhead and can significantly increase the number of learner updates per second.
 

docs/source/hope_jr.mdx

@@ -117,6 +117,14 @@ middle_dip           |   1484 |   1500 |   1547
 
 Once calibration is complete, the system will save the calibration to `/Users/your_username/.cache/huggingface/lerobot/calibration/teleoperators/homunculus_glove/red.json`
 
+#### Visualizing Teleoperator Glove
+
+After calibration, you can visualize the glove movements in real-time. Open the visualizer by navigating to the visualizer directory and opening the HTML file in your browser:
+
+```bash
+open examples/hopejr/visualizer/index.html
+```
+
 ### 1.3 Calibrate Robot Arm
 
 ```bash
@@ -232,7 +240,7 @@ lerobot-record \
     --dataset.private=true \
     --dataset.streaming_encoding=true \
     --dataset.encoder_threads=2 \
-    # --dataset.vcodec=auto \
+    # --dataset.rgb_encoder.vcodec=auto \
     --display_data=true
 ```
 
@@ -278,6 +286,6 @@ lerobot-record \
   --dataset.num_episodes=10 \
   --dataset.streaming_encoding=true \
   --dataset.encoder_threads=2 \
-  # --dataset.vcodec=auto \
+  # --dataset.rgb_encoder.vcodec=auto \
   --policy.path=outputs/train/hopejr_hand/checkpoints/last/pretrained_model
 ```

docs/source/il_robots.mdx

@@ -68,13 +68,13 @@ from lerobot.teleoperators.so_leader import SO101Leader, SO101LeaderConfig
 from lerobot.robots.so_follower import SO101Follower, SO101FollowerConfig
 
 robot_config = SO101FollowerConfig(
-    port="/dev/tty.usbmodem58760431541",
-    id="my_red_robot_arm",
+    port="/dev/tty.usbmodem5AB90687491",
+    id="my_follower_arm",
 )
 
 teleop_config = SO101LeaderConfig(
-    port="/dev/tty.usbmodem58760431551",
-    id="my_blue_leader_arm",
+    port="/dev/tty.usbmodem5AB90689011",
+    id="my_leader_arm",
 )
 
 robot = SO101Follower(robot_config)
@@ -108,13 +108,13 @@ With `rerun`, you can teleoperate again while simultaneously visualizing the cam
 <hfoption id="Command">
 ```bash
 lerobot-teleoperate \
-    --robot.type=koch_follower \
-    --robot.port=/dev/tty.usbmodem58760431541 \
-    --robot.id=my_awesome_follower_arm \
-    --robot.cameras="{ front: {type: opencv, index_or_path: 0, width: 1920, height: 1080, fps: 30}}" \
-    --teleop.type=koch_leader \
-    --teleop.port=/dev/tty.usbmodem58760431551 \
-    --teleop.id=my_awesome_leader_arm \
+    --robot.type=so101_follower \
+    --robot.port=/dev/tty.usbmodem5AB90687491 \
+    --robot.id=my_follower_arm \
+    --robot.cameras="{front: {type: opencv, index_or_path: 0, width: 640, height: 480, fps: 30}}" \
+    --teleop.type=so101_leader \
+    --teleop.port=/dev/tty.usbmodem5AB90689011 \
+    --teleop.id=my_leader_arm \
     --display_data=true
 ```
 </hfoption>
@@ -122,34 +122,48 @@ lerobot-teleoperate \
 
 <!-- prettier-ignore-start -->
 ```python
+import time
+from lerobot.teleoperators.so_leader import SO101Leader, SO101LeaderConfig
+from lerobot.robots.so_follower import SO101Follower, SO101FollowerConfig
 from lerobot.cameras.opencv import OpenCVCameraConfig
-from lerobot.teleoperators.koch_leader import KochLeader, KochLeaderConfig
-from lerobot.robots.koch_follower import KochFollower, KochFollowerConfig
+from lerobot.utils.visualization_utils import init_visualization, log_visualization_data, shutdown_visualization
 
-camera_config = {
-    "front": OpenCVCameraConfig(index_or_path=0, width=1920, height=1080, fps=30)
-}
-
-robot_config = KochFollowerConfig(
-    port="/dev/tty.usbmodem585A0076841",
-    id="my_red_robot_arm",
-    cameras=camera_config
+robot_config = SO101FollowerConfig(
+    port="/dev/tty.usbmodem5AB90687491",
+    id="my_follower_arm",
+    cameras={
+        "wrist": OpenCVCameraConfig(index_or_path=0, width=640, height=480, fps=30),
+        "top": OpenCVCameraConfig(index_or_path=1, width=640, height=480, fps=30)
+    }
 )
 
-teleop_config = KochLeaderConfig(
-    port="/dev/tty.usbmodem58760431551",
-    id="my_blue_leader_arm",
+teleop_config = SO101LeaderConfig(
+    port="/dev/tty.usbmodem5AB90689011",
+    id="my_leader_arm",
 )
 
-robot = KochFollower(robot_config)
-teleop_device = KochLeader(teleop_config)
+init_visualization("rerun", session_name="teleoperation")  # pass "foxglove" to stream to Foxglove instead
+
+robot = SO101Follower(robot_config)
+teleop_device = SO101Leader(teleop_config)
 robot.connect()
 teleop_device.connect()
 
+TARGET_HZ = 30
+TIME_PER_FRAME = 1.0 / TARGET_HZ
+
 while True:
+    start_time = time.perf_counter()
+
     observation = robot.get_observation()
     action = teleop_device.get_action()
     robot.send_action(action)
+    log_visualization_data("rerun", observation=observation, action=action)
+
+    elapsed_time = time.perf_counter() - start_time
+    sleep_time = TIME_PER_FRAME - elapsed_time
+    if sleep_time > 0:
+        time.sleep(sleep_time)
 ```
 <!-- prettier-ignore-end -->
 
@@ -193,7 +207,7 @@ lerobot-record \
     --dataset.num_episodes=5 \
     --dataset.single_task="Grab the black cube" \
     --dataset.streaming_encoding=true \
-    # --dataset.vcodec=auto \
+    # --dataset.rgb_encoder.vcodec=auto \
     --dataset.encoder_threads=2
 ```
 </hfoption>
@@ -202,13 +216,14 @@ lerobot-record \
 <!-- prettier-ignore-start -->
 ```python
 from lerobot.cameras.opencv import OpenCVCameraConfig
-from lerobot.datasets import LeRobotDataset
+from lerobot.datasets.lerobot_dataset import LeRobotDataset
 from lerobot.utils.feature_utils import hw_to_dataset_features
-from lerobot.robots.so_follower import SO100Follower, SO100FollowerConfig
-from lerobot.teleoperators.so_leader import SO100Leader, SO100LeaderConfig
+from lerobot.robots.so_follower import SO101Follower, SO101FollowerConfig
+from lerobot.teleoperators.so_leader.config_so_leader import SO101LeaderConfig
+from lerobot.teleoperators.so_leader.so_leader import SO101Leader
 from lerobot.common.control_utils import init_keyboard_listener
 from lerobot.utils.utils import log_say
-from lerobot.utils.visualization_utils import init_rerun
+from lerobot.utils.visualization_utils import init_visualization
 from lerobot.scripts.lerobot_record import record_loop
 from lerobot.processor import make_default_processors
 
@@ -218,71 +233,56 @@ EPISODE_TIME_SEC = 60
 RESET_TIME_SEC = 10
 TASK_DESCRIPTION = "My task description"
 
-# Create robot configuration
-robot_config = SO100FollowerConfig(
-    id="my_awesome_follower_arm",
-    cameras={
-        "front": OpenCVCameraConfig(index_or_path=0, width=640, height=480, fps=FPS) # Optional: fourcc="MJPG" for troubleshooting OpenCV async error.
-    },
-    port="/dev/tty.usbmodem58760434471",
-)
-
-teleop_config = SO100LeaderConfig(
-    id="my_awesome_leader_arm",
-    port="/dev/tty.usbmodem585A0077581",
-)
-
-# Initialize the robot and teleoperator
-robot = SO100Follower(robot_config)
-teleop = SO100Leader(teleop_config)
-
-# Configure the dataset features
-action_features = hw_to_dataset_features(robot.action_features, "action")
-obs_features = hw_to_dataset_features(robot.observation_features, "observation")
-dataset_features = {**action_features, **obs_features}
-
-# Create the dataset
-dataset = LeRobotDataset.create(
-    repo_id="<hf_username>/<dataset_repo_id>",
-    fps=FPS,
-    features=dataset_features,
-    robot_type=robot.name,
-    use_videos=True,
-    image_writer_threads=4,
-)
-
-# Initialize the keyboard listener and rerun visualization
-_, events = init_keyboard_listener()
-init_rerun(session_name="recording")
-
-# Connect the robot and teleoperator
-robot.connect()
-teleop.connect()
-
-# Create the required processors
-teleop_action_processor, robot_action_processor, robot_observation_processor = make_default_processors()
-
-episode_idx = 0
-while episode_idx < NUM_EPISODES and not events["stop_recording"]:
-    log_say(f"Recording episode {episode_idx + 1} of {NUM_EPISODES}")
-
-    record_loop(
-        robot=robot,
-        events=events,
-        fps=FPS,
-        teleop_action_processor=teleop_action_processor,
-        robot_action_processor=robot_action_processor,
-        robot_observation_processor=robot_observation_processor,
-        teleop=teleop,
-        dataset=dataset,
-        control_time_s=EPISODE_TIME_SEC,
-        single_task=TASK_DESCRIPTION,
-        display_data=True,
+def main():
+    # Create robot configuration
+    robot_config = SO101FollowerConfig(
+        port="/dev/tty.usbmodem5AB90687491",
+        id="my_follower_arm",
+        cameras={
+            "wrist": OpenCVCameraConfig(index_or_path=0, width=640, height=480, fps=30),
+            "top": OpenCVCameraConfig(index_or_path=1, width=640, height=480, fps=30)
+        }
     )
 
-    # Reset the environment if not stopping or re-recording
-    if not events["stop_recording"] and (episode_idx < NUM_EPISODES - 1 or events["rerecord_episode"]):
-        log_say("Reset the environment")
+    teleop_config = SO101LeaderConfig(
+        port="/dev/tty.usbmodem5AB90689011",
+        id="my_leader_arm",
+    )
+
+    # Initialize the robot and teleoperator
+    robot = SO101Follower(robot_config)
+    teleop = SO101Leader(teleop_config)
+
+    # Configure the dataset features
+    action_features = hw_to_dataset_features(robot.action_features, "action")
+    obs_features = hw_to_dataset_features(robot.observation_features, "observation")
+    dataset_features = {**action_features, **obs_features}
+
+    # Create the dataset
+    dataset = LeRobotDataset.create(
+        repo_id="<hf_username>/<dataset_repo_id>",
+        fps=FPS,
+        features=dataset_features,
+        robot_type=robot.name,
+        use_videos=True,
+        image_writer_threads=4,
+    )
+
+    # Initialize the keyboard listener and rerun visualization
+    _, events = init_keyboard_listener()
+    init_visualization("rerun", session_name="recording")
+
+    # Connect the robot and teleoperator
+    robot.connect()
+    teleop.connect()
+
+    # Create the required processors
+    teleop_action_processor, robot_action_processor, robot_observation_processor = make_default_processors()
+
+    episode_idx = 0
+    while episode_idx < NUM_EPISODES and not events["stop_recording"]:
+        log_say(f"Recording episode {episode_idx + 1} of {NUM_EPISODES}")
+
         record_loop(
             robot=robot,
             events=events,
@@ -291,26 +291,50 @@ while episode_idx < NUM_EPISODES and not events["stop_recording"]:
             robot_action_processor=robot_action_processor,
             robot_observation_processor=robot_observation_processor,
             teleop=teleop,
-            control_time_s=RESET_TIME_SEC,
+            dataset=dataset,
+            control_time_s=EPISODE_TIME_SEC,
             single_task=TASK_DESCRIPTION,
             display_data=True,
         )
 
-    if events["rerecord_episode"]:
-        log_say("Re-recording episode")
-        events["rerecord_episode"] = False
-        events["exit_early"] = False
-        dataset.clear_episode_buffer()
-        continue
+        # Reset the environment if not stopping or re-recording
+        if not events["stop_recording"] and (episode_idx < NUM_EPISODES - 1 or events["rerecord_episode"]):
+            log_say("Reset the environment")
+            record_loop(
+                robot=robot,
+                events=events,
+                fps=FPS,
+                teleop_action_processor=teleop_action_processor,
+                robot_action_processor=robot_action_processor,
+                robot_observation_processor=robot_observation_processor,
+                teleop=teleop,
+                control_time_s=RESET_TIME_SEC,
+                single_task=TASK_DESCRIPTION,
+                display_data=True,
+            )
 
-    dataset.save_episode()
-    episode_idx += 1
+        if events["rerecord_episode"]:
+            log_say("Re-recording episode")
+            events["rerecord_episode"] = False
+            events["exit_early"] = False
+            dataset.clear_episode_buffer()
+            continue
 
-# Clean up
-log_say("Stop recording")
-robot.disconnect()
-teleop.disconnect()
-dataset.push_to_hub()
+        dataset.save_episode()
+        episode_idx += 1
+
+    # finalize dataset
+    log_say("Finalizing dataset...")
+    dataset.finalize()
+    # Clean up
+    log_say("Stop recording")
+    robot.disconnect()
+    teleop.disconnect()
+    dataset.push_to_hub()
+
+
+if __name__ == "__main__":
+    main()
 ```
 <!-- prettier-ignore-end -->
 
@@ -348,7 +372,7 @@ The `record` function provides a suite of tools for capturing and managing data
 ##### 2. Checkpointing and Resuming
 
 - Checkpoints are automatically created during recording.
-- If an issue occurs, you can resume by re-running the same command with `--resume=true`. When resuming a recording, `--dataset.num_episodes` must be set to the **number of additional episodes to be recorded**, and not to the targeted total number of episodes in the dataset !
+- If an issue occurs or you want to record additional episodes in the same dataset, you can resume by re-running the same command with `--resume=true`. When resuming a recording, `--dataset.num_episodes` must be set to the **number of additional episodes to be recorded**, and not to the targeted total number of episodes in the dataset! Make sure that you also set `--dataset.root="local_path"`, it's a local path to save the new part of the dataset and is required to resume.
 - To start recording from scratch, **manually delete** the dataset directory.
 
 ##### 3. Recording Parameters
@@ -366,9 +390,17 @@ Set the flow of data recording using command-line arguments:
 
 Control the data recording flow using keyboard shortcuts:
 
-- Press **Right Arrow (`→`)**: Early stop the current episode or reset time and move to the next.
-- Press **Left Arrow (`←`)**: Cancel the current episode and re-record it.
-- Press **Escape (`ESC`)**: Immediately stop the session, encode videos, and upload the dataset.
+- Press **Right Arrow (`→`)** or **`n`**: Early stop the current episode or reset time and move to the next.
+- Press **Left Arrow (`←`)** or **`r`**: Cancel the current episode and re-record it.
+- Press **Escape (`ESC`)** or **`q`**: Immediately stop the session, encode videos, and upload the dataset.
+
+<Tip>
+
+These control-flow shortcuts work on **X11, Wayland, and headless/SSH** sessions. When a global keyboard backend isn't available (Wayland, a headless machine, or macOS without Accessibility permission), `lerobot-record` automatically reads the same keys from the terminal — launch it from an interactive terminal and keep it focused. You can also use the letter equivalents **`n`** (next, same as `→`), **`r`** (re-record, same as `←`) and **`q`** (quit, same as `ESC`). No `$DISPLAY` setup is required.
+
+This applies to the recording control flow only. Keyboard **teleoperation** (driving the robot with the keyboard) still needs a global key backend, so it works only on an X11 session, a Windows desktop, or macOS with Accessibility/Input Monitoring granted — not on Wayland or headless sessions.
+
+</Tip>
 
 #### Tips for gathering data
 
@@ -382,7 +414,7 @@ If you want to dive deeper into this important topic, you can check out the [blo
 
 #### Troubleshooting:
 
-- On Linux, if the left and right arrow keys and escape key don't have any effect during data recording, make sure you've set the `$DISPLAY` environment variable. See [pynput limitations](https://pynput.readthedocs.io/en/latest/limitations.html#linux).
+- On Linux, the recording control-flow keys (arrow keys, Escape) work on X11, Wayland, and headless/SSH sessions as long as `lerobot-record` runs in an interactive terminal — no `$DISPLAY` setup is needed. If the keys have no effect, make sure you are in an interactive (TTY) terminal, not a piped/non-TTY session, and that it is focused; the letter equivalents `n` / `r` / `q` also work. Keyboard _teleoperation_ (as opposed to the recording control flow) still requires a global key backend — an X11 session, a Windows desktop, or macOS with Accessibility/Input Monitoring granted — and is unavailable on Wayland or headless machines. See [pynput limitations](https://pynput.readthedocs.io/en/latest/limitations.html#linux).
 
 ## Visualize a dataset
 
@@ -422,7 +454,7 @@ from lerobot.utils.utils import log_say
 
 episode_idx = 0
 
-robot_config = SO100FollowerConfig(port="/dev/tty.usbmodem58760434471", id="my_awesome_follower_arm")
+robot_config = SO100FollowerConfig(port="/dev/tty.usbmodem5AB90687491", id="my_follower_arm")
 
 robot = SO100Follower(robot_config)
 robot.connect()
@@ -482,6 +514,12 @@ lerobot-train \
   --resume=true
 ```
 
+`--config_path` also accepts a **Hub repo id**: if a run pushed its checkpoints to the Hub (with `--save_checkpoint_to_hub=true`), you can resume straight from the repo — its latest checkpoint is downloaded and training continues, restoring the optimizer, scheduler, step counter and data order:
+
+```bash
+lerobot-train --config_path=${HF_USER}/my_policy --resume=true
+```
+
 If you do not want to push your model to the hub after training use `--policy.push_to_hub=false`.
 
 Additionally you can provide extra `tags` or specify a `license` for your model or make the model repo `private` by adding this: `--policy.private=true --policy.tags=\[ppo,rl\] --policy.license=mit`
@@ -490,6 +528,53 @@ Additionally you can provide extra `tags` or specify a `license` for your model
 
 If your local computer doesn't have a powerful GPU you could utilize Google Colab to train your model by following the [ACT training notebook](./notebooks#training-act).
 
+#### Train using Hugging Face Jobs
+
+Hugging Face jobs let's you easily select hardware and run the training in the cloud. So if you don't have a powerful GPU or you need more VRAM or just want to train a model much faster use HF Jobs! It's pay as you go and you simply pay for each second of use, you can see the pricing and additional information [here](https://huggingface.co/docs/hub/jobs).
+
+`lerobot-train` runs locally by default. To run on a HuggingFace GPU, pass `--job.target` with a hardware flavor name:
+
+```bash
+lerobot-train \
+  --dataset.repo_id=${HF_USER}/so101_test \
+  --policy.type=act \
+  --policy.repo_id=${HF_USER}/my_policy \
+  --job.target=a10g-small
+```
+
+List available flavors and pricing with `hf jobs hardware`. The run streams its logs to your terminal; press Ctrl-C to detach (the job keeps running in the cloud). Re-attach or cancel with:
+
+```bash
+hf jobs logs <job-id>
+hf jobs cancel <job-id>
+```
+
+If your dataset exists only locally (not yet on the Hub), it is automatically pushed to a **private** Hub repo so the job can download it by `repo_id` (nothing is made public). The trained model is pushed to the model repo at the end of the run. To also push every intermediate checkpoint to the Hub as it is saved (so you can monitor progress mid-run), add `--save_checkpoint_to_hub=true` — this requires a runtime image that includes this feature.
+
+Every job (and any dataset pushed by the run) is tagged `lerobot` so it's easy to find on the Hub. Add your own with `--job.tags '["my-tag"]'`.
+
+By default the job is capped at `2d` (48h) of wall-clock. Override it with an HF Jobs duration string, e.g. `--job.timeout=4h` to fail faster or `--job.timeout=7d` for a longer run.
+
+> **Note:** the model repo is created up front (it holds the staged training config the job runs from). If a run fails before the model is pushed, that repo is left on the Hub so you can inspect it — it is not deleted automatically, so repeated failures can leave empty repos behind. Remove one with `hf repo delete <repo-id>`.
+
+**Prerequisites:** run `hf auth login` before submitting. For Weights & Biases integration, run `wandb login` or set `WANDB_API_KEY` on your machine — the key is forwarded to the job automatically.
+
+**Resuming on a job.** Adding `--job.target` to a resume command runs the resume in the cloud — the same command works locally or remotely. The checkpoint repo is the source of truth, and new checkpoints continue the lineage in the same repo:
+
+```bash
+# resume a Hub run on a job (its checkpoints are already on the Hub)
+lerobot-train --config_path=${HF_USER}/my_policy --resume=true --job.target=a10g-small
+
+# resume a LOCAL run on a job — the checkpoint is uploaded to a private Hub repo first,
+# then the job resumes from it (a local-only dataset is uploaded the same way)
+lerobot-train \
+  --config_path=outputs/train/act_so101_test/checkpoints/last/pretrained_model/train_config.json \
+  --resume=true \
+  --job.target=a10g-small
+```
+
+Job settings come from the current command, so override `--job.target`, `--job.timeout`, etc. as needed; for the resumed run to itself be resumable later, keep `--save_checkpoint_to_hub=true`.
+
 #### Upload policy checkpoints
 
 Once training is done, upload the latest checkpoint with:
@@ -511,6 +596,8 @@ hf upload ${HF_USER}/act_so101_test${CKPT} \
 
 Use `lerobot-rollout` to deploy a trained policy on your robot. You can choose different strategies depending on your needs:
 
+The examples below load the model from `--policy.path`. To pin a specific pushed version — useful once `--save_checkpoint_to_hub=true` has committed several checkpoints — add `--policy.pretrained_revision` with a commit hash, branch, or tag. Each pushed checkpoint is tagged with its step (e.g. `--policy.pretrained_revision=010000`), so you can recover a checkpoint by step without looking up its commit sha.
+
 <hfoptions id="eval">
 <hfoption id="Base mode (no recording)">
 ```bash
@@ -546,5 +633,6 @@ The `--strategy.type` flag selects the execution mode:
 - `sentry`: Continuous recording with auto-upload (useful for large-scale evaluation)
 - `highlight`: Ring buffer recording with keystroke save (useful for capturing interesting events)
 - `dagger`: Human-in-the-loop data collection (see [HIL Data Collection](./hil_data_collection))
+- `episodic`: Episode-oriented policy recording with reset phases between episodes
 
 All strategies support `--inference.type=rtc` for smooth execution with slow VLA models (Pi0, Pi0.5, SmolVLA).

docs/source/inference.mdx

@@ -117,7 +117,7 @@ lerobot-rollout \
     --strategy.num_episodes=20 \
     --policy.path=outputs/pretrain/checkpoints/last/pretrained_model \
     --robot.type=bi_openarm_follower \
-    --teleop.type=openarm_mini \
+    --teleop.type=bi_openarm_mini \
     --dataset.repo_id=${HF_USER}/rollout_hil_data \
     --dataset.single_task="Fold the T-shirt"
 ```
@@ -157,6 +157,44 @@ Foot pedal input is also supported via `--strategy.input_device=pedal`. Configur
 | `--strategy.input_device`            | Input device: `keyboard` or `pedal` (default: keyboard) |
 | `--teleop.type`                      | **Required.** Teleoperator type                         |
 
+### Episodic (`--strategy.type=episodic`)
+
+Episode-oriented recording that mirrors the behavior of `lerobot-record`. The policy drives the robot for each episode; an optional teleoperator can drive the robot during the reset phase between episodes.
+
+```bash
+lerobot-rollout \
+    --strategy.type=episodic \
+    --policy.path=${HF_USER}/my_policy \
+    --robot.type=so100_follower \
+    --robot.port=/dev/ttyACM0 \
+    --teleop.type=so100_leader \
+    --teleop.port=/dev/ttyACM1 \
+    --dataset.repo_id=${HF_USER}/my_eval_data \
+    --dataset.num_episodes=20 \
+    --dataset.episode_time_s=30 \
+    --dataset.reset_time_s=10 \
+    --dataset.single_task="Pick up the red cube"
+```
+
+Teleop is optional — if omitted the robot holds its position during the reset phase.
+
+**Keyboard controls:**
+
+| Key         | Action                           |
+| ----------- | -------------------------------- |
+| `→` (right) | End the current episode early    |
+| `←` (left)  | Discard episode and re-record it |
+| `ESC`       | Stop the recording session       |
+
+| Flag                                            | Description                                                                |
+| ----------------------------------------------- | -------------------------------------------------------------------------- |
+| `--dataset.num_episodes`                        | Number of episodes to record                                               |
+| `--dataset.episode_time_s`                      | Duration of each recording episode in seconds                              |
+| `--dataset.reset_time_s`                        | Duration of the reset phase between episodes in seconds                    |
+| `--teleop.type`                                 | Optional. Teleoperator to drive the robot during resets                    |
+| `--strategy.reset_to_initial_position`          | Whether to reset the robot to its initial position between episodes        |
+| `--strategy.smooth_leader_to_follower_handover` | Whether to turn on or off the leader -> follower smooth handover behavior. |
+
 ---
 
 ## Inference Backends

docs/source/installation.mdx

@@ -207,6 +207,56 @@ pip install 'lerobot[feetech]'        # Feetech motor support
 
 _Multiple extras can be combined (e.g., `.[core_scripts,pi,pusht]`). For a full list of available extras, refer to `pyproject.toml`._
 
+### PyTorch CUDA variant (Linux only)
+
+On Linux, the install path determines which CUDA wheel you get. macOS and Windows installs use the PyPI default (MPS / CPU / CUDA-Windows wheel respectively) and can skip this section.
+
+<!-- prettier-ignore-start -->
+
+<hfoptions id="cuda_variant">
+<hfoption id="uv-source">
+
+**Source install via `uv` (`uv sync` or `uv pip install -e .`)**
+
+`torch` and `torchvision` are pinned by the project to the **CUDA 12.8** PyTorch index (`https://download.pytorch.org/whl/cu128`, driver floor **570.86**) — covers Ampere/Ada/Hopper/Blackwell GPUs. No action needed for typical NVIDIA setups.
+
+To override for a different CUDA variant:
+
+```bash
+uv pip install --force-reinstall torch torchvision \
+    --index-url https://download.pytorch.org/whl/cu126   # older drivers; or cu130 for Blackwell on driver ≥ 580
+```
+
+</hfoption>
+<hfoption id="pip-conda">
+
+**Source install via `pip`/`conda`, or `pip install lerobot` from PyPI**
+
+PyPI default torch wheel is currently a cu130-bundled Linux wheel, driver floor **580.65**.
+
+To pick a specific CUDA variant:
+
+**Using `pip` or `conda`** — install torch first with an explicit index, then lerobot:
+
+```bash
+pip install --index-url https://download.pytorch.org/whl/cu128 torch torchvision
+pip install -e ".[all]"          # source
+# — or —
+pip install lerobot              # from PyPI
+```
+
+**Using `uv` to install from PyPI** — one-liner via `--torch-backend` (uv ≥ 0.6):
+
+```bash
+uv pip install --torch-backend cu128 lerobot
+```
+
+Supported values include `auto`, `cpu`, `cu126`, `cu128`, `cu129`, `cu130`, plus various `rocm*` and `xpu`. Swap as needed for your driver.
+
+</hfoption>
+</hfoptions>
+<!-- prettier-ignore-end -->
+
 ### Troubleshooting
 
 If you encounter build errors, you may need to install additional system dependencies: `cmake`, `build-essential`, and `ffmpeg libs`.

docs/source/language_and_recipes.mdx

@@ -0,0 +1,147 @@
+# Language columns and recipes
+
+Most LeRobot datasets ship with a single `task` string per episode — fine for
+short, single-instruction skills, but not enough for the longer-horizon,
+multi-modal robot policies the field is moving toward (high-level planning,
+memory, interjections, VQA, tool use). To support those policies without
+forking the dataset format, LeRobot extends `LeRobotDataset` with two optional
+language columns and a small recipe layer that turns those rows into
+chat-style training samples on the fly.
+
+The design splits cleanly into three layers:
+
+1. **Data in the dataset** — language annotations stored next to frames in
+   `data/chunk-*/file-*.parquet` as two optional columns (`language_persistent`
+   and `language_events`). Datasets without these columns keep their existing
+   behavior.
+2. **Recipe** — a YAML file that declares which annotation rows to bind and
+   how to lay them out as chat turns (`role`, `content`, optional images,
+   optional tool calls). Recipes are pure config; no Python required to add a
+   new one.
+3. **Training format** — at sample time, `RenderMessagesStep` resolves the
+   recipe against the per-frame annotations and emits HF-style `messages` plus
+   LeRobot-specific sidecars (`message_streams`, `target_message_indices`)
+   that policy processors consume.
+
+This page describes each layer in turn.
+
+## Layer 1 — language columns in the dataset
+
+The two optional columns live next to frame data in
+`data/chunk-*/file-*.parquet`:
+
+- `language_persistent`: a list of rows broadcast across every frame in an episode for state that remains active, such as `subtask`, `plan`, and `memory`.
+- `language_events`: a list of rows only on the exact frame where an event was emitted, such as `interjection`, `vqa`, and speech tool calls.
+
+Both columns share the same row shape (event rows omit `timestamp` because the
+frame the row sits on already provides it):
+
+```text
+role: string
+content: string | null
+style: string | null
+timestamp: float32        # persistent rows only
+camera: string | null     # observation.images.* feature key, view-dependent rows only
+tool_calls: list[Json] | null
+```
+
+The `camera` field tags rows whose `content` is grounded in a specific camera
+view. Rows of view-dependent styles (`vqa` and `trace`) MUST set `camera` to
+the matching `observation.images.*` feature key. Rows of every other style —
+including `motion`, which describes robot-frame primitives in joint / Cartesian
+terms — MUST leave `camera` as `null`. Pipeline writers and the validator
+enforce this via `validate_camera_field(style, camera)`.
+
+`meta/tasks.parquet` remains the canonical source for the task. The special `${task}` recipe binding always reads that task string and does not depend on language annotations.
+
+### Architecture
+
+The language stack itself has three internal modules backing layer 1:
+
+1. `lerobot.datasets.language` defines the schema, style registry, and `column_for_style`.
+2. `lerobot.datasets.language_render` resolves rows and renders messages.
+3. `RenderMessagesStep` turns dataset samples into `messages`, `message_streams`, and `target_message_indices`.
+
+`LeRobotDataset` stays recipe-agnostic. It passes `language_persistent` and `language_events` through when present, and unannotated datasets keep their existing behavior.
+
+## Layer 2 — recipe anatomy
+
+Recipes are YAML files backed by `TrainingRecipe` and `MessageTurn`. They
+declare which annotation rows to pull (via `bindings`) and how to compose them
+into chat turns (`messages`).
+
+```yaml
+messages:
+  - { role: user, content: "${task}", stream: high_level }
+  - { role: assistant, content: "${subtask}", stream: low_level, target: true }
+```
+
+A recipe can also branch into a weighted **blend** of sub-recipes. At sample
+time, exactly one branch is selected deterministically from the sample index,
+so different frames train different objectives (e.g. memory updates vs.
+low-level execution vs. VQA) without any Python wiring.
+
+### Temporal semantics
+
+Persistent styles are active after emission until replaced:
+
+- `active_at(t, style=subtask)`
+- `nth_prev(style=memory, offset=1)`
+- `nth_next(style=subtask, offset=1)`
+
+Event styles only exist on their exact timestamp:
+
+- `emitted_at(t, style=interjection)`
+- `emitted_at(t, style=vqa, role=user, camera=observation.images.top)`
+- `emitted_at(t, role=assistant, tool_name=say)`
+
+Exact event matching has no tolerance window, so writers must stamp event rows with frame timestamps from the parquet data.
+
+### View-dependent resolution
+
+For view-dependent styles (`vqa` and `trace`), the resolver gains a
+`camera=` filter parallel to `role=` and `tool_name=`. Datasets with multiple
+cameras typically emit one (`vqa`, `user`) + (`vqa`, `assistant`) pair per
+camera at the same timestamp; without `camera=`, those resolvers see two
+matches and raise an ambiguity error. Recipes consume each camera through its
+own binding plus a matching image block, e.g.
+
+```yaml
+ask_vqa_top:
+  bindings:
+    vqa_query: "emitted_at(t, style=vqa, role=user, camera=observation.images.top)"
+    vqa: "emitted_at(t, style=vqa, role=assistant, camera=observation.images.top)"
+  messages:
+    - role: user
+      stream: high_level
+      if_present: vqa_query
+      content:
+        - { type: image, feature: observation.images.top }
+        - { type: text, text: "${vqa_query}" }
+    - {
+        role: assistant,
+        content: "${vqa}",
+        stream: high_level,
+        target: true,
+        if_present: vqa,
+      }
+```
+
+Add one such sub-recipe per camera the dataset records.
+
+## Layer 3 — training format
+
+Rendered samples use HF-style chat messages plus LeRobot sidecars:
+
+```python
+sample["messages"]
+sample["message_streams"]
+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.
+
+## Graceful absence
+
+If both language columns are missing, `None`, or empty, `RenderMessagesStep` is a no-op.
+If an event-scoped branch is selected on a frame without the required event row, rendering returns `None`, allowing a loader to retry another sample.

docs/source/lekiwi.mdx

@@ -319,7 +319,7 @@ If you want to dive deeper into this important topic, you can check out the [blo
 
 #### Troubleshooting:
 
-- On Linux, if the left and right arrow keys and escape key don't have any effect during data recording, make sure you've set the `$DISPLAY` environment variable. See [pynput limitations](https://pynput.readthedocs.io/en/latest/limitations.html#linux).
+- On Linux, the recording control-flow keys (arrow keys, Escape) work on X11, Wayland, and headless/SSH sessions as long as you run the recording from an interactive terminal (keep it focused) — no `$DISPLAY` setup is needed; the letter equivalents `n` / `r` / `q` also work. Note that **keyboard teleoperation of the LeKiwi base** is different: it relies on a global key backend and therefore works only on an X11 session, a Windows desktop, or macOS with Accessibility/Input Monitoring granted — not on Wayland or headless machines. See [pynput limitations](https://pynput.readthedocs.io/en/latest/limitations.html#linux).
 
 ## Replay an episode
 

docs/source/lelab.mdx

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

docs/source/lerobot-dataset-v3.mdx

@@ -10,6 +10,7 @@ This docs will guide you to:
 - Stream datasets without downloading using `StreamingLeRobotDataset`
 - Apply image transforms for data augmentation during training
 - Migrate existing `v2.1` datasets to `v3.0`
+- Experiment with other `LeRobotDataset` formats and implementations like Lance
 
 ## What’s new in `v3`
 
@@ -43,7 +44,7 @@ lerobot-record \
   --dataset.num_episodes=5 \
   --dataset.single_task="Grab the black cube" \
   --dataset.streaming_encoding=true \
-  # --dataset.vcodec=auto \
+  # --dataset.rgb_encoder.vcodec=auto \
   --dataset.encoder_threads=2
 ```
 
@@ -274,7 +275,7 @@ A converter aggregates per‑episode files into larger shards and writes episode
 pip install "https://github.com/huggingface/lerobot/archive/33cad37054c2b594ceba57463e8f11ee374fa93c.zip"
 
 # Convert an existing v2.1 dataset hosted on the Hub:
-python -m lerobot.datasets.v30.convert_dataset_v21_to_v30 --repo-id=<HF_USER/DATASET_ID>
+python -m lerobot.scripts.convert_dataset_v21_to_v30 --repo-id=<HF_USER/DATASET_ID>
 ```
 
 **What it does**
@@ -315,3 +316,39 @@ Dataset v3.0 uses incremental parquet writing with buffered metadata for efficie
 - Ensures the dataset is valid for loading
 
 Without calling `finalize()`, your parquet files will be incomplete and the dataset won't load properly.
+
+## Other formats and implementations
+
+### Lance
+
+Lance is a useful format for multimodal AI datasets, especially for large-scale training requiring high performance IO and random access.
+
+The `lerobot-lancedb` package implements `LeRobotLanceDataset` (for JPEG images) and `LeRobotLanceVideoDataset` (for mp4 videos).
+Those two storage layouts both subclass LeRobotDataset and can provide data loading speed ups.
+
+`LeRobotLanceDataset` is a drop-in replacement for `LeRobotDataset`:
+
+```python
+from lerobot.datasets import LeRobotDatasetMetadata
+from lerobot.policies.diffusion.configuration_diffusion import DiffusionConfig
+from lerobot_lancedb import LeRobotLanceDataset, LeRobotLanceVideoDataset
+
+cfg = DiffusionConfig(...)
+meta = LeRobotDatasetMetadata(root=local_dataset_path)  # or use repo_id=... to load metadata from the Hub
+delta_timestamps = {...}
+
+# Use LeRobotLanceDataset for image datasets
+dataset = LeRobotLanceDataset(
+    root=local_dataset_path,                            # or use repo_id=... to stream from the Hub
+    delta_timestamps=delta_timestamps,
+    return_uint8=True,
+)
+# Or use LeRobotLanceVideoDataset for video datasets:
+dataset = LeRobotLanceVideoDataset(
+    root=local_dataset_path,                            # or use repo_id=... to stream from the Hub
+    delta_timestamps=delta_timestamps,
+    return_uint8=True,
+)
+```
+
+Join the discussion on [Github](https://github.com/huggingface/lerobot/issues/3608) and explore the `lerobot-lancedb` documentation [here](https://lancedb.github.io/lerobot-lancedb/).

docs/source/libero.mdx

@@ -143,7 +143,7 @@ lerobot-train \
   --batch_size=4 \
   --eval.batch_size=1 \
   --eval.n_episodes=1 \
-  --eval_freq=1000
+  --env_eval_freq=1000
 ```
 
 ## Reproducing published results

docs/source/libero_plus.mdx

@@ -173,7 +173,7 @@ lerobot-train \
     --batch_size=4 \
     --eval.batch_size=1 \
     --eval.n_episodes=1 \
-    --eval_freq=1000
+    --env_eval_freq=1000
 ```
 
 ## Relationship to LIBERO

docs/source/lingbot_va.mdx

@@ -0,0 +1,187 @@
+# LingBot-VA
+
+LingBot-VA is an **autoregressive video-action world-model policy** built on the **Wan2.2**
+video-diffusion stack. It interleaves, in one autoregressive sequence, the prediction of
+future **video latents** and **robot actions** ("VA" = Video-Action). The LeRobot
+integration wires LingBot-VA into the standard training, evaluation and processor
+interfaces.
+
+## Model Overview
+
+LingBot-VA is a **dual-stream "mixture-of-transformers"**: a video/latent stream
+(`patch_embedding_mlp → blocks → proj_out`) and an action stream
+(`action_embedder → blocks → action_proj_out`) share the same 30 transformer blocks and
+text conditioning.
+
+| Component                | Class                   | Role                                                        |
+| ------------------------ | ----------------------- | ----------------------------------------------------------- |
+| DiT backbone (trainable) | `WanTransformer3DModel` | ~5B-param dual-stream transformer.                          |
+| VAE (frozen)             | `AutoencoderKLWan`      | Wan2.2 VAE, `z_dim=48`. Lazy-pulled from the source repo.   |
+| Text encoder (frozen)    | `UMT5EncoderModel`      | UMT5-XXL, `d_model=4096`. Lazy-pulled from the source repo. |
+
+At inference the policy runs an autoregressive loop per chunk: it denoises the video-latent
+stream (CFG, ~20 steps) and the action stream (~50 steps) with two independent
+flow-matching schedulers, maintaining a KV cache across chunks. Real observed keyframes are
+fed back into the KV cache as the chunk is executed (closed-loop world modeling).
+
+### What the LeRobot Integration Covers
+
+- Standard `policy.type=lingbot_va` configuration through LeRobot.
+- Ready-to-use LeRobot-format checkpoints on the Hub (converted from the released upstream ones).
+- Autoregressive dual-stream inference behind the standard `select_action` interface
+  (single-environment eval, `--eval.batch_size=1`).
+- Opt-in saving of the policy's **predicted (imagined) videos** during eval / training.
+- Evaluation with `lerobot-eval` on LIBERO and RoboTwin.
+- Training / fine-tuning via the dual-stream flow-matching loss (`policy.forward`), see below.
+
+## Installation
+
+1. Install LeRobot by following the [Installation Guide](./installation).
+2. Install the LingBot-VA extra:
+
+```bash
+pip install -e ".[lingbot_va]"
+```
+
+## Checkpoints
+
+The released upstream checkpoints have been converted to LeRobot format and pushed to the Hub:
+
+| Variant                | LeRobot checkpoint               |
+| ---------------------- | -------------------------------- |
+| LIBERO-Long post-train | `lerobot/lingbot_va_libero_long` |
+| RoboTwin post-train    | `lerobot/lingbot_va_robotwin`    |
+| Pretrained base        | `lerobot/lingbot_va_base`        |
+
+Only the trainable ~5B transformer is stored in the LeRobot
+`model.safetensors`. The frozen VAE + UMT5 + tokenizer (~20 GB) are pulled from
+`config.wan_pretrained_path` at load time (defaults to the source `robbyant/*` repo). The
+UMT5-XXL text encoder runs on CPU by default (`config.text_encoder_device`) so the 5B
+transformer + VAE fit on a single 24–32 GB GPU.
+
+## Evaluation (LIBERO)
+
+```bash
+lerobot-eval \
+    --policy.path=lerobot/lingbot_va_libero_long \
+    --policy.device=cuda \
+    --env.type=libero --env.task=libero_10 \
+    --env.observation_height=128 --env.observation_width=128 \
+    --eval.n_episodes=50 --eval.batch_size=1 \
+    --output_dir=outputs/eval/lingbot_va_libero
+```
+
+LingBot-VA's streaming inference (KV cache + observed-keyframe feedback) is implemented for
+single-environment eval; use `--eval.batch_size=1`.
+
+## Evaluation (RoboTwin)
+
+RoboTwin 2.0 needs the SAPIEN + CuRobo simulator stack. You can use the benchmark Docker image
+(`docker/Dockerfile.benchmark.robotwin`, which also needs `warp-lang==1.3.1` and CuRobo built
+with the GPU's compute capability in `TORCH_CUDA_ARCH_LIST`). RoboTwin uses **end-effector-pose
+control**, so run with `--env.action_mode=ee`: the policy predicts per-arm `xyz+quaternion+gripper`
+deltas (`robotwin_tshape` latent layout) that are composed onto the episode's initial eef pose and
+executed via CuRobo IK.
+
+```bash
+lerobot-eval \
+    --policy.path=lerobot/lingbot_va_robotwin \
+    --policy.device=cuda \
+    --env.type=robotwin --env.task=beat_block_hammer --env.action_mode=ee \
+    --eval.n_episodes=10 --eval.batch_size=1 \
+    --output_dir=outputs/eval/lingbot_va_robotwin
+```
+
+### Saving predicted (imagined) videos
+
+Set `--policy.save_predicted_video=true` to additionally VAE-decode the predicted video
+latents and write `pred_episode_*.mp4` next to the env-rendered `eval_episode_*.mp4` videos.
+The same flag works for the periodic eval during `lerobot-train`.
+
+## Training / fine-tuning
+
+`LingBotVAPolicy.forward(batch)` implements the dual-stream **flow-matching** loss
+(`latent_loss + action_loss`, timestep-weighted, action-masked) from the paper: it VAE-encodes
+the camera clips into video latents, UMT5-encodes the task, noises both streams, runs the
+transformer's block-causal training pass and returns `(loss, metrics)`. Optimizer preset is AdamW
+with a linear-warmup-then-constant schedule (matching upstream).
+
+Requirements:
+
+- The block-causal masks use PyTorch **flex-attention**, so build the policy with
+  `--policy.attn_mode=flex` for training (the default `torch` SDPA is inference-only).
+- The full 5B DiT does not fit a single 24–32 GB GPU under AdamW; fine-tune with **LoRA**
+  (`--policy.use_peft=true`) and/or optimizer offload. `get_optim_params` returns only the
+  trainable (e.g. adapter) parameters; the VAE + UMT5 text encoder stay frozen.
+
+```bash
+lerobot-train \
+  --policy.path=lerobot/lingbot_va_libero_long --policy.attn_mode=flex \
+  --policy.use_peft=true \
+  --dataset.repo_id=<your LeRobot-format dataset> \
+  --batch_size=1 --steps=... --output_dir=outputs/train/lingbot_va
+```
+
+The dataset must provide camera clips (a temporal window per camera, VAE-encoded to
+`frame_chunk_size` latent frames) and `frame_chunk_size * action_per_frame` action steps per item.
+
+## Data format (action channels & camera order)
+
+LingBot-VA is an **end-effector (Cartesian) pose** policy, it predicts EEF poses + gripper, not
+joint positions. Actions live in a fixed multi-embodiment **30-dim** layout; map your robot's
+action dimensions into these channels and pad the rest with `0` (`used_action_channel_ids` selects
+the channels a given checkpoint actually uses):
+
+| channels | meaning                                               |
+| -------- | ----------------------------------------------------- |
+| 0–6      | Left-arm end-effector pose                            |
+| 7–13     | Right-arm end-effector pose                           |
+| 14–20    | Left-arm joints (unused by the released checkpoints)  |
+| 21–27    | Right-arm joints (unused by the released checkpoints) |
+| 28       | Left gripper                                          |
+| 29       | Right gripper                                         |
+
+- **LIBERO** uses channels `0–6`: a 6-DoF EEF delta (xyz + rotation) + gripper (single arm).
+- **RoboTwin** uses channels `[0–6, 28, 7–13, 29]`: left EEF (xyz + quaternion) + left gripper +
+  right EEF + right gripper (16 dims). The env converts these poses to joint trajectories via
+  CuRobo IK — joints are never predicted.
+
+Joint-space datasets (or a different EEF convention) must be remapped into this schema before
+fine-tuning these checkpoints.
+
+**Camera order is fixed and order-sensitive**, per-camera latents are concatenated spatially in
+`obs_cam_keys` order, so the physical camera→slot mapping must match training:
+
+| benchmark | `obs_cam_keys` (in order)                                                                             | `camera_layout`                                                     |
+| --------- | ----------------------------------------------------------------------------------------------------- | ------------------------------------------------------------------- |
+| LIBERO    | `observation.images.image` (agentview / 3rd-person), `observation.images.image2` (eye-in-hand wrist)  | `width_concat` (latents concatenated on width)                      |
+| RoboTwin  | `observation.images.head_camera`, `observation.images.left_camera`, `observation.images.right_camera` | `robotwin_tshape` (full-res head below, two half-res wrists on top) |
+
+The first camera is the exterior/head view and the rest are wrist views.
+
+## Inference Hyperparameters (LIBERO)
+
+| Key                                    | Value                                                                             |
+| -------------------------------------- | --------------------------------------------------------------------------------- |
+| height × width                         | 128 × 128                                                                         |
+| cameras                                | `observation.images.image` (agentview), `observation.images.image2` (eye-in-hand) |
+| action channels used                   | 0–6 (7-DoF arm + gripper)                                                         |
+| action_per_frame / frame_chunk_size    | 4 / 4                                                                             |
+| attn_window                            | 30                                                                                |
+| video / action denoising steps         | 20 / 50                                                                           |
+| guidance_scale / action_guidance_scale | 5 / 1                                                                             |
+| snr_shift / action_snr_shift           | 5.0 / 0.05                                                                        |
+
+These are the defaults of `LingBotVAConfig`; override any of them via `--policy.<name>=...`.
+
+## Notes
+
+- **Attention backend:** inference uses the `torch` SDPA backend (always available). The
+  `flashattn` and `flex` backends are optional; `flex` is only needed for training.
+- **Model size:** the DiT is ~5B params and the frozen VAE+UMT5 add ~20 GB; inference needs
+  roughly 18–24 GB of VRAM.
+
+## License
+
+LingBot-VA is released under Apache-2.0. See the
+[upstream repository](https://github.com/Robbyant/lingbot-va).

docs/source/metaworld.mdx

@@ -120,11 +120,11 @@ lerobot-train \
   --batch_size=4 \
   --eval.batch_size=1 \
   --eval.n_episodes=1 \
-  --eval_freq=1000
+  --env_eval_freq=1000
 ```
 
 ## Practical tips
 
 - Use the one-hot task conditioning for multi-task training (MT10/MT50 conventions) so policies have explicit task context.
 - Inspect the dataset task descriptions and the `info["is_success"]` keys when writing post-processing or logging so your success metrics line up with the benchmark.
-- Adjust `batch_size`, `steps`, and `eval_freq` to match your compute budget.
+- Adjust `batch_size`, `steps`, and `env_eval_freq` to match your compute budget.

docs/source/molmoact2.mdx

@@ -0,0 +1,495 @@
+# MolmoAct2 Policy
+
+MolmoAct2 is the LeRobot policy implementation of
+[MolmoAct2](https://allenai.org/blog/molmoact2), ported into the LeRobot
+training, evaluation, checkpointing, and dataset interfaces for easier use with
+LeRobot datasets.
+
+This implementation currently supports training and evaluation for the regular
+MolmoAct2 model. MolmoAct2-Think, which supports adaptive depth reasoning, is
+not included in this LeRobot policy yet and is coming soon.
+
+For the original MolmoAct2 training code used for the experiments reported in
+the paper, see [allenai/molmoact2](https://github.com/allenai/molmoact2).
+
+## Installation Requirements
+
+Install LeRobot with the MolmoAct2 optional dependencies:
+
+```bash
+uv sync --locked --extra molmoact2
+```
+
+To run the models in this repository, you need an NVIDIA GPU. The measurements
+below were taken on a single NVIDIA H100 80GB with bf16 model loading, LIBERO with two RGB cameras. MolmoAct2 rows use `chunk_size=10`, action dim 7
+padded to `expected_max_action_dim=32`, and `num_flow_timesteps=8`. Training measurements use
+`gradient_checkpointing=true` and include the forward pass, backward pass,
+gradient clipping, optimizer step, and optimizer state allocation. Values are
+peak GPU memory sampled with `nvidia-smi`. Leave a few GiB of headroom for
+dataloader workers, CUDA context, and fragmentation.
+
+Multi-GPU training through `accelerate` increases throughput and global batch
+size, but this LeRobot port does not currently expose the original MolmoAct2
+`fsdp_devices` model-parallel training path. The current training script has
+not been tested for multi-node training.
+
+| Mode                                             | Peak Memory, bs=8 | Peak Memory, bs=16 | Peak Memory, bs=32 |
+| ------------------------------------------------ | ----------------: | -----------------: | -----------------: |
+| Inference, continuous, CUDA graph enabled (bs=1) |          12.1 GiB |                  - |                  - |
+| Fine-tuning, action expert only, continuous      |          16.5 GiB |           18.3 GiB |           21.4 GiB |
+| Fine-tuning, LoRA VLM, both action modes         |          20.2 GiB |           26.8 GiB |           41.3 GiB |
+| Fine-tuning, full model, both action modes       |          48.3 GiB |           49.8 GiB |           60.1 GiB |
+
+The repo has been tested with Ubuntu 22.04.
+
+## Usage
+
+To use MolmoAct2 in a LeRobot training config, set:
+
+```bash
+--policy.type=molmoact2
+```
+
+## Training
+
+MolmoAct2 can be fine-tuned from either the released MolmoAct2 Hugging Face
+checkpoint format or from a checkpoint already saved by LeRobot. Both routes use
+the same LeRobot training loop, dataset transforms, checkpoint saving, and
+logging. The difference is only how the initial policy weights and processor
+state are loaded.
+
+### Training With Original MolmoAct2 Weight
+
+Use `policy.checkpoint_path` when starting from a released MolmoAct2 checkpoint,
+for example `allenai/MolmoAct2` or `allenai/MolmoAct2-LIBERO`. LeRobot will load
+the original HF model files, then build its own policy processor from the
+dataset metadata and the policy options below.
+
+The command below shows full fine-tuning on the merged LIBERO dataset. It uses
+bf16 model loading, 8 flow timesteps, LeRobot dataset statistics, image
+augmentation, and LeRobot's checkpointing/logging path.
+
+```bash
+accelerate launch \
+  --num_processes=8 \
+  --mixed_precision=bf16 \
+  -m lerobot.scripts.lerobot_train \
+  --dataset.repo_id=allenai/MolmoAct2-LIBERO-Dataset \
+  --dataset.root=/path/to/lerobot/data/allenai/MolmoAct2-LIBERO-Dataset \
+  --dataset.video_backend=pyav \
+  --dataset.image_transforms.enable=true \
+  --policy.type=molmoact2 \
+  --policy.checkpoint_path=allenai/MolmoAct2-LIBERO \
+  --policy.device=cuda \
+  --policy.action_mode=both \
+  --policy.chunk_size=10 \
+  --policy.n_action_steps=10 \
+  --policy.setup_type="single franka robotic arm in libero" \
+  --policy.control_mode="delta end-effector pose" \
+  --policy.image_keys='["observation.images.image","observation.images.wrist_image"]' \
+  --policy.model_dtype=bfloat16 \
+  --policy.num_flow_timesteps=8 \
+  --policy.gradient_checkpointing=true \
+  --policy.freeze_embedding=true \
+  --policy.normalize_gripper=false \
+  --policy.enable_knowledge_insulation=false \
+  --policy.push_to_hub=false \
+  --wandb.enable=true \
+  --wandb.entity=<wandb_entity> \
+  --wandb.project=<wandb_project> \
+  --job_name=<job_name> \
+  --output_dir=outputs/<job_name> \
+  --steps=10000 \
+  --batch_size=32 \
+  --num_workers=4 \
+  --log_freq=20 \
+  --env_eval_freq=-1 \
+  --save_checkpoint=true \
+  --save_freq=2000
+```
+
+### Training With LeRobot MolmoAct2 Weight
+
+Use `policy.path` when starting from a MolmoAct2 checkpoint that was saved by
+LeRobot, either from a local `pretrained_model` directory or from the Hub. This
+restores the saved LeRobot policy config, model weights, processor, and
+normalization statistics. You can still override training-time options such as
+`batch_size`, `steps`, LoRA flags, or `policy.action_mode`.
+
+```bash
+accelerate launch \
+  --num_processes=8 \
+  --mixed_precision=bf16 \
+  -m lerobot.scripts.lerobot_train \
+  --dataset.repo_id=allenai/MolmoAct2-LIBERO-Dataset \
+  --dataset.root=/path/to/lerobot/data/allenai/MolmoAct2-LIBERO-Dataset \
+  --dataset.video_backend=pyav \
+  --dataset.image_transforms.enable=true \
+  --policy.path=/path/to/pretrained_model \
+  --policy.device=cuda \
+  --policy.action_mode=both \
+  --policy.chunk_size=10 \
+  --policy.n_action_steps=10 \
+  --policy.model_dtype=bfloat16 \
+  --policy.num_flow_timesteps=8 \
+  --policy.gradient_checkpointing=true \
+  --wandb.enable=true \
+  --wandb.entity=<wandb_entity> \
+  --wandb.project=<wandb_project> \
+  --job_name=<job_name> \
+  --output_dir=outputs/<job_name> \
+  --steps=10000 \
+  --batch_size=32 \
+  --num_workers=4 \
+  --log_freq=20 \
+  --env_eval_freq=-1 \
+  --save_checkpoint=true \
+  --save_freq=2000
+```
+
+### Common Practices
+
+For fine-tuning on a comparatively small dataset, such as a single LIBERO suite
+or a real-world dataset with less than 200 demonstrations, a global batch size of
+16 to 32 is a good starting point. In these settings, `policy.enable_lora_vlm=true` or `policy.train_action_expert_only=true` is also a practical choice. In both
+cases, we intentionally keep the action expert fully trainable, which we found
+to be crucial for model performance. For larger fine-tuning datasets, larger
+global batch sizes and full fine-tuning are usually preferred.
+
+### Common Policy Options
+
+- `policy.checkpoint_path`: original MolmoAct2 HF checkpoint to initialize from.
+  Use this for released MolmoAct2 weights.
+- `policy.path`: LeRobot checkpoint to initialize from. Use this for checkpoints
+  created by LeRobot training.
+- `policy.action_mode`: training target, one of `continuous`, `discrete`, or
+  `both`. `both` trains the flow-matching action expert and the discrete
+  action-token loss.
+- `policy.train_action_expert_only`: trains only parameters whose names contain
+  `action_expert`. It requires `policy.action_mode=continuous`.
+- `policy.enable_lora_vlm`: enables LoRA on VLM linear layers. Use
+  `policy.enable_lora_action_expert=true` only if LoRA should also cover action
+  expert linear layers. When `policy.enable_lora_action_expert=false`, the
+  action expert base weights remain fully trainable while the VLM is trained
+  through LoRA adapters. When `policy.enable_lora_action_expert=true`, the
+  action expert is also adapter-tuned instead of fully fine-tuned.
+- `policy.enable_knowledge_insulation`: when `true`, detaches action-expert
+  context K/V states before the action loss. The default is `false`.
+- `policy.chunk_size`: action horizon used by the policy. For LIBERO we use
+  `10`. This LeRobot port overrides the loaded checkpoint's
+  `max_action_horizon` with this value.
+- `policy.n_action_steps`: number of actions consumed from each predicted
+  chunk before querying the policy again. For LIBERO, set it to `chunk_size`.
+- `policy.setup_type`: text inserted into the prompt to describe the robot and
+  scene, e.g. `single franka robotic arm in libero`. More examples are listed
+  in the `metadata_by_tag` entries of
+  [`norm_stats.json`](https://huggingface.co/allenai/MolmoAct2/blob/main/norm_stats.json).
+- `policy.control_mode`: text inserted into the prompt to describe the action
+  space, e.g. `delta end-effector pose` or `absolute joint pose`.
+- `policy.image_keys`: ordered LeRobot image observation keys passed to the
+  processor.
+- `policy.model_dtype`: checkpoint/forward dtype, one of `float32`,
+  `bfloat16`, or `float16`. Use `bfloat16` for normal training.
+- `policy.num_flow_timesteps`: number of flow-matching timesteps sampled per
+  example during training. We use `8` for fine-tuning.
+- `policy.num_inference_steps`: optional override for continuous action
+  generation steps at inference time.
+- `policy.gradient_checkpointing`: enables checkpointing in the VLM/action path
+  to reduce activation memory.
+- `policy.freeze_embedding`: freezes input embeddings. The default is `true`.
+- `policy.normalize_gripper`: controls whether gripper dimensions are included
+  in state/action quantile normalization. The default is `false`.
+- `policy.normalize_language`: normalizes task strings before prompt
+  construction. The default is `true`.
+- `policy.mask_action_dim_padding`: masks padded dimensions in the flow loss.
+  Released checkpoints use `policy.expected_max_action_dim=32`.
+- `policy.max_sequence_length`: optional manual sequence cap. Leave unset to
+  infer it from images, state dimension, action dimension, action horizon, and
+  discrete-action mode.
+
+### Learning Rates
+
+MolmoAct2 uses parameter-group learning rates to match the original MolmoAct2
+fine-tuning experiments.
+
+- Full fine-tuning uses `policy.optimizer_lr=1e-5` for the VLM,
+  `policy.optimizer_vit_lr=5e-6` for the vision tower,
+  `policy.optimizer_connector_lr=5e-6` for image connector layers, and
+  `policy.optimizer_action_expert_lr=5e-5` for the action expert.
+- LoRA VLM fine-tuning sets the VLM, vision, and connector LoRA parameter
+  groups to `5e-5` when `policy.enable_lora_vlm=true`. By default,
+  `policy.enable_lora_action_expert=false`, so the action expert is still fully
+  fine-tuned with `policy.optimizer_action_expert_lr`. If
+  `policy.enable_lora_action_expert=true`, the action expert is trained through
+  LoRA adapters instead.
+- Action-expert-only fine-tuning trains only the action expert and uses
+  `policy.optimizer_action_expert_lr=5e-5`.
+
+You can override the full fine-tuning and action-expert learning rates with
+`policy.optimizer_lr`, `policy.optimizer_vit_lr`,
+`policy.optimizer_connector_lr`, and `policy.optimizer_action_expert_lr`.
+Scheduler settings can be changed with `policy.scheduler_warmup_steps`,
+`policy.scheduler_decay_steps`, and `policy.scheduler_decay_lr`.
+
+### Dataset Quantile Statistics
+
+MolmoAct2 defaults to quantile normalization for state and action features. If
+your dataset has not been converted with quantile statistics, you can add them
+with:
+
+```bash
+python src/lerobot/scripts/augment_dataset_quantile_stats.py \
+  --repo-id=your_dataset
+```
+
+Alternatively, train MolmoAct2 with mean/std normalization:
+
+```bash
+--policy.normalization_mapping='{"ACTION": "MEAN_STD", "STATE": "MEAN_STD", "VISUAL": "IDENTITY"}'
+```
+
+## Evaluation
+
+Evaluation also supports both LeRobot-saved checkpoints and original MolmoAct2
+HF checkpoints. For LIBERO replication, keep the EGL rendering environment
+fixed and use `policy.per_episode_seed=true`.
+
+**Important:** We found that `num_steps_wait=10` does not reliably let the
+LIBERO scene stabilize and can degrade measured success. All LIBERO evaluation
+results reported here use `num_steps_wait=50`.
+
+### Evaluation With LeRobot MolmoAct2 Weight
+
+Use `policy.path` for a checkpoint saved by LeRobot. The saved processor and
+normalization statistics are restored together with the model.
+
+```bash
+export MUJOCO_GL=egl
+export PYOPENGL_PLATFORM=egl
+export OMP_NUM_THREADS=1
+export MKL_NUM_THREADS=1
+
+lerobot-eval \
+  --policy.path=allenai/MolmoAct2-LIBERO-LeRobot \
+  --policy.inference_action_mode=continuous \
+  --policy.model_dtype=bfloat16 \
+  --policy.use_amp=true \
+  --policy.enable_inference_cuda_graph=true \
+  --policy.device=cuda \
+  --policy.per_episode_seed=true \
+  --policy.eval_seed=1000 \
+  --env.type=libero \
+  --env.task=libero_10,libero_goal,libero_object,libero_spatial \
+  --env.camera_name_mapping='{"agentview_image":"image","robot0_eye_in_hand_image":"wrist_image"}' \
+  --eval.batch_size=1 \
+  --eval.n_episodes=50 \
+  --seed=1000
+```
+
+### Evaluation With Original MolmoAct2 Weight
+
+You can evaluate a released Hugging Face checkpoint directly without first
+converting it to a LeRobot checkpoint. In this case, set
+`policy.checkpoint_path` to the HF model repo and provide `policy.norm_tag`.
+For LIBERO, `policy.norm_tag=libero` loads the LIBERO action/state
+normalization statistics, action horizon, prompt metadata, and image-key order
+from the checkpoint's `norm_stats.json`.
+
+To fully replicate the MolmoAct2 paper results with released Hugging Face
+checkpoints, we recommend using the v0.5.1-pinned
+[`allenai/lerobot` `molmoact2-hf-inference`](https://github.com/allenai/lerobot/tree/molmoact2-hf-inference)
+branch. That branch matches the original evaluation settings used for the
+reported numbers.
+
+```bash
+export MUJOCO_GL=egl
+export PYOPENGL_PLATFORM=egl
+export OMP_NUM_THREADS=1
+export MKL_NUM_THREADS=1
+
+lerobot-eval \
+  --policy.type=molmoact2 \
+  --policy.checkpoint_path=allenai/MolmoAct2-LIBERO \
+  --policy.norm_tag=libero \
+  --policy.inference_action_mode=continuous \
+  --policy.model_dtype=float32 \
+  --policy.use_amp=false \
+  --policy.enable_inference_cuda_graph=true \
+  --policy.device=cuda \
+  --policy.per_episode_seed=true \
+  --policy.eval_seed=1000 \
+  --env.type=libero \
+  --env.task=libero_goal \
+  --env.camera_name_mapping='{"agentview_image":"image","robot0_eye_in_hand_image":"wrist_image"}' \
+  --eval.batch_size=1 \
+  --eval.n_episodes=50 \
+  --seed=1000
+```
+
+Use `--env.task=libero_10,libero_goal,libero_object,libero_spatial` to run the
+full LIBERO suite. The same command works for other released MolmoAct2
+checkpoints as long as the requested `policy.norm_tag` exists in that
+checkpoint's `norm_stats.json`.
+
+### Common Evaluation Options
+
+- `policy.inference_action_mode`: required for rollout. Use `continuous` for
+  flow-matching inference or `discrete` for action-token inference. It must be
+  compatible with the training-time `policy.action_mode` saved in the
+  checkpoint.
+- `policy.path`: LeRobot checkpoint path or Hub repo. Use this for checkpoints
+  saved by LeRobot.
+- `policy.checkpoint_path`: original MolmoAct2 HF checkpoint path or Hub repo.
+  Use this with `policy.type=molmoact2` and `policy.norm_tag`.
+- `policy.norm_tag`: selects normalization statistics, prompt metadata,
+  image-key order, and action horizon from the original checkpoint's
+  `norm_stats.json`. It is required for direct original-HF checkpoint
+  evaluation.
+- `policy.model_dtype`: model load/forward dtype. Use `bfloat16` for normal
+  GPU evaluation. Use `float32` only when you explicitly want fp32 inference.
+- `policy.use_amp`: runs the policy forward under autocast during eval. For
+  `model_dtype=bfloat16`, keep this enabled.
+- `policy.enable_inference_cuda_graph`: enables the MolmoAct2 inference CUDA
+  graph path for faster repeated continuous-action rollout.
+- `policy.per_episode_seed` and `policy.eval_seed`: make stochastic continuous
+  action generation deterministic per episode for replication.
+- `env.task`: comma-separated LIBERO suites or a single suite. Use
+  `libero_10,libero_goal,libero_object,libero_spatial` for the full benchmark.
+- `env.camera_name_mapping`: maps LIBERO camera names to the image keys expected
+  by the policy processor.
+
+## Performance Results
+
+### LIBERO Benchmark Results
+
+MolmoAct2 has demonstrated strong performance on the LIBERO benchmark suite. To
+compare and test its LeRobot implementation, we fine-tuned
+[`allenai/MolmoAct2-LIBERO`](https://huggingface.co/allenai/MolmoAct2-LIBERO)
+for an additional 10k steps on the LIBERO dataset with per-GPU batch size 32 on
+8 H100 GPUs, then compared the results to the original MolmoAct2 reference
+results.
+
+The LeRobot fine-tuned checkpoint reported here is available at
+[`allenai/MolmoAct2-LIBERO-LeRobot`](https://huggingface.co/allenai/MolmoAct2-LIBERO-LeRobot)
+and was trained on
+[`allenai/MolmoAct2-LIBERO-Dataset`](https://huggingface.co/datasets/allenai/MolmoAct2-LIBERO-Dataset).
+
+| Benchmark      | LeRobot Implementation | MolmoAct2 Original |
+| -------------- | ---------------------: | -----------------: |
+| LIBERO Spatial |                  98.4% |              97.8% |
+| LIBERO Object  |                 100.0% |             100.0% |
+| LIBERO Goal    |                  98.0% |              97.8% |
+| LIBERO 10      |                  96.6% |              93.2% |
+| Average        |                 98.25% |             97.20% |
+
+These results demonstrate MolmoAct2's strong performance across diverse robotic
+manipulation tasks. To reproduce them, follow the instructions in the LIBERO
+evaluation section.
+
+## Hardware Deployment (lerobot-rollout)
+
+LeRobot-format checkpoints are available on the Hub for direct use with
+`lerobot-rollout`. Each checkpoint uses specific camera names that must
+match your robot's camera configuration.
+
+### Camera naming convention
+
+Each checkpoint expects specific `observation.images.*` keys.
+If your robot cameras have different names, use `--rename_map` to map them:
+
+| Checkpoint                    | Camera keys            | Description              |
+| ----------------------------- | ---------------------- | ------------------------ |
+| MolmoAct2-LIBERO-LeRobot      | `image`, `wrist_image` | LIBERO sim cameras       |
+| MolmoAct2-BimanualYAM-LeRobot | `top`, `left`, `right` | YAM 3-camera setup       |
+| MolmoAct2-DROID-LeRobot       | `cam0`, `cam1`         | External + wrist         |
+| MolmoAct2-SO100_101-LeRobot   | `cam0`, `cam1`         | Primary + secondary view |
+
+Example with an SO-100 robot using top and side cameras:
+
+```bash
+lerobot-rollout \
+  --policy.path=lerobot/MolmoAct2-SO100_101-LeRobot \
+  --rename_map='{"observation.images.top": "observation.images.cam0", "observation.images.side": "observation.images.cam1"}' \
+  --robot.type=so100_follower \
+  --robot.port=/dev/ttyACM0 \
+  --robot.cameras='{
+      top: {type: opencv, index_or_path: 0, width: 640, height: 480, fps: 30},
+      side: {type: opencv, index_or_path: 2, width: 640, height: 480, fps: 30}
+  }' \
+  --task="pick up the red cube" --duration=30
+```
+
+To use a wrist camera instead, just change the rename mapping:
+
+```bash
+--rename_map='{"observation.images.top": "observation.images.cam0", "observation.images.wrist": "observation.images.cam1"}'
+```
+
+### Joint frame transform (SO-100/101 zero-shot)
+
+<Tip warning={true}>
+The MolmoAct2-SO100_101 checkpoint was trained on data that uses a different
+joint calibration convention than LeRobot >= 0.5.0. Without a frame
+correction, the arm may move in the wrong direction.
+
+This affects both **zero-shot deployment** and **fine-tuning** from the
+original checkpoint. The pretrained weights expect the old convention, so
+all joint data (observations and actions) must be transformed to match.
+
+The converted LeRobot checkpoint (`lerobot/MolmoAct2-SO100_101-LeRobot`)
+already includes this correction in its processor pipeline. If you convert
+or fine-tune the checkpoint yourself, set the following in the policy config (`configuration_molmoact2.py`):
+
+- `joint_signs`: `[1, -1, 1, 1, 1, 1]` (flips shoulder_lift direction)
+- `joint_offsets`: `[0, 90, 90, 0, 0, 0]` (shifts shoulder_lift and elbow_flex by 90°)
+
+See the [backward compatibility guide](./backwardcomp) for details on the
+calibration change.
+
+</Tip>
+
+## Differences From the Original Implementation
+
+This LeRobot port is intended to match MolmoAct2 behavior while using LeRobot's
+dataset, training, evaluation, checkpoint, and logging infrastructure. The main
+differences from the original training repository are:
+
+- The original paper training stack loads the model in fp32 and trains under
+  mixed precision. This LeRobot port usually loads the checkpoint directly in
+  `policy.model_dtype=bfloat16` for lower memory use.
+- The original repository uses its own FSDP/model-parallel training path. The
+  LeRobot port uses the standard LeRobot/Accelerate training path and has not
+  been tested for multi-node training.
+- The original repository supports sequence packing. The LeRobot port trains on
+  one LeRobot sample per item and pads to an inferred fixed sequence budget.
+- The LeRobot port follows LeRobot's optimizer, scheduler, checkpoint saving,
+  dataset transforms, image augmentation, and Weights & Biases logging
+  conventions.
+- The original training path supports mixed action horizons by padding to
+  `max_action_horizon` and masking padded horizon slots in the action expert
+  self-attention. This is useful when training across datasets with different
+  control frequencies. The LeRobot port currently targets single-dataset
+  fine-tuning, so `policy.chunk_size` overrides the checkpoint
+  `max_action_horizon` and horizon masking is not implemented yet. Support for
+  this mixed-horizon path is planned.
+
+## Citation
+
+```bibtex
+@misc{fang2026molmoact2actionreasoningmodels,
+      title={MolmoAct2: Action Reasoning Models for Real-world Deployment},
+      author={Haoquan Fang and Jiafei Duan and Donovan Clay and Sam Wang and Shuo Liu and Weikai Huang and Xiang Fan and Wei-Chuan Tsai and Shirui Chen and Yi Ru Wang and Shanli Xing and Jaemin Cho and Jae Sung Park and Ainaz Eftekhar and Peter Sushko and Karen Farley and Angad Wadhwa and Cole Harrison and Winson Han and Ying-Chun Lee and Eli VanderBilt and Rose Hendrix and Suveen Ellawela and Lucas Ngoo and Joyce Chai and Zhongzheng Ren and Ali Farhadi and Dieter Fox and Ranjay Krishna},
+      year={2026},
+      eprint={2605.02881},
+      archivePrefix={arXiv},
+      primaryClass={cs.RO},
+      url={https://arxiv.org/abs/2605.02881},
+}
+```
+
+## License
+
+This model is licensed under Apache 2.0. It is intended for research and
+educational use in accordance with
+[Ai2's Responsible Use Guidelines](https://allenai.org/responsible-use),
+consistent with [allenai/molmoact2](https://github.com/allenai/molmoact2).

docs/source/multi_gpu_training.mdx

@@ -95,7 +95,7 @@ If you want to scale your hyperparameters when using multiple GPUs, you should d
 accelerate launch --num_processes=2 $(which lerobot-train) \
   --optimizer.lr=2e-4 \
   --dataset.repo_id=lerobot/pusht \
-  --policy=act
+  --policy.type=act
 ```
 
 **Training Steps Scaling:**
@@ -110,9 +110,64 @@ accelerate launch --num_processes=2 $(which lerobot-train) \
   --batch_size=8 \
   --steps=50000 \
   --dataset.repo_id=lerobot/pusht \
-  --policy=act
+  --policy.type=act
 ```
 
+## Training Large Models with FSDP
+
+DDP replicates the full model on every GPU, so a model that doesn't fit on one GPU won't fit under
+DDP either. For large models, use **FSDP** (Fully Sharded Data Parallel), which shards parameters,
+gradients, and optimizer state across GPUs. See the [accelerate FSDP guide](https://huggingface.co/docs/accelerate/usage_guides/fsdp) for background.
+
+An example on how to launch LeRobot training with FSDP across 4 GPUs (1 machine):
+
+```bash
+accelerate launch --config_file fsdp.yaml --num_processes=4 $(which lerobot-train) \
+  --dataset.repo_id=${HF_USER}/my_dataset \
+  --policy.type=<your_policy> \
+  --output_dir=outputs/train/my_policy_fsdp
+```
+
+A minimal `fsdp.yaml` (FSDP1; shards params/grads/optimizer — ZeRO-3-equivalent):
+
+```yaml
+compute_environment: LOCAL_MACHINE
+distributed_type: FSDP
+mixed_precision: bf16
+num_machines: 1
+num_processes: 4
+fsdp_config:
+  fsdp_version: 1
+  fsdp_sharding_strategy: FULL_SHARD # params + grads + optimizer (ZeRO-3)
+  fsdp_auto_wrap_policy: TRANSFORMER_BASED_WRAP
+  fsdp_transformer_layer_cls_to_wrap: <YourTransformerBlock> # repeated block class to shard
+  fsdp_use_orig_params: true # required: optimizer is built pre-prepare
+  fsdp_state_dict_type: FULL_STATE_DICT
+```
+
+Set `fsdp_transformer_layer_cls_to_wrap` to your model's repeated transformer-block class so each
+block is sharded as its own unit. `fsdp_use_orig_params: true` is required because LeRobot builds the
+optimizer before `accelerator.prepare()`.
+
+### FSDP checkpoints
+
+LeRobot gathers the full state dict across all ranks and the main process writes it as a single
+`model.safetensors`, loadable as usual with `Policy.from_pretrained(...)`. Two things to look out for:
+
+- **Checkpoints store fp32 weights.** Under mixed precision (`bf16`/`fp16`) FSDP keeps an fp32 master
+  copy, and the checkpoint saves it (~2× the bf16 size on disk) so training can resume consistently
+  with the fp32 optimizer state; `from_pretrained` casts back to the policy dtype on load. FSDP-specific
+  caveat: an fp32 checkpoint is materialized in full precision on the target device _before_ casting,
+  so loading it for inference on a tight GPU can OOM even when the bf16 model would fit — load on CPU
+  first, or cast `model.safetensors` to the deployment dtype offline.
+- The sharded optimizer state is gathered into a full (world-size-independent) state dict and saved
+  alongside the model in the same `optimizer_state.safetensors` / `optimizer_param_groups.json`
+  format as single-GPU training, so **resume-from-checkpoint is supported** with `--resume=true`.
+  Resume reshards both the model and the optimizer state to the _current_ FSDP topology, so you can
+  resume an FSDP checkpoint on a different number of GPUs. Note that the data sampler is only
+  sample-exact when the world size and batch size match the original run (a warning is logged
+  otherwise); the optimizer/model state itself is unaffected.
+
 ## Notes
 
 - The `--policy.use_amp` flag in `lerobot-train` is only used when **not** running with accelerate. When using accelerate, mixed precision is controlled by accelerate's configuration.

docs/source/multi_task_dit.mdx

@@ -314,7 +314,7 @@ lerobot-train \
   --steps=30000 \
   --save_freq=1000 \
   --log_freq=100 \
-  --eval_freq=1000 \
+  --env_eval_freq=1000 \
   --policy.type=multi_task_dit \
   --policy.device=cuda \
   --policy.horizon=32 \

docs/source/peft_training.mdx

@@ -28,13 +28,15 @@ lerobot-train \
  --steps=100000 \
  --batch_size=32 \
  --peft.method_type=LORA \
- --peft.r=64
+ --peft.r=64 \
+ --peft.lora_alpha=64
 ```
 
 Note the `--peft.method_type` parameter that let's you select which PEFT method to use. Here we use
 [LoRA](https://huggingface.co/docs/peft/main/en/package_reference/lora) (Low-Rank Adapter) which is probably the most
 popular fine-tuning method to date. Low-rank adaption means that we only fine-tune a matrix with comparably low rank
-instead of the full weight matrix. This rank can be specified using the `--peft.r` parameter. The higher the rank
+instead of the full weight matrix. This rank can be specified using the `--peft.r` parameter, and the LoRA scaling factor with
+`--peft.lora_alpha` (where `scaling = lora_alpha / r`). The higher the rank
 the closer you get to full fine-tuning
 
 There are more complex methods that have more parameters. These are not yet supported, feel free to raise an issue

docs/source/pi05.mdx

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

docs/source/pi0fast.mdx

@@ -96,7 +96,7 @@ lerobot-train \
     --policy.type=pi0_fast \
     --output_dir=./outputs/pi0fast_training \
     --job_name=pi0fast_training \
-    --policy.pretrained_path=lerobot/pi0_fast_base \
+    --policy.pretrained_path=lerobot/pi0fast-base \
     --policy.dtype=bfloat16 \
     --policy.gradient_checkpointing=true \
     --policy.chunk_size=10 \
@@ -187,7 +187,7 @@ lerobot-train \
   --dataset.repo_id=lerobot/libero \
   --output_dir=outputs/libero_pi0fast \
   --job_name=libero_pi0fast \
-  --policy.path=lerobot/pi0fast_base \
+  --policy.path=lerobot/pi0fast-base \
   --policy.dtype=bfloat16 \
   --steps=100000 \
   --save_freq=20000 \

docs/source/policy_evo1_README.md

@@ -0,0 +1,18 @@
+# EVO1
+
+EVO1 is a Vision-Language-Action policy for robot control. The LeRobot
+integration uses an InternVL3 vision-language backbone with a flow-matching
+action head, and supports staged training through the standard LeRobot policy
+APIs.
+
+The upstream EVO1 project is available at
+[MINT-SJTU/Evo-1](https://github.com/MINT-SJTU/Evo-1).
+
+```bibtex
+@misc{evo1,
+  title = {EVO1},
+  author = {{MINT-SJTU}},
+  year = {2025},
+  howpublished = {\url{https://github.com/MINT-SJTU/Evo-1}},
+}
+```

docs/source/policy_fastwam_README.md

@@ -0,0 +1,56 @@
+## Research Paper
+
+Paper: https://arxiv.org/abs/2603.16666
+
+## Repository
+
+Code: https://github.com/yuantianyuan01/FastWAM
+
+Project page: https://yuantianyuan01.github.io/FastWAM/
+
+## Citation
+
+```bibtex
+@article{yuan2026fastwam,
+  title = {Fast-WAM: Do World Action Models Need Test-time Future Imagination?},
+  author = {Tianyuan Yuan and Zibin Dong and Yicheng Liu and Hang Zhao},
+  journal = {arXiv preprint arXiv:2603.16666},
+  year = {2026},
+  url = {https://arxiv.org/abs/2603.16666}
+}
+```
+
+## Additional Resources
+
+Base video model: https://huggingface.co/Wan-AI/Wan2.2-TI2V-5B
+
+Released upstream checkpoints: https://huggingface.co/yuanty/fastwam
+
+## Results
+
+Evaluated on LIBERO with [`ZibinDong/fastwam_libero_uncond_2cam224`](https://huggingface.co/ZibinDong/fastwam_libero_uncond_2cam224):
+
+| Suite          | Success rate | n_episodes |
+| -------------- | -----------: | ---------: |
+| libero_spatial |        97.6% |        500 |
+| libero_object  |        99.0% |        500 |
+| libero_goal    |        95.0% |        500 |
+| libero_10      |        94.0% |        500 |
+| **average**    |    **96.4%** |       2000 |
+
+Reproduce: `lerobot-eval --policy.path=ZibinDong/fastwam_libero_uncond_2cam224 --policy.device=cuda --policy.torch_dtype=float32 --policy.n_action_steps=10 --env.type=libero --env.task=libero_spatial --env.observation_height=256 --env.observation_width=256 --eval.batch_size=1 --eval.n_episodes=50 --seed=0 --env.episode_length=300`.
+
+For LIBERO-10, use `--env.task=libero_10 --env.episode_length=600`:
+
+```bash
+lerobot-eval \
+    --policy.path=ZibinDong/fastwam_libero_uncond_2cam224 \
+    --policy.device=cuda \
+    --policy.torch_dtype=float32 \
+    --policy.n_action_steps=10 \
+    --env.type=libero \
+    --env.task=libero_10 --env.observation_height=256 --env.observation_width=256 \
+    --eval.batch_size=1 \
+    --eval.n_episodes=50 \
+    --seed=0 --env.episode_length=600
+```

docs/source/policy_groot_README.md

@@ -1,6 +1,13 @@
 ## Research Paper
 
-Paper: https://research.nvidia.com/labs/gear/gr00t-n1_5/
+GR00T N1 technical report (covers the GR00T N1.x family, including N1.7): https://arxiv.org/abs/2503.14734
+
+GR00T N1.7 model card: https://huggingface.co/nvidia/GR00T-N1.7-3B
+
+GR00T N1.5 research page (earlier version): https://research.nvidia.com/labs/gear/gr00t-n1_5/
+
+> GR00T N1.5 support was removed from LeRobot; the last release supporting it is `lerobot==0.5.1`.
+> Current releases support GR00T N1.7 only.
 
 ## Repository
 
@@ -24,4 +31,108 @@ Code: https://github.com/NVIDIA/Isaac-GR00T
 
 Blog: https://developer.nvidia.com/isaac/gr00t
 
-Hugging Face Model: https://huggingface.co/nvidia/GR00T-N1.5-3B
+Hugging Face Models:
+
+- GR00T N1.7: https://huggingface.co/nvidia/GR00T-N1.7-3B
+- GR00T N1.7 LIBERO checkpoints: https://huggingface.co/nvidia/GR00T-N1.7-LIBERO
+
+<details>
+<summary><b>Original-vs-LeRobot parity test</b></summary>
+
+## Original-vs-LeRobot parity test
+
+`tests/policies/groot/test_groot_vs_original.py` verifies this LeRobot
+reimplementation of GR00T N1.7 (Qwen3-VL backbone + flow-matching action head)
+against NVIDIA's original `gr00t` package with two comparisons, each parametrized
+over every embodiment tag present in the checkpoint:
+
+1. **Model parity** — given byte-identical pre-processed inputs and the same
+   flow-matching seed (recorded in each artifact), both implementations must produce
+   the **same raw model output** (`get_action(...)["action_pred"]`, the normalized
+   flow-matching prediction). Output shapes must match exactly; any action-horizon
+   or action-dim mismatch fails the test.
+2. **Preprocessor parity** — given the identical raw observations (per-camera
+   frames, state vectors, language instruction), LeRobot's own preprocessor pipeline
+   (real Qwen3-VL chat template / tokenizer / image packing + checkpoint-driven
+   state normalization, no mocks) must produce the **same collated model inputs**
+   (`input_ids`, `attention_mask`, `pixel_values`, `image_grid_thw`, `state`,
+   `embodiment_id`) as the original package's processor.
+
+### Why two environments
+
+The original `gr00t` package pins `transformers==4.57.3` (Python 3.10); this
+integration requires `transformers>=5.x` (Qwen3-VL). Under 5.x, `PretrainedConfig`
+is itself a defaulted dataclass, so the original config dataclasses fail to import
+(`non-default argument follows default argument`). The two implementations therefore
+**cannot be imported in the same Python process**.
+
+So the test uses a **producer / consumer** split across two venvs:
+
+1. **Producer** — `tests/policies/groot/utils/dump_original_n1_7.py`, run in the _original_
+   gr00t venv. For each embodiment it builds dummy inputs generically from the
+   checkpoint metadata (state dims from `statistics.json`; camera/language keys from
+   the processor modality configs), runs the original model, and saves to one `.npz`
+   per tag: the raw observations (`raw::` keys), the exact collated inputs
+   (`in::` keys), the seed, and the raw `action_pred`.
+2. **Consumer** — the pytest above, run in the _LeRobot_ venv. It discovers every
+   `.npz`; the model-parity case replays the byte-identical collated inputs through
+   the LeRobot model with the recorded seed and asserts the outputs match, and the
+   preprocessor-parity case replays the raw observations through LeRobot's full
+   preprocessor pipeline and asserts the collated tensors match.
+
+> Artifacts generated by older versions of the dump script contain no `raw::`
+> fields; the preprocessor-parity case then **skips** with a regeneration hint.
+> Re-run the producer to refresh them.
+
+### Fairness controls
+
+- **Same pre-processed inputs (model parity)** — the original processor's `input_ids`,
+  `pixel_values`, `image_grid_thw`, `attention_mask`, `state`, `embodiment_id` are
+  fed verbatim to the LeRobot model (no re-tokenization / re-normalization), so the
+  model comparison isolates the model. LeRobot's own tokenization / image packing is
+  covered separately by the preprocessor-parity case, which compares its output
+  against those same collated tensors from identical raw observations.
+- **Same precision + attention kernel** — both sides run **fp32 + SDPA**. The
+  original defaults to `use_flash_attention=True` (flash_attention_2 + bf16); the
+  producer forces SDPA + fp32. (With the defaults the gap is ~3e-2 — pure
+  kernel/rounding noise, not an implementation difference.)
+- **Same flow-matching seed** — fixed right before sampling on both sides; the
+  producer records it in each artifact (`--seed`, default 42) and the consumer
+  replays the recorded value.
+
+### How to run
+
+```bash
+# Resolve a local checkpoint (GR00T-N1.7-LIBERO / libero_10)
+CKPT=$(python - <<'PY'
+import os
+from huggingface_hub import snapshot_download
+print(os.path.join(snapshot_download("nvidia/GR00T-N1.7-LIBERO",
+      allow_patterns=["libero_10/*"]), "libero_10"))
+PY
+)
+
+# 1) Produce the original-side artifacts for all embodiments (original gr00t venv, CUDA)
+CUDA_VISIBLE_DEVICES=0 /path/to/Isaac-GR00T/.venv-original/bin/python \
+    tests/policies/groot/utils/dump_original_n1_7.py \
+    --ckpt "$CKPT" --out-dir tests/policies/groot/artifacts --device cuda --seed 42
+
+# 2) Run the parity test (LeRobot venv) — one parametrized case per embodiment
+CUDA_VISIBLE_DEVICES=0 GROOT_PARITY_DEVICE=cuda \
+    uv run pytest tests/policies/groot/test_groot_vs_original.py -v -s
+```
+
+The `.npz` artifacts are local-only (gitignored, ~6–10 MB each) and are regenerated by
+the producer; they are never committed. The tests **skip** (do not fail) on CI or
+when the checkpoint / artifacts are absent.
+
+#### Env knobs (all optional)
+
+| Var                                       | Default                          | Purpose                               |
+| ----------------------------------------- | -------------------------------- | ------------------------------------- |
+| `GROOT_N1_7_PARITY_DIR`                   | `tests/policies/groot/artifacts` | directory of per-tag `.npz` artifacts |
+| `GROOT_N1_7_LIBERO_CKPT`                  | auto (HF cache)                  | override checkpoint dir               |
+| `GROOT_PARITY_DEVICE`                     | `cuda` if available              | `cpu` or `cuda`                       |
+| `GROOT_PARITY_ATOL` / `GROOT_PARITY_RTOL` | `1e-3`                           | comparison tolerance                  |
+
+</details>

docs/source/policy_molmoact2_README.md

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

docs/source/policy_vla_jepa_README.md

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

docs/source/porting_datasets_v3.mdx

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

docs/source/reachy2.mdx

@@ -161,7 +161,7 @@ lerobot-record \
     --dataset.private=true \
     --dataset.streaming_encoding=true \
     --dataset.encoder_threads=2 \
-    # --dataset.vcodec=auto \
+    # --dataset.rgb_encoder.vcodec=auto \
     --display_data=true
 ```
 
@@ -203,7 +203,7 @@ lerobot-record \
     --dataset.private=true \
     --dataset.streaming_encoding=true \
     --dataset.encoder_threads=2 \
-    # --dataset.vcodec=auto \
+    # --dataset.rgb_encoder.vcodec=auto \
     --display_data=true
 ```
 

docs/source/rebot_b601.mdx

@@ -0,0 +1,186 @@
+# reBot B601-DM
+
+[reBot B601-DM](https://wiki.seeedstudio.com/rebot_arm_b601_dm_lerobot/) is an open-source, low-cost robot arm from Seeed Studio for embodied-AI and imitation learning. It comes as a **follower** arm (the `B601-DM`, a 6-DOF arm plus gripper driven by Damiao CAN motors) and a **leader** arm (the `StarArm102` / `reBot Arm 102`, driven by FashionStar UART smart servos) used to teleoperate it.
+
+This page covers **calibration** and **teleoperation** for both single-arm and bimanual (dual-arm) setups.
+
+<div style="display: flex; align-items: center; gap: 10px;">
+  <img
+    src="https://files.seeedstudio.com/wiki/robotics/projects/lerobot/b601dm_zeroposition.jpg"
+    alt="reBot B601-DM follower arm at its zero position"
+    width="48%"
+  />
+  <img
+    src="https://files.seeedstudio.com/wiki/robotics/projects/lerobot/102_zeroposition.jpg"
+    alt="reBot Arm 102 leader arm at its zero position"
+    width="48%"
+  />
+</div>
+
+_Left: the B601-DM follower at its zero position. Right: the reBot Arm 102 leader at its zero position. Images courtesy of [Seeed Studio](https://wiki.seeedstudio.com/rebot_arm_b601_dm_lerobot/)._
+
+## Install LeRobot 🤗
+
+Follow our [Installation Guide](./installation), then install the reBot support:
+
+```bash
+pip install -e ".[rebot]"
+```
+
+This pulls in `motorbridge` (CAN motor control for the B601-DM follower) and `motorbridge-smart-servo` (FashionStar UART servos for the reBot Arm 102 leader).
+
+## Registered device types
+
+| Type                     | Kind                                         |
+| ------------------------ | -------------------------------------------- |
+| `rebot_b601_follower`    | single-arm B601-DM follower robot            |
+| `bi_rebot_b601_follower` | bimanual (dual-arm) follower robot           |
+| `rebot_102_leader`       | single-arm reBot Arm 102 leader teleoperator |
+| `bi_rebot_102_leader`    | bimanual (dual-arm) leader teleoperator      |
+
+The bimanual types compose two single-arm instances and namespace each arm's
+observation/action keys with a `left_` / `right_` prefix. Per-arm settings are
+passed through nested `left_arm_config.*` / `right_arm_config.*` arguments.
+
+## Find the USB ports
+
+For each device, find the USB port associated with its motor bus using:
+
+```bash
+lerobot-find-port
+```
+
+<Tip warning={true}>
+  On Linux, remove `brltty` (`sudo apt remove brltty`) so it does not hold the
+  leader's USB serial port. You may also need to grant access to the serial
+  devices: `sudo chmod 666 /dev/ttyACM* /dev/ttyUSB*`.
+</Tip>
+
+## Calibration
+
+Neither arm stores a persistent hardware calibration: every time it connects, the motors are re-zeroed against the pose the arm is physically holding. Calibration simply records that zero pose. When prompted, **manually move the arm to its zero position** (the default sit-down pose shown above, gripper fully closed) and press <kbd>ENTER</kbd>.
+
+### Follower (B601-DM)
+
+<hfoptions id="calibrate-follower">
+<hfoption id="Single arm">
+
+```bash
+lerobot-calibrate \
+    --robot.type=rebot_b601_follower \
+    --robot.port=/dev/ttyACM0 \
+    --robot.id=follower \
+    --robot.can_adapter=damiao
+```
+
+</hfoption>
+<hfoption id="Dual arm">
+
+Connect the bimanual follower; calibration runs for the left arm, then the right arm.
+
+```bash
+lerobot-calibrate \
+    --robot.type=bi_rebot_b601_follower \
+    --robot.id=bi_follower \
+    --robot.left_arm_config.port=/dev/ttyACM0 \
+    --robot.left_arm_config.can_adapter=damiao \
+    --robot.right_arm_config.port=/dev/ttyACM1 \
+    --robot.right_arm_config.can_adapter=damiao
+```
+
+Per-arm calibration files are saved with `_left` / `_right` suffixes on the id.
+
+</hfoption>
+</hfoptions>
+
+### Leader (reBot Arm 102)
+
+<hfoptions id="calibrate-leader">
+<hfoption id="Single arm">
+
+```bash
+lerobot-calibrate \
+    --teleop.type=rebot_102_leader \
+    --teleop.port=/dev/ttyUSB0 \
+    --teleop.id=leader
+```
+
+</hfoption>
+<hfoption id="Dual arm">
+
+```bash
+lerobot-calibrate \
+    --teleop.type=bi_rebot_102_leader \
+    --teleop.id=bi_leader \
+    --teleop.left_arm_config.port=/dev/ttyUSB0 \
+    --teleop.right_arm_config.port=/dev/ttyUSB1
+```
+
+</hfoption>
+</hfoptions>
+
+## Teleoperation
+
+Once both arms are calibrated, drive the follower with the leader. The follower talks to its CAN bus through a Damiao serial bridge (`can_adapter=damiao`, the default) or a SocketCAN adapter (`can_adapter=socketcan`). See the [OpenArm page](./openarm) for more details on the SocketCAN adapter configuration.
+
+<hfoptions id="teleoperate">
+<hfoption id="Single arm">
+
+```bash
+lerobot-teleoperate \
+    --robot.type=rebot_b601_follower \
+    --robot.port=/dev/ttyACM0 \
+    --robot.id=follower \
+    --robot.can_adapter=damiao \
+    --teleop.type=rebot_102_leader \
+    --teleop.port=/dev/ttyUSB0 \
+    --teleop.id=leader
+```
+
+</hfoption>
+<hfoption id="Dual arm">
+
+The bimanual leader and follower reuse the single-arm classes; each arm is
+configured through nested `left_arm_config.*` / `right_arm_config.*` arguments,
+so a bimanual reBot Arm 102 leader drives a bimanual B601-DM follower.
+
+```bash
+lerobot-teleoperate \
+    --robot.type=bi_rebot_b601_follower \
+    --robot.id=bi_follower \
+    --robot.left_arm_config.port=/dev/ttyACM0 \
+    --robot.left_arm_config.can_adapter=damiao \
+    --robot.right_arm_config.port=/dev/ttyACM1 \
+    --robot.right_arm_config.can_adapter=damiao \
+    --teleop.type=bi_rebot_102_leader \
+    --teleop.id=bi_leader \
+    --teleop.left_arm_config.port=/dev/ttyUSB0 \
+    --teleop.right_arm_config.port=/dev/ttyUSB1
+```
+
+</hfoption>
+</hfoptions>
+
+<Tip>
+  The leader and follower share the same joint names (`shoulder_pan,
+  shoulder_lift, elbow_flex, wrist_flex, wrist_yaw, wrist_roll, gripper`), so
+  leader actions map directly onto the follower.
+</Tip>
+
+If the motion of a joint is reversed, flip its sign in the leader's `joint_directions` (the gripper also carries a scale to widen its range to the follower):
+
+```bash
+lerobot-teleoperate \
+    --robot.type=rebot_b601_follower \
+    --robot.port=/dev/ttyACM0 \
+    --robot.can_adapter=damiao \
+    --teleop.type=rebot_102_leader \
+    --teleop.port=/dev/ttyUSB0 \
+    --teleop.joint_directions='{"shoulder_pan":-1,"shoulder_lift":-1,"elbow_flex":1,"wrist_flex":1,"wrist_yaw":1,"wrist_roll":-1,"gripper":-6}'
+```
+
+## Recording datasets
+
+Swap `lerobot-teleoperate` for `lerobot-record` (with the same `--robot.*` / `--teleop.*` arguments, plus `--dataset.*`) to record demonstrations for training. See [Imitation Learning for Robots](./il_robots) for the full workflow.
+
+For hardware assembly and wiring, see the [Seeed Studio reBot wiki](https://wiki.seeedstudio.com/rebot_arm_b601_dm_lerobot/).

docs/source/robocasa.mdx

@@ -166,7 +166,7 @@ lerobot-train \
   --output_dir=./outputs/smolvla_robocasa_CloseFridge \
   --steps=100000 \
   --batch_size=4 \
-  --eval_freq=5000 \
+  --env_eval_freq=5000 \
   --eval.batch_size=1 \
   --eval.n_episodes=5 \
   --save_freq=10000

docs/source/robometer.mdx

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

docs/source/smolvla.mdx

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

docs/source/so101.mdx

@@ -122,7 +122,7 @@ The video below shows the sequence of steps for setting the motor ids.
 
 #### Follower
 
-Connect the usb cable from your computer and the power supply to the follower arm's controller board. Then, run the following command or run the API example with the port you got from the previous step. You'll also need to give your leader arm a name with the `id` parameter.
+Connect the usb cable from your computer and the power supply to the follower arm's controller board. Then, run the following command or run the API example with the port you got from the previous step. You'll also need to give your follower arm a name with the `id` parameter.
 
 <hfoptions id="setup_motors">
 <hfoption id="Command">

docs/source/streaming_video_encoding.mdx

@@ -17,9 +17,9 @@ This makes `save_episode()` near-instant (the video is already encoded by the ti
 | Parameter               | CLI Flag                          | Type          | Default       | Description                                                       |
 | ----------------------- | --------------------------------- | ------------- | ------------- | ----------------------------------------------------------------- |
 | `streaming_encoding`    | `--dataset.streaming_encoding`    | `bool`        | `True`        | Enable real-time encoding during capture                          |
-| `vcodec`                | `--dataset.vcodec`                | `str`         | `"libsvtav1"` | Video codec. `"auto"` detects best HW encoder                     |
+| `vcodec`                | `--dataset.rgb_encoder.vcodec`    | `str`         | `"libsvtav1"` | Video codec. `"auto"` detects best HW encoder                     |
 | `encoder_threads`       | `--dataset.encoder_threads`       | `int \| None` | `None` (auto) | Threads per encoder instance. `None` will leave the vcoded decide |
-| `encoder_queue_maxsize` | `--dataset.encoder_queue_maxsize` | `int`         | `60`          | Max buffered frames per camera (~2s at 30fps). Consumes RAM       |
+| `encoder_queue_maxsize` | `--dataset.encoder_queue_maxsize` | `int`         | `30`          | Max buffered frames per camera (~1s at 30fps). Consumes RAM       |
 
 ## 3. Performance Considerations
 
@@ -48,7 +48,7 @@ This parameter controls how many threads each encoder instance uses internally:
 
 ### Backpressure and Frame Dropping
 
-Each camera has a bounded queue (`encoder_queue_maxsize`, default 60 frames). When the encoder can't keep up:
+Each camera has a bounded queue (`encoder_queue_maxsize`, default 30 frames). When the encoder can't keep up:
 
 1. The queue fills up (consuming RAM)
 2. New frames are **dropped** (not blocked) — the capture loop continues uninterrupted
@@ -82,15 +82,15 @@ Use HW encoding when:
 
 ### Available HW Encoders
 
-| Encoder             | Platform      | Hardware                                                                                         | CLI Value                            |
-| ------------------- | ------------- | ------------------------------------------------------------------------------------------------ | ------------------------------------ |
-| `h264_videotoolbox` | macOS         | Apple Silicon / Intel                                                                            | `--dataset.vcodec=h264_videotoolbox` |
-| `hevc_videotoolbox` | macOS         | Apple Silicon / Intel                                                                            | `--dataset.vcodec=hevc_videotoolbox` |
-| `h264_nvenc`        | Linux/Windows | NVIDIA GPU                                                                                       | `--dataset.vcodec=h264_nvenc`        |
-| `hevc_nvenc`        | Linux/Windows | NVIDIA GPU                                                                                       | `--dataset.vcodec=hevc_nvenc`        |
-| `h264_vaapi`        | Linux         | Intel/AMD GPU                                                                                    | `--dataset.vcodec=h264_vaapi`        |
-| `h264_qsv`          | Linux/Windows | Intel Quick Sync                                                                                 | `--dataset.vcodec=h264_qsv`          |
-| `auto`              | Any           | Probes the system for available HW encoders. Falls back to `libsvtav1` if no HW encoder is found | `--dataset.vcodec=auto`              |
+| Encoder             | Platform      | Hardware                                                                                         | CLI Value                                        |
+| ------------------- | ------------- | ------------------------------------------------------------------------------------------------ | ------------------------------------------------ |
+| `h264_videotoolbox` | macOS         | Apple Silicon / Intel                                                                            | `--dataset.rgb_encoder.vcodec=h264_videotoolbox` |
+| `hevc_videotoolbox` | macOS         | Apple Silicon / Intel                                                                            | `--dataset.rgb_encoder.vcodec=hevc_videotoolbox` |
+| `h264_nvenc`        | Linux/Windows | NVIDIA GPU                                                                                       | `--dataset.rgb_encoder.vcodec=h264_nvenc`        |
+| `hevc_nvenc`        | Linux/Windows | NVIDIA GPU                                                                                       | `--dataset.rgb_encoder.vcodec=hevc_nvenc`        |
+| `h264_vaapi`        | Linux         | Intel/AMD GPU                                                                                    | `--dataset.rgb_encoder.vcodec=h264_vaapi`        |
+| `h264_qsv`          | Linux/Windows | Intel Quick Sync                                                                                 | `--dataset.rgb_encoder.vcodec=h264_qsv`          |
+| `auto`              | Any           | Probes the system for available HW encoders. Falls back to `libsvtav1` if no HW encoder is found | `--dataset.rgb_encoder.vcodec=auto`              |
 
 > [!NOTE]
 > In order to use the HW accelerated encoders you might need to upgrade your GPU drivers.
@@ -100,15 +100,15 @@ Use HW encoding when:
 
 ## 5. Troubleshooting
 
-| Symptom                                                            | Likely Cause                                 | Fix                                                                                                                                                                                                                                                                                  |
-| ------------------------------------------------------------------ | -------------------------------------------- | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ |
-| System freezes or choppy robot movement or Rerun visualization lag | CPU starved (100% load usage)                | Close other apps, reduce encoding throughput, lower `encoder_threads`, use `h264`, use `display_data=False`. If the CPU continues to be at 100% then it might be insufficient for your setup, consider `--dataset.streaming_encoding=false` or HW encoding (`--dataset.vcodec=auto`) |
-| "Encoder queue full" warnings or dropped frames in dataset         | Encoder can't keep up (Queue overflow)       | If CPU is not at 100%: Increase `encoder_threads`, increase `encoder_queue_maxsize` or use HW encoding (`--dataset.vcodec=auto`).                                                                                                                                                    |
-| High RAM usage                                                     | Queue filling faster than encoding           | `encoder_threads` too low or CPU insufficient. Reduce `encoder_queue_maxsize` or use HW encoding                                                                                                                                                                                     |
-| Large video files                                                  | Using HW encoder or H.264                    | Expected trade-off. Switch to `libsvtav1` if CPU allows                                                                                                                                                                                                                              |
-| `save_episode()` still slow                                        | `streaming_encoding` is `False`              | Set `--dataset.streaming_encoding=true`                                                                                                                                                                                                                                              |
-| Encoder thread crash                                               | Codec not available or invalid settings      | Check `vcodec` is installed, try `--dataset.vcodec=auto`                                                                                                                                                                                                                             |
-| Recorded dataset is missing frames                                 | CPU/GPU starvation or occasional load spikes | If ~5% of frames are missing, your system is likely overloaded — follow the recommendations above. If fewer frames are missing (~2%), they are probably due to occasional transient load spikes (often at startup) and can be considered expected.                                   |
+| Symptom                                                            | Likely Cause                                 | Fix                                                                                                                                                                                                                                                                                              |
+| ------------------------------------------------------------------ | -------------------------------------------- | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ |
+| System freezes or choppy robot movement or Rerun visualization lag | CPU starved (100% load usage)                | Close other apps, reduce encoding throughput, lower `encoder_threads`, use `h264`, use `display_data=False`. If the CPU continues to be at 100% then it might be insufficient for your setup, consider `--dataset.streaming_encoding=false` or HW encoding (`--dataset.rgb_encoder.vcodec=auto`) |
+| "Encoder queue full" warnings or dropped frames in dataset         | Encoder can't keep up (Queue overflow)       | If CPU is not at 100%: Increase `encoder_threads`, increase `encoder_queue_maxsize` or use HW encoding (`--dataset.rgb_encoder.vcodec=auto`).                                                                                                                                                    |
+| High RAM usage                                                     | Queue filling faster than encoding           | `encoder_threads` too low or CPU insufficient. Reduce `encoder_queue_maxsize` or use HW encoding                                                                                                                                                                                                 |
+| Large video files                                                  | Using HW encoder or H.264                    | Expected trade-off. Switch to `libsvtav1` if CPU allows                                                                                                                                                                                                                                          |
+| `save_episode()` still slow                                        | `streaming_encoding` is `False`              | Set `--dataset.streaming_encoding=true`                                                                                                                                                                                                                                                          |
+| Encoder thread crash                                               | Codec not available or invalid settings      | Check `vcodec` is installed, try `--dataset.rgb_encoder.vcodec=auto`                                                                                                                                                                                                                             |
+| Recorded dataset is missing frames                                 | CPU/GPU starvation or occasional load spikes | If ~5% of frames are missing, your system is likely overloaded — follow the recommendations above. If fewer frames are missing (~2%), they are probably due to occasional transient load spikes (often at startup) and can be considered expected.                                               |
 
 ## 6. Recommended Configurations
 
@@ -146,7 +146,7 @@ On very constrained systems, streaming encoding may compete too heavily with the
 # 2camsx 640x480x3 @30fps: Requires some tuning.
 
 # Use H.264, disable streaming, consider batching encoding
-lerobot-record --dataset.vcodec=h264 --dataset.streaming_encoding=false ...
+lerobot-record --dataset.rgb_encoder.vcodec=h264 --dataset.streaming_encoding=false ...
 ```
 
 ## 7. Closing note

docs/source/tools.mdx

@@ -0,0 +1,210 @@
+# Tools
+
+LeRobot v3.1 supports **tool calls** in policies — assistant messages can
+emit structured invocations like `say(text="OK, starting now")` that the
+runtime dispatches to a real implementation (TTS, controller, logger, …).
+
+This page covers:
+
+1. Where the tool catalog lives.
+2. How the annotation pipeline produces tool-call atoms.
+3. How to add your own tool.
+
+## Where tools are declared
+
+Two layers.
+
+**The catalog** — a list of OpenAI-style function schemas — lives at
+`meta/info.json["tools"]` on each dataset. Example:
+
+```json
+{
+  "features": { "...": "..." },
+  "tools": [
+    {
+      "type": "function",
+      "function": {
+        "name": "say",
+        "description": "Speak a short utterance to the user via the TTS executor.",
+        "parameters": {
+          "type": "object",
+          "properties": {
+            "text": {
+              "type": "string",
+              "description": "The verbatim text to speak."
+            }
+          },
+          "required": ["text"]
+        }
+      }
+    }
+  ]
+}
+```
+
+Read it via the dataset metadata accessor:
+
+```python
+from lerobot.datasets.dataset_metadata import LeRobotDatasetMetadata
+
+meta = LeRobotDatasetMetadata(repo_id="pepijn/super_poulain_final_annotations")
+tools = meta.tools     # list[dict] — OpenAI tool schemas
+```
+
+If the dataset's `info.json` doesn't declare any tools, `meta.tools`
+returns `DEFAULT_TOOLS` from `lerobot.datasets.language` — currently a
+single-entry list with the canonical `say` schema. So unannotated
+datasets and chat-template consumers keep working without any
+configuration:
+
+```python
+prompt_str = tokenizer.apply_chat_template(
+    sample["messages"],
+    tools=meta.tools,                 # works either way
+    add_generation_prompt=False,
+    tokenize=False,
+)
+```
+
+**The implementations** — runnable Python — will live under
+`src/lerobot/tools/`, one file per tool. The runtime dispatcher and
+the canonical `say` implementation (wrapping Kyutai's pocket-tts) are
+not part of the catalog layer described here; today this layer ships
+only the schema storage and the `DEFAULT_TOOLS` fallback constant.
+
+## Per-row tool _invocations_
+
+The catalog above describes _what can be called_. The actual _call_ — the
+function name plus the argument values — is stored per-row, on the
+assistant atoms in `language_events`:
+
+```python
+{
+  "role": "assistant",
+  "content": null,
+  "style": null,
+  "timestamp": 12.4,
+  "camera": null,
+  "tool_calls": [
+    { "type": "function",
+      "function": { "name": "say", "arguments": { "text": "On it." } } }
+  ]
+}
+```
+
+Recipes splice these into rendered messages via `tool_calls_from`:
+
+```yaml
+user_interjection_response:
+  bindings:
+    speech: "emitted_at(t, role=assistant, tool_name=say)"
+  messages:
+    - { role: user, content: "${task}", stream: high_level }
+    - {
+        role: assistant,
+        content: "${current_plan}",
+        stream: high_level,
+        target: true,
+        tool_calls_from: speech,
+      }
+```
+
+The model's training target is one assistant turn that carries both the
+plan text _and_ the `say` tool call. At inference, the runtime parses
+the generated text back into structured `tool_calls` and dispatches to
+the matching implementation.
+
+## How to add your own tool
+
+> **Note:** Steps 2 and 3 below describe the runtime layer
+> (`src/lerobot/tools/`, the `Tool` protocol, `TOOL_REGISTRY`,
+> `get_tools(meta)`) which is not part of the catalog layer shipped
+> today — those modules don't yet exist in the tree. Step 1 alone is
+> enough to make the tool visible to the chat template via
+> `meta.tools` so the model can learn to _generate_ the call;
+> executing the call at inference requires the runtime layer.
+
+Three steps. Concrete example: a `record_observation` tool the policy
+can call to capture an extra observation outside the regular control
+loop.
+
+### Step 1 — declare the schema
+
+Add an entry under `meta/info.json["tools"]`. Either edit the file
+directly on disk _before_ running the annotation pipeline (it'll be
+preserved) or hand it to `lerobot-annotate` via a config flag.
+
+```json
+{
+  "tools": [
+    { "type": "function", "function": { "name": "say", "...": "..." } },
+    {
+      "type": "function",
+      "function": {
+        "name": "record_observation",
+        "description": "Capture a high-resolution still image for the user.",
+        "parameters": {
+          "type": "object",
+          "properties": {
+            "label": {
+              "type": "string",
+              "description": "Short label for the saved image."
+            }
+          },
+          "required": ["label"]
+        }
+      }
+    }
+  ]
+}
+```
+
+The schema follows OpenAI's function-calling convention exactly, so the
+chat template can render it natively.
+
+### Step 2 — implement the call
+
+Create `src/lerobot/tools/record_observation.py`:
+
+```python
+from .base import Tool
+from typing import Any
+
+RECORD_OBSERVATION_SCHEMA: dict[str, Any] = { "...": "..." }   # mirrors the JSON above
+
+
+class RecordObservationTool:
+    name = "record_observation"
+    schema = RECORD_OBSERVATION_SCHEMA
+
+    def __init__(self, schema: dict | None = None, output_dir: str = "."):
+        self.output_dir = output_dir
+
+    def call(self, arguments: dict) -> str:
+        label = arguments["label"]
+        # ... save the latest camera frame to <output_dir>/<label>.png ...
+        return f"saved {label}.png"
+```
+
+One file per tool keeps dependencies isolated — `record_observation`
+might pull `pillow`, while `say` pulls `pocket-tts`. Users installing
+only the tools they need avoid heavy transitive deps.
+
+### Step 3 — register it
+
+Add to `src/lerobot/tools/registry.py`:
+
+```python
+from .record_observation import RecordObservationTool
+
+TOOL_REGISTRY["record_observation"] = RecordObservationTool
+```
+
+That's it. At runtime `get_tools(meta)` looks up each schema in
+`meta.tools`, instantiates the matching registered class, and returns
+a name → instance dict the dispatcher can route into.
+
+If you want to use a tool _without_ writing an implementation (e.g. for
+training-time chat-template formatting only), step 1 alone is enough —
+the model still learns to _generate_ the call. Steps 2 and 3 are only
+needed to actually _execute_ it at inference.

docs/source/topreward.mdx

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

docs/source/using_dataset_tools.mdx

@@ -11,8 +11,9 @@ LeRobot provides several utilities for manipulating datasets:
 3. **Merge Datasets** - Combine multiple datasets into one. The datasets must have identical features, and episodes are concatenated in the order specified in `repo_ids`
 4. **Add Features** - Add new features to a dataset
 5. **Remove Features** - Remove features from a dataset
-6. **Convert to Video** - Convert image-based datasets to video format for efficient storage
-7. **Show the Info of Datasets** - Show the summary of datasets information such as number of episode etc.
+6. **Convert to Video** - Convert image-based datasets to video format for efficient storage (RGB and depth cameras are encoded with separate encoders)
+7. **Re-encode Videos** - Re-encode an existing video dataset's RGB and/or depth streams with new encoder settings
+8. **Show the Info of Datasets** - Show the summary of datasets information such as number of episode etc.
 
 The core implementation is in `lerobot.datasets.dataset_tools`.
 An example script detailing how to use the tools API is available in `examples/dataset/use_dataset_tools.py`.
@@ -117,10 +118,19 @@ lerobot-edit-dataset \
     --repo_id lerobot/pusht_image \
     --operation.type convert_image_to_video \
     --operation.output_dir outputs/pusht_video \
-    --operation.vcodec libsvtav1 \
-    --operation.pix_fmt yuv420p \
-    --operation.g 2 \
-    --operation.crf 30
+    --operation.rgb_encoder.vcodec libsvtav1 \
+    --operation.rgb_encoder.pix_fmt yuv420p \
+    --operation.rgb_encoder.g 2 \
+    --operation.rgb_encoder.crf 30
+
+# Convert a dataset that includes depth maps, customizing the depth encoder
+lerobot-edit-dataset \
+    --repo_id lerobot/pusht_image \
+    --operation.type convert_image_to_video \
+    --operation.output_dir outputs/pusht_video \
+    --operation.depth_encoder.depth_min 0.01 \
+    --operation.depth_encoder.depth_max 10.0 \
+    --operation.depth_encoder.use_log true
 
 # Convert only specific episodes
 lerobot-edit-dataset \
@@ -147,15 +157,42 @@ lerobot-edit-dataset \
 **Parameters:**
 
 - `output_dir`: Custom output directory (optional - by default uses `new_repo_id` or `{repo_id}_video`)
-- `vcodec`: Video codec to use - options: `h264`, `hevc`, `libsvtav1` (default: `libsvtav1`)
-- `pix_fmt`: Pixel format - options: `yuv420p`, `yuv444p` (default: `yuv420p`)
-- `g`: Group of pictures (GOP) size - lower values give better quality but larger files (default: 2)
-- `crf`: Constant rate factor - lower values give better quality but larger files, 0 is lossless (default: 30)
-- `fast_decode`: Fast decode tuning option (default: 0)
+- `rgb_encoder`: Video encoder settings applied to RGB cameras — all sub-fields accessible via `--operation.rgb_encoder.<field>`. See [Video Encoding Parameters](./video_encoding_parameters) for more details.
+- `depth_encoder`: Video encoder settings applied to depth-map cameras (e.g. from an Intel RealSense). In addition to the standard encoder fields it exposes the depth quantization knobs (`depth_min`, `depth_max`, `shift`, `use_log`), accessible via `--operation.depth_encoder.<field>`. These quantization settings are persisted to the dataset metadata so depth can be dequantized back to physical units on load. See the [Depth streams](./video_encoding_parameters#depth-streams) section for details.
 - `episode_indices`: List of specific episodes to convert (default: all episodes)
 - `num_workers`: Number of parallel workers for processing (default: 4)
 
-**Note:** The resulting dataset will be a proper LeRobotDataset with all cameras encoded as videos in the `videos/` directory, with parquet files containing only metadata (no raw image data). All episodes, stats, and tasks are preserved.
+**Note:** The resulting dataset will be a proper LeRobotDataset with all cameras encoded as videos in the `videos/` directory, with parquet files containing only metadata (no raw image data). Depth-map cameras are detected automatically and routed to the `depth_encoder`, while RGB cameras use the `rgb_encoder`. All episodes, stats, and tasks are preserved.
+
+#### Re-encode Videos
+
+Re-encode the videos of an existing video dataset with different encoder settings, without going back to raw frames. RGB videos use the `rgb_encoder` and depth videos use the `depth_encoder`. Provide only the encoder(s) you want to re-encode; the other stream type is left untouched.
+
+```bash
+# Re-encode all RGB videos with new settings (saves to lerobot/pusht_reencoded by default)
+lerobot-edit-dataset \
+    --repo_id lerobot/pusht \
+    --operation.type reencode_videos \
+    --operation.rgb_encoder.vcodec h264 \
+    --operation.rgb_encoder.pix_fmt yuv420p \
+    --operation.rgb_encoder.crf 23
+
+# Re-encode both RGB and depth videos in a dataset with depth maps
+lerobot-edit-dataset \
+    --repo_id lerobot/pusht_depth \
+    --operation.type reencode_videos \
+    --operation.rgb_encoder.vcodec h264 \
+    --operation.depth_encoder.crf 50
+```
+
+**Parameters:**
+
+- `rgb_encoder`: Encoder settings applied to every RGB video. Omit to skip re-encoding RGB videos.
+- `depth_encoder`: Encoder settings applied to every depth video. Omit to skip re-encoding depth videos.
+- `num_workers`: Number of parallel workers for processing.
+
+> [!NOTE]
+> When re-encoding depth videos, the existing depth quantization parameters (`depth_min`, `depth_max`, `shift`, `use_log`) and the `is_depth_map` flag are **preserved** — re-encoding only changes the codec/quality of the stored stream, not how depth is dequantized on load.
 
 ### Show the information of datasets
 
@@ -228,6 +265,8 @@ lerobot-dataset-viz \
 
 Once executed, the tool opens `rerun.io` and displays the camera streams, robot states, and actions for the selected episode.
 
+To use [Foxglove](https://foxglove.dev) instead of Rerun, install the extra add `--display-mode foxglove`. This starts a WebSocket server (connect the Foxglove app to `ws://127.0.0.1:8765`) that serves the episode as a seekable timeline you can play/pause and scrub.
+
 For advanced usage—including visualizing datasets stored on a remote server—run:
 
 ```bash

docs/source/video_encoding_parameters.mdx

@@ -0,0 +1,188 @@
+# Video encoding parameters
+
+When video storage is enabled, LeRobot stores each camera stream as an **MP4** file instead of saving one image file per timestep. Video encoding compresses across time, which usually cuts dataset size and I/O compared to a pile of PNG, while keeping MP4 — a format every player and loader understands.
+
+Encoding frames into an MP4 is a full FFmpeg pipeline: choice of encoder, pixel format, GOP/keyframes, quality vs. speed, and optional extra encoder flags. Most of these knobs are user-tunable through `rgb_encoder`, a nested `RGBEncoderConfig` (`lerobot.configs.video.RGBEncoderConfig`) passed through PyAV.
+
+You can set these parameters from the CLI with `--dataset.rgb_encoder.<field>` (e.g. with `lerobot-record` or `lerobot-rollout`). The same block applies to every camera video stream in that run.
+
+<Tip>
+  Video storage must be on for `rgb_encoder` to have any effect —
+  `use_videos=True` in Python APIs, or `--dataset.video=true` on the CLI (the
+  recording default). With video off, inputs stay as images and `rgb_encoder` is
+  ignored.
+</Tip>
+
+For details on **when** frames are written vs. encoded (streaming vs. post-episode), queues, and other top-level `--dataset.*` switches, see [Streaming Video Encoding](./streaming_video_encoding). For an encoding-parameter comparison and experiments, see the [video-benchmark Space](https://huggingface.co/spaces/lerobot/video-benchmark).
+
+---
+
+## Example
+
+```bash
+lerobot-record \
+    --robot.type=so100_follower \
+    --robot.port=/dev/tty.usbmodem58760431541 \
+    --robot.cameras="{laptop: {type: opencv, index_or_path: 0, width: 640, height: 480, fps: 30}}" \
+    --robot.id=black \
+    --teleop.type=so100_leader \
+    --teleop.port=/dev/tty.usbmodem58760431551 \
+    --teleop.id=blue \
+    --dataset.repo_id=<my_username>/<my_dataset_name> \
+    --dataset.num_episodes=2 \
+    --dataset.single_task="Grab the cube" \
+    --dataset.streaming_encoding=true \
+    --dataset.encoder_threads=2 \
+    --dataset.rgb_encoder.vcodec=h264 \
+    --dataset.rgb_encoder.preset=fast \
+    --dataset.rgb_encoder.extra_options={"tune": "film", "profile:v": "high", "bf": 2} \
+    --display_data=true
+```
+
+---
+
+## Tuning parameters
+
+<Tip warning={true}>
+The defaults are tuned to balance **compression ratio**, **visual quality**, and **decoding/seek speed** for typical robotics datasets. Changing them can affect both recording (CPU load, frame drops) and training (decoding throughput, image quality).
+
+Only override these parameters if you have a specific reason to, and measure the impact on your pipeline before relying on the new settings.
+
+</Tip>
+
+All flags below are prefixed with `--dataset.rgb_encoder.` on the CLI.
+
+| Parameter       | Type             | Default       | Description                                                                                                                                                                            |
+| --------------- | ---------------- | ------------- | -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
+| `vcodec`        | `str`            | `"libsvtav1"` | Video codec name. `"auto"` picks the first available hardware encoder from a fixed preference list, falling back to `libsvtav1`.                                                       |
+| `pix_fmt`       | `str`            | `"yuv420p"`   | Output pixel format. Must be supported by the chosen codec in your FFmpeg build.                                                                                                       |
+| `g`             | `int`            | `2`           | GOP size — a keyframe every `g` frames. Emitted as FFmpeg option `g`.                                                                                                                  |
+| `crf`           | `int` or `float` | `30`          | Abstract quality value, mapped per codec (see the [mapping](#mapping-videoencoderconfig--ffmpeg-options) below). Lower → higher quality / larger output where the mapping is monotone. |
+| `preset`        | `int` or `str`   | `12` \*       | Encoder speed preset; meaning depends on the codec. <br/>\* When unset and `vcodec=libsvtav1`, LeRobot defaults to `12`.                                                               |
+| `fast_decode`   | `int`            | `0`           | `libsvtav1`: `0–2`, passed via `svtav1-params`. <br/>`h264` / `hevc` (software): if `>0`, sets `tune=fastdecode`. <br/>Other codecs: usually unused.                                   |
+| `video_backend` | `str`            | `"pyav"`      | Only `"pyav"` is currently implemented for video encoding.                                                                                                                             |
+| `extra_options` | `dict`           | `{}`          | Extra FFmpeg or codec specific options merged after the structured fields above. Cannot override keys already set by those fields.                                                     |
+
+---
+
+## Depth streams
+
+Depth maps (Intel RealSense, Reachy 2) are stored as their **own video streams** alongside the RGB streams. Raw depth (`uint16` millimetres or `float32` metres) can't survive an 8-bit codec, so LeRobot **quantizes** each map to a 12-bit code (`[0, 4095]`) — logarithmically by default, to match the `1/depth` error profile of depth sensors — then packs it into a high-bit-depth pixel format (`gray12le`) and encodes it with a 12-bit codec.
+
+```mermaid
+flowchart LR
+    A["Raw depth (uint16 mm / float32 m)"] --> B["Clip to depth_min, depth_max"]
+    B --> C["Quantize to 12-bit code 0–4095 (log or linear)"]
+    C --> D["Pack into gray12le"]
+    D --> E["Encode video (hevc Main 12)"]
+    E --> F[("MP4 + metadata: depth_min/max, shift, use_log")]
+    F -. "load time (depth_output_unit)" .-> G["Dequantize to mm or m"]
+
+    classDef input fill:#e3f2fd,stroke:#1565c0,color:#0d47a1;
+    classDef encode fill:#ede7f6,stroke:#5e35b1,color:#311b92;
+    classDef store fill:#fff8e1,stroke:#f9a825,color:#e65100;
+    classDef load fill:#e8f5e9,stroke:#2e7d32,color:#1b5e20;
+
+    class A input;
+    class B,C,D,E encode;
+    class F store;
+    class G load;
+```
+
+Configure the depth pipeline through a parallel **`depth_encoder`** block (`DepthEncoderConfig`). It shares every `RGBEncoderConfig` field (`vcodec`, `pix_fmt`, `crf`, …) and adds four quantizer knobs, set via `--dataset.depth_encoder.<field>`:
+
+```bash
+lerobot-record \
+    ... \
+    --dataset.depth_encoder.vcodec=hevc \
+    --dataset.depth_encoder.depth_min=0.05 \
+    --dataset.depth_encoder.depth_max=5.0 \
+    --dataset.depth_encoder.use_log=true
+```
+
+| Parameter       | Type    | Default                         | Description                                                                                                                            |
+| --------------- | ------- | ------------------------------- | -------------------------------------------------------------------------------------------------------------------------------------- |
+| `vcodec`        | `str`   | `"hevc"`                        | HEVC Main 12 (a 12-bit-capable codec, MP4-compatible).                                                                                 |
+| `extra_options` | `dict`  | `{"x265-params": "lossless=1"}` | **Depth defaults to lossless** (exact round-trip); `crf` is ignored. Pass `extra_options={}` and set `crf` for a smaller lossy stream. |
+| `pix_fmt`       | `str`   | `"gray12le"`                    | Single-channel 12-bit pixel format used to carry the quantized codes.                                                                  |
+| `depth_min`     | `float` | `0.01`                          | Depth in metres mapped to quantum `0`. Values below are clipped on decode.                                                             |
+| `depth_max`     | `float` | `10.0`                          | Depth in metres mapped to quantum `4095`. Values above are clipped on decode.                                                          |
+| `shift`         | `float` | `3.5`                           | Pre-log offset (metres) used in logarithmic quantization for numerical stability near zero. Must satisfy `depth_min + shift > 0`.      |
+| `use_log`       | `bool`  | `True`                          | If `true`, quantize in log-space (recommended for typical depth sensors). Set to `false` for uniform/linear quantization.              |
+
+> [!TIP]
+> `depth_min`, `depth_max`, and `shift` are always interpreted in **metres**, regardless of the input depth's unit. Inputs are auto-detected: integer arrays (e.g. `uint16` millimetres straight from a RealSense) are treated as millimetres, floating arrays as metres.
+> Pick `depth_min` / `depth_max` to bracket the actual working range of your sensor — quanta outside that range saturate, which can crush detail at the boundaries.
+
+Depth features are flagged with `"is_depth_map": true` in `meta/info.json`, and their quantizer settings (`video.depth_min`, `video.depth_max`, `video.shift`, `video.use_log`) are persisted — which is what lets depth be **dequantized back to physical units** on load.
+
+### Output unit at load time
+
+`depth_encoder` is a **record-time** concern. The unit that depth maps are dequantized to on _load_ (e.g. during training) is set separately by the read-time flag `--dataset.depth_output_unit`:
+
+```bash
+lerobot-train \
+    --dataset.repo_id=<my_username>/<my_dataset_name> \
+    --dataset.depth_output_unit=m \
+    --policy.type=act
+```
+
+| Parameter           | Type  | Default | Description                                                                                  |
+| ------------------- | ----- | ------- | -------------------------------------------------------------------------------------------- |
+| `depth_output_unit` | `str` | `"mm"`  | Physical unit depth maps are dequantized to on load: `"mm"` (millimetres) or `"m"` (metres). |
+
+> [!TIP]
+> This is purely a decode-time presentation choice — it does **not** alter the stored video or its metadata, so the same dataset can be read as `mm` or `m` without re-encoding. It has no effect on datasets without depth cameras.
+
+---
+
+## Persistence in dataset metadata
+
+After the first episode of a video stream is encoded, the encoder configuration is **persisted into the dataset metadata** (`meta/info.json`) under each video feature, alongside the values probed from the file itself. For a video feature `observation.images.<camera>`, the layout in `info.json` is:
+
+```json
+{
+  "features": {
+    "observation.images.laptop": {
+      "dtype": "video",
+      "shape": [480, 640, 3],
+      "info": {
+        "video.height": 480,
+        "video.width": 640,
+        "video.codec": "h264",
+        "video.pix_fmt": "yuv420p",
+        "video.fps": 30,
+        "video.channels": 3,
+        "is_depth_map": false,
+        "video.g": 2,
+        "video.crf": 30,
+        "video.preset": "fast",
+        "video.fast_decode": 0,
+        "video.video_backend": "pyav",
+        "video.extra_options": { "tune": "film", "profile:v": "high", "bf": 2 }
+      }
+    }
+  }
+}
+```
+
+Two sources contribute to the `info` block:
+
+- **Stream-derived** (read back from the encoded MP4 with PyAV): `video.height`, `video.width`, `video.codec`, `video.pix_fmt`, `video.fps`, `video.channels`, `is_depth_map`, plus `audio.*` if an audio stream is present.
+- **Encoder-derived** (taken from `RGBEncoderConfig` or `DepthEncoderConfig`): `video.g`, `video.crf`, `video.preset`, `video.fast_decode`, `video.video_backend`, `video.extra_options`.
+
+<Tip>
+  This block is populated **once**, from the **first** episode. It assumes every
+  episode in the dataset was encoded with the same `rgb_encoder`. Changing
+  encoder settings partway through a recording is not supported — the
+  `info.json` will only reflect the parameters used for the first episode.
+</Tip>
+
+---
+
+## Merging datasets
+
+When aggregating datasets with `merge_datasets`, video files are concatenated as-is (no re-encoding), and encoder fields in `info.json` are merged per-key:
+
+- **Stream-derived fields must match** across sources: `video.codec`, `video.pix_fmt`, `video.height`, `video.width`, `video.fps`. Otherwise FFmpeg's concat demuxer fails.
+- **Encoder-tuning fields are merged loosely**: `video.g`, `video.crf`, `video.preset`, `video.fast_decode`, `video.extra_options`. If every source agrees, the value is kept; if not, it's set to `null` (or `{}` for `video.extra_options`) and a warning is logged.

docs/source/vla_jepa.mdx

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

docs/source/vlabench.mdx

@@ -165,7 +165,7 @@ lerobot-train \
   --output_dir=./outputs/smolvla_vlabench_primitive \
   --steps=100000 \
   --batch_size=4 \
-  --eval_freq=5000 \
+  --env_eval_freq=5000 \
   --eval.batch_size=1 \
   --eval.n_episodes=1 \
   --save_freq=10000

examples/annotations/run_hf_job.py

@@ -0,0 +1,77 @@
+#!/usr/bin/env python
+
+# Copyright 2026 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""Launch ``lerobot-annotate`` on a Hugging Face job (vllm + Qwen3.6-27B VLM).
+
+Spawns one single-GPU ``h200`` job that:
+
+  1. installs ``lerobot`` from ``main`` plus the annotation extras,
+  2. boots one vllm server with Qwen3.6-27B (dense VLM),
+  3. runs the plan / interjections / vqa modules across the dataset
+     in free-form mode (each episode generates its own subtasks +
+     memory),
+  4. uploads the annotated dataset to ``--new_repo_id`` (when set)
+     or back to ``--repo_id``.
+
+Usage:
+
+    HF_TOKEN=hf_... uv run python examples/annotations/run_hf_job.py
+
+Adjust ``CMD`` (dataset, model, hub repo) and ``flavor`` below for your
+run. For larger datasets, scale to ``h200x4`` and raise
+``--vlm.parallel_servers`` / ``--vlm.num_gpus`` to match.
+"""
+
+import os
+
+from huggingface_hub import get_token, run_job
+
+token = os.environ.get("HF_TOKEN") or get_token()
+if not token:
+    raise RuntimeError("No HF token. Run `huggingface-cli login` or `export HF_TOKEN=hf_...`")
+
+CMD = (
+    "apt-get update -qq && apt-get install -y -qq git ffmpeg && "
+    "pip install --no-deps "
+    "'lerobot @ git+https://github.com/huggingface/lerobot.git@main' && "
+    "pip install --upgrade-strategy only-if-needed "
+    "datasets pyarrow av jsonlines draccus gymnasium torchcodec mergedeep pyyaml-include toml typing-inspect "
+    "openai && "
+    "export VLLM_MEMORY_PROFILER_ESTIMATE_CUDAGRAPHS=0 && "
+    "export VLLM_VIDEO_BACKEND=pyav && "
+    "lerobot-annotate "
+    "--repo_id=pepijn223/robocasa_pretrain_human300_v4 "
+    "--new_repo_id=pepijn223/robocasa_pretrain_human300_v4_annotated "
+    "--push_to_hub=true "
+    "--vlm.backend=openai "
+    "--vlm.model_id=Qwen/Qwen3.6-27B "
+    "--vlm.num_gpus=1 "
+    '--vlm.serve_command="vllm serve Qwen/Qwen3.6-27B '
+    "--tensor-parallel-size 1 --max-model-len 32768 "
+    '--gpu-memory-utilization 0.8 --uvicorn-log-level warning --port {port}" '
+    "--vlm.serve_ready_timeout_s=1800 "
+    # Qwen3.6 ships with thinking on; annotation wants plain JSON answers.
+    "--vlm.chat_template_kwargs='{\"enable_thinking\": false}'"
+)
+
+job = run_job(
+    image="vllm/vllm-openai:latest",
+    command=["bash", "-c", CMD],
+    flavor="h200",
+    secrets={"HF_TOKEN": token},
+    timeout="2h",
+)
+print(f"Job URL: {job.url}")
+print(f"Job ID:  {job.id}")

examples/dataset/create_progress_videos.py

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

examples/hopejr/visualizer/index.html

@@ -0,0 +1,182 @@
+<!DOCTYPE html>
+<html lang="en">
+<head>
+    <meta charset="UTF-8">
+    <meta name="viewport" content="width=device-width, initial-scale=1.0">
+    <title>3D Hand Joint Visualizer</title>
+    <style>
+        body {
+            font-family: Arial, sans-serif;
+            margin: 0;
+            padding: 0;
+            overflow: hidden;
+            display: flex;
+            flex-direction: column;
+            height: 100vh;
+        }
+
+        .controls {
+            padding: 15px;
+            background-color: #f5f5f5;
+            z-index: 100;
+        }
+
+        .status {
+            padding: 10px;
+            border-radius: 5px;
+            margin: 10px 0;
+        }
+
+        .connected {
+            background-color: #d4edda;
+            color: #155724;
+        }
+
+        .disconnected {
+            background-color: #f8d7da;
+            color: #721c24;
+        }
+
+        button {
+            padding: 8px 16px;
+            background-color: #4CAF50;
+            color: white;
+            border: none;
+            border-radius: 4px;
+            cursor: pointer;
+            margin-right: 10px;
+        }
+
+        button:hover {
+            background-color: #45a049;
+        }
+
+        button:disabled {
+            background-color: #cccccc;
+            cursor: not-allowed;
+        }
+
+        .container {
+            display: flex;
+            flex: 1;
+            overflow: hidden;
+        }
+
+        #canvas-container {
+            flex: 3;
+            position: relative;
+        }
+
+        #sidebar {
+            flex: 1;
+            padding: 15px;
+            background-color: #f8f9fa;
+            overflow-y: auto;
+            max-width: 300px;
+            border-left: 1px solid #ddd;
+        }
+
+        .joint-info {
+            margin-bottom: 10px;
+            padding: 8px;
+            border: 1px solid #ddd;
+            border-radius: 4px;
+        }
+
+        .joint-name {
+            font-weight: bold;
+        }
+
+        .joint-value {
+            font-family: monospace;
+        }
+
+        .bar-container {
+            width: 100%;
+            background-color: #e0e0e0;
+            height: 10px;
+            border-radius: 5px;
+            overflow: hidden;
+            margin-top: 5px;
+        }
+
+        .bar {
+            height: 100%;
+            background-color: #4CAF50;
+            width: 0%;
+            transition: width 0.2s ease-in-out;
+        }
+
+        .log-container {
+            margin-top: 20px;
+            border: 1px solid #ddd;
+            border-radius: 5px;
+            padding: 10px;
+            height: 150px;
+            overflow-y: auto;
+            font-family: monospace;
+            background-color: #f8f9fa;
+        }
+
+        .view-controls {
+            position: absolute;
+            bottom: 10px;
+            left: 10px;
+            z-index: 10;
+        }
+
+        .view-button {
+            background-color: rgba(0, 0, 0, 0.5);
+            color: white;
+            border: none;
+            padding: 5px 10px;
+            margin-right: 5px;
+            border-radius: 3px;
+            cursor: pointer;
+        }
+    </style>
+</head>
+<body>
+    <div class="controls">
+        <button id="connectButton">Connect to Device</button>
+        <button id="disconnectButton" disabled>Disconnect</button>
+        <select id="baudRate">
+            <option value="9600">9600</option>
+            <option value="19200">19200</option>
+            <option value="38400">38400</option>
+            <option value="57600">57600</option>
+            <option value="115200" selected>115200</option>
+        </select>
+        <span id="statusIndicator" class="status disconnected">Status: Disconnected</span>
+    </div>
+
+    <div class="container">
+        <div id="canvas-container">
+            <!-- 3D canvas will be inserted here -->
+            <div class="view-controls">
+                <button class="view-button" id="frontView">Front</button>
+                <button class="view-button" id="sideView">Side</button>
+                <button class="view-button" id="topView">Top</button>
+                <button class="view-button" id="resetView">Reset</button>
+            </div>
+        </div>
+
+        <div id="sidebar">
+            <h3>Joint Values</h3>
+            <div id="jointsContainer">
+                <!-- Joint info will be added here -->
+            </div>
+
+            <div class="log-container" id="logContainer">
+                <!-- Log messages will be added here -->
+            </div>
+        </div>
+    </div>
+
+    <!-- Import Three.js -->
+    <script src="https://cdn.jsdelivr.net/npm/three@0.132.2/build/three.min.js"></script>
+    <script src="https://cdn.jsdelivr.net/npm/three@0.132.2/examples/js/controls/OrbitControls.js"></script>
+
+    <script src="script.js"></script>
+</body>
+</html>

examples/hopejr/visualizer/script.js

@@ -0,0 +1,669 @@
+// === Hand Visualizer with Pre-Connect Sliders + Per-Joint Angle Limits ===
+// Assumes your HTML already has elements with the following IDs:
+// connectButton, disconnectButton, baudRate, statusIndicator, jointsContainer, logContainer,
+// canvas-container, frontView, sideView, topView, resetView
+// Requires Three.js + OrbitControls loaded on the page.
+
+// -------------------- Config --------------------
+const MAX_JOINTS = 16;
+const RAW_MIN = 0, RAW_MAX = 4096;
+const RAW_CENTER = (RAW_MIN + RAW_MAX) / 2;
+const DEG = Math.PI / 180;
+const UI_DEG_MIN = -90, UI_DEG_MAX = 90; // UI sliders for angle limits
+
+// -------------------- State --------------------
+let port;
+let reader;
+let keepReading = false;
+let isConnected = false;
+const decoder = new TextDecoder();
+let inputBuffer = '';
+
+let jointValues = new Array(MAX_JOINTS).fill(RAW_CENTER);
+
+// Auto-calibration: track observed min/max per joint
+let observedMin = new Array(MAX_JOINTS).fill(Infinity);
+let observedMax = new Array(MAX_JOINTS).fill(-Infinity);
+let calibrationEnabled = true;
+
+// Three.js
+let scene, camera, renderer, controls;
+let hand = { palm: null, fingers: [] };
+
+// DOM
+const connectButton     = document.getElementById('connectButton');
+const disconnectButton  = document.getElementById('disconnectButton');
+const baudRateSelect    = document.getElementById('baudRate');
+const statusIndicator   = document.getElementById('statusIndicator');
+const jointsContainer   = document.getElementById('jointsContainer');
+const logContainer      = document.getElementById('logContainer');
+const canvasContainer   = document.getElementById('canvas-container');
+const frontViewBtn      = document.getElementById('frontView');
+const sideViewBtn       = document.getElementById('sideView');
+const topViewBtn        = document.getElementById('topView');
+const resetViewBtn      = document.getElementById('resetView');
+
+// Helpers
+const clamp = (x, a, b) => Math.max(a, Math.min(b, x));
+const invLerp = (a, b, x) => clamp((x - a) / (b - a), 0, 1);
+
+// -------------------- Joint Map with per-joint angle limits --------------------
+const fingerJointMap = [
+  // Thumb (4)
+  { finger:0, joint:0, type:'CMC_ABDUCTION',  min:RAW_MIN, max:RAW_MAX, inverted:true },
+  { finger:0, joint:1, type:'CMC_FLEXION',    min:RAW_MIN, max:RAW_MAX, inverted:true },
+  { finger:0, joint:2, type:'MCP_FLEXION',    min:RAW_MIN, max:RAW_MAX, inverted:true }, // +45° only
+  { finger:0, joint:3, type:'IP_FLEXION',     min:RAW_MIN, max:RAW_MAX, inverted:true }, // +45° only
+
+  // Index (3)
+  { finger:1, joint:0, type:'MCP_ABDUCTION',  min:RAW_MIN, max:RAW_MAX, inverted:true },
+  { finger:1, joint:1, type:'MCP_FLEXION',    min:RAW_MIN, max:RAW_MAX, inverted:false },
+  { finger:1, joint:2, type:'PIP_FLEXION',    min:RAW_MIN, max:RAW_MAX, inverted:true }, // +45° only
+
+  // Middle (3)
+  { finger:2, joint:0, type:'MCP_ABDUCTION',  min:RAW_MIN, max:RAW_MAX, inverted:true },
+  { finger:2, joint:1, type:'MCP_FLEXION',    min:RAW_MIN, max:RAW_MAX, inverted:true },
+  { finger:2, joint:2, type:'PIP_FLEXION',    min:RAW_MIN, max:RAW_MAX, inverted:true }, // +45° only
+
+  // Ring (3)
+  { finger:3, joint:0, type:'MCP_ABDUCTION',  min:RAW_MIN, max:RAW_MAX, inverted:true },
+  { finger:3, joint:1, type:'MCP_FLEXION',    min:RAW_MIN, max:RAW_MAX, inverted:false },
+  { finger:3, joint:2, type:'PIP_FLEXION',    min:RAW_MIN, max:RAW_MAX, inverted:false }, // +45° only
+
+  // Pinky (3)
+  { finger:4, joint:0, type:'MCP_ABDUCTION',  min:RAW_MIN, max:RAW_MAX, inverted:false },
+  { finger:4, joint:1, type:'MCP_FLEXION',    min:RAW_MIN, max:RAW_MAX, inverted:false },
+  { finger:4, joint:2, type:'PIP_FLEXION',    min:RAW_MIN, max:RAW_MAX, inverted:false }  // +45° only
+];
+
+// Assign angle limits (radians) per joint (default ±45°, exceptions: +45° only)
+for (const j of fingerJointMap) {
+  const isThumb = j.finger === 0;
+  const isPIP   = j.type === 'PIP_FLEXION';
+  let minA = -45 * DEG, maxA = +45 * DEG;
+  if ((isThumb && (j.type === 'MCP_FLEXION' || j.type === 'IP_FLEXION')) || (!isThumb && isPIP)) {
+    minA = 0;
+    maxA = +45 * DEG;
+  }
+  j.angleMin = minA;
+  j.angleMax = maxA;
+}
+
+// -------------------- UI: Joint Panel --------------------
+const uiRefs = []; // per joint: { valueLabel, bar, barWrap, slider, invertChk, minDeg, maxDeg }
+
+function initializeJointElements() {
+  jointsContainer.innerHTML = '';
+  uiRefs.length = 0;
+
+  for (let i = 0; i < MAX_JOINTS; i++) {
+    const wrap = document.createElement('div');
+    wrap.className = 'joint-info';
+
+    const fingerIndex = i < 4 ? 0 : Math.floor((i - 4) / 3) + 1;
+    const jointInfo = fingerJointMap[i];
+    const jointType = jointInfo?.type || 'Unknown';
+    const fingerName = ['Thumb', 'Index', 'Middle', 'Ring', 'Pinky'][fingerIndex];
+
+    // Header
+    const nameEl = document.createElement('div');
+    nameEl.className = 'joint-name';
+    nameEl.textContent = `${fingerName} – ${jointType}`;
+
+    // Value + bar
+    const valueEl = document.createElement('div');
+    valueEl.className = 'joint-value';
+    valueEl.textContent = `Value: ${jointValues[i]}`;
+
+    const barWrap = document.createElement('div');
+    barWrap.className = 'bar-container';
+    const barEl = document.createElement('div');
+    barEl.className = 'bar';
+    barWrap.appendChild(barEl);
+
+    // Slider for pre-connect manual control
+    const slider = document.createElement('input');
+    slider.type = 'range';
+    slider.min = String(RAW_MIN);
+    slider.max = String(RAW_MAX);
+    slider.value = String(jointValues[i]);
+    slider.step = '1';
+    slider.className = 'joint-slider';
+
+    slider.addEventListener('input', () => {
+      if (isConnected) return; // ignore while connected
+      let v = parseInt(slider.value, 10);
+      if (jointInfo?.inverted) v = (jointInfo.min + jointInfo.max) - v;
+      jointValues[i] = clamp(jointInfo ? v : 0, RAW_MIN, RAW_MAX);
+      updateJointDisplay(i, jointValues[i]);
+      updateHandModel();
+    });
+
+    // Invert checkbox
+    const invertLbl = document.createElement('label');
+    invertLbl.className = 'invert-toggle';
+    const invertChk = document.createElement('input');
+    invertChk.type = 'checkbox';
+    invertChk.checked = !!jointInfo?.inverted;
+    invertChk.addEventListener('change', () => {
+      if (jointInfo) jointInfo.inverted = invertChk.checked;
+      addLogMessage(`${fingerName} ${jointType} inversion ${invertChk.checked ? 'enabled' : 'disabled'}`);
+    });
+    invertLbl.appendChild(invertChk);
+    invertLbl.appendChild(document.createTextNode('Invert Values'));
+
+    // Angle limits (deg) controls
+    const limitsRow = document.createElement('div');
+    limitsRow.className = 'limits-row';
+
+    const minDeg = document.createElement('input');
+    minDeg.type = 'number';
+    minDeg.min = String(UI_DEG_MIN);
+    minDeg.max = String(UI_DEG_MAX);
+    minDeg.step = '1';
+    minDeg.value = String(Math.round((jointInfo.angleMin || 0) / DEG));
+    minDeg.className = 'limit-num';
+
+    const maxDeg = document.createElement('input');
+    maxDeg.type = 'number';
+    maxDeg.min = String(UI_DEG_MIN);
+    maxDeg.max = String(UI_DEG_MAX);
+    maxDeg.step = '1';
+    maxDeg.value = String(Math.round((jointInfo.angleMax || 0) / DEG));
+    maxDeg.className = 'limit-num';
+
+    const minLbl = document.createElement('span'); minLbl.textContent = 'min°';
+    const maxLbl = document.createElement('span'); maxLbl.textContent = 'max°';
+    minLbl.className = 'limit-label'; maxLbl.className = 'limit-label';
+
+    function syncLimits() {
+      let mn = parseFloat(minDeg.value);
+      let mx = parseFloat(maxDeg.value);
+      if (isNaN(mn)) mn = -45;
+      if (isNaN(mx)) mx = +45;
+      if (mn > mx) [mn, mx] = [mx, mn];
+      jointInfo.angleMin = clamp(mn, UI_DEG_MIN, UI_DEG_MAX) * DEG;
+      jointInfo.angleMax = clamp(mx, UI_DEG_MIN, UI_DEG_MAX) * DEG;
+      minDeg.value = String(Math.round(jointInfo.angleMin / DEG));
+      maxDeg.value = String(Math.round(jointInfo.angleMax / DEG));
+      updateHandModel();
+    }
+    minDeg.addEventListener('change', syncLimits);
+    maxDeg.addEventListener('change', syncLimits);
+
+    limitsRow.appendChild(minLbl);
+    limitsRow.appendChild(minDeg);
+    limitsRow.appendChild(maxLbl);
+    limitsRow.appendChild(maxDeg);
+
+         // Calibration controls
+     const calibRow = document.createElement('div');
+     calibRow.className = 'calib-row';
+
+     const resetCalibBtn = document.createElement('button');
+     resetCalibBtn.textContent = 'Reset Calib';
+     resetCalibBtn.className = 'calib-btn';
+     resetCalibBtn.addEventListener('click', () => {
+       observedMin[i] = Infinity;
+       observedMax[i] = -Infinity;
+       addLogMessage(`Reset calibration for ${fingerName} ${jointType}`);
+     });
+
+     const calibStatus = document.createElement('span');
+     calibStatus.className = 'calib-status';
+     calibStatus.textContent = `Range: --`;
+
+     calibRow.appendChild(resetCalibBtn);
+     calibRow.appendChild(calibStatus);
+
+     // Compose
+     wrap.appendChild(nameEl);
+     wrap.appendChild(valueEl);
+     wrap.appendChild(barWrap);
+     wrap.appendChild(slider);
+     wrap.appendChild(invertLbl);
+     wrap.appendChild(limitsRow);
+     wrap.appendChild(calibRow);
+
+     jointsContainer.appendChild(wrap);
+
+     uiRefs[i] = { valueLabel: valueEl, bar: barEl, barWrap, slider, invertChk, minDeg, maxDeg, nameEl, calibStatus };
+  }
+
+  setConnectedUI(false); // initial state: sliders active
+}
+
+ // Toggle UI between pre-connect SLIDERS vs post-connect BARS
+ function setConnectedUI(connected) {
+   isConnected = connected;
+   for (let i = 0; i < uiRefs.length; i++) {
+     const ui = uiRefs[i];
+     if (!ui) continue;
+     // Show bars when connected; sliders disabled/hidden
+     ui.barWrap.style.display = connected ? '' : 'none';
+     ui.slider.disabled = connected;
+     ui.slider.style.display = connected ? 'none' : '';
+   }
+
+   // Reset calibration when connecting
+   if (connected) {
+     observedMin.fill(Infinity);
+     observedMax.fill(-Infinity);
+     addLogMessage('Calibration reset - move joints through full range for best results');
+   }
+ }
+
+// Update joint display (value text + bar color/width + slider position if needed)
+function updateJointDisplay(jointIndex, value) {
+  const ui = uiRefs[jointIndex];
+  const info = fingerJointMap[jointIndex];
+  if (!ui || !info) return;
+
+  ui.valueLabel.textContent = `Value: ${value}`;
+
+  // bar
+  const min = info.min, max = info.max;
+  const pct = clamp((value - min) / (max - min), 0, 1) * 100;
+  ui.bar.style.width = `${pct}%`;
+  const hue = Math.floor(pct * 1.2); // 0..120
+  ui.bar.style.backgroundColor = `hsl(${hue}, 80%, 50%)`;
+
+  // slider (only meaningful when not connected; keep in sync anyway)
+  const rawForSlider = info.inverted ? (info.min + info.max) - value : value;
+  if (!isConnected) ui.slider.value = String(clamp(Math.round(rawForSlider), RAW_MIN, RAW_MAX));
+}
+
+// -------------------- Serial I/O --------------------
+async function readSerialData() {
+  while (port?.readable && keepReading) {
+    reader = port.readable.getReader();
+    try {
+      while (true) {
+        const { value, done } = await reader.read();
+        if (done) break;
+        if (value) processData(decoder.decode(value));
+      }
+    } catch (err) {
+      console.error('Error reading:', err);
+      addLogMessage(`Error: ${err.message}`);
+      break;
+    } finally {
+      reader.releaseLock();
+    }
+  }
+}
+
+ function processData(chunk) {
+   inputBuffer += chunk;
+   let idx;
+   while ((idx = inputBuffer.indexOf('\n')) !== -1) {
+     const line = inputBuffer.slice(0, idx).trim();
+     inputBuffer = inputBuffer.slice(idx + 1);
+
+     const vals = line.split(/\s+/).map(v => parseInt(v, 10));
+     if (vals.length === MAX_JOINTS && vals.every(v => Number.isFinite(v))) {
+       for (let i = 0; i < MAX_JOINTS; i++) {
+         const info = fingerJointMap[i];
+         if (!info) continue;
+
+         let rawValue = vals[i];
+
+         // Update calibration tracking
+         if (calibrationEnabled) {
+           observedMin[i] = Math.min(observedMin[i], rawValue);
+           observedMax[i] = Math.max(observedMax[i], rawValue);
+
+           // Update calibration display
+           const ui = uiRefs[i];
+           if (ui && ui.calibStatus) {
+             if (observedMin[i] !== Infinity && observedMax[i] !== -Infinity) {
+               ui.calibStatus.textContent = `Range: ${observedMin[i]}-${observedMax[i]}`;
+             }
+           }
+
+           // Remap observed range to target range
+           if (observedMin[i] !== Infinity && observedMax[i] !== -Infinity && observedMax[i] > observedMin[i]) {
+             const observedRange = observedMax[i] - observedMin[i];
+             const targetRange = info.max - info.min;
+             const normalizedValue = (rawValue - observedMin[i]) / observedRange;
+             rawValue = info.min + (normalizedValue * targetRange);
+           }
+         }
+
+         let v = clamp(rawValue, info.min, info.max);
+         if (info.inverted) v = (info.min + info.max) - v;
+         jointValues[i] = v;
+         updateJointDisplay(i, v);
+       }
+       updateHandModel();
+     } else {
+       addLogMessage(`Received: ${line}`);
+     }
+   }
+ }
+
+async function connectToDevice() {
+  try {
+    port = await navigator.serial.requestPort();
+    const baudRate = parseInt(baudRateSelect.value, 10) || 115200;
+    await port.open({ baudRate });
+
+    keepReading = true;
+    setConnectedUI(true);
+
+    statusIndicator.textContent = 'Status: Connected';
+    statusIndicator.className = 'status connected';
+    connectButton.disabled = true;
+    disconnectButton.disabled = false;
+    baudRateSelect.disabled = true;
+
+    addLogMessage(`Connected at ${baudRate} baud`);
+    readSerialData();
+  } catch (e) {
+    console.error('Connect error:', e);
+    addLogMessage(`Connection error: ${e.message}`);
+  }
+}
+
+async function disconnectFromDevice() {
+  try {
+    keepReading = false;
+    if (reader) {
+      try { reader.cancel(); } catch {}
+    }
+    if (port) {
+      await port.close();
+      port = null;
+    }
+  } catch (e) {
+    console.error('Disconnect error:', e);
+    addLogMessage(`Disconnection error: ${e.message}`);
+  } finally {
+    setConnectedUI(false);
+    statusIndicator.textContent = 'Status: Disconnected';
+    statusIndicator.className = 'status disconnected';
+    connectButton.disabled = false;
+    disconnectButton.disabled = true;
+    baudRateSelect.disabled = false;
+    addLogMessage('Disconnected');
+  }
+}
+
+// -------------------- Three.js Scene --------------------
+function initThreeJS() {
+  scene = new THREE.Scene();
+  scene.background = new THREE.Color(0xf0f0f0);
+
+  camera = new THREE.PerspectiveCamera(
+    75,
+    canvasContainer.clientWidth / canvasContainer.clientHeight,
+    0.1, 1000
+  );
+  camera.position.set(0, 15, 15);
+  camera.lookAt(0, 0, 0);
+
+  renderer = new THREE.WebGLRenderer({ antialias: true });
+  renderer.setSize(canvasContainer.clientWidth, canvasContainer.clientHeight);
+  renderer.setPixelRatio(window.devicePixelRatio);
+  canvasContainer.appendChild(renderer.domElement);
+
+  controls = new THREE.OrbitControls(camera, renderer.domElement);
+  controls.enableDamping = true;
+  controls.dampingFactor = 0.25;
+
+  const ambientLight = new THREE.AmbientLight(0x404040);
+  scene.add(ambientLight);
+  const dir1 = new THREE.DirectionalLight(0xffffff, 0.5);
+  dir1.position.set(1, 1, 1);
+  scene.add(dir1);
+  const dir2 = new THREE.DirectionalLight(0xffffff, 0.3);
+  dir2.position.set(-1, 1, -1);
+  scene.add(dir2);
+
+  const gridHelper = new THREE.GridHelper(20, 20);
+  scene.add(gridHelper);
+
+  createHandModel();
+  window.addEventListener('resize', onWindowResize);
+  animate();
+}
+
+function createHandModel() {
+  const palmMaterial   = new THREE.MeshPhongMaterial({ color: 0xf5c396 });
+  const fingerMaterial = new THREE.MeshPhongMaterial({ color: 0xf5c396 });
+  const jointMaterial  = new THREE.MeshPhongMaterial({ color: 0xe3a977 });
+
+     const palmGeometry = new THREE.BoxGeometry(7, 1, 8);
+   hand.palm = new THREE.Mesh(palmGeometry, palmMaterial);
+   hand.palm.position.set(0, 0, 0);
+   hand.palm.rotation.x = Math.PI / 2; // hand vertical, palm facing forward
+   scene.add(hand.palm);
+
+  const fingerWidth = 1, fingerHeight = 0.8;
+  const fingerSegmentLengths = [3, 2, 1.5];
+  const thumbSegmentLengths  = [2, 2, 1.5];
+
+  const fingerBasePositions = [
+    [ 3,   0,  -2],  // Thumb
+    [ 1.5,-0.5,-4],  // Index
+    [ 0,  -0.5,-4],  // Middle
+    [-1.5,-0.5,-4],  // Ring
+    [-3,  -0.5,-4],  // Pinky
+  ];
+  const fingerBaseRot = [
+    { x:0, y:-Math.PI/3,  z: Math.PI/3 },  // Thumb
+    { x:0, y:-Math.PI/48, z: 0 },
+    { x:0, y: Math.PI/48, z: 0 },
+    { x:0, y: Math.PI/32, z: 0 },
+    { x:0, y: Math.PI/24, z: 0 }
+  ];
+
+  for (let fIdx = 0; fIdx < 5; fIdx++) {
+    const finger = { name:['Thumb','Index','Middle','Ring','Pinky'][fIdx], segments:[], joints:[] };
+    const isThumb = fIdx === 0;
+    const segLens = isThumb ? thumbSegmentLengths : fingerSegmentLengths;
+
+    finger.group = new THREE.Group();
+    finger.group.position.set(...fingerBasePositions[fIdx]);
+    finger.group.rotation.x = fingerBaseRot[fIdx].x;
+    finger.group.rotation.y = fingerBaseRot[fIdx].y;
+    finger.group.rotation.z = fingerBaseRot[fIdx].z;
+    finger.group.userData.baseRot = {
+      x:finger.group.rotation.x,
+      y:finger.group.rotation.y,
+      z:finger.group.rotation.z
+    };
+    hand.palm.add(finger.group);
+
+    let parent = finger.group;
+    for (let s = 0; s < segLens.length; s++) {
+      const segGroup = new THREE.Group();
+
+      const jGeom = new THREE.SphereGeometry(fingerWidth * 0.6, 8, 8);
+      const joint = new THREE.Mesh(jGeom, jointMaterial);
+      segGroup.add(joint);
+
+      const segGeom = new THREE.BoxGeometry(fingerWidth, fingerHeight, segLens[s]);
+      const seg = new THREE.Mesh(segGeom, fingerMaterial);
+      seg.position.z = -segLens[s] / 2;
+      segGroup.add(seg);
+
+      parent.add(segGroup);
+
+      finger.segments.push(segGroup);
+      finger.joints.push(joint);
+
+      if (s < segLens.length - 1) {
+        const connector = new THREE.Group();
+        connector.position.z = -segLens[s];
+        segGroup.add(connector);
+        parent = connector;
+      }
+    }
+
+    hand.fingers.push(finger);
+  }
+
+  addFingerLabels();
+  addHandLabel();
+}
+
+function addFingerLabels() {
+  const names = ['Thumb','Index','Middle','Ring','Pinky'];
+  for (let i = 0; i < hand.fingers.length; i++) {
+    const finger = hand.fingers[i];
+    const canvas = document.createElement('canvas');
+    const ctx = canvas.getContext('2d');
+    canvas.width = 128; canvas.height = 32;
+    ctx.fillStyle = '#ffffff'; ctx.fillRect(0,0,canvas.width,canvas.height);
+    ctx.font = 'bold 16px Arial';
+    ctx.fillStyle = '#000000';
+    ctx.textAlign = 'center'; ctx.textBaseline = 'middle';
+    ctx.fillText(names[i], canvas.width/2, canvas.height/2);
+
+    const texture = new THREE.CanvasTexture(canvas);
+    const geom = new THREE.PlaneGeometry(2, 0.5);
+    const mat = new THREE.MeshBasicMaterial({ map:texture, transparent:true, side:THREE.DoubleSide });
+    const label = new THREE.Mesh(geom, mat);
+    label.position.set(0, -1.5, -2);
+    label.rotation.x = Math.PI / 2;
+    finger.group.add(label);
+  }
+}
+
+function addHandLabel() {
+  const canvas = document.createElement('canvas');
+  const ctx = canvas.getContext('2d');
+  canvas.width = 256; canvas.height = 64;
+  ctx.fillStyle = '#ffffff'; ctx.fillRect(0,0,canvas.width,canvas.height);
+  ctx.font = 'bold 24px Arial';
+  ctx.fillStyle = '#000000';
+  ctx.textAlign = 'center'; ctx.textBaseline = 'middle';
+     ctx.fillText('RIGHT HAND (VERTICAL)', canvas.width/2, canvas.height/2);
+
+  const texture = new THREE.CanvasTexture(canvas);
+  const geom = new THREE.PlaneGeometry(7, 1.75);
+  const mat = new THREE.MeshBasicMaterial({ map:texture, transparent:true, side:THREE.DoubleSide });
+  const label = new THREE.Mesh(geom, mat);
+  label.position.set(0, -2, 0);
+  label.rotation.x = Math.PI / 2;
+  scene.add(label);
+}
+
+function updateHandModel() {
+  for (let i = 0; i < MAX_JOINTS; i++) {
+    const info = fingerJointMap[i];
+    if (!info) continue;
+    const { finger, joint, type, min, max, angleMin, angleMax } = info;
+    const raw = jointValues[i];
+    const f = hand.fingers[finger];
+    if (!f) continue;
+
+    const center = (min + max) / 2;
+    let angle = 0;
+
+         if (type.includes('ABDUCTION')) {
+       // symmetric around neutral
+       const k = clamp((raw - center) / ((max - min) / 2), -1, 1);
+       angle = angleMin + (k + 1) * 0.5 * (angleMax - angleMin);
+
+       const base = f.group.userData.baseRot || {x:0,y:0,z:0};
+       if (finger === 0 && joint === 0) {
+         // Thumb: abduction about Z (toward/away from palm)
+         f.group.rotation.z = base.z + angle;
+       } else {
+         // Other fingers: side-to-side about Y
+         f.group.rotation.y = base.y + angle;
+       }
+     } else if (type.includes('FLEXION')) {
+       const isThumb = finger === 0;
+       const isMCP = type === 'MCP_FLEXION';
+       const isPIP = type === 'PIP_FLEXION';
+       const positiveOnly = (isThumb && (type === 'MCP_FLEXION' || type === 'IP_FLEXION')) || (!isThumb && isPIP);
+
+       if (positiveOnly) {
+         const t = raw <= center ? 0 : invLerp(center, max, raw); // 0..1
+         angle = angleMin + t * (angleMax - angleMin);           // 0..+limit
+       } else {
+         const k = clamp((raw - center) / ((max - min) / 2), -1, 1);
+         angle = angleMin + (k + 1) * 0.5 * (angleMax - angleMin);
+       }
+
+       if (isMCP) {
+         // MCP flexion applies to the finger base group (same as abduction)
+         const base = f.group.userData.baseRot || {x:0,y:0,z:0};
+         f.group.rotation.x = base.x + angle;
+       } else if (f.segments[joint]) {
+         // PIP/DIP/IP flexion applies to individual segments
+         f.segments[joint].rotation.x = angle;
+       }
+     }
+  }
+}
+
+// -------------------- Render Loop --------------------
+function onWindowResize() {
+  camera.aspect = canvasContainer.clientWidth / canvasContainer.clientHeight;
+  camera.updateProjectionMatrix();
+  renderer.setSize(canvasContainer.clientWidth, canvasContainer.clientHeight);
+}
+
+function animate() {
+  requestAnimationFrame(animate);
+  controls.update();
+  renderer.render(scene, camera);
+}
+
+// -------------------- Misc UI --------------------
+function addLogMessage(msg) {
+  const el = document.createElement('div');
+  el.textContent = msg;
+  logContainer.appendChild(el);
+  logContainer.scrollTop = logContainer.scrollHeight;
+  while (logContainer.children.length > 100) {
+    logContainer.removeChild(logContainer.firstChild);
+  }
+}
+
+// Camera view controls
+frontViewBtn?.addEventListener('click', () => { camera.position.set(0, 0, 20); camera.lookAt(0,0,0); controls.update(); });
+sideViewBtn?.addEventListener('click',  () => { camera.position.set(20, 0, 0); camera.lookAt(0,0,0); controls.update(); });
+topViewBtn?.addEventListener('click',   () => { camera.position.set(0, 20, 0); camera.lookAt(0,0,0); controls.update(); });
+resetViewBtn?.addEventListener('click', () => { camera.position.set(10,10,10); camera.lookAt(0,0,0); controls.update(); });
+
+// Serial connect buttons
+connectButton?.addEventListener('click', connectToDevice);
+disconnectButton?.addEventListener('click', disconnectFromDevice);
+
+// Web Serial support check
+if (!navigator.serial) {
+  statusIndicator.textContent = 'Status: Web Serial API not supported in this browser';
+  connectButton.disabled = true;
+  addLogMessage('ERROR: Web Serial API is not supported in this browser. Try Chrome or Edge.');
+}
+
+// -------------------- Boot --------------------
+initThreeJS();
+initializeJointElements();
+
+ // -------------------- Styles (inline) --------------------
+ const styleElement = document.createElement('style');
+ styleElement.textContent = `
+ .joint-info { border-bottom: 1px solid #eee; padding: 8px 0; }
+ .joint-name { font-weight: 600; margin-bottom: 4px; }
+ .joint-value { font-size: 12px; color: #333; margin-bottom: 4px; }
+ .bar-container { width: 100%; height: 8px; background: #ddd; border-radius: 4px; overflow: hidden; }
+ .bar { height: 100%; width: 0%; background: #4caf50; }
+ .joint-slider { width: 100%; margin: 6px 0; }
+ .invert-toggle { display: inline-flex; align-items: center; gap: 6px; margin-top: 4px; font-size: 12px; color: #555; }
+ .limits-row { display: flex; align-items: center; gap: 6px; margin-top: 6px; flex-wrap: wrap; }
+ .limit-label { font-size: 11px; color: #666; }
+ .limit-num { width: 60px; }
+ .calib-row { display: flex; align-items: center; gap: 8px; margin-top: 4px; }
+ .calib-btn { padding: 2px 6px; font-size: 11px; background: #f44336; color: white; border: none; border-radius: 3px; cursor: pointer; }
+ .calib-btn:hover { background: #d32f2f; }
+ .calib-status { font-size: 11px; color: #666; }
+ .status.connected { color: #0a0; }
+ .status.disconnected { color: #a00; }
+ `;
+ document.head.appendChild(styleElement);

examples/lekiwi/evaluate.py

@@ -17,7 +17,7 @@
 import logging
 import time
 
-from lerobot.common.control_utils import init_keyboard_listener, predict_action
+from lerobot.common.control_utils import predict_action
 from lerobot.datasets import LeRobotDataset
 from lerobot.policies import make_pre_post_processors
 from lerobot.policies.act import ACTPolicy
@@ -26,6 +26,7 @@ from lerobot.processor import make_default_processors
 from lerobot.robots.lekiwi import LeKiwiClient, LeKiwiClientConfig
 from lerobot.utils.constants import ACTION, OBS_STR
 from lerobot.utils.feature_utils import build_dataset_frame, hw_to_dataset_features
+from lerobot.utils.keyboard_input import init_keyboard_listener
 from lerobot.utils.robot_utils import precise_sleep
 from lerobot.utils.utils import log_say
 from lerobot.utils.visualization_utils import init_rerun, log_rerun_data

examples/lekiwi/record.py

@@ -14,7 +14,6 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-from lerobot.common.control_utils import init_keyboard_listener
 from lerobot.datasets import LeRobotDataset
 from lerobot.processor import make_default_processors
 from lerobot.robots.lekiwi import LeKiwiClient, LeKiwiClientConfig
@@ -23,6 +22,7 @@ from lerobot.teleoperators.keyboard import KeyboardTeleop, KeyboardTeleopConfig
 from lerobot.teleoperators.so_leader import SO100Leader, SO100LeaderConfig
 from lerobot.utils.constants import ACTION, OBS_STR
 from lerobot.utils.feature_utils import hw_to_dataset_features
+from lerobot.utils.keyboard_input import init_keyboard_listener
 from lerobot.utils.utils import log_say
 from lerobot.utils.visualization_utils import init_rerun
 

examples/notebooks/quickstart.ipynb

@@ -80,7 +80,7 @@
     "}\n",
     "\n",
     "# Dataset\n",
-    "HF_USER = \"your_hf_username\"  # `huggingface-cli whoami` to find your username\n",
+    "HF_USER = \"your_hf_username\"  # `hf auth whoami` to find your username\n",
     "DATASET_NAME = \"my_so101_dataset\"\n",
     "TASK_DESCRIPTION = \"pick and place the block\"\n",
     "NUM_EPISODES = 10\n",
@@ -291,7 +291,34 @@
     "\n",
     "Uses `POLICY_PATH` from the Configuration cell (defaults to the Hub repo ID). You can also put there the `LAST_CHECKPOINT_PATH`.\n",
     "\n",
-    "See the [inference docs](https://huggingface.co/docs/lerobot/il_robots#run-inference-and-evaluate-your-policy) for details."
+    "See the [inference docs](https://huggingface.co/docs/lerobot/il_robots#run-inference-and-evaluate-your-policy) for details.\n",
+    "\n",
+    "Recently ```lerobot-rollout``` was introduced, you can [read more about it here](https://huggingface.co/docs/lerobot/main/en/il_robots?eval=Base+mode+%28no+recording%29#run-inference-and-evaluate-your-policy)."
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "print_cmd(\n",
+    "    \"lerobot-rollout\",\n",
+    "    \"--strategy.type=base\",\n",
+    "    f\"--policy.path={POLICY_PATH}\",\n",
+    "    f\"--robot.type={ROBOT_TYPE}\",\n",
+    "    f\"--robot.port={ROBOT_PORT}\",\n",
+    "    CAMERAS_FLAG,\n",
+    "    f'--task=\"{TASK_DESCRIPTION}\"',\n",
+    "    \"--duration=60\",\n",
+    ")"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "if you are using the V0.5.1 release you should use ```lerobot-record``` instead of rollout"
    ]
   },
   {

examples/omx/README.md

@@ -0,0 +1,136 @@
+# OMX Follower — Cube Pick And Place Example
+
+This is an example of what is possible to do with LeRobot on a physical setup.
+It is a WIP and being used internally at LeRobot and specific to our setup, but we hope it can be a useful reference for how to use LeRobot APIs and CLIs.
+
+It includes an end-to-end example for the **OMX Follower** robot arm: pick and place a cube dataset, train a policy, and deploy it autonomously.
+
+## Hardware
+
+| Component | Value                                |
+| --------- | ------------------------------------ |
+| Robot     | OMX Follower                         |
+| Cameras   | 2× OpenCV cameras (wrist + top-down) |
+
+## Scripts
+
+| Script                 | Purpose                                                         |
+| ---------------------- | --------------------------------------------------------------- |
+| `reset_environment.py` | Standalone utility: sweep workspace, grab cube, place cube      |
+| `record_grab.py`       | Automated data collection: reset → place → record grab episodes |
+
+## Setup
+
+Make sure you have LeRobot installed in your env. (See [the installation guide](https://huggingface.co/docs/lerobot/installation))
+
+Next, we will declare some environment variables for convenience. Adjust the camera indices and robot port to match your system configuration.
+
+```bash
+export ROBOT_PORT=/dev/ttyACM0
+export TELEOP_PORT=/dev/ttyACM1
+export HF_USERNAME=<your_hf_username>
+export ROBOT_CAMERAS="{ wrist: {type: opencv, index_or_path: 0, width: 640, height: 480, fps: 30, fourcc: MJPG}, top: {type: opencv, index_or_path: 2, width: 640, height: 480, fps: 30, fourcc: MJPG} }"
+```
+
+## Step 1 — Collect Data
+
+```bash
+lerobot-record \
+    --robot.type=omx_follower \
+    --robot.port=$ROBOT_PORT \
+    --robot.id=omx_follower \
+    --robot.cameras="$ROBOT_CAMERAS" \
+    --teleop.type=omx_leader \
+    --teleop.port=$TELEOP_PORT \
+    --teleop.id=omx_leader \
+    --dataset.repo_id=$HF_USERNAME/omx_pickandplace \
+    --dataset.root=data/omx_pickandplace \
+    --dataset.num_episodes=50 \
+    --dataset.single_task="Pick the cube and place it in the blue square" \
+    --dataset.streaming_encoding=true \
+    --dataset.push_to_hub=true
+```
+
+### Bonus Auto-Collect script
+
+/!\ This is specific to our setup and the task of picking and placing a cube. It is not a general-purpose data collection script. As you may notice, it doesn't require a teleop.
+
+```bash
+python -m examples.omx.record_grab \
+    --robot.type=omx_follower \
+    --robot.port=$ROBOT_PORT \
+    --robot.id=omx_follower \
+    --robot.cameras="$ROBOT_CAMERAS" \
+    --dataset.repo_id=$HF_USERNAME/omx_pickandplace \
+    --dataset.root=data/omx_pickandplace \
+    --dataset.num_episodes=50 \
+    --dataset.single_task="Pick the cube and place it in the blue square" \
+    --dataset.streaming_encoding=true \
+    --dataset.push_to_hub=true
+```
+
+Each episode:
+
+1. The arm grabs the cube from the center of the workspace and places it at a random position.
+2. The arm returns to HOME.
+3. A targeted grab is recorded: HOME → approach raised → lower onto cube → grasp → lift → carry → drop → HOME.
+
+A dataset is already available here [`maximellerbach/omx_pickandplace`](https://huggingface.co/datasets/maximellerbach/omx_pickandplace), so you can skip directly to training if you want.
+
+## Step 2 — Train
+
+To train a simple `ACT` policy on the collected dataset, you can use the `lerobot-train` CLI:
+
+```bash
+lerobot-train \
+    --dataset.repo_id=$HF_USERNAME/omx_pickandplace \
+    --policy.type=act \
+    --output_dir=outputs/train/omx_pickandplace_act \
+    --policy.device=cuda \
+    --policy.repo_id=$HF_USERNAME/omx_pickandplace_act \
+    --steps=20000 \
+    --wandb.enable=true
+```
+
+A pretrained `ACT` policy is already available here [`maximellerbach/omx_pickandplace_act`](https://huggingface.co/maximellerbach/omx_pickandplace_act).
+
+## Step 3 — Rollout
+
+Use the `lerobot-rollout` CLI with base strategy:
+
+```bash
+lerobot-rollout \
+    --strategy.type=base \
+    --robot.type=omx_follower \
+    --robot.port=$ROBOT_PORT \
+    --robot.id=omx_follower \
+    --robot.cameras="$ROBOT_CAMERAS" \
+    --policy.path=$HF_USERNAME/omx_pickandplace_act \
+```
+
+For continuous recording with automatic upload (sentry mode):
+
+```bash
+lerobot-rollout \
+    --strategy.type=sentry \
+    --strategy.upload_every_n_episodes=10 \
+    --robot.type=omx_follower \
+    --robot.port=$ROBOT_PORT \
+    --robot.id=omx_follower \
+    --robot.cameras="$ROBOT_CAMERAS" \
+    --policy.path=$HF_USERNAME/omx_pickandplace_act \
+    --dataset.repo_id=$HF_USERNAME/rollout_omx_pickandplace_act \
+```
+
+## Environment Reset Utility
+
+Those are specific to this particular physical setup. Those are scripts that execute hardcoded sequences of actions on the robot to reset the environment, which is useful for data collection and evaluation. They are not general-purpose scripts.
+
+`reset_environment.py` can be run standalone to prepare the workspace:
+
+```bash
+# Grab cube + place it at a random position on the left side
+python -m examples.omx.reset_environment --port $ROBOT_PORT --mode grab_and_place
+```
+
+It also exposes `grab_cube(robot)` and `place_cube(robot)` for use in custom scripts.

examples/omx/record_grab.py

@@ -0,0 +1,422 @@
+#!/usr/bin/env python3
+"""
+Auto-record grab episodes for the OMX robot arm.
+
+Each episode cycle:
+  1. grab_and_place  — grab cube from workspace center and place at a random (pan, reach) position
+  2. HOME            — return arm to home with gripper open
+  3. record_grab     — execute a targeted grab to the stored position while recording
+                       observations + actions to a LeRobotDataset
+
+Usage (run from repo root):
+    python -m examples.omx.record_grab \\
+        --robot.type=omx_follower \\
+        --robot.port=/dev/ttyACM0 \\
+        --robot.id=omx_follower \\
+        --robot.cameras="{ wrist: {type: opencv, index_or_path: 6, width: 640, height: 480, fps: 30, fourcc: MJPG}, top: {type: opencv, index_or_path: 4, width: 640, height: 480, fps: 30, fourcc: MJPG} }" \\
+        --dataset.repo_id=<hf_username>/<dataset_name> \\
+        --dataset.root=data/omx_grab \\
+        --dataset.num_episodes=50 \\
+        --dataset.single_task="Grab the cube" \\
+        --dataset.streaming_encoding=true
+"""
+
+import logging
+from dataclasses import dataclass
+from pprint import pformat
+
+import numpy as np
+
+from lerobot.cameras import CameraConfig  # noqa: F401
+from lerobot.cameras.opencv import OpenCVCameraConfig  # noqa: F401
+from lerobot.configs import parser
+from lerobot.configs.dataset import DatasetRecordConfig
+from lerobot.datasets import (
+    LeRobotDataset,
+    VideoEncodingManager,
+    aggregate_pipeline_dataset_features,
+    create_initial_features,
+)
+from lerobot.processor import make_default_processors
+from lerobot.robots import RobotConfig, make_robot_from_config
+from lerobot.robots.omx_follower import OmxFollower
+from lerobot.utils.constants import ACTION, OBS_STR
+from lerobot.utils.feature_utils import build_dataset_frame, combine_feature_dicts
+from lerobot.utils.robot_utils import precise_sleep
+
+from .reset_environment import (
+    APPROACH_SPEED,
+    GRIPPER_CLOSE_POS,
+    HOME_POSE,
+    PUSH_END_ELBOW_FLEX,
+    PUSH_END_SHOULDER_LIFT,
+    PUSH_START_ELBOW_FLEX,
+    PUSH_START_SHOULDER_LIFT,
+    array_to_pose,
+    grab_cube,
+    horizontal_wrist_flex,
+    move_to_pose,
+    place_cube,
+    pose_to_array,
+)
+
+# ── Grab-episode motion parameters ────────────────────────────────────────────
+
+# Shoulder-lift offset for the raised approach phase (subtracted from the target sl, arm is higher).
+GRAB_RAISE_SL_OFFSET = 20.0
+GRAB_LOWER_SPEED = 20.0
+RECORD_SPEED = 30.0
+
+# Pose the arm travels to after closing the gripper (cube held).
+GRAB_CARRY_POSE = {
+    "shoulder_pan.pos": -23.0,
+    "shoulder_lift.pos": 5.0,
+    "elbow_flex.pos": 18.0,
+    "wrist_flex.pos": -14.0,
+    "wrist_roll.pos": 0.0,
+    "gripper.pos": GRIPPER_CLOSE_POS,
+}
+
+# Per-joint jitter limits (degrees) applied to transit waypoints for human-like variation.
+# Cube-approach and carry poses are never jittered to preserve precision.
+_JITTER_LIMITS: dict[str, float] = {
+    "shoulder_pan.pos": 5.0,
+    "shoulder_lift.pos": 4.0,
+    "elbow_flex.pos": 4.0,
+    "wrist_flex.pos": 3.0,
+    "wrist_roll.pos": 2.0,
+    "gripper.pos": 0.0,
+}
+
+
+def _jitter_pose(pose: dict, rng: np.random.Generator) -> dict:
+    """Return a copy of pose with independent per-joint random perturbations."""
+    return {
+        k: v + rng.uniform(-_JITTER_LIMITS.get(k, 0.0), _JITTER_LIMITS.get(k, 0.0)) for k, v in pose.items()
+    }
+
+
+def _random_stuck_pose(rng: np.random.Generator) -> dict:
+    """Return a physically plausible stuck pose (failed grasp), gripper closed.
+
+    ef bounds are piecewise-linear in sl so the arm stays in a reachable,
+    table-safe envelope across the full sl range:
+      sl=-50 → ef ∈ [  0,  50]   (arm raised, can be bent forward)
+      sl=  0 → ef ∈ [-25,  25]   (mid reach)
+      sl= 30 → ef ∈ [-20,   0]   (arm extended, little room to flex)
+    wrist_flex is randomly offset from the horizontal value.
+    """
+    pan = float(rng.uniform(-5.0, 35.0))
+    sl = float(rng.uniform(-50.0, 30.0))
+
+    if sl <= 0.0:
+        alpha = (sl + 50.0) / 50.0  # 0 at sl=-50, 1 at sl=0
+        ef_lo = alpha * -25.0  # 0 → -25
+        ef_hi = 50.0 + alpha * -25.0  # 50 → 25
+    else:
+        alpha = sl / 30.0  # 0 at sl=0, 1 at sl=30
+        ef_lo = -25.0 + alpha * 5.0  # -25 → -20
+        ef_hi = 25.0 + alpha * -25.0  # 25 → 0
+
+    ef = float(rng.uniform(ef_lo, ef_hi))
+    wf = horizontal_wrist_flex(sl, ef) + float(rng.uniform(-15.0, 15.0))
+    return {
+        "shoulder_pan.pos": pan,
+        "shoulder_lift.pos": sl,
+        "elbow_flex.pos": ef,
+        "wrist_flex.pos": wf,
+        "wrist_roll.pos": float(rng.uniform(-15.0, 15.0)),
+        "gripper.pos": GRIPPER_CLOSE_POS,
+    }
+
+
+logger = logging.getLogger(__name__)
+
+
+@dataclass
+class OmxRecordGrabConfig:
+    robot: RobotConfig
+    dataset: DatasetRecordConfig
+    # Resume recording on an existing dataset.
+    resume: bool = False
+    # Fraction of episodes that start from a random stuck pose (gripper closed) to
+    # generate recovery data.  0.0 = disabled, 1.0 = all episodes are recovery starts.
+    recovery_prob: float = 0.5
+
+
+def record_episode_spline(
+    robot: OmxFollower,
+    waypoints: list[dict],
+    speeds: list[float],
+    dataset: LeRobotDataset,
+    task: str,
+) -> None:
+    """Execute a Catmull-Rom-style spline through waypoints, recording each frame.
+
+    Segment durations are parameterized from the maximum absolute joint delta
+    between consecutive waypoints divided by the requested segment speed,
+    producing non-uniform timing in joint space. Interior tangents are derived
+    from the adjacent per-segment velocities, with clamped (zero-velocity)
+    endpoints so the arm starts and stops smoothly. Each segment is cubic
+    Hermite, giving C1 continuity at every waypoint.
+    """
+    pts = [pose_to_array(w) for w in waypoints]
+    n = len(pts)
+
+    # Steps and duration per segment
+    n_steps_list = []
+    timestamps = []
+    for i in range(n - 1):
+        max_dist = float(np.max(np.abs(pts[i + 1] - pts[i])))
+        ns = max(1, int(max_dist / speeds[i] * dataset.fps)) if max_dist >= 0.5 else 0
+        n_steps_list.append(ns)
+        timestamps.append(ns / dataset.fps)
+
+    # Velocity tangents (deg/sec) — clamped at endpoints, Catmull-Rom for interior
+    vels = [np.zeros_like(pts[0])]
+    for i in range(1, n - 1):
+        v_prev = (pts[i] - pts[i - 1]) / timestamps[i - 1] if timestamps[i - 1] > 0 else np.zeros_like(pts[0])
+        v_next = (pts[i + 1] - pts[i]) / timestamps[i] if timestamps[i] > 0 else np.zeros_like(pts[0])
+        vels.append(0.5 * (v_prev + v_next))
+    vels.append(np.zeros_like(pts[0]))
+
+    dt = 1.0 / dataset.fps
+    for seg in range(n - 1):
+        ns = n_steps_list[seg]
+        if ns == 0:
+            continue
+        p0, p1 = pts[seg], pts[seg + 1]
+        # Scale velocity (deg/sec) to t-space tangent (deg/t-unit, where t: 0→1 over ns steps)
+        m0 = vels[seg] * timestamps[seg]
+        m1 = vels[seg + 1] * timestamps[seg]
+
+        for step in range(1, ns + 1):
+            t = step / ns
+            h00 = 2 * t**3 - 3 * t**2 + 1
+            h10 = t**3 - 2 * t**2 + t
+            h01 = -2 * t**3 + 3 * t**2
+            h11 = t**3 - t**2
+            commanded = h00 * p0 + h10 * m0 + h01 * p1 + h11 * m1
+
+            action = array_to_pose(commanded)
+            robot.send_action(action)
+            obs = robot.get_observation()
+            obs_frame = build_dataset_frame(dataset.features, obs, prefix=OBS_STR)
+            action_frame = build_dataset_frame(dataset.features, action, prefix=ACTION)
+            dataset.add_frame({**obs_frame, **action_frame, "task": task})
+            precise_sleep(dt)
+
+
+def record_grab_episode(
+    robot: OmxFollower,
+    dataset: LeRobotDataset,
+    pan: float,
+    t: float,
+    task: str,
+    recovery_start: bool = False,
+) -> None:
+    """Execute a targeted grab to the stored (pan, t) position, recording every frame.
+
+    Normal sequence (initial HOME move is NOT recorded):
+      HOME → raised approach above cube → lower → close gripper
+           → raise [jittered] → retract [jittered] → GRAB_CARRY_POSE → drop → HOME
+
+    Recovery sequence (recovery_start=True): arm is moved to a random stuck pose
+    (gripper closed) without recording, then recording begins from there:
+      stuck_pose → raised approach above cube → [normal grab sequence from there]
+
+    All segments are joined by a Catmull-Rom spline (C1-continuous velocities).
+    """
+    sl = PUSH_START_SHOULDER_LIFT + t * (PUSH_END_SHOULDER_LIFT - PUSH_START_SHOULDER_LIFT)
+    ef = PUSH_START_ELBOW_FLEX + t * (PUSH_END_ELBOW_FLEX - PUSH_START_ELBOW_FLEX)
+    sl_raised = sl - GRAB_RAISE_SL_OFFSET
+    wf_horizontal = horizontal_wrist_flex(sl, ef)
+
+    rng = np.random.default_rng()
+
+    if recovery_start:
+        stuck_pose = _random_stuck_pose(rng)
+        logger.info(f"Recovery start: {stuck_pose}")
+        move_to_pose(robot, stuck_pose, APPROACH_SPEED)
+        first_waypoints = [stuck_pose]
+        first_speeds = []
+    else:
+        jittery_start = _jitter_pose(HOME_POSE, rng)
+        move_to_pose(robot, jittery_start, APPROACH_SPEED)
+        first_waypoints = [jittery_start]
+        first_speeds = []
+
+    waypoints = first_waypoints + [
+        {  # raised approach: arm above cube
+            "shoulder_pan.pos": pan,
+            "shoulder_lift.pos": sl_raised,
+            "elbow_flex.pos": ef,
+            "wrist_flex.pos": horizontal_wrist_flex(sl_raised, ef),
+            "wrist_roll.pos": 0.0,
+            "gripper.pos": 60.0,
+        },
+        {  # lower onto cube — no jitter: precision needed
+            "shoulder_pan.pos": pan,
+            "shoulder_lift.pos": sl,
+            "elbow_flex.pos": ef,
+            "wrist_flex.pos": wf_horizontal,
+            "wrist_roll.pos": 0.0,
+            "gripper.pos": 60.0,
+        },
+        {  # close gripper — no jitter: precision needed
+            "shoulder_pan.pos": pan,
+            "shoulder_lift.pos": sl,
+            "elbow_flex.pos": ef,
+            "wrist_flex.pos": wf_horizontal,
+            "wrist_roll.pos": 0.0,
+            "gripper.pos": GRIPPER_CLOSE_POS,
+        },
+        _jitter_pose(
+            {  # raise with cube
+                "shoulder_pan.pos": pan,
+                "shoulder_lift.pos": sl_raised,
+                "elbow_flex.pos": ef,
+                "wrist_flex.pos": horizontal_wrist_flex(sl_raised, ef),
+                "wrist_roll.pos": 0.0,
+                "gripper.pos": GRIPPER_CLOSE_POS,
+            },
+            rng,
+        ),
+        _jitter_pose(
+            {  # retract: fold arm toward HOME before sweeping to carry zone
+                "shoulder_pan.pos": pan * 0.25,
+                "shoulder_lift.pos": HOME_POSE["shoulder_lift.pos"] + 5.0,
+                "elbow_flex.pos": HOME_POSE["elbow_flex.pos"] - 5.0,
+                "wrist_flex.pos": 0.0,
+                "wrist_roll.pos": 0.0,
+                "gripper.pos": GRIPPER_CLOSE_POS,
+            },
+            rng,
+        ),
+        GRAB_CARRY_POSE,  # no jitter: target drop zone
+        {**GRAB_CARRY_POSE, "gripper.pos": 60.0},  # drop cube
+        HOME_POSE,
+    ]
+    speeds = first_speeds + [
+        RECORD_SPEED,  # (HOME →) raised approach
+        GRAB_LOWER_SPEED,  # raised approach → lower
+        GRAB_LOWER_SPEED,  # lower → close gripper
+        RECORD_SPEED,  # close gripper → raise
+        RECORD_SPEED,  # raise → retract
+        RECORD_SPEED,  # retract → carry pose
+        RECORD_SPEED,  # carry pose → drop
+        RECORD_SPEED,  # drop → HOME
+    ]
+
+    record_episode_spline(robot, waypoints, speeds, dataset, task)
+
+    # Dwell at HOME for ~0.5 s before next episode
+    home_action = build_dataset_frame(dataset.features, HOME_POSE, prefix=ACTION)
+    dt = 1.0 / dataset.fps
+    for _ in range(int(dataset.fps * 0.5)):
+        robot.send_action(HOME_POSE)
+        obs = robot.get_observation()
+        obs_frame = build_dataset_frame(dataset.features, obs, prefix=OBS_STR)
+        dataset.add_frame({**obs_frame, **home_action, "task": task})
+        precise_sleep(dt)
+
+
+@parser.wrap()
+def record_grab(cfg: OmxRecordGrabConfig) -> LeRobotDataset:
+    logging.basicConfig(level=logging.INFO, format="%(levelname)s: %(message)s")
+    logger.info(pformat(cfg))
+
+    robot = make_robot_from_config(cfg.robot)
+    use_videos = cfg.dataset.video
+
+    teleop_action_processor, _, robot_obs_processor = make_default_processors()
+
+    dataset_features = combine_feature_dicts(
+        aggregate_pipeline_dataset_features(
+            pipeline=teleop_action_processor,
+            initial_features=create_initial_features(action=robot.action_features),
+            use_videos=use_videos,
+        ),
+        aggregate_pipeline_dataset_features(
+            pipeline=robot_obs_processor,
+            initial_features=create_initial_features(observation=robot.observation_features),
+            use_videos=use_videos,
+        ),
+    )
+
+    num_cameras = len(robot.cameras) if hasattr(robot, "cameras") else 0
+    dataset = None
+
+    try:
+        if cfg.resume:
+            dataset = LeRobotDataset.resume(
+                cfg.dataset.repo_id,
+                root=cfg.dataset.root,
+                streaming_encoding=cfg.dataset.streaming_encoding,
+                batch_encoding_size=cfg.dataset.video_encoding_batch_size,
+                vcodec=cfg.dataset.vcodec,
+                encoder_threads=cfg.dataset.encoder_threads,
+                image_writer_processes=cfg.dataset.num_image_writer_processes if num_cameras > 0 else 0,
+                image_writer_threads=cfg.dataset.num_image_writer_threads_per_camera * num_cameras
+                if num_cameras > 0
+                else 0,
+            )
+        else:
+            cfg.dataset.stamp_repo_id()
+            dataset = LeRobotDataset.create(
+                cfg.dataset.repo_id,
+                cfg.dataset.fps,
+                root=cfg.dataset.root,
+                robot_type=robot.name,
+                features=dataset_features,
+                use_videos=use_videos,
+                streaming_encoding=cfg.dataset.streaming_encoding,
+                batch_encoding_size=cfg.dataset.video_encoding_batch_size,
+                vcodec=cfg.dataset.vcodec,
+                encoder_threads=cfg.dataset.encoder_threads,
+                image_writer_processes=cfg.dataset.num_image_writer_processes if num_cameras > 0 else 0,
+                image_writer_threads=cfg.dataset.num_image_writer_threads_per_camera * num_cameras
+                if num_cameras > 0
+                else 0,
+            )
+
+        robot.connect(calibrate=True)
+
+        rng = np.random.default_rng()
+        with VideoEncodingManager(dataset):
+            for episode_idx in range(cfg.dataset.num_episodes):
+                logger.info(f"=== Episode {episode_idx + 1}/{cfg.dataset.num_episodes} ===")
+
+                logger.info("Step 1: grabbing and placing cube...")
+                grab_cube(robot)
+                pan, t = place_cube(robot)
+                logger.info(f"Cube placed at pan={pan:.1f}, reach={t:.2f}")
+
+                recovery_start = cfg.recovery_prob > 0 and float(rng.random()) < cfg.recovery_prob
+                logger.info(f"Step 2: recording {'recovery ' if recovery_start else ''}grab episode...")
+                record_grab_episode(
+                    robot,
+                    dataset,
+                    pan,
+                    t,
+                    cfg.dataset.single_task,
+                    recovery_start=recovery_start,
+                )
+
+                dataset.save_episode()
+                logger.info(f"Episode {episode_idx + 1} saved.")
+
+    finally:
+        if dataset:
+            dataset.finalize()
+        if robot.is_connected:
+            robot.disconnect()
+
+    if cfg.dataset.push_to_hub and dataset and dataset.num_episodes > 0:
+        dataset.push_to_hub(tags=cfg.dataset.tags, private=cfg.dataset.private)
+
+    return dataset
+
+
+if __name__ == "__main__":
+    record_grab()

examples/omx/reset_environment.py

@@ -0,0 +1,267 @@
+#!/usr/bin/env python3
+"""
+Auto-reset and cube-grab utility for the OMX robot arm.
+
+Provides:
+  - grab_cube(robot): sweep workspace, center cube, close gripper
+  - place_cube(robot): carry cube to a random position, release
+
+Standalone usage (run from repo root):
+    python -m examples.omx.reset_environment --port /dev/ttyACM1 --mode grab
+    python -m examples.omx.reset_environment --port /dev/ttyACM1 --mode grab_and_place
+
+Joint range: -100 to 100 for arm joints; gripper: 50 = closed, 80 = open.
+
+To read current joint values for calibration, add after robot.connect():
+    obs = robot.get_observation()
+    print({k: round(obs[k], 1) for k in JOINT_NAMES})
+    robot.disconnect(); raise SystemExit
+
+Parallel-to-ground IK: wrist_flex = WRIST_HORIZONTAL_OFFSET - shoulder_lift - elbow_flex.
+Linear interpolation preserves this constraint between any two poses that satisfy it.
+"""
+
+import argparse
+import logging
+
+import numpy as np
+
+from lerobot.robots.omx_follower import OmxFollower, OmxFollowerConfig
+from lerobot.robots.robot import Robot
+from lerobot.utils.robot_utils import precise_sleep
+
+logger = logging.getLogger(__name__)
+
+# ── Poses ─────────────────────────────────────────────────────────────────────
+
+HOME_POSE = {
+    "shoulder_pan.pos": 0.0,
+    "shoulder_lift.pos": -50.0,
+    "elbow_flex.pos": 50.0,
+    "wrist_flex.pos": 0.0,
+    "wrist_roll.pos": 0.0,
+    "gripper.pos": 60.0,
+}
+
+SWEEP_WAYPOINTS = [
+    {
+        "shoulder_pan.pos": -60.0,
+        "shoulder_lift.pos": 50.0,
+        "elbow_flex.pos": -60.0,
+        "wrist_flex.pos": -20.0,
+        "wrist_roll.pos": 0.0,
+        "gripper.pos": 60.0,
+    },
+    {
+        "shoulder_pan.pos": -30.0,
+        "shoulder_lift.pos": 50.0,
+        "elbow_flex.pos": -60.0,
+        "wrist_flex.pos": -5.0,
+        "wrist_roll.pos": 0.0,
+        "gripper.pos": 60.0,
+    },
+    {
+        "shoulder_pan.pos": 20.0,
+        "shoulder_lift.pos": 50.0,
+        "elbow_flex.pos": -55.0,
+        "wrist_flex.pos": -5.0,
+        "wrist_roll.pos": 0.0,
+        "gripper.pos": 60.0,
+    },
+]
+
+# ── Motion parameters ─────────────────────────────────────────────────────────
+
+CONTROL_HZ = 30
+APPROACH_SPEED = 50.0
+SWEEP_SPEED = 40.0
+
+# ── Grab-sequence parameters ──────────────────────────────────────────────────
+
+GRAB_PAN = 0.0
+SWEEP_LEFT_PAN = -60.0
+SWEEP_RIGHT_PAN = 60.0
+SWEEP_END_OFFSET = 5.0  # stop before center so the cube isn't pushed past GRAB_PAN
+SWEEP_END_PAN_RANGE = (15.0, 20.0)
+
+SWEEP_LOW_SHOULDER_LIFT = 50.0
+SWEEP_LOW_ELBOW_FLEX_START = -60.0
+SWEEP_LOW_ELBOW_FLEX_END = -55.0
+
+SWEEP_HIGH_WRIST_FLEX = -20.0  # wrist tilted up during high approach to clear obstacles
+
+PUSH_START_SHOULDER_LIFT = 0.0
+PUSH_START_ELBOW_FLEX = 45.0
+PUSH_END_SHOULDER_LIFT = 50.0
+PUSH_END_ELBOW_FLEX = -50.0
+# Subtracted from shoulder_lift during the push sweep to clear the platform surface.
+# Does not affect the grab-target interpolation in record_grab.py.
+PUSH_RAISE_OFFSET = 5.0
+
+WRIST_HORIZONTAL_OFFSET = 0.0  # tune if gripper tilts during push: + tilts nose up, - down
+GRIPPER_CLOSE_POS = 50.0
+
+PLACE_LEFT_PAN_RANGE = (5.0, 30.0)  # random pan range for cube placement on the left side
+PLACE_REACH_RANGE = (0.1, 0.7)  # 0 = arm retracted (PUSH_START), 1 = fully extended (PUSH_END)
+
+JOINT_NAMES = [
+    "shoulder_pan.pos",
+    "shoulder_lift.pos",
+    "elbow_flex.pos",
+    "wrist_flex.pos",
+    "wrist_roll.pos",
+    "gripper.pos",
+]
+
+# ── Helpers ───────────────────────────────────────────────────────────────────
+
+
+def pose_to_array(pose: dict) -> np.ndarray:
+    return np.array([pose[k] for k in JOINT_NAMES])
+
+
+def array_to_pose(arr: np.ndarray) -> dict:
+    return {k: float(arr[i]) for i, k in enumerate(JOINT_NAMES)}
+
+
+def horizontal_wrist_flex(shoulder_lift: float, elbow_flex: float) -> float:
+    return WRIST_HORIZONTAL_OFFSET - shoulder_lift - elbow_flex
+
+
+def _low_sweep_pose(pan: float, elbow_flex: float, wrist_flex: float | None = None) -> dict:
+    sl = SWEEP_LOW_SHOULDER_LIFT
+    return {
+        "shoulder_pan.pos": pan,
+        "shoulder_lift.pos": sl,
+        "elbow_flex.pos": elbow_flex,
+        "wrist_flex.pos": horizontal_wrist_flex(sl, elbow_flex) if wrist_flex is None else wrist_flex,
+        "wrist_roll.pos": 0.0,
+        "gripper.pos": 60.0,
+    }
+
+
+def _high_sweep_pose(pan: float) -> dict:
+    return {**HOME_POSE, "shoulder_pan.pos": pan, "wrist_flex.pos": SWEEP_HIGH_WRIST_FLEX}
+
+
+def _push_pose(shoulder_lift: float, elbow_flex: float, pan: float = GRAB_PAN, gripper: float = 70.0) -> dict:
+    return {
+        "shoulder_pan.pos": pan,
+        "shoulder_lift.pos": shoulder_lift,
+        "elbow_flex.pos": elbow_flex,
+        "wrist_flex.pos": horizontal_wrist_flex(shoulder_lift, elbow_flex),
+        "wrist_roll.pos": 0.0,
+        "gripper.pos": gripper,
+    }
+
+
+def move_to_pose(robot: Robot, target: dict, speed: float) -> None:
+    """Interpolate from current position to target at the given speed (units/s)."""
+    obs = robot.get_observation()
+    current = np.array([obs[k] for k in JOINT_NAMES])
+    goal = pose_to_array(target)
+
+    max_distance = float(np.max(np.abs(goal - current)))
+    if max_distance < 0.5:
+        return
+
+    n_steps = max(1, int(max_distance / speed * CONTROL_HZ))
+    dt = 1.0 / CONTROL_HZ
+    for step in range(1, n_steps + 1):
+        t = step / n_steps
+        robot.send_action(array_to_pose(current + t * (goal - current)))
+        precise_sleep(dt)
+
+
+# ── Sequences ─────────────────────────────────────────────────────────────────
+
+
+def grab_cube(robot: Robot) -> None:
+    """Left sweep → right sweep → extend arm parallel to ground → close gripper."""
+    move_to_pose(robot, HOME_POSE, APPROACH_SPEED)
+
+    for pan, end_pan in [
+        (SWEEP_LEFT_PAN, GRAB_PAN - SWEEP_END_OFFSET),
+        (SWEEP_RIGHT_PAN, GRAB_PAN + SWEEP_END_OFFSET),
+    ]:
+        logger.info(f"Sweeping {'left' if pan < 0 else 'right'} → center...")
+        move_to_pose(robot, _high_sweep_pose(pan), APPROACH_SPEED)
+        move_to_pose(
+            robot, _low_sweep_pose(pan, SWEEP_LOW_ELBOW_FLEX_START, wrist_flex=-20.0), APPROACH_SPEED
+        )
+        move_to_pose(robot, _low_sweep_pose(end_pan, SWEEP_LOW_ELBOW_FLEX_END, wrist_flex=0.0), SWEEP_SPEED)
+        move_to_pose(robot, HOME_POSE, APPROACH_SPEED)
+
+    logger.info("Extending to push cube into gripper...")
+    move_to_pose(
+        robot,
+        _push_pose(PUSH_START_SHOULDER_LIFT - PUSH_RAISE_OFFSET, PUSH_START_ELBOW_FLEX),
+        APPROACH_SPEED,
+    )
+    move_to_pose(
+        robot,
+        _push_pose(PUSH_END_SHOULDER_LIFT - PUSH_RAISE_OFFSET, PUSH_END_ELBOW_FLEX),
+        SWEEP_SPEED,
+    )
+
+    logger.info("Closing gripper...")
+    move_to_pose(
+        robot,
+        _push_pose(PUSH_END_SHOULDER_LIFT, PUSH_END_ELBOW_FLEX, gripper=GRIPPER_CLOSE_POS),
+        APPROACH_SPEED,
+    )
+
+    logger.info("Grab complete.")
+
+
+def place_cube(robot: Robot) -> tuple[float, float]:
+    """Carry the cube (gripper closed) to a random position on the left side, then release.
+
+    Returns:
+        (pan, t): pan angle and reach scalar [0, 1] of the placement position.
+    """
+    pan = float(np.random.uniform(*PLACE_LEFT_PAN_RANGE))
+    t = float(np.random.uniform(*PLACE_REACH_RANGE))
+    sl = PUSH_START_SHOULDER_LIFT + t * (PUSH_END_SHOULDER_LIFT - PUSH_START_SHOULDER_LIFT)
+    ef = PUSH_START_ELBOW_FLEX + t * (PUSH_END_ELBOW_FLEX - PUSH_START_ELBOW_FLEX)
+    logger.info(f"Placing cube at pan={pan:.1f}, reach={t:.2f}...")
+
+    move_to_pose(robot, {**HOME_POSE, "gripper.pos": GRIPPER_CLOSE_POS}, APPROACH_SPEED)
+    move_to_pose(
+        robot, {**HOME_POSE, "shoulder_pan.pos": pan, "gripper.pos": GRIPPER_CLOSE_POS}, APPROACH_SPEED
+    )
+    move_to_pose(robot, _push_pose(sl, ef, pan=pan, gripper=GRIPPER_CLOSE_POS), APPROACH_SPEED)
+    move_to_pose(robot, _push_pose(sl, ef, pan=pan, gripper=80.0), APPROACH_SPEED)
+    move_to_pose(robot, HOME_POSE, APPROACH_SPEED)
+    logger.info("Place complete.")
+    return pan, t
+
+
+# ── Entry point ───────────────────────────────────────────────────────────────
+
+
+def main():
+    parser = argparse.ArgumentParser(description="OMX arm reset / grab script")
+    parser.add_argument("--port", default="/dev/ttyACM1")
+    parser.add_argument("--robot_id", default="omx_follower")
+    parser.add_argument("--mode", choices=["grab", "grab_and_place"], default="grab_and_place")
+    args = parser.parse_args()
+
+    logging.basicConfig(level=logging.INFO, format="%(levelname)s: %(message)s")
+
+    robot = OmxFollower(OmxFollowerConfig(port=args.port, id=args.robot_id))
+    robot.connect(calibrate=True)
+
+    try:
+        if args.mode == "grab":
+            grab_cube(robot)
+        elif args.mode == "grab_and_place":
+            grab_cube(robot)
+            place_cube(robot)
+
+    finally:
+        robot.disconnect()
+
+
+if __name__ == "__main__":
+    main()

examples/phone_to_so100/evaluate.py

@@ -18,7 +18,7 @@ import logging
 import time
 
 from lerobot.cameras.opencv import OpenCVCameraConfig
-from lerobot.common.control_utils import init_keyboard_listener, predict_action
+from lerobot.common.control_utils import predict_action
 from lerobot.configs import FeatureType, PolicyFeature
 from lerobot.datasets import LeRobotDataset, aggregate_pipeline_dataset_features, create_initial_features
 from lerobot.model.kinematics import RobotKinematics
@@ -41,6 +41,7 @@ from lerobot.robots.so_follower.robot_kinematic_processor import (
 from lerobot.types import RobotAction, RobotObservation
 from lerobot.utils.constants import ACTION, OBS_STR
 from lerobot.utils.feature_utils import build_dataset_frame, combine_feature_dicts
+from lerobot.utils.keyboard_input import init_keyboard_listener
 from lerobot.utils.robot_utils import precise_sleep
 from lerobot.utils.utils import log_say
 from lerobot.utils.visualization_utils import init_rerun, log_rerun_data

examples/phone_to_so100/record.py

@@ -15,7 +15,6 @@
 # limitations under the License.
 
 from lerobot.cameras.opencv import OpenCVCameraConfig
-from lerobot.common.control_utils import init_keyboard_listener
 from lerobot.datasets import LeRobotDataset, aggregate_pipeline_dataset_features, create_initial_features
 from lerobot.model.kinematics import RobotKinematics
 from lerobot.processor import (
@@ -39,6 +38,7 @@ from lerobot.teleoperators.phone.config_phone import PhoneOS
 from lerobot.teleoperators.phone.phone_processor import MapPhoneActionToRobotAction
 from lerobot.types import RobotAction, RobotObservation
 from lerobot.utils.feature_utils import combine_feature_dicts
+from lerobot.utils.keyboard_input import init_keyboard_listener
 from lerobot.utils.utils import log_say
 from lerobot.utils.visualization_utils import init_rerun
 

examples/so100_to_so100_EE/evaluate.py

@@ -18,7 +18,7 @@ import logging
 import time
 
 from lerobot.cameras.opencv import OpenCVCameraConfig
-from lerobot.common.control_utils import init_keyboard_listener, predict_action
+from lerobot.common.control_utils import predict_action
 from lerobot.configs import FeatureType, PolicyFeature
 from lerobot.datasets import LeRobotDataset, aggregate_pipeline_dataset_features, create_initial_features
 from lerobot.model.kinematics import RobotKinematics
@@ -41,6 +41,7 @@ from lerobot.robots.so_follower.robot_kinematic_processor import (
 from lerobot.types import RobotAction, RobotObservation
 from lerobot.utils.constants import ACTION, OBS_STR
 from lerobot.utils.feature_utils import build_dataset_frame, combine_feature_dicts
+from lerobot.utils.keyboard_input import init_keyboard_listener
 from lerobot.utils.robot_utils import precise_sleep
 from lerobot.utils.utils import log_say
 from lerobot.utils.visualization_utils import init_rerun, log_rerun_data

examples/so100_to_so100_EE/record.py

@@ -16,7 +16,6 @@
 
 
 from lerobot.cameras.opencv import OpenCVCameraConfig
-from lerobot.common.control_utils import init_keyboard_listener
 from lerobot.datasets import LeRobotDataset, aggregate_pipeline_dataset_features, create_initial_features
 from lerobot.model.kinematics import RobotKinematics
 from lerobot.processor import (
@@ -36,6 +35,7 @@ from lerobot.scripts.lerobot_record import record_loop
 from lerobot.teleoperators.so_leader import SO100Leader, SO100LeaderConfig
 from lerobot.types import RobotAction, RobotObservation
 from lerobot.utils.feature_utils import combine_feature_dicts
+from lerobot.utils.keyboard_input import init_keyboard_listener
 from lerobot.utils.utils import log_say
 from lerobot.utils.visualization_utils import init_rerun
 

examples/tutorial/rl/hilserl_example.py

@@ -4,13 +4,13 @@ from pathlib import Path
 from queue import Empty, Full
 
 import torch
-import torch.optim as optim
 
 from lerobot.datasets import LeRobotDataset
 from lerobot.envs.configs import HILSerlProcessorConfig, HILSerlRobotEnvConfig
-from lerobot.policies import SACConfig
-from lerobot.policies.sac.modeling_sac import SACPolicy
+from lerobot.policies import GaussianActorConfig
+from lerobot.policies.gaussian_actor.modeling_gaussian_actor import GaussianActorPolicy
 from lerobot.rewards.classifier.modeling_classifier import Classifier
+from lerobot.rl.algorithms.sac import SACAlgorithm, SACAlgorithmConfig
 from lerobot.rl.buffer import ReplayBuffer
 from lerobot.rl.gym_manipulator import make_robot_env
 from lerobot.robots.so_follower import SO100FollowerConfig
@@ -28,7 +28,7 @@ def run_learner(
     transitions_queue: mp.Queue,
     parameters_queue: mp.Queue,
     shutdown_event: mp.Event,
-    policy_learner: SACPolicy,
+    policy_learner: GaussianActorPolicy,
     online_buffer: ReplayBuffer,
     offline_buffer: ReplayBuffer,
     lr: float = 3e-4,
@@ -40,8 +40,9 @@ def run_learner(
     policy_learner.train()
     policy_learner.to(device)
 
-    # Create Adam optimizer from scratch - simple and clean
-    optimizer = optim.Adam(policy_learner.parameters(), lr=lr)
+    algo_config = SACAlgorithmConfig.from_policy_config(policy_learner.config)
+    algorithm = SACAlgorithm(policy=policy_learner, config=algo_config)
+    algorithm.make_optimizers_and_scheduler()
 
     print(f"[LEARNER] Online buffer capacity: {online_buffer.capacity}")
     print(f"[LEARNER] Offline buffer capacity: {offline_buffer.capacity}")
@@ -83,24 +84,26 @@ def run_learner(
                 else:
                     batch[key] = online_batch[key]
 
-            loss, _ = policy_learner.forward(batch)
+            def batch_iter(b=batch):
+                while True:
+                    yield b
 
-            optimizer.zero_grad()
-            loss.backward()
-            optimizer.step()
+            stats = algorithm.update(batch_iter())
             training_step += 1
 
             if training_step % LOG_EVERY == 0:
+                log_dict = stats.to_log_dict()
                 print(
-                    f"[LEARNER] Training step {training_step}, Loss: {loss.item():.4f}, "
+                    f"[LEARNER] Training step {training_step}, "
+                    f"critic_loss: {log_dict.get('critic', 'N/A'):.4f}, "
                     f"Buffers: Online={len(online_buffer)}, Offline={len(offline_buffer)}"
                 )
 
             # Send updated parameters to actor every 10 training steps
             if training_step % SEND_EVERY == 0:
                 try:
-                    state_dict = {k: v.cpu() for k, v in policy_learner.state_dict().items()}
-                    parameters_queue.put_nowait(state_dict)
+                    weights = algorithm.get_weights()
+                    parameters_queue.put_nowait(weights)
                     print("[LEARNER] Sent updated parameters to actor")
                 except Full:
                     # Missing write due to queue not being consumed (should happen rarely)
@@ -113,7 +116,7 @@ def run_actor(
     transitions_queue: mp.Queue,
     parameters_queue: mp.Queue,
     shutdown_event: mp.Event,
-    policy_actor: SACPolicy,
+    policy_actor: GaussianActorPolicy,
     reward_classifier: Classifier,
     env_cfg: HILSerlRobotEnvConfig,
     device: torch.device = "mps",
@@ -144,15 +147,15 @@ def run_actor(
 
             while step < MAX_STEPS_PER_EPISODE and not shutdown_event.is_set():
                 try:
-                    new_params = parameters_queue.get_nowait()
-                    policy_actor.load_state_dict(new_params)
+                    new_weights = parameters_queue.get_nowait()
+                    policy_actor.load_state_dict(new_weights)
                     print("[ACTOR] Updated policy parameters from learner")
                 except Empty:  # No new updated parameters available from learner, waiting
                     pass
 
-                # Get action from policy
+                # Get action from policy (returns full action: continuous + discrete)
                 policy_obs = make_policy_obs(obs, device=device)
-                action_tensor = policy_actor.select_action(policy_obs)  # predicts a single action
+                action_tensor = policy_actor.select_action(policy_obs)
                 action = action_tensor.squeeze(0).cpu().numpy()
 
                 # Step environment
@@ -261,14 +264,14 @@ def main():
     action_features = hw_to_dataset_features(env.robot.action_features, "action")
 
     # Create SAC policy for action selection
-    policy_cfg = SACConfig(
+    policy_cfg = GaussianActorConfig(
         device=device,
         input_features=obs_features,
         output_features=action_features,
     )
 
-    policy_actor = SACPolicy(policy_cfg)
-    policy_learner = SACPolicy(policy_cfg)
+    policy_actor = GaussianActorPolicy(policy_cfg)
+    policy_learner = GaussianActorPolicy(policy_cfg)
 
     demonstrations_repo_id = "lerobot/example_hil_serl_dataset"
     offline_dataset = LeRobotDataset(repo_id=demonstrations_repo_id)

pyproject.toml

@@ -59,8 +59,8 @@ keywords = ["lerobot", "huggingface", "robotics",  "machine learning", "artifici
 
 dependencies = [
     # Core ML
-    "torch>=2.7,<2.11.0",
-    "torchvision>=0.22.0,<0.26.0",
+    "torch>=2.7,<2.12.0",
+    "torchvision>=0.22.0,<0.27.0",
     "numpy>=2.0.0,<2.3.0", # NOTE: Explicitly listing numpy helps the resolver converge faster. Upper bound imposed by opencv-python-headless.
     "opencv-python-headless>=4.9.0,<4.14.0",
     "Pillow>=10.0.0,<13.0.0",
@@ -95,17 +95,28 @@ dependencies = [
 
 # ── Feature-scoped extras ──────────────────────────────────
 dataset = [
-    "datasets>=4.0.0,<5.0.0",
+    "datasets>=4.7.0,<5.0.0",
     "pandas>=2.0.0,<3.0.0", # NOTE: Transitive dependency of datasets
     "pyarrow>=21.0.0,<30.0.0", # NOTE: Transitive dependency of datasets
     "lerobot[av-dep]",
-    "torchcodec>=0.3.0,<0.11.0; sys_platform != 'win32' and (sys_platform != 'linux' or (platform_machine != 'aarch64' and platform_machine != 'arm64' and platform_machine != 'armv7l')) and (sys_platform != 'darwin' or platform_machine != 'x86_64')", # NOTE: Windows support starts at version 0.7 (needs torch==2.8), ffmpeg>=8 support starts at version 0.8.1 (needs torch==2.9), system-wide ffmpeg support starts at version 0.10 (needs torch==2.10).
+
+    # NOTE: torchcodec wheel availability matrix (PyPI):
+    #   - linux x86_64/amd64 + macOS arm64 : wheels since 0.3.0 (the historic supported set).
+    #   - win32 x86_64                     : wheels since 0.7.0  (needs torch>=2.8).
+    #   - linux aarch64/arm64              : wheels since 0.11.0 (needs torch>=2.11).
+    #   - macOS x86_64 (Intel) and linux armv7l: no wheels in any released version -> fall through to the PyAV decoder.
+    # Each platform gets its own line so the resolver picks the minimum version that has a wheel for it.
+
+    # Other torch/torchcodec pairings (informational): 0.8.1 = ffmpeg>=8 support, 0.10 = system-wide ffmpeg support, 0.12 needs torch==2.12.
+    "torchcodec>=0.3.0,<0.12.0; (sys_platform == 'linux' and (platform_machine == 'x86_64' or platform_machine == 'AMD64')) or (sys_platform == 'darwin' and platform_machine == 'arm64')",
+    "torchcodec>=0.7.0,<0.12.0; sys_platform == 'win32'",
+    "torchcodec>=0.11.0,<0.12.0; sys_platform == 'linux' and (platform_machine == 'aarch64' or platform_machine == 'arm64')",
     "jsonlines>=4.0.0,<5.0.0",
 ]
 training = [
     "lerobot[dataset]",
-    "accelerate>=1.10.0,<2.0.0",
-    "wandb>=0.24.0,<0.25.0",
+    "wandb>=0.24.0,<0.28.0",
+    "lerobot[accelerate-dep]",
 ]
 hardware = [
     "lerobot[pynput-dep]",
@@ -113,7 +124,8 @@ hardware = [
     "lerobot[deepdiff-dep]",
 ]
 viz = [
-    "rerun-sdk>=0.24.0,<0.27.0",
+    "rerun-sdk>=0.24.0,<0.34.0",
+    "foxglove-sdk>=0.25.1,<0.26.0",
 ]
 # ── User-facing composite extras (map to CLI scripts) ─────
 # lerobot-record, lerobot-replay, lerobot-calibrate, lerobot-teleoperate, etc.
@@ -127,9 +139,19 @@ dataset_viz = ["lerobot[dataset]", "lerobot[viz]"]
 # Common
 av-dep = ["av>=15.0.0,<16.0.0"]
 pygame-dep = ["pygame>=2.5.1,<2.7.0"]
-placo-dep = ["placo>=0.9.6,<0.9.17"]
+# NOTE: 0.9.16 links against liburdfdom_sensor.so.4, which is unavailable on Ubuntu 24.04
+# (noble ships urdfdom 3.x). Cap below 0.9.16 until system urdfdom 4.x is broadly available.
+#
+# NOTE: placo pulls in pin (Pinocchio), whose binary wheels dlopen specific cmeel sonames
+# (liburdfdom_sensor.so.4.0, libtinyxml2.so.10) but declare only `>=` floors on their cmeel
+# packages. The 2026-05-21 major bumps (cmeel-urdfdom 6.0.0 -> .so.6, cmeel-tinyxml2 11.0.0
+# -> .so.11) ship newer sonames, so left unpinned the resolver grabs them and `import placo`
+# fails at load with "liburdfdom_sensor.so.4.0: cannot open shared object file" (see #3755).
+# 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"]
-grpcio-dep = ["grpcio==1.73.1", "protobuf>=6.31.1,<6.32.0"]
+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"]
 peft-dep = ["peft>=0.18.0,<1.0.0"]
 scipy-dep = ["scipy>=1.14.0,<2.0.0"]
@@ -140,6 +162,9 @@ pyserial-dep = ["pyserial>=3.5,<4.0"]
 deepdiff-dep = ["deepdiff>=7.0.1,<9.0.0"]
 pynput-dep = ["pynput>=1.7.8,<1.9.0"]
 pyzmq-dep = ["pyzmq>=26.2.1,<28.0.0"]
+motorbridge-dep = ["motorbridge>=0.3.2,<0.4.0"]
+motorbridge-smart-servo-dep = ["motorbridge-smart-servo>=0.0.4,<0.1.0"]
+timm-dep = ["timm>=1.0.0,<1.1.0"]
 
 # Motors
 feetech = ["feetech-servo-sdk>=1.0.0,<2.0.0", "lerobot[pyserial-dep]", "lerobot[deepdiff-dep]"]
@@ -162,7 +187,15 @@ unitree_g1 = [
     "lerobot[matplotlib-dep]",
     "lerobot[pygame-dep]",
 ]
-reachy2 = ["reachy2_sdk>=1.0.15,<1.1.0"]
+# reachy2-sdk caps grpcio<=1.73.1 and protobuf<=6.32.0; quarantined here so downstream users aren't held back. reachy2-sdk is unlikely to release new versions.
+reachy2 = [
+    "reachy2_sdk>=1.0.15,<1.1.0",
+    "grpcio<=1.73.1",
+    "protobuf<=6.32.0",
+]
+# Seeed Studio reBot B601-DM follower (motorbridge / CAN) + StarArm102 / reBot Arm 102
+# leader (motorbridge-smart-servo / FashionStar UART servos).
+rebot = ["lerobot[motorbridge-dep]", "lerobot[motorbridge-smart-servo-dep]"]
 kinematics = ["lerobot[placo-dep]"]
 intelrealsense = [
     "pyrealsense2>=2.55.1.6486,<2.57.0 ; sys_platform != 'darwin'",
@@ -180,38 +213,62 @@ wallx = [
     "lerobot[qwen-vl-utils-dep]",
 ]
 pi = ["lerobot[transformers-dep]", "lerobot[scipy-dep]"]
-smolvla = ["lerobot[transformers-dep]", "num2words>=0.5.14,<0.6.0", "accelerate>=1.7.0,<2.0.0"]
+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]"]
 groot = [
     "lerobot[transformers-dep]",
     "lerobot[peft-dep]",
     "lerobot[diffusers-dep]",
+    "lerobot[dataset]", # NOTE: processor_groot builds a LeRobotDataset for relative-action training stats
     "dm-tree>=0.1.8,<1.0.0",
-    "timm>=1.0.0,<1.1.0",
+    "lerobot[timm-dep]",
     "decord>=0.6.0,<1.0.0; (platform_machine == 'AMD64' or platform_machine == 'x86_64')",
-    "ninja>=1.11.1,<2.0.0",
-    "flash-attn>=2.5.9,<3.0.0 ; sys_platform != 'darwin'"
 ]
 sarm = ["lerobot[transformers-dep]", "pydantic>=2.0.0,<3.0.0", "faker>=33.0.0,<35.0.0", "lerobot[matplotlib-dep]", "lerobot[qwen-vl-utils-dep]"]
+robometer = ["lerobot[transformers-dep]", "lerobot[qwen-vl-utils-dep]", "lerobot[peft-dep]"]
+topreward = ["lerobot[transformers-dep]"]
 xvla = ["lerobot[transformers-dep]"]
 eo1 = ["lerobot[transformers-dep]", "lerobot[qwen-vl-utils-dep]"]
-hilserl = ["lerobot[transformers-dep]", "gym-hil>=0.1.13,<0.2.0", "lerobot[grpcio-dep]", "lerobot[placo-dep]"]
+fastwam = [
+    "lerobot[transformers-dep]",
+    "lerobot[diffusers-dep]",
+]
+evo1 = ["lerobot[transformers-dep]"]
+hilserl = ["lerobot[transformers-dep]", "lerobot[dataset]", "gym-hil>=0.1.14,<0.2.0", "lerobot[grpcio-dep]", "lerobot[placo-dep]"]
+vla_jepa = ["lerobot[transformers-dep]", "lerobot[diffusers-dep]", "lerobot[qwen-vl-utils-dep]"]
+lingbot_va = ["lerobot[transformers-dep]", "lerobot[diffusers-dep]", "lerobot[accelerate-dep]"]
 
 # Features
 async = ["lerobot[grpcio-dep]", "lerobot[matplotlib-dep]"]
 peft = ["lerobot[transformers-dep]", "lerobot[peft-dep]"]
 
+# Annotation pipeline (lerobot-annotate). The only backend is ``openai``,
+# which talks to any OpenAI-compatible server (``vllm serve`` /
+# ``transformers serve`` / hosted). Distributed runs use Hugging Face Jobs
+# (see examples/annotations/run_hf_job.py).
+annotations = [
+    "lerobot[dataset]",
+    "lerobot[transformers-dep]",
+    "openai>=1.40,<2.0",
+    # ``vllm`` is intentionally NOT a hard dep: it pins an older torch, and
+    # uv's single unified lock would then cap ``torch`` for every extra
+    # (e.g. forcing 2.8 while ``torchcodec`` in [dataset] needs 2.11 -> ABI
+    # break in CI). The HF Jobs image (``vllm/vllm-openai``) provides vLLM;
+    # install it locally only if you run your own ``vllm serve``.
+]
+
 # Development
-dev = ["pre-commit>=3.7.0,<5.0.0", "debugpy>=1.8.1,<1.9.0", "lerobot[grpcio-dep]", "grpcio-tools==1.73.1", "mypy>=1.19.1", "ruff>=0.14.1", "lerobot[notebook]"]
+dev = ["pre-commit>=3.7.0,<5.0.0", "debugpy>=1.8.1,<1.9.0", "lerobot[grpcio-dep]", "grpcio-tools>=1.73.1,<2.0.0", "mypy>=1.19.1", "ruff>=0.14.1", "lerobot[notebook]"]
 notebook = ["jupyter>=1.0.0,<2.0.0", "ipykernel>=6.0.0,<7.0.0"]
 test = ["pytest>=8.1.0,<9.0.0", "pytest-timeout>=2.4.0,<3.0.0", "pytest-cov>=5.0.0,<8.0.0", "mock-serial>=0.0.1,<0.1.0 ; sys_platform != 'win32'"]
 video_benchmark = ["scikit-image>=0.23.2,<0.26.0", "pandas>=2.2.2,<2.4.0"]
 
 # Simulation
 # NOTE: Explicitly listing scipy helps flatten the dependecy tree.
-aloha = ["lerobot[dataset]", "gym-aloha>=0.1.2,<0.2.0", "lerobot[scipy-dep]"]
+aloha = ["lerobot[dataset]", "gym-aloha>=0.1.4,<0.2.0", "lerobot[scipy-dep]"]
 pusht = ["lerobot[dataset]", "gym-pusht>=0.1.5,<0.2.0", "pymunk>=6.6.0,<7.0.0"] # TODO: Fix pymunk version in gym-pusht instead
-libero = ["lerobot[dataset]", "lerobot[transformers-dep]", "hf-libero>=0.1.3,<0.2.0; sys_platform == 'linux'", "lerobot[scipy-dep]"]
+libero = ["lerobot[dataset]", "lerobot[transformers-dep]", "hf-libero>=0.1.4,<0.2.0; sys_platform == 'linux'", "lerobot[scipy-dep]"]
 metaworld = ["lerobot[dataset]", "metaworld==3.0.0", "lerobot[scipy-dep]"]
 # NOTE: vlabench is NOT exposed as a `lerobot` extra. Its only distribution
 # is the OpenMOSS/VLABench GitHub repo (package name `VLABench`, no PyPI
@@ -249,16 +306,22 @@ all = [
     "lerobot[lekiwi]",
     "lerobot[openarms]",
     "lerobot[reachy2]",
+    "lerobot[rebot]",
     "lerobot[kinematics]",
     "lerobot[intelrealsense]",
     "lerobot[diffusion]",
     "lerobot[multi_task_dit]",
     "lerobot[wallx]",
     "lerobot[pi]",
+    "lerobot[molmoact2]",
     "lerobot[smolvla]",
-    # "lerobot[groot]", TODO(Steven): Gr00t requires specific installation instructions for flash-attn
+    "lerobot[fastwam]",
+    "lerobot[groot]",
     "lerobot[xvla]",
+    "lerobot[evo1]",
     "lerobot[hilserl]",
+    "lerobot[vla_jepa]",
+    "lerobot[lingbot_va]",
     "lerobot[async]",
     "lerobot[dev]",
     "lerobot[test]",
@@ -269,6 +332,8 @@ all = [
     "lerobot[libero]; sys_platform == 'linux'",
     "lerobot[metaworld]",
     "lerobot[sarm]",
+    "lerobot[robometer]",
+    "lerobot[topreward]",
     "lerobot[peft]",
     # "lerobot[unitree_g1]", TODO: Unitree requires specific installation instructions for unitree_sdk2
 ]
@@ -290,11 +355,26 @@ lerobot-find-joint-limits="lerobot.scripts.lerobot_find_joint_limits:main"
 lerobot-imgtransform-viz="lerobot.scripts.lerobot_imgtransform_viz:main"
 lerobot-edit-dataset="lerobot.scripts.lerobot_edit_dataset:main"
 lerobot-setup-can="lerobot.scripts.lerobot_setup_can:main"
+lerobot-annotate="lerobot.scripts.lerobot_annotate:main"
 lerobot-rollout="lerobot.scripts.lerobot_rollout:main"
 
 # ---------------- Tool Configurations ----------------
+
+# cu128 wheels keep broad hardware reach; the driver floor is 570.86.
+# To use a different CUDA variant, reinstall torch with an explicit index, e.g.:
+#   uv pip install --force-reinstall torch torchvision \
+#       --index-url https://download.pytorch.org/whl/cu130
+[[tool.uv.index]]
+name = "pytorch-cu128"
+url = "https://download.pytorch.org/whl/cu128"
+explicit = true
+
+[tool.uv.sources]
+torch = [{ index = "pytorch-cu128", marker = "sys_platform == 'linux'" }]
+torchvision = [{ index = "pytorch-cu128", marker = "sys_platform == 'linux'" }]
+
 [tool.setuptools.package-data]
-lerobot = ["envs/*.json"]
+lerobot = ["envs/*.json", "annotations/steerable_pipeline/prompts/*.txt"]
 
 [tool.setuptools.packages.find]
 where = ["src"]
@@ -370,8 +450,12 @@ default.extend-ignore-identifiers-re = [
     "ein",
     "thw",
     "inpt",
+    "arange",
+    "is_compileable",
     "ROBOTIS",
-    "OT_VALUE"
+    "OT_VALUE",
+    "VanderBilt",
+    "seperated_timestep",
 ]
 
 # TODO: Uncomment when ready to use

requirements-macos.txt

@@ -1,729 +0,0 @@
-#
-# This file is autogenerated by pip-compile with Python 3.12
-# by the following command:
-#
-#    pip-compile --output-file=requirements-macos.txt requirements.in
-#
--e .[all]
-    # via -[all]
-absl-py==2.4.0
-    # via
-    #   dm-control
-    #   dm-env
-    #   dm-tree
-    #   labmaze
-    #   mujoco
-accelerate==1.13.0
-    # via
-    #   lerobot
-    #   peft
-aiohappyeyeballs==2.6.1
-    # via aiohttp
-aiohttp==3.13.3
-    # via fsspec
-aiosignal==1.4.0
-    # via aiohttp
-annotated-doc==0.0.4
-    # via
-    #   fastapi
-    #   typer
-annotated-types==0.7.0
-    # via pydantic
-anyio==4.12.1
-    # via
-    #   httpx
-    #   starlette
-    #   watchfiles
-asttokens==3.0.1
-    # via stack-data
-attrs==25.4.0
-    # via
-    #   aiohttp
-    #   dm-tree
-    #   jsonlines
-    #   rerun-sdk
-av==15.1.0
-    # via
-    #   lerobot
-    #   qwen-vl-utils
-certifi==2026.2.25
-    # via
-    #   httpcore
-    #   httpx
-    #   requests
-    #   sentry-sdk
-cffi==2.0.0
-    # via pymunk
-cfgv==3.5.0
-    # via pre-commit
-charset-normalizer==3.4.5
-    # via requests
-click==8.3.1
-    # via
-    #   typer
-    #   uvicorn
-    #   wandb
-cloudpickle==3.1.2
-    # via gymnasium
-cmake==4.1.3
-    # via lerobot
-cmeel==0.59.0
-    # via
-    #   cmeel-assimp
-    #   cmeel-boost
-    #   cmeel-console-bridge
-    #   cmeel-octomap
-    #   cmeel-qhull
-    #   cmeel-tinyxml2
-    #   cmeel-urdfdom
-    #   cmeel-zlib
-    #   coal-library
-    #   eigenpy
-    #   eiquadprog
-    #   pin
-    #   placo
-    #   rhoban-cmeel-jsoncpp
-cmeel-assimp==5.4.3.1
-    # via coal-library
-cmeel-boost==1.87.0.1
-    # via
-    #   coal-library
-    #   eigenpy
-    #   eiquadprog
-    #   pin
-cmeel-console-bridge==1.0.2.3
-    # via cmeel-urdfdom
-cmeel-octomap==1.10.0
-    # via coal-library
-cmeel-qhull==8.0.2.1
-    # via coal-library
-cmeel-tinyxml2==10.0.0
-    # via cmeel-urdfdom
-cmeel-urdfdom==4.0.1
-    # via pin
-cmeel-zlib==1.3.1
-    # via cmeel-assimp
-coal-library==3.0.1
-    # via pin
-contourpy==1.3.3
-    # via
-    #   lerobot
-    #   matplotlib
-coverage[toml]==7.13.4
-    # via pytest-cov
-cycler==0.12.1
-    # via matplotlib
-datasets==4.6.1
-    # via lerobot
-debugpy==1.8.20
-    # via lerobot
-decorator==5.2.1
-    # via ipython
-deepdiff==8.6.1
-    # via lerobot
-diffusers==0.35.2
-    # via lerobot
-dill==0.4.0
-    # via
-    #   datasets
-    #   multiprocess
-distlib==0.4.0
-    # via virtualenv
-dm-control==1.0.37
-    # via gym-aloha
-dm-env==1.6
-    # via dm-control
-dm-tree==0.1.9
-    # via
-    #   dm-control
-    #   dm-env
-docopt==0.6.2
-    # via num2words
-draccus==0.10.0
-    # via lerobot
-dynamixel-sdk==3.8.4
-    # via lerobot
-eigenpy==3.10.3
-    # via coal-library
-einops==0.8.2
-    # via lerobot
-eiquadprog==1.2.9
-    # via placo
-etils[epath,epy]==1.14.0
-    # via mujoco
-executing==2.2.1
-    # via stack-data
-faker==34.0.2
-    # via lerobot
-farama-notifications==0.0.4
-    # via gymnasium
-fastapi==0.135.1
-    # via
-    #   lerobot
-    #   teleop
-feetech-servo-sdk==1.0.0
-    # via lerobot
-filelock==3.25.0
-    # via
-    #   datasets
-    #   diffusers
-    #   huggingface-hub
-    #   python-discovery
-    #   torch
-    #   virtualenv
-fonttools==4.61.1
-    # via matplotlib
-frozenlist==1.8.0
-    # via
-    #   aiohttp
-    #   aiosignal
-fsspec[http]==2026.2.0
-    # via
-    #   datasets
-    #   etils
-    #   huggingface-hub
-    #   torch
-gitdb==4.0.12
-    # via gitpython
-gitpython==3.1.46
-    # via wandb
-glfw==2.10.0
-    # via
-    #   dm-control
-    #   mujoco
-grpcio==1.73.1
-    # via
-    #   grpcio-tools
-    #   lerobot
-    #   reachy2-sdk
-    #   reachy2-sdk-api
-grpcio-tools==1.73.1
-    # via
-    #   lerobot
-    #   reachy2-sdk-api
-gym-aloha==0.1.3
-    # via lerobot
-gym-hil==0.1.13
-    # via lerobot
-gym-pusht==0.1.6
-    # via lerobot
-gymnasium==1.2.3
-    # via
-    #   gym-aloha
-    #   gym-hil
-    #   gym-pusht
-    #   lerobot
-    #   metaworld
-h11==0.16.0
-    # via
-    #   httpcore
-    #   uvicorn
-hebi-py==2.11.0
-    # via lerobot
-hf-xet==1.3.2
-    # via huggingface-hub
-hidapi==0.14.0.post4
-    # via
-    #   gym-hil
-    #   lerobot
-httpcore==1.0.9
-    # via httpx
-httptools==0.7.1
-    # via uvicorn
-httpx==0.28.1
-    # via
-    #   datasets
-    #   huggingface-hub
-huggingface-hub==1.6.0
-    # via
-    #   accelerate
-    #   datasets
-    #   diffusers
-    #   lerobot
-    #   peft
-    #   tokenizers
-    #   transformers
-identify==2.6.17
-    # via pre-commit
-idna==3.11
-    # via
-    #   anyio
-    #   httpx
-    #   requests
-    #   yarl
-imageio[ffmpeg]==2.37.2
-    # via
-    #   gym-aloha
-    #   gym-hil
-    #   lerobot
-    #   metaworld
-    #   scikit-image
-imageio-ffmpeg==0.6.0
-    # via imageio
-importlib-metadata==8.7.1
-    # via diffusers
-iniconfig==2.3.0
-    # via pytest
-ipython==9.11.0
-    # via meshcat
-ipython-pygments-lexers==1.1.1
-    # via ipython
-ischedule==1.2.7
-    # via placo
-jedi==0.19.2
-    # via ipython
-jinja2==3.1.6
-    # via torch
-jsonlines==4.0.0
-    # via lerobot
-kiwisolver==1.4.9
-    # via matplotlib
-labmaze==1.0.6
-    # via dm-control
-lazy-loader==0.5
-    # via scikit-image
-librt==0.8.1
-    # via mypy
-lxml==6.0.2
-    # via dm-control
-markdown-it-py==4.0.0
-    # via rich
-markupsafe==3.0.3
-    # via jinja2
-matplotlib==3.10.8
-    # via lerobot
-matplotlib-inline==0.2.1
-    # via ipython
-mdurl==0.1.2
-    # via markdown-it-py
-mergedeep==1.3.4
-    # via draccus
-meshcat==0.3.2
-    # via placo
-metaworld==3.0.0
-    # via lerobot
-mock-serial==0.0.1
-    # via lerobot
-mpmath==1.3.0
-    # via sympy
-mujoco==3.5.0
-    # via
-    #   dm-control
-    #   gym-aloha
-    #   gym-hil
-    #   metaworld
-multidict==6.7.1
-    # via
-    #   aiohttp
-    #   yarl
-multiprocess==0.70.18
-    # via datasets
-mypy==1.19.1
-    # via lerobot
-mypy-extensions==1.1.0
-    # via
-    #   mypy
-    #   typing-inspect
-networkx==3.6.1
-    # via
-    #   scikit-image
-    #   torch
-nodeenv==1.10.0
-    # via pre-commit
-num2words==0.5.14
-    # via lerobot
-numpy==2.2.6
-    # via
-    #   accelerate
-    #   cmeel-boost
-    #   contourpy
-    #   datasets
-    #   diffusers
-    #   dm-control
-    #   dm-env
-    #   dm-tree
-    #   gymnasium
-    #   hebi-py
-    #   imageio
-    #   labmaze
-    #   lerobot
-    #   matplotlib
-    #   meshcat
-    #   metaworld
-    #   mujoco
-    #   opencv-python
-    #   opencv-python-headless
-    #   pandas
-    #   peft
-    #   pyquaternion
-    #   reachy2-sdk
-    #   rerun-sdk
-    #   scikit-image
-    #   scipy
-    #   shapely
-    #   teleop
-    #   tifffile
-    #   torchvision
-    #   transformers
-    #   transforms3d
-opencv-python==4.13.0.92
-    # via
-    #   gym-pusht
-    #   reachy2-sdk
-opencv-python-headless==4.12.0.88
-    # via lerobot
-orderly-set==5.5.0
-    # via deepdiff
-packaging==25.0
-    # via
-    #   accelerate
-    #   datasets
-    #   huggingface-hub
-    #   lazy-loader
-    #   lerobot
-    #   matplotlib
-    #   peft
-    #   pytest
-    #   qwen-vl-utils
-    #   reachy2-sdk
-    #   scikit-image
-    #   transformers
-    #   wandb
-pandas==2.3.3
-    # via
-    #   datasets
-    #   lerobot
-parso==0.8.6
-    # via jedi
-pathspec==1.0.4
-    # via mypy
-peft==0.18.1
-    # via lerobot
-pexpect==4.9.0
-    # via ipython
-pillow==12.1.1
-    # via
-    #   diffusers
-    #   imageio
-    #   matplotlib
-    #   meshcat
-    #   qwen-vl-utils
-    #   rerun-sdk
-    #   scikit-image
-    #   torchvision
-pin==3.4.0
-    # via placo
-placo==0.9.16
-    # via lerobot
-platformdirs==4.9.4
-    # via
-    #   python-discovery
-    #   virtualenv
-    #   wandb
-pluggy==1.6.0
-    # via
-    #   pytest
-    #   pytest-cov
-pre-commit==4.5.1
-    # via lerobot
-prompt-toolkit==3.0.52
-    # via ipython
-propcache==0.4.1
-    # via
-    #   aiohttp
-    #   yarl
-protobuf==6.31.1
-    # via
-    #   dm-control
-    #   grpcio-tools
-    #   lerobot
-    #   reachy2-sdk
-    #   reachy2-sdk-api
-    #   wandb
-psutil==7.2.2
-    # via
-    #   accelerate
-    #   imageio
-    #   peft
-ptyprocess==0.7.0
-    # via pexpect
-pure-eval==0.2.3
-    # via stack-data
-pyarrow==23.0.1
-    # via
-    #   datasets
-    #   rerun-sdk
-pycparser==3.0
-    # via cffi
-pydantic==2.12.5
-    # via
-    #   fastapi
-    #   wandb
-pydantic-core==2.41.5
-    # via pydantic
-pygame==2.6.1
-    # via
-    #   gym-hil
-    #   gym-pusht
-    #   lerobot
-pygments==2.19.2
-    # via
-    #   ipython
-    #   ipython-pygments-lexers
-    #   pytest
-    #   rich
-pymunk==6.11.1
-    # via
-    #   gym-pusht
-    #   lerobot
-pyngrok==7.5.1
-    # via meshcat
-pynput==1.8.1
-    # via
-    #   gym-hil
-    #   lerobot
-pyobjc-core==12.1
-    # via
-    #   pyobjc-framework-applicationservices
-    #   pyobjc-framework-cocoa
-    #   pyobjc-framework-coretext
-    #   pyobjc-framework-quartz
-pyobjc-framework-applicationservices==12.1
-    # via pynput
-pyobjc-framework-cocoa==12.1
-    # via
-    #   pyobjc-framework-applicationservices
-    #   pyobjc-framework-coretext
-    #   pyobjc-framework-quartz
-pyobjc-framework-coretext==12.1
-    # via pyobjc-framework-applicationservices
-pyobjc-framework-quartz==12.1
-    # via
-    #   pynput
-    #   pyobjc-framework-applicationservices
-    #   pyobjc-framework-coretext
-pyopengl==3.1.10
-    # via
-    #   dm-control
-    #   mujoco
-pyparsing==3.3.2
-    # via
-    #   dm-control
-    #   matplotlib
-pyquaternion==0.9.9
-    # via reachy2-sdk
-pyrealsense2-macosx==2.56.5
-    # via lerobot
-pyserial==3.5
-    # via
-    #   dynamixel-sdk
-    #   feetech-servo-sdk
-    #   lerobot
-pytest==8.4.2
-    # via
-    #   lerobot
-    #   pytest-cov
-    #   pytest-timeout
-    #   teleop
-pytest-cov==7.0.0
-    # via lerobot
-pytest-timeout==2.4.0
-    # via lerobot
-python-dateutil==2.9.0.post0
-    # via
-    #   faker
-    #   matplotlib
-    #   pandas
-python-discovery==1.1.1
-    # via virtualenv
-python-dotenv==1.2.2
-    # via uvicorn
-pytz==2026.1.post1
-    # via pandas
-pyyaml==6.0.3
-    # via
-    #   accelerate
-    #   datasets
-    #   draccus
-    #   hebi-py
-    #   huggingface-hub
-    #   peft
-    #   pre-commit
-    #   pyngrok
-    #   pyyaml-include
-    #   transformers
-    #   uvicorn
-    #   wandb
-pyyaml-include==1.4.1
-    # via draccus
-pyzmq==27.1.0
-    # via
-    #   lerobot
-    #   meshcat
-qwen-vl-utils==0.0.14
-    # via lerobot
-reachy2-sdk==1.0.15
-    # via lerobot
-reachy2-sdk-api==1.0.21
-    # via reachy2-sdk
-regex==2026.2.28
-    # via
-    #   diffusers
-    #   transformers
-requests==2.32.5
-    # via
-    #   datasets
-    #   diffusers
-    #   dm-control
-    #   qwen-vl-utils
-    #   teleop
-    #   wandb
-rerun-sdk==0.26.2
-    # via lerobot
-rhoban-cmeel-jsoncpp==1.9.4.9
-    # via placo
-rich==14.3.3
-    # via typer
-safetensors==0.7.0
-    # via
-    #   accelerate
-    #   diffusers
-    #   lerobot
-    #   peft
-    #   transformers
-scikit-image==0.25.2
-    # via
-    #   gym-pusht
-    #   lerobot
-scipy==1.17.1
-    # via
-    #   dm-control
-    #   lerobot
-    #   metaworld
-    #   scikit-image
-    #   torchdiffeq
-sentry-sdk==2.54.0
-    # via wandb
-shapely==2.1.2
-    # via gym-pusht
-shellingham==1.5.4
-    # via typer
-six==1.17.0
-    # via
-    #   pynput
-    #   python-dateutil
-smmap==5.0.3
-    # via gitdb
-stack-data==0.6.3
-    # via ipython
-starlette==0.52.1
-    # via fastapi
-sympy==1.14.0
-    # via torch
-teleop==0.1.4
-    # via lerobot
-termcolor==3.3.0
-    # via lerobot
-tifffile==2026.3.3
-    # via scikit-image
-tokenizers==0.22.2
-    # via transformers
-toml==0.10.2
-    # via draccus
-torch==2.10.0
-    # via
-    #   accelerate
-    #   lerobot
-    #   peft
-    #   torchdiffeq
-    #   torchvision
-torchcodec==0.10.0
-    # via lerobot
-torchdiffeq==0.2.5
-    # via lerobot
-torchvision==0.25.0
-    # via lerobot
-tornado==6.5.4
-    # via meshcat
-tqdm==4.67.3
-    # via
-    #   datasets
-    #   dm-control
-    #   huggingface-hub
-    #   peft
-    #   transformers
-traitlets==5.14.3
-    # via
-    #   ipython
-    #   matplotlib-inline
-transformers==5.3.0
-    # via
-    #   lerobot
-    #   peft
-transforms3d==0.4.2
-    # via teleop
-typer==0.24.1
-    # via
-    #   huggingface-hub
-    #   transformers
-typing-extensions==4.15.0
-    # via
-    #   aiosignal
-    #   anyio
-    #   etils
-    #   faker
-    #   fastapi
-    #   gymnasium
-    #   huggingface-hub
-    #   mypy
-    #   pydantic
-    #   pydantic-core
-    #   rerun-sdk
-    #   starlette
-    #   torch
-    #   typing-inspect
-    #   typing-inspection
-    #   wandb
-typing-inspect==0.9.0
-    # via draccus
-typing-inspection==0.4.2
-    # via
-    #   fastapi
-    #   pydantic
-tzdata==2025.3
-    # via pandas
-u-msgpack-python==2.8.0
-    # via meshcat
-urllib3==2.6.3
-    # via
-    #   requests
-    #   sentry-sdk
-uvicorn[standard]==0.41.0
-    # via teleop
-uvloop==0.22.1
-    # via uvicorn
-virtualenv==21.1.0
-    # via pre-commit
-wandb==0.24.2
-    # via lerobot
-watchfiles==1.1.1
-    # via uvicorn
-wcwidth==0.6.0
-    # via prompt-toolkit
-websocket-client==1.9.0
-    # via teleop
-websockets==16.0
-    # via uvicorn
-wrapt==2.1.2
-    # via dm-tree
-xxhash==3.6.0
-    # via datasets
-yarl==1.23.0
-    # via aiohttp
-zipp==3.23.0
-    # via
-    #   etils
-    #   importlib-metadata
-
-# The following packages are considered to be unsafe in a requirements file:
-# setuptools

requirements-ubuntu.txt

@@ -1,882 +0,0 @@
-#
-# This file is autogenerated by pip-compile with Python 3.12
-# by the following command:
-#
-#    pip-compile --output-file=requirements-ubuntu.txt requirements.in
-#
--e .[all]
-    # via -[all]
-absl-py==2.4.0
-    # via
-    #   dm-control
-    #   dm-env
-    #   dm-tree
-    #   labmaze
-    #   mujoco
-    #   tensorboard
-accelerate==1.13.0
-    # via
-    #   lerobot
-    #   peft
-aiohappyeyeballs==2.6.1
-    # via aiohttp
-aiohttp==3.13.3
-    # via fsspec
-aiosignal==1.4.0
-    # via aiohttp
-annotated-doc==0.0.4
-    # via
-    #   fastapi
-    #   typer
-annotated-types==0.7.0
-    # via pydantic
-antlr4-python3-runtime==4.9.3
-    # via
-    #   hydra-core
-    #   omegaconf
-anyio==4.12.1
-    # via
-    #   httpx
-    #   starlette
-    #   watchfiles
-asttokens==3.0.1
-    # via stack-data
-attrs==25.4.0
-    # via
-    #   aiohttp
-    #   dm-tree
-    #   jsonlines
-    #   jsonschema
-    #   referencing
-    #   rerun-sdk
-av==15.1.0
-    # via
-    #   lerobot
-    #   qwen-vl-utils
-bddl==1.0.1
-    # via hf-libero
-certifi==2026.2.25
-    # via
-    #   httpcore
-    #   httpx
-    #   requests
-    #   sentry-sdk
-cffi==2.0.0
-    # via pymunk
-cfgv==3.5.0
-    # via pre-commit
-charset-normalizer==3.4.5
-    # via requests
-click==8.3.1
-    # via
-    #   typer
-    #   uvicorn
-    #   wandb
-cloudpickle==3.1.2
-    # via
-    #   gymnasium
-    #   hf-libero
-cmake==4.1.3
-    # via lerobot
-cmeel==0.59.0
-    # via
-    #   cmeel-assimp
-    #   cmeel-boost
-    #   cmeel-console-bridge
-    #   cmeel-octomap
-    #   cmeel-qhull
-    #   cmeel-tinyxml2
-    #   cmeel-urdfdom
-    #   cmeel-zlib
-    #   coal-library
-    #   eigenpy
-    #   eiquadprog
-    #   pin
-    #   placo
-    #   rhoban-cmeel-jsoncpp
-cmeel-assimp==5.4.3.1
-    # via coal-library
-cmeel-boost==1.87.0.1
-    # via
-    #   coal-library
-    #   eigenpy
-    #   eiquadprog
-    #   pin
-cmeel-console-bridge==1.0.2.3
-    # via cmeel-urdfdom
-cmeel-octomap==1.10.0
-    # via coal-library
-cmeel-qhull==8.0.2.1
-    # via coal-library
-cmeel-tinyxml2==10.0.0
-    # via cmeel-urdfdom
-cmeel-urdfdom==4.0.1
-    # via pin
-cmeel-zlib==1.3.1
-    # via cmeel-assimp
-coal-library==3.0.1
-    # via pin
-contourpy==1.3.3
-    # via
-    #   lerobot
-    #   matplotlib
-coverage[toml]==7.13.4
-    # via pytest-cov
-cuda-bindings==12.9.4
-    # via torch
-cuda-pathfinder==1.4.1
-    # via cuda-bindings
-cycler==0.12.1
-    # via matplotlib
-datasets==4.6.1
-    # via lerobot
-debugpy==1.8.20
-    # via lerobot
-decorator==5.2.1
-    # via ipython
-deepdiff==8.6.1
-    # via lerobot
-diffusers==0.35.2
-    # via lerobot
-dill==0.4.0
-    # via
-    #   datasets
-    #   multiprocess
-distlib==0.4.0
-    # via virtualenv
-dm-control==1.0.37
-    # via gym-aloha
-dm-env==1.6
-    # via dm-control
-dm-tree==0.1.9
-    # via
-    #   dm-control
-    #   dm-env
-docopt==0.6.2
-    # via num2words
-draccus==0.10.0
-    # via lerobot
-dynamixel-sdk==3.8.4
-    # via lerobot
-easydict==1.13
-    # via hf-libero
-egl-probe==1.0.2
-    # via robomimic
-eigenpy==3.10.3
-    # via coal-library
-einops==0.8.2
-    # via
-    #   hf-libero
-    #   lerobot
-eiquadprog==1.2.9
-    # via placo
-etils[epath,epy]==1.14.0
-    # via mujoco
-evdev==1.9.3
-    # via pynput
-executing==2.2.1
-    # via stack-data
-faker==34.0.2
-    # via lerobot
-farama-notifications==0.0.4
-    # via gymnasium
-fastapi==0.135.1
-    # via
-    #   lerobot
-    #   teleop
-fastjsonschema==2.21.2
-    # via nbformat
-feetech-servo-sdk==1.0.0
-    # via lerobot
-filelock==3.25.0
-    # via
-    #   datasets
-    #   diffusers
-    #   huggingface-hub
-    #   python-discovery
-    #   torch
-    #   virtualenv
-fonttools==4.61.1
-    # via matplotlib
-frozenlist==1.8.0
-    # via
-    #   aiohttp
-    #   aiosignal
-fsspec[http]==2026.2.0
-    # via
-    #   datasets
-    #   etils
-    #   huggingface-hub
-    #   torch
-future==1.0.0
-    # via hf-libero
-gitdb==4.0.12
-    # via gitpython
-gitpython==3.1.46
-    # via wandb
-glfw==2.10.0
-    # via
-    #   dm-control
-    #   mujoco
-grpcio==1.73.1
-    # via
-    #   grpcio-tools
-    #   lerobot
-    #   reachy2-sdk
-    #   reachy2-sdk-api
-    #   tensorboard
-grpcio-tools==1.73.1
-    # via
-    #   lerobot
-    #   reachy2-sdk-api
-gym-aloha==0.1.3
-    # via lerobot
-gym-hil==0.1.13
-    # via lerobot
-gym-pusht==0.1.6
-    # via lerobot
-gymnasium==1.2.3
-    # via
-    #   gym-aloha
-    #   gym-hil
-    #   gym-pusht
-    #   hf-libero
-    #   lerobot
-    #   metaworld
-h11==0.16.0
-    # via
-    #   httpcore
-    #   uvicorn
-h5py==3.16.0
-    # via robomimic
-hebi-py==2.11.0
-    # via lerobot
-hf-egl-probe==1.0.2
-    # via hf-libero
-hf-libero==0.1.3
-    # via lerobot
-hf-xet==1.3.2
-    # via huggingface-hub
-hidapi==0.14.0.post4
-    # via
-    #   gym-hil
-    #   lerobot
-httpcore==1.0.9
-    # via httpx
-httptools==0.7.1
-    # via uvicorn
-httpx==0.28.1
-    # via
-    #   datasets
-    #   huggingface-hub
-huggingface-hub==1.6.0
-    # via
-    #   accelerate
-    #   datasets
-    #   diffusers
-    #   lerobot
-    #   peft
-    #   tokenizers
-    #   transformers
-hydra-core==1.3.2
-    # via hf-libero
-identify==2.6.17
-    # via pre-commit
-idna==3.11
-    # via
-    #   anyio
-    #   httpx
-    #   requests
-    #   yarl
-imageio[ffmpeg]==2.37.2
-    # via
-    #   gym-aloha
-    #   gym-hil
-    #   lerobot
-    #   metaworld
-    #   robomimic
-    #   scikit-image
-imageio-ffmpeg==0.6.0
-    # via
-    #   imageio
-    #   robomimic
-importlib-metadata==8.7.1
-    # via diffusers
-iniconfig==2.3.0
-    # via pytest
-ipython==9.11.0
-    # via meshcat
-ipython-pygments-lexers==1.1.1
-    # via ipython
-ischedule==1.2.7
-    # via placo
-jedi==0.19.2
-    # via ipython
-jinja2==3.1.6
-    # via torch
-jsonlines==4.0.0
-    # via lerobot
-jsonschema==4.26.0
-    # via nbformat
-jsonschema-specifications==2025.9.1
-    # via jsonschema
-jupyter-core==5.9.1
-    # via nbformat
-jupytext==1.19.1
-    # via bddl
-kiwisolver==1.4.9
-    # via matplotlib
-labmaze==1.0.6
-    # via dm-control
-lazy-loader==0.5
-    # via scikit-image
-librt==0.8.1
-    # via mypy
-llvmlite==0.46.0
-    # via numba
-lxml==6.0.2
-    # via dm-control
-markdown==3.10.2
-    # via tensorboard
-markdown-it-py==4.0.0
-    # via
-    #   jupytext
-    #   mdit-py-plugins
-    #   rich
-markupsafe==3.0.3
-    # via
-    #   jinja2
-    #   werkzeug
-matplotlib==3.10.8
-    # via
-    #   hf-libero
-    #   lerobot
-matplotlib-inline==0.2.1
-    # via ipython
-mdit-py-plugins==0.5.0
-    # via jupytext
-mdurl==0.1.2
-    # via markdown-it-py
-mergedeep==1.3.4
-    # via draccus
-meshcat==0.3.2
-    # via placo
-metaworld==3.0.0
-    # via lerobot
-mock-serial==0.0.1
-    # via lerobot
-mpmath==1.3.0
-    # via sympy
-mujoco==3.5.0
-    # via
-    #   dm-control
-    #   gym-aloha
-    #   gym-hil
-    #   hf-libero
-    #   metaworld
-    #   robosuite
-multidict==6.7.1
-    # via
-    #   aiohttp
-    #   yarl
-multiprocess==0.70.18
-    # via datasets
-mypy==1.19.1
-    # via lerobot
-mypy-extensions==1.1.0
-    # via
-    #   mypy
-    #   typing-inspect
-nbformat==5.10.4
-    # via jupytext
-networkx==3.6.1
-    # via
-    #   bddl
-    #   scikit-image
-    #   torch
-nodeenv==1.10.0
-    # via pre-commit
-num2words==0.5.14
-    # via lerobot
-numba==0.64.0
-    # via robosuite
-numpy==2.2.6
-    # via
-    #   accelerate
-    #   bddl
-    #   cmeel-boost
-    #   contourpy
-    #   datasets
-    #   diffusers
-    #   dm-control
-    #   dm-env
-    #   dm-tree
-    #   gymnasium
-    #   h5py
-    #   hebi-py
-    #   hf-libero
-    #   imageio
-    #   labmaze
-    #   lerobot
-    #   matplotlib
-    #   meshcat
-    #   metaworld
-    #   mujoco
-    #   numba
-    #   opencv-python
-    #   opencv-python-headless
-    #   pandas
-    #   peft
-    #   pyquaternion
-    #   reachy2-sdk
-    #   rerun-sdk
-    #   robomimic
-    #   robosuite
-    #   scikit-image
-    #   scipy
-    #   shapely
-    #   teleop
-    #   tensorboard
-    #   tensorboardx
-    #   tifffile
-    #   torchvision
-    #   transformers
-    #   transforms3d
-nvidia-cublas-cu12==12.8.4.1
-    # via
-    #   nvidia-cudnn-cu12
-    #   nvidia-cusolver-cu12
-    #   torch
-nvidia-cuda-cupti-cu12==12.8.90
-    # via torch
-nvidia-cuda-nvrtc-cu12==12.8.93
-    # via torch
-nvidia-cuda-runtime-cu12==12.8.90
-    # via torch
-nvidia-cudnn-cu12==9.10.2.21
-    # via torch
-nvidia-cufft-cu12==11.3.3.83
-    # via torch
-nvidia-cufile-cu12==1.13.1.3
-    # via torch
-nvidia-curand-cu12==10.3.9.90
-    # via torch
-nvidia-cusolver-cu12==11.7.3.90
-    # via torch
-nvidia-cusparse-cu12==12.5.8.93
-    # via
-    #   nvidia-cusolver-cu12
-    #   torch
-nvidia-cusparselt-cu12==0.7.1
-    # via torch
-nvidia-nccl-cu12==2.27.5
-    # via torch
-nvidia-nvjitlink-cu12==12.8.93
-    # via
-    #   nvidia-cufft-cu12
-    #   nvidia-cusolver-cu12
-    #   nvidia-cusparse-cu12
-    #   torch
-nvidia-nvshmem-cu12==3.4.5
-    # via torch
-nvidia-nvtx-cu12==12.8.90
-    # via torch
-omegaconf==2.3.0
-    # via hydra-core
-opencv-python==4.13.0.92
-    # via
-    #   gym-pusht
-    #   hf-libero
-    #   reachy2-sdk
-    #   robosuite
-opencv-python-headless==4.12.0.88
-    # via lerobot
-orderly-set==5.5.0
-    # via deepdiff
-packaging==25.0
-    # via
-    #   accelerate
-    #   datasets
-    #   huggingface-hub
-    #   hydra-core
-    #   jupytext
-    #   lazy-loader
-    #   lerobot
-    #   matplotlib
-    #   peft
-    #   pytest
-    #   qwen-vl-utils
-    #   reachy2-sdk
-    #   scikit-image
-    #   tensorboard
-    #   tensorboardx
-    #   transformers
-    #   wandb
-pandas==2.3.3
-    # via
-    #   datasets
-    #   lerobot
-parso==0.8.6
-    # via jedi
-pathspec==1.0.4
-    # via mypy
-peft==0.18.1
-    # via lerobot
-pexpect==4.9.0
-    # via ipython
-pillow==12.1.1
-    # via
-    #   diffusers
-    #   imageio
-    #   matplotlib
-    #   meshcat
-    #   qwen-vl-utils
-    #   rerun-sdk
-    #   robosuite
-    #   scikit-image
-    #   tensorboard
-    #   torchvision
-pin==3.4.0
-    # via placo
-placo==0.9.16
-    # via lerobot
-platformdirs==4.9.4
-    # via
-    #   jupyter-core
-    #   python-discovery
-    #   virtualenv
-    #   wandb
-pluggy==1.6.0
-    # via
-    #   pytest
-    #   pytest-cov
-pre-commit==4.5.1
-    # via lerobot
-prompt-toolkit==3.0.52
-    # via ipython
-propcache==0.4.1
-    # via
-    #   aiohttp
-    #   yarl
-protobuf==6.31.1
-    # via
-    #   dm-control
-    #   grpcio-tools
-    #   lerobot
-    #   reachy2-sdk
-    #   reachy2-sdk-api
-    #   tensorboard
-    #   tensorboardx
-    #   wandb
-psutil==7.2.2
-    # via
-    #   accelerate
-    #   imageio
-    #   peft
-    #   robomimic
-ptyprocess==0.7.0
-    # via pexpect
-pure-eval==0.2.3
-    # via stack-data
-pyarrow==23.0.1
-    # via
-    #   datasets
-    #   rerun-sdk
-pycparser==3.0
-    # via cffi
-pydantic==2.12.5
-    # via
-    #   fastapi
-    #   wandb
-pydantic-core==2.41.5
-    # via pydantic
-pygame==2.6.1
-    # via
-    #   gym-hil
-    #   gym-pusht
-    #   lerobot
-pygments==2.19.2
-    # via
-    #   ipython
-    #   ipython-pygments-lexers
-    #   pytest
-    #   rich
-pymunk==6.11.1
-    # via
-    #   gym-pusht
-    #   lerobot
-pyngrok==7.5.1
-    # via meshcat
-pynput==1.8.1
-    # via
-    #   gym-hil
-    #   lerobot
-pyopengl==3.1.10
-    # via
-    #   dm-control
-    #   mujoco
-pyparsing==3.3.2
-    # via
-    #   dm-control
-    #   matplotlib
-pyquaternion==0.9.9
-    # via reachy2-sdk
-pyrealsense2==2.56.5.9235
-    # via lerobot
-pyserial==3.5
-    # via
-    #   dynamixel-sdk
-    #   feetech-servo-sdk
-    #   lerobot
-pytest==8.4.2
-    # via
-    #   bddl
-    #   lerobot
-    #   pytest-cov
-    #   pytest-timeout
-    #   teleop
-pytest-cov==7.0.0
-    # via lerobot
-pytest-timeout==2.4.0
-    # via lerobot
-python-dateutil==2.9.0.post0
-    # via
-    #   faker
-    #   matplotlib
-    #   pandas
-python-discovery==1.1.1
-    # via virtualenv
-python-dotenv==1.2.2
-    # via uvicorn
-python-xlib==0.33
-    # via pynput
-pytz==2026.1.post1
-    # via pandas
-pyyaml==6.0.3
-    # via
-    #   accelerate
-    #   datasets
-    #   draccus
-    #   hebi-py
-    #   huggingface-hub
-    #   jupytext
-    #   omegaconf
-    #   peft
-    #   pre-commit
-    #   pyngrok
-    #   pyyaml-include
-    #   transformers
-    #   uvicorn
-    #   wandb
-pyyaml-include==1.4.1
-    # via draccus
-pyzmq==27.1.0
-    # via
-    #   lerobot
-    #   meshcat
-qwen-vl-utils==0.0.14
-    # via lerobot
-reachy2-sdk==1.0.15
-    # via lerobot
-reachy2-sdk-api==1.0.21
-    # via reachy2-sdk
-referencing==0.37.0
-    # via
-    #   jsonschema
-    #   jsonschema-specifications
-regex==2026.2.28
-    # via
-    #   diffusers
-    #   transformers
-requests==2.32.5
-    # via
-    #   datasets
-    #   diffusers
-    #   dm-control
-    #   qwen-vl-utils
-    #   teleop
-    #   wandb
-rerun-sdk==0.26.2
-    # via lerobot
-rhoban-cmeel-jsoncpp==1.9.4.9
-    # via placo
-rich==14.3.3
-    # via typer
-robomimic==0.2.0
-    # via hf-libero
-robosuite==1.4.0
-    # via hf-libero
-rpds-py==0.30.0
-    # via
-    #   jsonschema
-    #   referencing
-safetensors==0.7.0
-    # via
-    #   accelerate
-    #   diffusers
-    #   lerobot
-    #   peft
-    #   transformers
-scikit-image==0.25.2
-    # via
-    #   gym-pusht
-    #   lerobot
-scipy==1.17.1
-    # via
-    #   dm-control
-    #   lerobot
-    #   metaworld
-    #   robosuite
-    #   scikit-image
-    #   torchdiffeq
-sentry-sdk==2.54.0
-    # via wandb
-shapely==2.1.2
-    # via gym-pusht
-shellingham==1.5.4
-    # via typer
-six==1.17.0
-    # via
-    #   pynput
-    #   python-dateutil
-    #   python-xlib
-smmap==5.0.3
-    # via gitdb
-stack-data==0.6.3
-    # via ipython
-starlette==0.52.1
-    # via fastapi
-sympy==1.14.0
-    # via torch
-teleop==0.1.4
-    # via lerobot
-tensorboard==2.20.0
-    # via robomimic
-tensorboard-data-server==0.7.2
-    # via tensorboard
-tensorboardx==2.6.4
-    # via robomimic
-termcolor==3.3.0
-    # via
-    #   lerobot
-    #   robomimic
-thop==0.1.1.post2209072238
-    # via hf-libero
-tifffile==2026.3.3
-    # via scikit-image
-tokenizers==0.22.2
-    # via transformers
-toml==0.10.2
-    # via draccus
-torch==2.10.0
-    # via
-    #   accelerate
-    #   lerobot
-    #   peft
-    #   robomimic
-    #   thop
-    #   torchdiffeq
-    #   torchvision
-torchcodec==0.10.0
-    # via lerobot
-torchdiffeq==0.2.5
-    # via lerobot
-torchvision==0.25.0
-    # via
-    #   lerobot
-    #   robomimic
-tornado==6.5.4
-    # via meshcat
-tqdm==4.67.3
-    # via
-    #   datasets
-    #   dm-control
-    #   huggingface-hub
-    #   peft
-    #   robomimic
-    #   transformers
-traitlets==5.14.3
-    # via
-    #   ipython
-    #   jupyter-core
-    #   matplotlib-inline
-    #   nbformat
-transformers==5.3.0
-    # via
-    #   hf-libero
-    #   lerobot
-    #   peft
-transforms3d==0.4.2
-    # via teleop
-triton==3.6.0
-    # via torch
-typer==0.24.1
-    # via
-    #   huggingface-hub
-    #   transformers
-typing-extensions==4.15.0
-    # via
-    #   aiosignal
-    #   anyio
-    #   etils
-    #   faker
-    #   fastapi
-    #   gymnasium
-    #   huggingface-hub
-    #   mypy
-    #   pydantic
-    #   pydantic-core
-    #   referencing
-    #   rerun-sdk
-    #   starlette
-    #   torch
-    #   typing-inspect
-    #   typing-inspection
-    #   wandb
-typing-inspect==0.9.0
-    # via draccus
-typing-inspection==0.4.2
-    # via
-    #   fastapi
-    #   pydantic
-tzdata==2025.3
-    # via pandas
-u-msgpack-python==2.8.0
-    # via meshcat
-urllib3==2.6.3
-    # via
-    #   requests
-    #   sentry-sdk
-uvicorn[standard]==0.41.0
-    # via teleop
-uvloop==0.22.1
-    # via uvicorn
-virtualenv==21.1.0
-    # via pre-commit
-wandb==0.24.2
-    # via
-    #   hf-libero
-    #   lerobot
-watchfiles==1.1.1
-    # via uvicorn
-wcwidth==0.6.0
-    # via prompt-toolkit
-websocket-client==1.9.0
-    # via teleop
-websockets==16.0
-    # via uvicorn
-werkzeug==3.1.6
-    # via tensorboard
-wrapt==2.1.2
-    # via dm-tree
-xxhash==3.6.0
-    # via datasets
-yarl==1.23.0
-    # via aiohttp
-zipp==3.23.0
-    # via
-    #   etils
-    #   importlib-metadata
-
-# The following packages are considered to be unsafe in a requirements file:
-# setuptools

requirements.in

@@ -1,9 +0,0 @@
-# requirements.in
-
-# requirements-macos.txt was generated on macOS and is platform-specific (macOS 26.3.1 25D2128 arm64).
-# Darwin MacBook-Pro.local 25.3.0 Darwin Kernel Version 25.3.0: Wed Jan 28 20:54:55 PST 2026; root:xnu-12377.91.3~2/RELEASE_ARM64_T8132 arm64
-
-# requirements-ubuntu.txt was generated on Linux and is platform-specific (Ubuntu 24.04.4 LTS x86_64).
-# Linux lerobot-linux 6.17.0-14-generic #14~24.04.1-Ubuntu SMP PREEMPT_DYNAMIC Thu Jan 15 15:52:10 UTC 2 x86_64 x86_64 x86_64 GNU/Linux
-
--e .[all]

src/lerobot/annotations/__init__.py

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

src/lerobot/annotations/steerable_pipeline/__init__.py

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

src/lerobot/annotations/steerable_pipeline/config.py

@@ -0,0 +1,211 @@
+#!/usr/bin/env python
+
+# Copyright 2026 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from __future__ import annotations
+
+from dataclasses import dataclass, field
+from pathlib import Path
+from typing import Any
+
+
+@dataclass
+class PlanConfig:
+    """``plan`` module: subtasks + plan + memory + task augmentation."""
+
+    enabled: bool = True
+
+    # ``task_aug`` rephrasings at t=0 (renderer rotates ${task} among them); 0 disables.
+    n_task_rephrasings: int = 10
+
+    # Derive the task from video instead of episode_task: off / if_short / always.
+    # Affects prompts only; ``meta/tasks.parquet`` is untouched.
+    derive_task_from_video: str = "if_short"
+    derive_task_min_words: int = 3
+
+    # --- Frame input: timestamped contact sheets (always on) ---------------
+    # The subtask describe/segment passes ALWAYS render the episode as
+    # macrodata/refiner-style contact sheets: sampled frames packed into JPEG
+    # grids with each frame's timestamp burned into its corner, so the VLM
+    # cites the exact source time of a boundary directly. This is far cheaper
+    # in vision tokens than one image per frame (≈2× faster subtask generation
+    # in practice), which is why the sampling is dense by default.
+    #
+    # ``frames_per_second`` is the sampling rate: 2.0 = one frame every 0.5s.
+    frames_per_second: float = 2.0
+    # Frame budget per VLM call (= columns × rows × sheets). When a whole
+    # episode sampled at ``frames_per_second`` exceeds this, the episode is
+    # AUTOMATICALLY split into consecutive windows of
+    # ``max_frames_per_prompt`` frames each (one describe→segment call per
+    # window, still at the full ``frames_per_second`` density), and the
+    # per-window spans are merged + stitched into one contiguous cover. So an
+    # episode of any length is always covered at the full sampling density.
+    max_frames_per_prompt: int = 60
+    contact_sheet_columns: int = 5
+    contact_sheet_frames_per_sheet: int = 20
+    contact_sheet_frame_width: int = 224
+    contact_sheet_quality: int = 84
+
+    min_subtask_seconds: float = 1.5
+    plan_max_steps: int = 8
+
+    # Narrate-only grounding pass before segmenting — best defense against subtasks
+    # invented from the task text (+1 VLM call/episode).
+    subtask_describe_first: bool = True
+
+    # Emit ``style="plan"`` rows at each boundary; False = subtasks + memory only.
+    emit_plan: bool = True
+
+    # Emit ``style="memory"`` rows at each boundary; False = subtasks (+ plan) only.
+    # Symmetric counterpart of ``emit_plan``.
+    emit_memory: bool = True
+
+    # (subtask spans are always stitched to a contiguous full-episode cover; not configurable.)
+
+    # Optional EgoMimic-style 5-axis task augmentation; replaces n_task_rephrasings.
+    task_aug_axes: TaskAugAxesConfig = field(default_factory=lambda: TaskAugAxesConfig())
+
+
+@dataclass
+class TaskAugAxesConfig:
+    """5-axis t=0 task augmentation (EgoMimic-style): synonym / omit_arm /
+    omit_orientation / omit_grasp_method / combined. Replaces n_task_rephrasings
+    when enabled; each variant becomes a ``task_aug`` row. Axes with nothing to
+    omit emit fewer entries. Defaults (3+3+2+2+2) match EgoMimic."""
+
+    enabled: bool = False
+
+    synonym_paraphrase: int = 3
+    omit_arm: int = 3
+    omit_orientation: int = 2
+    omit_grasp_method: int = 2
+    combined_omissions: int = 2
+
+
+@dataclass
+class InterjectionsConfig:
+    """``interjections`` module: interjections + paired speech."""
+
+    enabled: bool = True
+
+    # Each emits a paired (interjection, speech) row + a plan refresh at that ts.
+    max_interjections_per_episode: int = 3
+    interjection_min_t: float = 2.0
+
+    # Frame window centered on the timestamp so the VLM sees motion, not one frame.
+    interjection_window_seconds: float = 2.0
+    interjection_window_frames: int = 4
+
+
+@dataclass
+class VqaConfig:
+    """``vqa`` module: general VQA."""
+
+    enabled: bool = True
+    vqa_emission_hz: float = 1.0
+    K: int = 1
+    """Consecutive frames per emission tick. The VLM grounds on the FIRST frame,
+    so K>1 smears stale labels onto moved frames. Default 1 (no smear)."""
+    question_types: tuple[str, ...] = ("bbox", "keypoint", "count", "attribute", "spatial")
+
+    # True: ground VQA only on --vlm.camera_key (default: every camera).
+    restrict_to_default_camera: bool = False
+
+
+@dataclass
+class VlmConfig:
+    """Shared Qwen-VL client configuration."""
+
+    # Only ``openai`` (OpenAI-compatible vLLM server, auto-spawned when
+    # auto_serve=True); ``stub`` is for tests.
+    backend: str = "openai"
+    model_id: str = "Qwen/Qwen3.6-27B"
+
+    # OpenAI-compatible endpoint; ``EMPTY`` key works for local servers.
+    api_base: str = "http://localhost:8000/v1"
+    api_key: str = "EMPTY"
+
+    # Spawn a server if none answers api_base; False = fail fast on a remote.
+    auto_serve: bool = True
+    serve_port: int = 8000
+    # Override the auto-serve command; ``{port}`` substituted per replica.
+    serve_command: str | None = None
+
+    # Independent servers for round-robin routing (one per GPU). num_gpus=0 = one each.
+    parallel_servers: int = 1
+    num_gpus: int = 0
+    client_concurrency: int = 16
+    serve_ready_timeout_s: float = 600.0
+
+    max_new_tokens: int = 512
+    temperature: float = 0.2
+
+    # Auto-serve context length (None → 32768); other vLLM flags go in serve_command.
+    max_model_len: int | None = None
+
+    # Camera for keyframes; None → first ``observation.images.*`` key.
+    camera_key: str | None = None
+    # Forwarded as extra_body.chat_template_kwargs (e.g. {"enable_thinking": false}).
+    chat_template_kwargs: dict[str, Any] | None = None
+
+
+@dataclass
+class ExecutorConfig:
+    """Executor settings (intra-process episode concurrency; distribution via HF Jobs)."""
+
+    # Episodes processed concurrently per phase; main knob for saturating the servers.
+    episode_parallelism: int = 16
+
+
+@dataclass
+class AnnotationPipelineConfig:
+    """Top-level config for ``lerobot-annotate`` (rewrites data shards in place)."""
+
+    # Hub dataset: download source when ``root`` unset; push target when push_to_hub
+    # is on and ``new_repo_id`` unset.
+    repo_id: str | None = None
+
+    # Separate push target (matches the LeRobot edit tools). Unset → push in place.
+    new_repo_id: str | None = None
+
+    root: Path | None = None
+
+    # Defaults to ``<root>/.annotate_staging/``.
+    staging_dir: Path | None = None
+
+    seed: int = 1729
+
+    plan: PlanConfig = field(default_factory=PlanConfig)
+    interjections: InterjectionsConfig = field(default_factory=InterjectionsConfig)
+    vqa: VqaConfig = field(default_factory=VqaConfig)
+
+    vlm: VlmConfig = field(default_factory=VlmConfig)
+    executor: ExecutorConfig = field(default_factory=ExecutorConfig)
+
+    skip_validation: bool = False
+    only_episodes: tuple[int, ...] | None = None
+
+    # Keyframe decode backend forwarded to ``decode_video_frames``. None →
+    # library default (torchcodec when available, else PyAV). Or pin
+    # ``"torchcodec"`` / ``"pyav"`` explicitly.
+    video_backend: str | None = None
+
+    # Upload to the Hub (new_repo_id if set, else repo_id; one must be set).
+    push_to_hub: bool = False
+    push_private: bool = False
+    push_commit_message: str | None = None
+
+    def resolved_staging_dir(self, root: Path) -> Path:
+        return self.staging_dir if self.staging_dir is not None else root / ".annotate_staging"

src/lerobot/annotations/steerable_pipeline/executor.py

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

src/lerobot/annotations/steerable_pipeline/frames.py

@@ -0,0 +1,483 @@
+#!/usr/bin/env python
+
+# Copyright 2026 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""Keyframe extraction for the annotation pipeline.
+
+Modules attach decoded camera frames to their VLM prompts so the model can
+ground subtask decomposition, interjection scenarios, and VQA in actual
+visual content. The pipeline shares one provider across modules and one
+episode at a time, with a small per-episode cache so multiple modules
+querying the same timestamp pay decode cost once.
+"""
+
+from __future__ import annotations
+
+import io
+import logging
+import math
+import threading
+from collections.abc import Sequence
+from dataclasses import dataclass, field
+from pathlib import Path
+from typing import Any, Protocol
+
+import PIL.Image
+import torch
+
+from lerobot.configs import RGBEncoderConfig
+from lerobot.datasets.video_utils import decode_video_frames, reencode_video
+
+from .reader import EpisodeRecord, snap_to_frame
+
+logger = logging.getLogger(__name__)
+
+
+class FrameProvider(Protocol):
+    """Decodes camera frames at episode-relative timestamps."""
+
+    @property
+    def camera_keys(self) -> list[str]:
+        """All ``observation.images.*`` feature keys this provider can decode."""
+
+    def frames_at(
+        self,
+        record: EpisodeRecord,
+        timestamps: list[float],
+        camera_key: str | None = None,
+    ) -> list[Any]:
+        """Return one decoded frame per timestamp from ``camera_key`` (or default).
+
+        Frames are ``torch.Tensor`` (``C, H, W`` uint8) — the shape
+        :func:`lerobot.datasets.video_utils.decode_video_frames` returns.
+        :func:`to_image_blocks` converts them to PIL only at the VLM-message
+        boundary.
+
+        Empty list if the camera is unavailable. ``camera_key=None`` falls back
+        to the provider's default camera so existing single-camera callers
+        (the ``plan`` and ``interjections`` modules) keep working unchanged.
+        """
+
+    def video_for_episode(
+        self,
+        record: EpisodeRecord,
+        max_frames: int,
+        camera_key: str | None = None,
+    ) -> list[Any]:
+        """Return up to ``max_frames`` decoded frames covering the whole episode.
+
+        Sampling is uniform across the episode duration. Frames are
+        ``torch.Tensor`` (``C, H, W`` uint8); :func:`to_video_block` wraps
+        them into one ``{"type":"video", "video":<list>}`` block for a
+        Qwen-VL-compatible model that pools temporally itself. Empty list if
+        no camera available.
+        """
+
+
+@dataclass
+class _NullProvider:
+    """No-op provider used when the dataset has no video keys or in tests."""
+
+    @property
+    def camera_keys(self) -> list[str]:
+        return []
+
+    def frames_at(
+        self,
+        record: EpisodeRecord,
+        timestamps: list[float],
+        camera_key: str | None = None,
+    ) -> list[Any]:
+        return []
+
+    def video_for_episode(
+        self,
+        record: EpisodeRecord,
+        max_frames: int,
+        camera_key: str | None = None,
+    ) -> list[Any]:
+        return []
+
+
+def null_provider() -> FrameProvider:
+    return _NullProvider()
+
+
+@dataclass
+class VideoFrameProvider:
+    """Decodes frames from the dataset's ``observation.images.*`` streams.
+
+    By default the *first* camera key is used for the ``plan`` module
+    (subtask decomposition) and the ``interjections`` module (interjection
+    scenarios) — those prompts care about *what is happening*, not which
+    angle. The ``vqa`` module instead iterates over every camera in
+    :attr:`camera_keys` so each frame's
+    grounded answer (bbox/keypoint/...) is tagged with the camera it was
+    grounded against.
+
+    ``camera_key`` overrides the default-camera choice but does not restrict
+    :attr:`camera_keys`. Pass ``camera_key`` explicitly to ``frames_at`` /
+    ``video_for_episode`` to read a non-default stream.
+
+    Caches up to ``cache_size`` decoded frames per process to keep
+    co-timestamped ``interjections`` + ``plan`` plan-update calls cheap.
+    """
+
+    root: Path
+    camera_key: str | None = None
+    tolerance_s: float = 1e-2
+    cache_size: int = 256
+    # Keyframe decode backend forwarded to
+    # :func:`lerobot.datasets.video_utils.decode_video_frames`. ``None``
+    # uses the library default (torchcodec when available, else PyAV).
+    video_backend: str | None = None
+    _meta: Any = field(default=None, init=False, repr=False)
+    _cache: dict = field(default_factory=dict, init=False, repr=False)
+    _camera_keys: list[str] = field(default_factory=list, init=False, repr=False)
+    # Pipeline runs the three module phases under a ThreadPoolExecutor (see
+    # ``ExecutorConfig.episode_parallelism``); guard the dict cache and the
+    # one-shot warn flag against concurrent updates from worker threads.
+    _lock: threading.Lock = field(default_factory=threading.Lock, init=False, repr=False)
+    # Serializes decode_video_frames calls: torchcodec hands out one
+    # ``VideoDecoder`` per file from a process-wide cache, and the decoder
+    # is not safe to drive from multiple threads at once.
+    _decode_lock: threading.Lock = field(default_factory=threading.Lock, init=False, repr=False)
+    _warned_decode_fail: bool = field(default=False, init=False, repr=False)
+
+    def __post_init__(self) -> None:
+        from lerobot.datasets.dataset_metadata import LeRobotDatasetMetadata  # noqa: PLC0415
+
+        self._meta = LeRobotDatasetMetadata(repo_id="local", root=self.root)
+        # Only ``video_keys`` are decodable here: the clip/decode paths read
+        # ``videos/<key>/from_timestamp`` from episode metadata, which exists
+        # only for video-stored cameras. Image-stored cameras (also in
+        # ``camera_keys``) would KeyError, so restrict the list — and the
+        # default — to video keys.
+        # Depth cameras are excluded from the annotation pipeline for now.
+        depth_keys = set(self._meta.depth_keys)
+        keys = [key for key in self._meta.video_keys if key not in depth_keys]
+        # Last-resort fallback: if metadata didn't surface any video keys but
+        # the caller explicitly named a camera (``--vlm.camera_key=...``),
+        # trust them — the key is by definition known to exist on the dataset.
+        if not keys and self.camera_key:
+            keys = [self.camera_key]
+        self._camera_keys = keys
+        if self.camera_key is None:
+            self.camera_key = keys[0] if keys else None
+
+    @property
+    def camera_keys(self) -> list[str]:
+        """All ``observation.images.*`` keys available on this dataset."""
+        return list(self._camera_keys)
+
+    def frames_at(
+        self,
+        record: EpisodeRecord,
+        timestamps: list[float],
+        camera_key: str | None = None,
+    ) -> list[Any]:
+        target = camera_key if camera_key is not None else self.camera_key
+        if not timestamps or target is None:
+            return []
+        # Snap each request to the nearest real frame timestamp: callers
+        # sample uniform grids whose points land mid-frame, and
+        # ``decode_video_frames`` rejects queries farther than
+        # ``tolerance_s`` from a decodable frame. Snapping also dedupes
+        # repeat queries through the cache.
+        if record.frame_timestamps:
+            timestamps = [snap_to_frame(float(ts), record.frame_timestamps) for ts in timestamps]
+
+        out: list[Any] = []
+        misses: list[float] = []
+        miss_indices: list[int] = []
+        with self._lock:
+            for i, ts in enumerate(timestamps):
+                key = (record.episode_index, target, round(float(ts), 6))
+                cached = self._cache.get(key)
+                if cached is not None:
+                    out.append(cached)
+                else:
+                    out.append(None)
+                    misses.append(float(ts))
+                    miss_indices.append(i)
+
+        if misses:
+            decoded = self._decode(record.episode_index, misses, target)
+            # ``_decode`` returns exactly one frame per requested timestamp,
+            # or an empty list if decoding failed wholesale. A partial list
+            # would mean a frame/timestamp misalignment, so only pair them up
+            # when the counts match (``strict=True`` then guards regressions).
+            if len(decoded) == len(miss_indices):
+                with self._lock:
+                    for i, frame in zip(miss_indices, decoded, strict=True):
+                        out[i] = frame
+                        key = (record.episode_index, target, round(float(timestamps[i]), 6))
+                        if len(self._cache) >= self.cache_size:
+                            self._cache.pop(next(iter(self._cache)))
+                        self._cache[key] = frame
+        # filter out any None left over from decode failures
+        return [frame for frame in out if frame is not None]
+
+    def video_for_episode(
+        self,
+        record: EpisodeRecord,
+        max_frames: int,
+        camera_key: str | None = None,
+    ) -> list[Any]:
+        """Return up to ``max_frames`` frames uniformly sampled across the episode.
+
+        The whole episode duration is covered; the model picks subtask
+        boundaries from the temporal pooling it does internally. Frames are
+        ``torch.Tensor`` (see :meth:`frames_at`).
+        """
+        target = camera_key if camera_key is not None else self.camera_key
+        if max_frames <= 0 or target is None or not record.frame_timestamps:
+            return []
+        n_frames = min(max_frames, len(record.frame_timestamps))
+        if n_frames == len(record.frame_timestamps):
+            timestamps = list(record.frame_timestamps)
+        else:
+            t0 = record.frame_timestamps[0]
+            t_last = record.frame_timestamps[-1]
+            if t_last <= t0:
+                timestamps = [float(t0)] * n_frames
+            else:
+                step = (t_last - t0) / (n_frames - 1) if n_frames > 1 else 0.0
+                timestamps = [float(t0 + i * step) for i in range(n_frames)]
+        return self.frames_at(record, timestamps, camera_key=target)
+
+    def episode_clip_path(self, record: EpisodeRecord, cache_dir: Path) -> Path | None:
+        """Extract the episode's subclip to ``cache_dir/ep_{idx:06d}.mp4``.
+
+        Returns ``None`` if the dataset has no video tracks or extraction
+        failed. Skips re-extract when the cached clip already exists.
+        Re-encodes to H.264 via
+        :func:`lerobot.datasets.video_utils.reencode_video` so the resulting
+        mp4 is decodable by every downstream video processor — stream-copy
+        would inherit the source codec (often AV1 in modern LeRobot
+        datasets), which vllm's libav build cannot decode.
+        """
+        if self.camera_key is None:
+            return None
+        cache_dir.mkdir(parents=True, exist_ok=True)
+        out_path = cache_dir / f"ep_{record.episode_index:06d}.mp4"
+        if out_path.exists() and out_path.stat().st_size > 0:
+            return out_path
+        ep = self._meta.episodes[record.episode_index]
+        from_timestamp = float(ep[f"videos/{self.camera_key}/from_timestamp"])
+        to_timestamp = float(ep[f"videos/{self.camera_key}/to_timestamp"])
+        src = self.root / self._meta.get_video_file_path(record.episode_index, self.camera_key)
+        encoder = RGBEncoderConfig(vcodec="h264", pix_fmt="yuv420p", g=None, crf=23, preset="ultrafast")
+        try:
+            reencode_video(
+                src,
+                out_path,
+                video_encoder=encoder,
+                overwrite=True,
+                start_time_s=from_timestamp,
+                end_time_s=to_timestamp,
+            )
+        except Exception:
+            logger.warning(
+                "clip extraction failed for episode %s (%s)", record.episode_index, src, exc_info=True
+            )
+            return None
+        return out_path if out_path.exists() and out_path.stat().st_size > 0 else None
+
+    def _decode(self, episode_index: int, timestamps: list[float], camera_key: str) -> list[Any]:
+        """Decode ``timestamps`` from the episode's video as ``(C, H, W)`` tensors.
+
+        Delegates to :func:`lerobot.datasets.video_utils.decode_video_frames`
+        (torchcodec when available, PyAV otherwise; ``video_backend`` pins
+        one explicitly). Returns one frame per requested timestamp, or ``[]``
+        if decoding failed — callers treat ``[]`` as "no frames available".
+        """
+        ep = self._meta.episodes[episode_index]
+        from_timestamp = ep[f"videos/{camera_key}/from_timestamp"]
+        shifted = [from_timestamp + ts for ts in timestamps]
+        video_path = self.root / self._meta.get_video_file_path(episode_index, camera_key)
+
+        try:
+            # The module phases decode under a ThreadPoolExecutor (see
+            # ``ExecutorConfig.episode_parallelism``) but torchcodec's cached
+            # per-file decoder is single-threaded, so serialize decodes on a
+            # dedicated lock. Frame extraction is a small fraction of episode
+            # wall time (VLM calls dominate), so the contention is cheap.
+            with self._decode_lock:
+                # Stacked ``(N, C, H, W)`` uint8 tensor; one row per timestamp.
+                decoded = decode_video_frames(
+                    video_path, shifted, self.tolerance_s, backend=self.video_backend, return_uint8=True
+                )
+            return list(decoded)
+        except Exception as exc:
+            # Log loudly the first time so a silent vqa-module no-op (every
+            # prompt skipped because frames_at returned []) is debuggable from
+            # the job log instead of post-hoc parquet inspection. Subsequent
+            # failures stay quiet.
+            with self._lock:
+                already_warned = self._warned_decode_fail
+                if not already_warned:
+                    self._warned_decode_fail = True
+            if not already_warned:
+                logger.warning(
+                    "VideoFrameProvider._decode failed for episode=%s camera=%s video_path=%s backend=%s: %s",
+                    episode_index,
+                    camera_key,
+                    video_path,
+                    self.video_backend,
+                    exc,
+                    exc_info=exc,
+                )
+            return []
+
+
+def make_frame_provider(
+    root: Path, camera_key: str | None = None, video_backend: str | None = None
+) -> FrameProvider:
+    """Build a :class:`VideoFrameProvider` if videos are present, else null."""
+    try:
+        provider = VideoFrameProvider(root=root, camera_key=camera_key, video_backend=video_backend)
+    except Exception:
+        return null_provider()
+    if provider.camera_key is None:
+        return null_provider()
+    return provider
+
+
+def _frame_to_pil(frame: Any) -> Any:
+    """Materialise a decoded frame as a ``PIL.Image`` for the VLM message.
+
+    Frames flow through the provider as ``torch.Tensor`` (``C, H, W`` uint8,
+    straight from :func:`decode_video_frames`); PIL is only created here, at
+    the VLM-message boundary, because the chat backends expect PIL images /
+    data URLs. Non-tensor inputs (e.g. test stubs) pass through untouched.
+    """
+    if not isinstance(frame, torch.Tensor):
+        return frame
+    array = frame.detach().cpu()
+    if array.ndim == 3 and array.shape[0] in (1, 3):
+        array = array.permute(1, 2, 0)  # (C, H, W) -> (H, W, C)
+    if array.shape[-1] == 1:
+        array = array.squeeze(-1)
+    return PIL.Image.fromarray(array.to(torch.uint8).numpy())
+
+
+def to_image_blocks(frames: list[Any]) -> list[dict[str, Any]]:
+    """Convert decoded frames to Qwen-VL-compatible image content blocks."""
+    return [{"type": "image", "image": _frame_to_pil(frame)} for frame in frames]
+
+
+def to_video_block(frames: list[Any]) -> list[dict[str, Any]]:
+    """Wrap a list of decoded frames as one Qwen-VL video block.
+
+    Returns ``[]`` when the list is empty, so the caller can splat the result
+    into a content array without a separate emptiness check.
+    """
+    if not frames:
+        return []
+    return [{"type": "video", "video": [_frame_to_pil(frame) for frame in frames]}]
+
+
+def to_video_url_block(url: str | None, fps: float = 2.0) -> list[dict[str, Any]]:
+    """Wrap a video file URL as one ``video_url`` block.
+
+    Used by the ``openai`` backend (transformers serve / vllm serve /
+    ktransformers serve), where the server handles frame sampling.
+    Returns ``[]`` when ``url`` is ``None`` so the caller can splat.
+    """
+    if not url:
+        return []
+    return [{"type": "video_url", "video_url": {"url": url}, "fps": fps}]
+
+
+def _draw_timestamp_badge(image: PIL.Image.Image, timestamp: float) -> PIL.Image.Image:
+    """Burn ``timestamp`` (seconds) into the top-left corner of ``image``.
+
+    A solid black badge with white text, so a VLM reading a contact sheet can
+    cite the exact source time of each tile (e.g. ``012.50s``) directly,
+    instead of the caller having to map tile position back to time. Mirrors
+    the macrodata/refiner contact-sheet convention.
+    """
+    from PIL import ImageDraw, ImageFont
+
+    result = image.copy()
+    draw = ImageDraw.Draw(result)
+    font = ImageFont.load_default()
+    label = f"{timestamp:06.2f}s"
+    left, top, right, bottom = draw.textbbox((0, 0), label, font=font)
+    text_w, text_h = right - left, bottom - top
+    pad = max(3, round(min(image.width, image.height) * 0.018))
+    draw.rectangle((0, 0, text_w + pad * 2, text_h + pad * 2), fill=(0, 0, 0))
+    draw.text((pad - left, pad - top), label, fill=(255, 255, 255), font=font)
+    return result
+
+
+def to_contact_sheet_blocks(
+    frames: Sequence[Any],
+    timestamps: Sequence[float],
+    *,
+    columns: int = 5,
+    frames_per_sheet: int = 20,
+    frame_width: int = 224,
+    quality: int = 84,
+) -> list[dict[str, Any]]:
+    """Pack decoded frames into timestamped JPEG contact-sheet image blocks.
+
+    Each frame is resized to ``frame_width`` wide, stamped with its
+    episode-relative timestamp, and tiled row-major into grids of
+    ``frames_per_sheet`` (``columns`` wide). One ``{"type":"image", ...}``
+    block is returned per grid; many frames collapse into a few images, so a
+    long episode's temporal coverage stays dense at a fraction of the vision
+    tokens N separate frames would cost. ``frames`` and ``timestamps`` must be
+    aligned and equal length. Returns ``[]`` for empty input.
+    """
+    from PIL import Image
+
+    if not frames:
+        return []
+    columns = max(1, columns)
+    frames_per_sheet = max(1, frames_per_sheet)
+    rows_per_sheet = math.ceil(frames_per_sheet / columns)
+
+    tiles: list[PIL.Image.Image] = []
+    for ts, frame in zip(timestamps, frames, strict=False):
+        img = _frame_to_pil(frame)
+        if not isinstance(img, PIL.Image.Image):
+            continue
+        img = img.convert("RGB")
+        if img.width != frame_width:
+            height = max(1, round(img.height * frame_width / img.width))
+            img = img.resize((frame_width, height), resample=Image.Resampling.BILINEAR)
+        tiles.append(_draw_timestamp_badge(img, float(ts)))
+    if not tiles:
+        return []
+
+    blocks: list[dict[str, Any]] = []
+    for start in range(0, len(tiles), frames_per_sheet):
+        chunk = tiles[start : start + frames_per_sheet]
+        cell_w = max(tile.width for tile in chunk)
+        cell_h = max(tile.height for tile in chunk)
+        sheet = Image.new("RGB", (cell_w * columns, cell_h * rows_per_sheet), color=(0, 0, 0))
+        for i, tile in enumerate(chunk):
+            x = (i % columns) * cell_w
+            y = (i // columns) * cell_h
+            sheet.paste(tile, (x, y))
+        # JPEG round-trip at ``quality`` to match the refiner convention and
+        # shrink the wire payload; vision-token count is set by resolution, so
+        # the real saving is the grid packing, not the codec.
+        buf = io.BytesIO()
+        sheet.save(buf, format="JPEG", quality=quality)
+        buf.seek(0)
+        blocks.append({"type": "image", "image": Image.open(buf).convert("RGB")})
+    return blocks

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

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

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

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

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

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

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

@@ -0,0 +1,780 @@
+#!/usr/bin/env python
+
+# Copyright 2026 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""``plan`` module: subtask decomposition + plan + memory (PERSISTENT styles)."""
+
+from __future__ import annotations
+
+import logging
+from collections.abc import Sequence
+from dataclasses import dataclass, field
+from typing import Any
+
+from ..config import PlanConfig
+from ..frames import (
+    FrameProvider,
+    null_provider,
+    to_contact_sheet_blocks,
+)
+from ..prompts import load as load_prompt
+from ..reader import EpisodeRecord, reconstruct_subtask_spans, snap_to_frame
+from ..staging import EpisodeStaging
+from ..vlm_client import VlmClient
+
+logger = logging.getLogger(__name__)
+
+
+# Prepended to every describe / segment prompt so the VLM knows the images are
+# timestamped contact-sheet grids, not a single video, and reads the burned-in
+# per-tile timestamp when choosing boundaries.
+def _contact_sheet_preamble(columns: int) -> str:
+    return (
+        "CONTACT SHEETS — how to read the images below:\n"
+        f"- Each image is a grid of sampled video frames, {columns} per row, "
+        "with time running left-to-right then top-to-bottom (row-major).\n"
+        "- Each frame has its timestamp burned into the top-left corner, e.g. "
+        '"012.50s". Use that printed timestamp (not the tile position) when you '
+        "choose start/end times; boundaries should land on or near a printed "
+        "timestamp.\n"
+        "- Frames continue across grids: an action may span the end of one sheet "
+        "and the start of the next, so do not place a boundary just because a new "
+        "image begins.\n\n"
+    )
+
+
+# Appended to every describe (and segment) prompt. A visual, causal definition
+# of where one event ends and the next begins — adapted from macrodata/refiner —
+# to sharpen cut points while the existing prompt keeps owning the imperative
+# phrasing.
+_CAUSAL_BOUNDARY_RULES = (
+    "EVENT BOUNDARIES — where one event ends and the next begins:\n"
+    "- Start a new event whenever the world state changes: an object becomes "
+    "held (the gripper closes on it), an object is released (the gripper opens "
+    "and it stays put), an object reaches a new location, a lid/door/drawer "
+    "changes open/closed state, a tool starts or stops affecting a surface, or "
+    "contents visibly move (e.g. poured).\n"
+    "- If a single action changes the same state gradually and continuously, "
+    "keep it as ONE event — do not split it.\n"
+    "- If the same action repeats on different objects or target locations, "
+    "treat each repetition as a separate event.\n"
+    "- Do NOT create boundaries for idle time, camera motion, hesitation, or "
+    "tiny hand adjustments."
+)
+
+
+@dataclass
+class PlanSubtasksMemoryModule:
+    """Generate subtask spans, plan, and memory rows.
+
+    All output is persistent (lives in ``language_persistent``):
+
+    - ``subtask`` rows: one per span, stamped at the span's *start* timestamp
+      (snapped to an exact frame).
+    - ``plan`` rows: emitted at ``t=0``; refreshed at every interjection
+      timestamp via :meth:`run_plan_updates` (called by the executor after
+      the ``interjections`` module completes).
+    - ``memory`` rows: emitted at each subtask boundary (= subtask start
+      timestamp from the second subtask onward).
+    """
+
+    vlm: VlmClient
+    config: PlanConfig
+    frame_provider: FrameProvider = field(default_factory=null_provider)
+
+    @property
+    def enabled(self) -> bool:
+        return self.config.enabled
+
+    def run_episode(self, record: EpisodeRecord, staging: EpisodeStaging) -> None:
+        rows: list[dict[str, Any]] = []
+        # Task driving every plan-module prompt: canonical episode_task, or a
+        # video-derived one when it's empty/placeholder (see derive_task_*).
+        effective_task = self._resolve_effective_task(record)
+        # task_aug rows at t=0: phrasings the renderer rotates ${task} through.
+        # Either the structured 5-axis taxonomy (task_aug_axes.enabled) or
+        # free-form n_task_rephrasings; the effective task is always emitted
+        # first so the rotation covers the source-of-truth phrasing.
+        t0 = float(record.frame_timestamps[0]) if record.frame_timestamps else 0.0
+        variants: list[str] | None = None
+        if self.config.task_aug_axes.enabled and effective_task:
+            variants = self._generate_task_aug_by_axes(effective_task, self.config.task_aug_axes)
+        elif self.config.n_task_rephrasings > 0 and effective_task:
+            variants = self._generate_task_rephrasings(effective_task, n=self.config.n_task_rephrasings)
+        if variants is not None:
+            rows.extend(self._task_aug_rows([effective_task, *variants], t0))
+
+        subtask_spans = self._generate_subtasks(record, task=effective_task)
+
+        # subtask rows
+        for span in subtask_spans:
+            rows.append(
+                {
+                    "role": "assistant",
+                    "content": span["text"],
+                    "style": "subtask",
+                    "timestamp": snap_to_frame(span["start"], record.frame_timestamps),
+                    "tool_calls": None,
+                }
+            )
+        # Plan rows at every subtask boundary (incl. t=0). The plan is a
+        # numbered list of still-todo subtasks, so re-emitting at each
+        # boundary makes it shrink as work progresses — ${plan} at frame t is
+        # exactly what's left to do.
+        if self.config.emit_plan:
+            for span in subtask_spans:
+                boundary_t = snap_to_frame(span["start"], record.frame_timestamps)
+                plan_text = self._generate_plan(
+                    record, subtask_spans, refresh_t=boundary_t, task=effective_task
+                )
+                if plan_text is not None:
+                    rows.append(
+                        {
+                            "role": "assistant",
+                            "content": plan_text,
+                            "style": "plan",
+                            "timestamp": float(boundary_t),
+                            "tool_calls": None,
+                        }
+                    )
+        # memory rows at every subtask boundary except the very first start;
+        # skipped entirely when ``emit_memory`` is False (subtasks-only / plan-only).
+        prior_memory = ""
+        memory_boundaries = enumerate(subtask_spans[1:], start=1) if self.config.emit_memory else []
+        for i, span in memory_boundaries:
+            completed = subtask_spans[i - 1]["text"]
+            remaining = [s["text"] for s in subtask_spans[i:]]
+            mem_text = self._generate_memory(record, prior_memory, completed, remaining, task=effective_task)
+            if mem_text:
+                ts = snap_to_frame(span["start"], record.frame_timestamps)
+                rows.append(
+                    {
+                        "role": "assistant",
+                        "content": mem_text,
+                        "style": "memory",
+                        "timestamp": ts,
+                        "tool_calls": None,
+                    }
+                )
+                prior_memory = mem_text
+        staging.write("plan", rows)
+
+    # ------------------------------------------------------------------
+    # Task derivation + rephrasings
+    # ------------------------------------------------------------------
+
+    _PLACEHOLDER_TASKS: frozenset[str] = frozenset(
+        {
+            "debug",
+            "test",
+            "tbd",
+            "todo",
+            "n/a",
+            "na",
+            "untitled",
+            "unnamed",
+            "default",
+            "placeholder",
+        }
+    )
+
+    def _resolve_effective_task(self, record: EpisodeRecord) -> str:
+        """Decide which task string drives the ``plan`` module for this episode.
+
+        Returns the user-supplied ``record.episode_task`` unless
+        ``derive_task_from_video`` says otherwise (see config docstring).
+        Falls back gracefully to the canonical task if video derivation
+        fails.
+        """
+        canonical = (record.episode_task or "").strip()
+        mode = (self.config.derive_task_from_video or "off").strip().lower()
+        if mode == "always":
+            derived = self._derive_task_from_video(record)
+            return derived or canonical
+        if mode == "if_short" and self._task_seems_bad(canonical):
+            derived = self._derive_task_from_video(record)
+            if derived:
+                return derived
+        return canonical
+
+    def _task_seems_bad(self, task: str) -> bool:
+        if not task:
+            return True
+        if len(task.split()) < int(self.config.derive_task_min_words):
+            return True
+        return task.lower() in self._PLACEHOLDER_TASKS
+
+    @staticmethod
+    def _task_aug_rows(phrasings: Sequence[str], t0: float) -> list[dict[str, Any]]:
+        """Build deduplicated ``task_aug`` rows (role=user) at ``t0``."""
+        seen: set[str] = set()
+        rows: list[dict[str, Any]] = []
+        for phrasing in phrasings:
+            key = phrasing.strip()
+            if not key or key in seen:
+                continue
+            seen.add(key)
+            rows.append(
+                {"role": "user", "content": key, "style": "task_aug", "timestamp": t0, "tool_calls": None}
+            )
+        return rows
+
+    # ------------------------------------------------------------------
+    # VLM call helpers — every plan-module prompt follows the same shape:
+    # build messages → single VLM call → pull a named field.
+    # ------------------------------------------------------------------
+
+    def _vlm_field(self, messages: list[dict[str, Any]], field: str) -> Any:
+        """Run a single VLM call and return ``result[field]`` or ``None``.
+
+        Centralizes the ``vlm.generate_json([m])[0]`` + ``isinstance(dict)``
+        dance every prompt-call site needs.
+        """
+        result = self.vlm.generate_json([messages])[0]
+        if isinstance(result, dict):
+            return result.get(field)
+        return None
+
+    @staticmethod
+    def _text_message(text: str) -> list[dict[str, Any]]:
+        """One-shot text-only user message wrapped for ``generate_json``."""
+        return [{"role": "user", "content": [{"type": "text", "text": text}]}]
+
+    def _video_message(
+        self,
+        record: EpisodeRecord,
+        prompt: str,
+        window: tuple[float, float] | None = None,
+    ) -> list[dict[str, Any]]:
+        """User message combining the (optionally windowed) contact sheets with ``prompt``.
+
+        The prompt is always prefixed with a short explanation of how to read
+        the timestamped grids, so the model treats them as one ordered
+        sequence of frames rather than unrelated images.
+        """
+        prompt = _contact_sheet_preamble(self.config.contact_sheet_columns) + prompt
+        content = [*self._episode_video_block(record, window=window), {"type": "text", "text": prompt}]
+        return [{"role": "user", "content": content}]
+
+    def _derive_task_from_video(self, record: EpisodeRecord) -> str | None:
+        """Ask the VLM "what is this video about" with no task hint at all."""
+        text = self._vlm_field(self._video_message(record, load_prompt("plan_video_task")), "task")
+        return text.strip() if isinstance(text, str) and text.strip() else None
+
+    def _generate_task_rephrasings(self, base_task: str, *, n: int) -> list[str]:
+        """Generate ``n`` text-only paraphrases of ``base_task``."""
+        if n <= 0 or not base_task:
+            return []
+        prompt = load_prompt("plan_task_rephrasings").format(base_task=base_task, n=n)
+        raw = self._vlm_field(self._text_message(prompt), "rephrasings")
+        if not isinstance(raw, list):
+            return []
+        out = [item.strip().strip('"').strip("'") for item in raw if isinstance(item, str)]
+        return [s for s in out if s][:n]
+
+    # ------------------------------------------------------------------
+    # Structured 5-axis task augmentation (EgoMimic-style taxonomy)
+    # ------------------------------------------------------------------
+
+    def _generate_task_aug_by_axes(self, base_task: str, axes_cfg: Any) -> list[str]:
+        """One VLM call → variants along the 5-axis taxonomy.
+
+        Variants from all axes are flattened into a single list (the
+        downstream pipeline doesn't need to know about the per-axis
+        bucketing — every variant becomes a ``task_aug`` row). Order
+        is preserved for reproducibility: synonym_paraphrase first,
+        then omit_arm, then omit_orientation, then omit_grasp_method,
+        then combined_omissions.
+        """
+        if not base_task:
+            return []
+        prompt = load_prompt("plan_task_aug_axes").format(
+            base_task=base_task,
+            n_synonym=axes_cfg.synonym_paraphrase,
+            n_omit_arm=axes_cfg.omit_arm,
+            n_omit_orientation=axes_cfg.omit_orientation,
+            n_omit_grasp_method=axes_cfg.omit_grasp_method,
+            n_combined=axes_cfg.combined_omissions,
+        )
+        result = self.vlm.generate_json([self._text_message(prompt)])[0]
+        if not isinstance(result, dict):
+            return []
+        ordered_axes = (
+            "synonym_paraphrase",
+            "omit_arm",
+            "omit_orientation",
+            "omit_grasp_method",
+            "combined_omissions",
+        )
+        flat: list[str] = []
+        seen: set[str] = set()
+        for axis in ordered_axes:
+            entries = result.get(axis)
+            if not isinstance(entries, list):
+                continue
+            for item in entries:
+                if not isinstance(item, str):
+                    continue
+                key = item.strip().strip('"').strip("'")
+                if not key or key in seen:
+                    continue
+                seen.add(key)
+                flat.append(key)
+        return flat
+
+    def _episode_video_block(
+        self, record: EpisodeRecord, window: tuple[float, float] | None = None
+    ) -> list[dict[str, Any]]:
+        """Timestamped contact sheets for the describe / segmentation prompts.
+
+        Always renders the (optionally windowed) episode as contact sheets:
+        frames sampled at ``frames_per_second`` and packed into timestamped
+        JPEG grids. ``max_frames_per_prompt`` caps the frame count; whole
+        episodes that exceed it are windowed upstream in
+        :meth:`_generate_subtasks` so each call stays within budget while the
+        full episode keeps its sampling density.
+
+        When ``window=(w0, w1)`` is given the badges are WINDOW-RELATIVE
+        (``ts - w0``) to match the window-relative time frame the
+        segmentation prompt works in (spans are offset back to absolute time
+        afterwards).
+        """
+        if not record.frame_timestamps:
+            return []
+        if window is not None:
+            w0, w1 = float(window[0]), float(window[1])
+            dur = max(0.0, w1 - w0)
+            n = max(1, int(round(dur * self.config.frames_per_second)) + 1)
+            n = min(n, self.config.max_frames_per_prompt)
+            if n <= 1 or dur <= 0.0:
+                timestamps = [0.5 * (w0 + w1)]
+            else:
+                step = dur / (n - 1)
+                timestamps = [w0 + i * step for i in range(n)]
+            frames = self.frame_provider.frames_at(record, timestamps)
+            rel = [ts - w0 for ts in timestamps[: len(frames)]]
+            return self._contact_sheet_blocks(frames, rel)
+        episode_duration = record.frame_timestamps[-1] - record.frame_timestamps[0]
+        n = max(1, int(round(episode_duration * self.config.frames_per_second)) + 1)
+        n = min(n, self.config.max_frames_per_prompt)
+        timestamps = self._uniform_episode_timestamps(record, n)
+        frames = self.frame_provider.frames_at(record, timestamps)
+        return self._contact_sheet_blocks(frames, timestamps[: len(frames)])
+
+    @staticmethod
+    def _uniform_episode_timestamps(record: EpisodeRecord, n: int) -> list[float]:
+        """``n`` episode-relative timestamps spanning ``[t0, t_last]`` uniformly."""
+        ts = record.frame_timestamps
+        if n >= len(ts):
+            return [float(t) for t in ts]
+        t0, t_last = float(ts[0]), float(ts[-1])
+        if t_last <= t0 or n <= 1:
+            return [t0] * max(1, n)
+        step = (t_last - t0) / (n - 1)
+        return [t0 + i * step for i in range(n)]
+
+    def _contact_sheet_blocks(self, frames: list[Any], timestamps: list[float]) -> list[dict[str, Any]]:
+        """Build timestamped contact-sheet image blocks from decoded frames."""
+        return to_contact_sheet_blocks(
+            frames,
+            timestamps,
+            columns=self.config.contact_sheet_columns,
+            frames_per_sheet=self.config.contact_sheet_frames_per_sheet,
+            frame_width=self.config.contact_sheet_frame_width,
+            quality=self.config.contact_sheet_quality,
+        )
+
+    def run_plan_updates(
+        self,
+        record: EpisodeRecord,
+        staging: EpisodeStaging,
+        interjection_times: Sequence[float],
+        interjection_texts: Sequence[str] | None = None,
+    ) -> None:
+        """Append additional ``plan`` rows at every interjection timestamp.
+
+        Plans refresh ONLY on user interjections (event-driven). The
+        interjection text is forwarded into the prompt so the refreshed plan
+        reflects the user's correction.
+        """
+        if not self.config.emit_plan:
+            return
+        existing = staging.read("plan")
+        # Pass the last frame timestamp so the final span is closed (else its
+        # end == start, zero duration, and a refresh inside it is missed).
+        episode_end_t = float(record.frame_timestamps[-1]) if record.frame_timestamps else None
+        spans = reconstruct_subtask_spans(existing, episode_end_t=episode_end_t)
+        already_planned: set[float] = {float(r["timestamp"]) for r in existing if r.get("style") == "plan"}
+        new_rows = list(existing)
+
+        texts: list[str | None] = (
+            [None] * len(interjection_times)
+            if interjection_texts is None
+            else [str(t) if t else None for t in interjection_texts]
+        )
+        for raw_t, inter_text in zip(interjection_times, texts, strict=True):
+            t = snap_to_frame(raw_t, record.frame_timestamps)
+            if t in already_planned:
+                continue
+            already_planned.add(t)
+            plan_text = self._generate_plan(record, spans, refresh_t=t, interjection=inter_text)
+            if plan_text is not None:
+                new_rows.append(
+                    {
+                        "role": "assistant",
+                        "content": plan_text,
+                        "style": "plan",
+                        "timestamp": t,
+                        "tool_calls": None,
+                    }
+                )
+        staging.write("plan", new_rows)
+
+    def _generate_subtasks(self, record: EpisodeRecord, *, task: str | None = None) -> list[dict[str, Any]]:
+        """Generate subtask spans, optionally via a multi-call quality chain.
+
+        Single call (default): watch video → emit subtask JSON.
+
+        Multi-call (opt-in, higher quality, more VLM calls):
+          1. ``subtask_describe_first`` — a grounding pass that narrates
+             ONLY what is visible (no JSON commitment to subtasks yet);
+             its description is injected into the segmentation prompt so
+             the model segments its own grounded observations instead of
+             pattern-matching the task text.
+          2. segmentation — emit subtask JSON (as before).
+        """
+        if record.row_count == 0 or not record.frame_timestamps:
+            return []
+        episode_duration = record.frame_timestamps[-1] - record.frame_timestamps[0]
+        effective_task = task if task is not None else record.episode_task
+
+        # ---- Auto-windowing (keeps the full sampling density) --------
+        # Contact sheets are cheap, but a whole long episode sampled at
+        # ``frames_per_second`` can still exceed ``max_frames_per_prompt``.
+        # When it does, split into consecutive windows of exactly that many
+        # frames (one describe→segment call each, still at the full sampling
+        # density), then merge + stitch — so an episode of any length is
+        # covered at full density rather than subsampled into one sparse call.
+        fps = max(1e-6, float(self.config.frames_per_second))
+        n_whole = int(round(episode_duration * fps)) + 1
+        if n_whole > self.config.max_frames_per_prompt:
+            window_s = self.config.max_frames_per_prompt / fps
+            return self._generate_subtasks_windowed(record, effective_task, window_s)
+
+        # ---- Pass 1 (optional): grounding description ----------------
+        observation_block = ""
+        if getattr(self.config, "subtask_describe_first", False):
+            description = self._describe_episode(record, effective_task)
+            if description:
+                observation_block = (
+                    "You watched this video and described, chronologically, "
+                    "ONLY what the robot actually does:\n"
+                    f'"""{description}"""\n\n'
+                    "Segment THAT grounded description (cross-checked against "
+                    "the video) into atomic subtasks. Do not introduce any "
+                    "action that is not in your description above.\n\n"
+                )
+
+        # ---- Pass 2: segmentation ------------------------------------
+        prompt = self._with_causal_rules(
+            load_prompt("plan_subtasks").format(
+                episode_task=effective_task,
+                min_subtask_seconds=self.config.min_subtask_seconds,
+                max_steps=self.config.plan_max_steps,
+                episode_duration=f"{episode_duration:.3f}",
+                observation_block=observation_block,
+            )
+        )
+        spans = self._vlm_field(self._video_message(record, prompt), "subtasks")
+        cleaned = self._clean_spans(spans, record)
+        if not cleaned:
+            return []
+
+        # ---- Full-episode coverage stitch ----------------------------
+        # The VLM can start after t0 or leave gaps, so frames fall through
+        # with no active subtask. Always stitch into a contiguous
+        # [t0, t_last] cover.
+        cleaned = self._stitch_full_coverage(cleaned, record)
+
+        return cleaned
+
+    def _generate_subtasks_windowed(
+        self, record: EpisodeRecord, task: str, window_s: float
+    ) -> list[dict[str, Any]]:
+        """Subtask generation in fixed-length windows at constant fps.
+
+        Splits ``[t0, t_last]`` into consecutive windows of ``window_s``
+        seconds, runs the describe -> segment chain on each window's own
+        frames (sampled at ``frames_per_second``), offsets
+        each window's spans back to absolute episode time, then merges +
+        stitches into a contiguous whole-episode cover.
+        """
+        t0 = float(record.frame_timestamps[0])
+        t_last = float(record.frame_timestamps[-1])
+        all_spans: list[dict[str, Any]] = []
+        w0 = t0
+        n_windows = 0
+        while w0 < t_last - 1e-6:
+            w1 = min(w0 + window_s, t_last)
+            all_spans.extend(self._subtasks_for_window(record, task, w0, w1))
+            n_windows += 1
+            w0 = w1
+        logger.info(
+            "episode %d: windowed subtask gen over %d window(s) of %.1fs -> %d raw spans",
+            record.episode_index,
+            n_windows,
+            window_s,
+            len(all_spans),
+        )
+        # Merge across windows: clamp to the absolute episode, sort, and
+        # frame-snap to distinct starts (handles any boundary collisions).
+        cleaned = self._clean_spans(all_spans, record)
+        if not cleaned:
+            return []
+        return self._stitch_full_coverage(cleaned, record)
+
+    def _subtasks_for_window(
+        self, record: EpisodeRecord, task: str, w0: float, w1: float
+    ) -> list[dict[str, Any]]:
+        """Run describe -> segment on one ``[w0, w1]`` window.
+
+        The model works in window-RELATIVE time ``[0, L]`` (it perceives
+        the window as a clip starting at 0); spans are offset back to
+        absolute ``[w0, w1]`` before returning.
+        """
+        window = (w0, w1)
+        win_len = max(0.0, w1 - w0)
+
+        observation_block = ""
+        if getattr(self.config, "subtask_describe_first", False):
+            description = self._describe_episode(record, task, window=window)
+            if description:
+                observation_block = (
+                    "You watched this video clip and described, chronologically, "
+                    "ONLY what the robot actually does:\n"
+                    f'"""{description}"""\n\n'
+                    "Segment THAT grounded description (cross-checked against "
+                    "the clip) into atomic subtasks. Do not introduce any "
+                    "action that is not in your description above.\n\n"
+                )
+
+        prompt = self._with_causal_rules(
+            load_prompt("plan_subtasks").format(
+                episode_task=task,
+                min_subtask_seconds=self.config.min_subtask_seconds,
+                max_steps=self.config.plan_max_steps,
+                episode_duration=f"{win_len:.3f}",
+                observation_block=observation_block,
+            )
+        )
+        spans = self._vlm_field(self._video_message(record, prompt, window=window), "subtasks")
+        # Window-relative clamp; no frame-snap dedupe yet (done on the
+        # merged absolute set).
+        cleaned = self._clean_spans(spans, record, bounds=(0.0, win_len), dedupe=False)
+        if not cleaned:
+            return []
+
+        # Offset window-relative spans back to absolute episode time.
+        for s in cleaned:
+            s["start"] = w0 + float(s["start"])
+            s["end"] = w0 + float(s["end"])
+        return cleaned
+
+    def _stitch_full_coverage(
+        self, spans: list[dict[str, Any]], record: EpisodeRecord
+    ) -> list[dict[str, Any]]:
+        """Make subtask spans tile the full episode with no gaps.
+
+        * The first subtask starts at the episode's first frame ``t0``
+          (any idle / approach before the first labelled action is folded
+          into it), so every early frame has an active subtask.
+        * Each subtask's ``end`` is snapped to the next subtask's
+          ``start`` (gaps between spans are closed), and the final
+          subtask's ``end`` extends to the last frame ``t_last``.
+
+        Starts are otherwise left as the (already frame-snapped, distinct)
+        values the VLM produced — only the FIRST start is pulled
+        back to ``t0``, which can't collide with a later span because it
+        was already the earliest. Purely deterministic; runs after the
+        VLM passes.
+        """
+        if not spans or not record.frame_timestamps:
+            return spans
+        t0 = float(record.frame_timestamps[0])
+        t_last = float(record.frame_timestamps[-1])
+        spans = sorted(spans, key=lambda s: float(s["start"]))
+        spans[0]["start"] = t0
+        for i in range(len(spans) - 1):
+            spans[i]["end"] = float(spans[i + 1]["start"])
+        spans[-1]["end"] = t_last
+        for s in spans:
+            if float(s["end"]) < float(s["start"]):
+                s["end"] = float(s["start"])
+        return spans
+
+    @staticmethod
+    def _with_causal_rules(prompt: str) -> str:
+        """Append the causal event-boundary rules to a describe/segment prompt."""
+        return f"{prompt}\n\n{_CAUSAL_BOUNDARY_RULES}"
+
+    def _clean_spans(
+        self,
+        spans: Any,
+        record: EpisodeRecord,
+        bounds: tuple[float, float] | None = None,
+        dedupe: bool = True,
+    ) -> list[dict[str, Any]]:
+        """Clamp / sort / (optionally) dedupe raw VLM subtask spans into valid rows.
+
+        ``bounds`` overrides the clamp range — pass the window's
+        ``(w_lo, w_hi)`` when cleaning window-relative spans, or leave
+        ``None`` to clamp to the whole episode ``[t0, t_last]``.
+        ``dedupe`` runs the frame-snap distinct-start step; skip it for
+        window-relative spans (frame snapping is done once on the merged,
+        absolute-time set).
+        """
+        if not spans:
+            return []
+        if bounds is not None:
+            lo, hi = float(bounds[0]), float(bounds[1])
+        else:
+            lo = record.frame_timestamps[0]
+            hi = record.frame_timestamps[-1]
+        cleaned: list[dict[str, Any]] = []
+        for span in spans:
+            try:
+                start = float(span["start"])
+                end = float(span["end"])
+                text = str(span["text"]).strip()
+            except (KeyError, ValueError, TypeError):
+                continue
+            start = max(lo, min(start, hi))
+            end = max(lo, min(end, hi))
+            if end < start:
+                start, end = end, start
+            if not text:
+                continue
+            cleaned.append({"text": text, "start": start, "end": end})
+        cleaned.sort(key=lambda s: s["start"])
+        if dedupe:
+            return self._dedupe_starts_to_distinct_frames(cleaned, record)
+        return cleaned
+
+    def _describe_episode(
+        self, record: EpisodeRecord, task: str, window: tuple[float, float] | None = None
+    ) -> str:
+        """Grounding pass: free-form chronological description of the (windowed) video."""
+        prompt = self._with_causal_rules(load_prompt("plan_subtask_describe").format(episode_task=task))
+        text = self._vlm_field(self._video_message(record, prompt, window=window), "description")
+        return text.strip() if isinstance(text, str) and text.strip() else ""
+
+    @staticmethod
+    def _dedupe_starts_to_distinct_frames(
+        spans: list[dict[str, Any]], record: EpisodeRecord
+    ) -> list[dict[str, Any]]:
+        """Bump same-frame subtask starts onto distinct frames.
+
+        Two consecutive VLM spans whose ``start`` rounds to the same
+        source frame (after :func:`snap_to_frame`) would otherwise emit
+        two ``style=subtask`` rows at the identical persistent
+        timestamp. The training-time renderer's ``active_at(t,
+        style=subtask)`` resolver can't disambiguate that and raises
+        ``Ambiguous resolver for style='subtask'``.
+
+        Walk the (sorted-by-start) spans, snap each to its frame, and
+        if the snapped frame is already taken push the span onto the
+        next unused frame so both subtasks survive on distinct
+        timestamps. If the episode ends before a free frame is found,
+        the trailing span is dropped with a warning — better than
+        poisoning the render.
+        """
+        if not spans:
+            return spans
+        frames = record.frame_timestamps
+        if not frames:
+            return spans
+        used: set[float] = set()
+        out: list[dict[str, Any]] = []
+        for span in spans:
+            ts = snap_to_frame(span["start"], frames)
+            if ts in used:
+                next_ts = next((f for f in frames if f > ts and f not in used), None)
+                if next_ts is None:
+                    logger.warning(
+                        "episode %d: subtask %r snapped to occupied frame "
+                        "%.3f and no free later frame exists — dropping",
+                        record.episode_index,
+                        span.get("text"),
+                        ts,
+                    )
+                    continue
+                ts = next_ts
+            used.add(ts)
+            new_span = {**span, "start": ts}
+            if float(new_span.get("end", ts)) < ts:
+                new_span["end"] = ts
+            out.append(new_span)
+        return out
+
+    def _generate_plan(
+        self,
+        record: EpisodeRecord,  # noqa: ARG002  (kept for signature stability)
+        subtask_spans: Sequence[dict[str, Any]],
+        *,
+        refresh_t: float | None = None,
+        interjection: str | None = None,  # noqa: ARG002
+        task: str | None = None,  # noqa: ARG002
+    ) -> str | None:
+        """Deterministic plan = numbered list of *still-todo* subtasks.
+
+        No VLM call: a plain numbered list keeps the plan aligned with the
+        upcoming subtasks (the old VLM "compact hierarchical plan" prompt
+        cost a round-trip per episode/refresh and could diverge).
+
+            1. <subtask 1>
+            2. <subtask 2>
+
+        On a refresh at ``refresh_t`` (from ``run_plan_updates`` on
+        interjections, and ``run_episode`` at each boundary), only subtasks
+        starting at or after ``refresh_t`` are included — so it always
+        describes what's left.
+        """
+        if not subtask_spans:
+            return None
+        remaining = [
+            s for s in subtask_spans if refresh_t is None or float(s.get("start", 0.0)) >= float(refresh_t)
+        ]
+        if not remaining:
+            # Past the last subtask boundary on a late refresh — nothing
+            # left to plan; emit None so the caller skips the row.
+            return None
+        return "\n".join(f"{i}. {span.get('text', '').strip()}" for i, span in enumerate(remaining, start=1))
+
+    def _generate_memory(
+        self,
+        record: EpisodeRecord,
+        prior_memory: str,
+        completed: str,
+        remaining: Sequence[str],
+        *,
+        task: str | None = None,
+    ) -> str:
+        prompt = load_prompt("plan_memory").format(
+            episode_task=(task if task is not None else record.episode_task),
+            prior_memory=prior_memory or "(none)",
+            completed_subtask=completed,
+            remaining_subtasks=", ".join(remaining) if remaining else "(none)",
+        )
+        memory = self._vlm_field(self._text_message(prompt), "memory")
+        return memory.strip() if isinstance(memory, str) else ""

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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