add trim_episode_start dataset edit operation

Add a dataset edit operation to trim the first N seconds from episodes while rebuilding frame and episode indices and metadata consistently. Skip episodes that are too short to trim and cover parsing plus metadata invariants with focused tests.

Made-with: Cursor

.github/workflows/full_tests.yml

@@ -173,8 +173,6 @@ jobs:
         shell: bash
         working-directory: /lerobot
     steps:
-      - name: Fix ptxas permissions
-        run: chmod +x /lerobot/.venv/lib/python3.10/site-packages/triton/backends/nvidia/bin/ptxas
       - name: Run pytest on GPU
         run: pytest tests -vv --maxfail=10
       - name: Run end-to-end tests

AI_POLICY.md

@@ -1,25 +0,0 @@
-# AI Usage Policy
-
-The LeRobot project welcomes contributions from everyone, and we have a few guidelines regarding AI usage to ensure high code quality, clear communication, and a healthy open-source ecosystem:
-
-- **Please disclose significant AI assistance.** If you used AI tools (e.g., Copilot, Claude, Cursor, ChatGPT) to generate a substantial portion of your code or text, let us know in your PR description. Transparency helps us review your changes more effectively.
-- **Own your code (The Human-in-the-Loop).** You must fully understand all the changes you are proposing. If you cannot explain what your AI-assisted code does or how it interacts with LeRobot's broader architecture, please take the time to learn and test it before submitting.
-- **Keep issues and discussions focused.** You are welcome to use AI to help draft issues or PR descriptions, but please review and edit them carefully before posting. AI can often be overly verbose; trimming the noise and getting straight to the point helps our maintainers address your needs faster.
-
-Our core maintainers also use AI tools to aid their workflows, but they do so while bringing deep contextual knowledge of the LeRobot codebase to validate the output. We ask all contributors to apply that same level of rigor.
-
-## Remember the Human Maintainers
-
-Please remember that LeRobot is maintained by a dedicated team of humans.
-
-Every discussion, issue, and pull request is read and reviewed by real people. While AI tools can generate thousands of lines of code in seconds, reviewing that code still takes human time and energy. Submitting unverified or low-effort AI output puts an unfair burden on our maintainers.
-
-Today, the quality of the AI output still heavily depends on the developer driving the tool. We ask that you respect our maintainers' time by thoroughly vetting, testing, and refining your submissions.
-
-## AI is Welcome Here
-
-LeRobot operates at the cutting edge of AI and robotics, and many of our maintainers actively embrace AI coding assistants as valuable productivity tools. We are a pro-AI project!
-
-Our reason for having an AI policy is not an anti-AI stance. Rather, it exists to ensure that AI is used to enhance human contributions, not replace them with unverified noise. It's about how the tools are used, not the tools themselves.
-
-We value the unique human insight you bring to the LeRobot community. Let AI empower your workflow, but always let your own judgment take the wheel.

CONTRIBUTING.md

@@ -2,7 +2,7 @@
 
 Everyone is welcome to contribute, and we value everybody's contribution. Code is not the only way to help the community. Answering questions, helping others, reaching out, and improving the documentation are immensely valuable.
 
-Whichever way you choose to contribute, please be mindful to respect our [code of conduct](./CODE_OF_CONDUCT.md) and our [AI policy](./AI_POLICY.md).
+Whichever way you choose to contribute, please be mindful to respect our [code of conduct](./CODE_OF_CONDUCT.md).
 
 ## Ways to Contribute
 

MANIFEST.in

@@ -1,3 +1,2 @@
 include src/lerobot/templates/lerobot_modelcard_template.md
 include src/lerobot/datasets/card_template.md
-include src/lerobot/envs/metaworld_config.json

benchmarks/video/README.md

@@ -28,9 +28,9 @@ We don't expect the same optimal settings for a dataset of images from a simulat
 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.
+- `aliberts/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.
+- `aliberts/paris_street`: (720 x 1280 pixels) real-world outdoor, moving camera.
-- `lerobot/kitchen`: (1080 x 1920 pixels) real-world indoor, fixed camera.
+- `aliberts/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.
 
@@ -179,7 +179,7 @@ python benchmark/video/run_video_benchmark.py \
     --output-dir outputs/video_benchmark \
     --repo-ids \
         lerobot/pusht_image \
-        lerobot/aloha_mobile_shrimp_image \
+        aliberts/aloha_mobile_shrimp_image \
     --vcodec libx264 libx265 \
     --pix-fmt yuv444p yuv420p \
     --g 2 20 None \
@@ -203,9 +203,9 @@ python benchmark/video/run_video_benchmark.py \
     --output-dir outputs/video_benchmark \
     --repo-ids \
         lerobot/pusht_image \
-        lerobot/aloha_mobile_shrimp_image \
+        aliberts/aloha_mobile_shrimp_image \
-        lerobot/paris_street \
+        aliberts/paris_street \
-        lerobot/kitchen \
+        aliberts/kitchen \
     --vcodec libx264 libx265 \
     --pix-fmt yuv444p yuv420p \
     --g 1 2 3 4 5 6 10 15 20 40 None \
@@ -221,9 +221,9 @@ python benchmark/video/run_video_benchmark.py \
     --output-dir outputs/video_benchmark \
     --repo-ids \
         lerobot/pusht_image \
-        lerobot/aloha_mobile_shrimp_image \
+        aliberts/aloha_mobile_shrimp_image \
-        lerobot/paris_street \
+        aliberts/paris_street \
-        lerobot/kitchen \
+        aliberts/kitchen \
     --vcodec libsvtav1 \
     --pix-fmt yuv420p \
     --g 1 2 3 4 5 6 10 15 20 40 None \
@@ -252,37 +252,37 @@ Since we're using av1 encoding, we're choosing the `pyav` decoder as `video_read
 
 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 |           |           |           |
+| video_images_size_ratio            | vcodec     | pix_fmt |           |           |           |
-| --------------------------------- | ---------- | ------- | --------- | --------- | --------- |
+| ---------------------------------- | ---------- | ------- | --------- | --------- | --------- |
-|                                   | libx264    |         | libx265   |           | libsvtav1 |
+|                                    | libx264    |         | libx265   |           | libsvtav1 |
-| repo_id                           | yuv420p    | yuv444p | yuv420p   | yuv444p   | yuv420p   |
+| repo_id                            | yuv420p    | yuv444p | yuv420p   | yuv444p   | yuv420p   |
-| lerobot/pusht_image               | **16.97%** | 17.58%  | 18.57%    | 18.86%    | 22.06%    |
+| 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%     |
+| aliberts/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%     |
+| aliberts/paris_street              | 2.12%      | 2.13%   | **1.54%** | **1.54%** | 4.43%     |
-| lerobot/kitchen                   | 1.40%      | 1.39%   | **1.00%** | **1.00%** | 2.52%     |
+| aliberts/kitchen                   | 1.40%      | 1.39%   | **1.00%** | **1.00%** | 2.52%     |
 
-| video_images_load_time_ratio      | vcodec  | pix_fmt |          |         |           |
+| video_images_load_time_ratio       | vcodec  | pix_fmt |          |         |           |
-| --------------------------------- | ------- | ------- | -------- | ------- | --------- |
+| ---------------------------------- | ------- | ------- | -------- | ------- | --------- |
-|                                   | libx264 |         | libx265  |         | libsvtav1 |
+|                                    | libx264 |         | libx265  |         | libsvtav1 |
-| repo_id                           | yuv420p | yuv444p | yuv420p  | yuv444p | yuv420p   |
+| repo_id                            | yuv420p | yuv444p | yuv420p  | yuv444p | yuv420p   |
-| lerobot/pusht_image               | 6.45    | 5.19    | **1.90** | 2.12    | 2.47      |
+| 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**  |
+| aliberts/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**  |
+| aliberts/paris_street              | 2.21    | 2.05    | 0.36     | 0.49    | **0.30**  |
-| lerobot/kitchen                   | 1.46    | 1.46    | 0.28     | 0.51    | **0.26**  |
+| aliberts/kitchen                   | 1.46    | 1.46    | 0.28     | 0.51    | **0.26**  |
 
-|                                   |          | vcodec   | pix_fmt      |          |           |              |
+|                                    |          | vcodec   | pix_fmt      |          |           |              |
-| --------------------------------- | -------- | -------- | ------------ | -------- | --------- | ------------ |
+| ---------------------------------- | -------- | -------- | ------------ | -------- | --------- | ------------ |
-|                                   |          | libx264  |              | libx265  |           | libsvtav1    |
+|                                    |          | libx264  |              | libx265  |           | libsvtav1    |
-| repo_id                           | metric   | yuv420p  | yuv444p      | yuv420p  | yuv444p   | yuv420p      |
+| 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     |
+| 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_psnr | 35.44    | 37.07        | 35.49    | **37.30** | 37.20        |
-|                                   | avg_ssim | 98.28%   | **98.85%**   | 98.31%   | 98.84%    | 98.72%       |
+|                                    | 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** |
+| aliberts/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_psnr | 35.91    | 36.21        | 35.88    | 36.09     | **40.17**    |
-|                                   | avg_ssim | 95.19%   | 95.18%       | 95.00%   | 95.05%    | **97.73%**   |
+|                                    | 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** |
+| aliberts/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_psnr | 33.48    | 33.68        | 32.05    | 32.15     | **35.40**    |
-|                                   | avg_ssim | 93.76%   | 93.75%       | 89.46%   | 89.46%    | **95.46%**   |
+|                                    | 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** |
+| aliberts/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_psnr | 36.73    | 37.33        | 36.56    | 36.75     | **39.12**    |
-|                                   | avg_ssim | 95.47%   | 95.58%       | 95.52%   | 95.53%    | **96.82%**   |
+|                                    | avg_ssim | 95.47%   | 95.58%       | 95.52%   | 95.53%    | **96.82%**   |

docker/Dockerfile.internal

@@ -85,8 +85,6 @@ RUN if [ "$UNBOUND_DEPS" = "true" ]; then \
 
 RUN uv pip install --no-cache ".[all]"
 
-RUN chmod +x /lerobot/.venv/lib/python${PYTHON_VERSION}/site-packages/triton/backends/nvidia/bin/ptxas
-
 # Copy the rest of the application source code
 # Make sure to have the git-LFS files for testing
 COPY --chown=user_lerobot:user_lerobot . .

docs/source/_toctree.yml

@@ -29,8 +29,6 @@
     title: Using the Dataset Tools
   - local: dataset_subtask
     title: Using Subtasks in the Dataset
-  - local: streaming_video_encoding
-    title: Streaming Video Encoding
   title: "Datasets"
 - sections:
   - local: act

docs/source/act.mdx

@@ -88,8 +88,5 @@ lerobot-record \
   --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
 ```

docs/source/async.mdx

@@ -48,7 +48,7 @@ python -m lerobot.async_inference.robot_client \
     --task="dummy" \ # POLICY: The task to run the policy on (`Fold my t-shirt`). Not necessarily defined for all policies, such as `act`
     --policy_type=your_policy_type \ # POLICY: the type of policy to run (smolvla, act, etc)
     --pretrained_name_or_path=user/model \ # POLICY: the model name/path on server to the checkpoint to run (e.g., lerobot/smolvla_base)
-    --policy_device=mps \ # POLICY: the device to run the policy on, on the server (cuda, mps, xpu, cpu)
+    --policy_device=mps \ # POLICY: the device to run the policy on, on the server
     --actions_per_chunk=50 \ # POLICY: the number of actions to output at once
     --chunk_size_threshold=0.5 \ # CLIENT: the threshold for the chunk size before sending a new observation to the server
     --aggregate_fn_name=weighted_average \ # CLIENT: the function to aggregate actions on overlapping portions

docs/source/earthrover_mini_plus.mdx

@@ -170,13 +170,13 @@ Once you can drive the robot well, you can start recording data to train AI mode
 We use Hugging Face to store your data online. First, log in with your token from [Hugging Face settings](https://huggingface.co/settings/tokens):
 
 ```bash
-hf auth login --token ${HUGGINGFACE_TOKEN} --add-to-git-credential
+huggingface-cli login --token ${HUGGINGFACE_TOKEN} --add-to-git-credential
 ```
 
 Store your Hugging Face username:
 
 ```bash
-HF_USER=$(hf auth whoami | awk -F': *' 'NR==1 {print $2}')
+HF_USER=$(huggingface-cli whoami | head -n 1)
 echo $HF_USER
 ```
 
@@ -185,16 +185,13 @@ echo $HF_USER
 Use the standard recording command:
 
 ```bash
-lerobot-record \
+python src/lerobot/scripts/lerobot_record.py \
     --robot.type=earthrover_mini_plus \
     --teleop.type=keyboard_rover \
     --dataset.repo_id=your_username/dataset_name \
     --dataset.num_episodes=2 \
     --dataset.fps=10 \
     --dataset.single_task="Navigate around obstacles" \
-    --dataset.streaming_encoding=true \
-    --dataset.encoder_threads=2 \
-    # --dataset.vcodec=auto \
     --display_data=true
 ```
 

docs/source/envhub.mdx

@@ -155,10 +155,10 @@ Upload your repository to Hugging Face:
 pip install huggingface_hub
 
 # Login to Hugging Face
-hf auth login
+huggingface-cli login
 
 # Create a new repository
-hf repo create my-org/my-custom-env
+huggingface-cli repo create my-custom-env --type space --org my-org
 
 # Initialize git and push
 git init

docs/source/groot.mdx

@@ -120,12 +120,9 @@ lerobot-record \
   --display_data=true \
   --dataset.repo_id=<user>/eval_groot-bimanual  \
   --dataset.num_episodes=10 \
-  --dataset.single_task="Grab and handover the red cube to the other arm" \
+  --dataset.single_task="Grab and handover the red cube to the other arm"
-  --dataset.streaming_encoding=true \
+  --policy.path=<user>/groot-bimanual # your trained model
-  --dataset.encoder_threads=2 \
+  --dataset.episode_time_s=30
-  # --dataset.vcodec=auto \
-  --policy.path=<user>/groot-bimanual \ # your trained model
-  --dataset.episode_time_s=30 \
   --dataset.reset_time_s=10
 ```
 

docs/source/hope_jr.mdx

@@ -224,15 +224,12 @@ lerobot-record \
     --teleop.port=/dev/tty.usbmodem1201 \
     --teleop.id=right \
     --teleop.side=right \
-    --dataset.repo_id=<USER>/hand_record_test_with_video_data \
+    --dataset.repo_id=nepyope/hand_record_test_with_video_data \
     --dataset.single_task="Hand recording test with video data" \
     --dataset.num_episodes=1 \
     --dataset.episode_time_s=5 \
     --dataset.push_to_hub=true \
     --dataset.private=true \
-    --dataset.streaming_encoding=true \
-    --dataset.encoder_threads=2 \
-    # --dataset.vcodec=auto \
     --display_data=true
 ```
 
@@ -244,7 +241,7 @@ lerobot-replay \
     --robot.port=/dev/tty.usbmodem58760432281 \
     --robot.id=right \
     --robot.side=right \
-    --dataset.repo_id=<USER>/hand_record_test_with_camera \
+    --dataset.repo_id=nepyope/hand_record_test_with_camera \
     --dataset.episode=0
 ```
 
@@ -252,13 +249,13 @@ lerobot-replay \
 
 ```bash
 lerobot-train \
-  --dataset.repo_id=<USER>/hand_record_test_with_video_data \
+  --dataset.repo_id=nepyope/hand_record_test_with_video_data \
   --policy.type=act \
   --output_dir=outputs/train/hopejr_hand \
   --job_name=hopejr \
   --policy.device=mps \
   --wandb.enable=true \
-  --policy.repo_id=<USER>/hand_test_policy
+  --policy.repo_id=nepyope/hand_test_policy
 ```
 
 ### Evaluate
@@ -273,11 +270,8 @@ lerobot-record \
   --robot.side=right \
   --robot.cameras='{"main": {"type": "opencv", "index_or_path": 0, "width": 640, "height": 480, "fps": 30}}' \
   --display_data=false \
-  --dataset.repo_id=<USER>/eval_hopejr \
+  --dataset.repo_id=nepyope/eval_hopejr \
   --dataset.single_task="Evaluate hopejr hand policy" \
   --dataset.num_episodes=10 \
-  --dataset.streaming_encoding=true \
-  --dataset.encoder_threads=2 \
-  # --dataset.vcodec=auto \
   --policy.path=outputs/train/hopejr_hand/checkpoints/last/pretrained_model
 ```

docs/source/il_robots.mdx

@@ -159,13 +159,13 @@ We use the Hugging Face hub features for uploading your dataset. If you haven't
 Add your token to the CLI by running this command:
 
 ```bash
-hf auth login --token ${HUGGINGFACE_TOKEN} --add-to-git-credential
+huggingface-cli login --token ${HUGGINGFACE_TOKEN} --add-to-git-credential
 ```
 
 Then store your Hugging Face repository name in a variable:
 
 ```bash
-HF_USER=$(hf auth whoami | awk -F': *' 'NR==1 {print $2}')
+HF_USER=$(hf auth whoami | head -n 1)
 echo $HF_USER
 ```
 
@@ -185,10 +185,7 @@ lerobot-record \
     --display_data=true \
     --dataset.repo_id=${HF_USER}/record-test \
     --dataset.num_episodes=5 \
-    --dataset.single_task="Grab the black cube" \
+    --dataset.single_task="Grab the black cube"
-    --dataset.streaming_encoding=true \
-    # --dataset.vcodec=auto \
-    --dataset.encoder_threads=2
 ```
 </hfoption>
 <hfoption id="API example">
@@ -327,7 +324,7 @@ You can look for other LeRobot datasets on the hub by searching for `LeRobot` [t
 You can also push your local dataset to the Hub manually, running:
 
 ```bash
-hf upload ${HF_USER}/record-test ~/.cache/huggingface/lerobot/{repo-id} --repo-type dataset
+huggingface-cli upload ${HF_USER}/record-test ~/.cache/huggingface/lerobot/{repo-id} --repo-type dataset
 ```
 
 #### Record function
@@ -491,7 +488,7 @@ If your local computer doesn't have a powerful GPU you could utilize Google Cola
 Once training is done, upload the latest checkpoint with:
 
 ```bash
-hf upload ${HF_USER}/act_so101_test \
+huggingface-cli upload ${HF_USER}/act_so101_test \
   outputs/train/act_so101_test/checkpoints/last/pretrained_model
 ```
 
@@ -499,7 +496,7 @@ You can also upload intermediate checkpoints with:
 
 ```bash
 CKPT=010000
-hf upload ${HF_USER}/act_so101_test${CKPT} \
+huggingface-cli upload ${HF_USER}/act_so101_test${CKPT} \
   outputs/train/act_so101_test/checkpoints/${CKPT}/pretrained_model
 ```
 
@@ -518,9 +515,6 @@ lerobot-record  \
   --display_data=false \
   --dataset.repo_id=${HF_USER}/eval_so100 \
   --dataset.single_task="Put lego brick into the transparent box" \
-  --dataset.streaming_encoding=true \
-  --dataset.encoder_threads=2 \
-  # --dataset.vcodec=auto \
   # <- Teleop optional if you want to teleoperate in between episodes \
   # --teleop.type=so100_leader \
   # --teleop.port=/dev/ttyACM0 \

docs/source/installation.mdx

@@ -40,13 +40,6 @@ conda install ffmpeg -c conda-forge
 >
 > - _[On Linux only]_ If you want to bring your own ffmpeg: Install [ffmpeg build dependencies](https://trac.ffmpeg.org/wiki/CompilationGuide/Ubuntu#GettheDependencies) and [compile ffmpeg from source with libsvtav1](https://trac.ffmpeg.org/wiki/CompilationGuide/Ubuntu#libsvtav1), and make sure you use the corresponding ffmpeg binary to your install with `which ffmpeg`.
 
-> [!NOTE]
-> When installing LeRobot inside WSL (Windows Subsystem for Linux), make sure to install `evdev` with the following command:
->
-> ```bash
-> conda install evdev -c conda-forge
-> ```
-
 ## Step 3: Install LeRobot 🤗
 
 ### From Source

docs/source/lekiwi.mdx

@@ -279,13 +279,13 @@ We use the Hugging Face hub features for uploading your dataset. If you haven't
 Add your token to the CLI by running this command:
 
 ```bash
-hf auth login --token ${HUGGINGFACE_TOKEN} --add-to-git-credential
+huggingface-cli login --token ${HUGGINGFACE_TOKEN} --add-to-git-credential
 ```
 
 Then store your Hugging Face repository name in a variable:
 
 ```bash
-HF_USER=$(hf auth whoami | awk -F': *' 'NR==1 {print $2}')
+HF_USER=$(huggingface-cli whoami | head -n 1)
 echo $HF_USER
 ```
 

docs/source/lerobot-dataset-v3.mdx

@@ -41,10 +41,7 @@ lerobot-record \
   --display_data=true \
   --dataset.repo_id=${HF_USER}/record-test \
   --dataset.num_episodes=5 \
-  --dataset.single_task="Grab the black cube" \
+  --dataset.single_task="Grab the black cube"
-  --dataset.streaming_encoding=true \
-  # --dataset.vcodec=auto \
-  --dataset.encoder_threads=2
 ```
 
 See the [recording guide](./il_robots#record-a-dataset) for more details.

docs/source/phone_teleop.mdx

@@ -66,13 +66,12 @@ Run on of the examples scripts to teleoperate, record a dataset, replay a datase
 
 All scripts assume you configured your robot (e.g., SO-100 follower) and set the correct serial port.
 
-Additionally you need to **copy the URDF of the robot into the examples folder**. For the examples in this tutorial (using SO100/SO101), copy the `SO101` folder from the [SO-ARM100 repo](https://github.com/TheRobotStudio/SO-ARM100/blob/main/Simulation/SO101) into the `examples/phone_to_so100/` directory, so that the URDF file path becomes `examples/phone_to_so100/SO101/so101_new_calib.urdf`.
+Additionally you need to **copy the urdf of the robot to the examples folder**. For the examples in this tutorial (Using SO100/SO101) it is highly recommended to use the urdf in the [SO-ARM100 repo](https://github.com/TheRobotStudio/SO-ARM100/blob/main/Simulation/SO101/so101_new_calib.urdf)
 
 - Run this example to teleoperate:
 
   ```bash
-  cd examples/phone_to_so100
+  python examples/phone_to_so100/teleoperate.py
-  python teleoperate.py
   ```
 
 After running the example:
@@ -85,22 +84,19 @@ Additionally you can customize mapping or safety limits by editing the processor
 - Run this example to record a dataset, which saves absolute end effector observations and actions:
 
   ```bash
-  cd examples/phone_to_so100
+  python examples/phone_to_so100/record.py
-  python record.py
   ```
 
 - Run this example to replay recorded episodes:
 
   ```bash
-  cd examples/phone_to_so100
+  python examples/phone_to_so100/replay.py
-  python replay.py
   ```
 
 - Run this example to evaluate a pretrained policy:
 
   ```bash
-  cd examples/phone_to_so100
+  python examples/phone_to_so100/evaluate.py
-  python evaluate.py
   ```
 
 ### Important pipeline steps and options

docs/source/pi0.mdx

@@ -60,7 +60,7 @@ policy.type=pi0
 For training π₀, you can use the standard LeRobot training script with the appropriate configuration:
 
 ```bash
-lerobot-train \
+python src/lerobot/scripts/lerobot_train.py \
     --dataset.repo_id=your_dataset \
     --policy.type=pi0 \
     --output_dir=./outputs/pi0_training \

docs/source/pi05.mdx

@@ -56,7 +56,7 @@ policy.type=pi05
 Here's a complete training command for finetuning the base π₀.₅ model on your own dataset:
 
 ```bash
-lerobot-train \
+python src/lerobot/scripts/lerobot_train.py\
     --dataset.repo_id=your_dataset \
     --policy.type=pi05 \
     --output_dir=./outputs/pi05_training \

docs/source/reachy2.mdx

@@ -159,9 +159,6 @@ lerobot-record \
     --dataset.fps=15 \
     --dataset.push_to_hub=true \
     --dataset.private=true \
-    --dataset.streaming_encoding=true \
-    --dataset.encoder_threads=2 \
-    # --dataset.vcodec=auto \
     --display_data=true
 ```
 
@@ -201,9 +198,6 @@ lerobot-record \
     --dataset.fps=15 \
     --dataset.push_to_hub=true \
     --dataset.private=true \
-    --dataset.streaming_encoding=true \
-    --dataset.encoder_threads=2 \
-    # --dataset.vcodec=auto \
     --display_data=true
 ```
 

docs/source/sarm.mdx

@@ -269,7 +269,7 @@ This generates visualizations showing video frames with subtask boundaries overl
 Train with **no annotations** - uses linear progress from 0 to 1:
 
 ```bash
-lerobot-train \
+python src/lerobot/scripts/lerobot_train.py \
   --dataset.repo_id=your-username/your-dataset \
   --policy.type=sarm \
   --policy.annotation_mode=single_stage \
@@ -288,7 +288,7 @@ lerobot-train \
 Train with **dense annotations only** (sparse auto-generated):
 
 ```bash
-lerobot-train \
+python src/lerobot/scripts/lerobot_train.py \
   --dataset.repo_id=your-username/your-dataset \
   --policy.type=sarm \
   --policy.annotation_mode=dense_only \
@@ -307,7 +307,7 @@ lerobot-train \
 Train with **both sparse and dense annotations**:
 
 ```bash
-lerobot-train \
+python src/lerobot/scripts/lerobot_train.py \
   --dataset.repo_id=your-username/your-dataset \
   --policy.type=sarm \
   --policy.annotation_mode=dual \
@@ -468,7 +468,7 @@ This script:
 Once you have the progress file, train your policy with RA-BC weighting. The progress file is auto-detected from the dataset path (`sarm_progress.parquet`). Currently PI0, PI0.5 and SmolVLA are supported with RA-BC:
 
 ```bash
-lerobot-train \
+python src/lerobot/scripts/lerobot_train.py \
   --dataset.repo_id=your-username/your-dataset \
   --policy.type=pi0 \
   --use_rabc=true \

docs/source/smolvla.mdx

@@ -106,9 +106,6 @@ lerobot-record \
   --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 \
   # <- Teleop optional if you want to teleoperate in between episodes \
   # --teleop.type=so100_leader \
   # --teleop.port=/dev/ttyACM0 \

docs/source/streaming_video_encoding.mdx

@@ -1,155 +0,0 @@
-# Streaming Video Encoding Guide
-
-## 1. Overview
-
-Streaming video encoding eliminates the traditional PNG round-trip during video dataset recording. Instead of:
-
-1. Capture frame -> write PNG to disk -> (at episode end) read PNG's -> encode to MP4 -> delete PNG's
-
-Frames can be encoded in real-time during capture:
-
-1. Capture frame -> queue to encoder thread -> encode to MP4 directly
-
-This makes `save_episode()` near-instant (the video is already encoded by the time the episode ends) and removes the blocking wait that previously occurred between episodes, especially with multiple cameras in long episodes.
-
-## 2. Tuning Parameters
-
-| 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                     |
-| `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       |
-
-## 3. Performance Considerations
-
-Streaming encoding means the CPU is encoding video **during** the capture loop, not after. This creates a CPU budget that must be shared between:
-
-- **Control loop** (reading cameras, control the robot, writing non-video data)
-- **Encoder threads** (one pool per camera)
-- **Rerun visualization** (if enabled)
-- **OS and other processes**
-
-### Resolution & Number of Cameras Impact
-
-| Setup                     | Throughput (px/sec) | CPU Encoding Load | Notes                          |
-| ------------------------- | ------------------- | ----------------- | ------------------------------ |
-| 2camsx 640x480x3 @30fps   | 55M                 | Low               | Works on most systems          |
-| 2camsx 1280x720x3 @30fps  | 165M                | Moderate          | Comfortable on modern systems  |
-| 2camsx 1920x1080x3 @30fps | 373M                | High              | Requires powerful high-end CPU |
-
-### `encoder_threads` Tuning
-
-This parameter controls how many threads each encoder instance uses internally:
-
-- **Higher values** (e.g., 4-5): Faster encoding, but uses more CPU cores per camera. Good for high-end systems with many cores.
-- **Lower values** (e.g., 1-2): Less CPU per camera, freeing cores for capture and visualization. Good for low-res images and capable CPUs.
-- **`None` (default)**: Lets the codec decide. Information available in the codec logs.
-
-### Backpressure and Frame Dropping
-
-Each camera has a bounded queue (`encoder_queue_maxsize`, default 60 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
-3. A warning is logged: `"Encoder queue full for {camera}, dropped N frame(s)"`
-4. At episode end, total dropped frames per camera are reported
-
-### Symptoms of Encoder Falling Behind
-
-- **System feels laggy and freezes**: all CPUs are at 100%
-- **Dropped frame warnings** in the log or lower frames/FPS than expected in the recorded dataset
-- **Choppy robot movement**: If CPU is severely overloaded, even the capture loop may be affected
-- **Accumulated rerun lag**: Visualization falls behind real-time
-
-## 4. Hardware-Accelerated Encoding
-
-### When to Use
-
-Use HW encoding when:
-
-- CPU is the bottleneck (dropped frames, choppy robot, rerun lag)
-- You have compatible hardware (GPU or dedicated encoder)
-- You're recording at high throughput (high resolution or with many cameras)
-
-### Choosing a Codec
-
-| Codec                 | CPU Usage | File Size      | Quality | Notes                                                            |
-| --------------------- | --------- | -------------- | ------- | ---------------------------------------------------------------- |
-| `libsvtav1` (default) | High      | Smallest       | Best    | Default. Best compression but most CPU-intensive                 |
-| `h264`                | Medium    | ~30-50% larger | Good    | Software H.264. Lower CPU                                        |
-| HW encoders           | Very Low  | Largest        | Good    | Offloads to dedicated hardware. Best for CPU-constrained systems |
-
-### 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`              |
-
-> [!NOTE]
-> In order to use the HW accelerated encoders you might need to upgrade your GPU drivers.
-
-> [!NOTE]
-> `libsvtav1` is the default because it provides the best training performance; other vcodecs can reduce CPU usage and be faster, but they typically produce larger files and may affect training time.
-
-## 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.                                   |
-
-## 6. Recommended Configurations
-
-These estimates are conservative; we recommend testing them on your setup—start with a low load and increase it gradually.
-
-### High-End Systems: modern 12+ cores (24+ threads)
-
-A throughput between ~250-500M px/sec should be comfortable in CPU. For even better results try HW encoding if available.
-
-```bash
-# 3camsx 1280x720x3 @30fps: Defaults work well. Optionally increase encoder parallelism.
-# 2camsx 1920x1080x3 @30fps: Defaults work well. Optionally increase encoder parallelism.
-lerobot-record --dataset.encoder_threads=5 ...
-
-# 3camsx 1920x1080x3 @30fps: Might require some tuning.
-```
-
-### Mid-Range Systems: modern 8+ cores (16+ threads) or Apple Silicon
-
-A throughput between ~80-300M px/sec should be possible in CPU.
-
-```bash
-# 3camsx 640x480x3 @30fps: Defaults work well. Optionally decrease encoder parallelism.
-# 2camsx 1280x720x3 @30fps: Defaults work well. Optionally decrease encoder parallelism.
-lerobot-record --dataset.encoder_threads=2 ...
-
-# 2camsx 1920x1080x3 @30fps: Might require some tuning.
-```
-
-### Low-Resource Systems: modern 4+ cores (8+ threads) or Raspberry Pi 5
-
-On very constrained systems, streaming encoding may compete too heavily with the capture loop. Disabling it falls back to the PNG-based approach where encoding happens between episodes (blocking, but doesn't interfere with capture). Alternatively, record at a lower throughput to reduce both capture and encoding load. Consider also changing codec to `h264` and using batch encoding.
-
-```bash
-# 2camsx 640x480x3 @30fps: Requires some tuning.
-
-# Use H.264, disable streaming, consider batching encoding
-lerobot-record --dataset.vcodec=h264 --dataset.streaming_encoding=false ...
-```
-
-## 7. Closing note
-
-Performance ultimately depends on your exact setup — frames-per-second, resolution, CPU cores and load, available memory, episode length, and the encoder you choose. Always test with your target workload, be mindful about your CPU & system capabilities and tune `encoder_threads`, `encoder_queue_maxsize`, and
-`vcodec` reasonably. That said, a common practical configuration (for many applications) is three cameras at 640×480x3 @30fps; this usually runs fine with the default streaming video encoding settings in modern systems. Always verify your recorded dataset is healthy by comparing the video duration to the CLI episode duration and confirming the row count equals FPS × CLI duration.

docs/source/unitree_g1.mdx

@@ -216,7 +216,7 @@ lerobot-teleoperate \
 ### Record Dataset in Simulation
 
 ```bash
-lerobot-record \
+python -m lerobot.scripts.lerobot_record \
     --robot.type=unitree_g1 \
     --robot.is_simulation=true \
     --robot.cameras='{"global_view": {"type": "zmq", "server_address": "localhost", "port": 5555, "camera_name": "head_camera", "width": 640, "height": 480, "fps": 30}}' \
@@ -229,10 +229,7 @@ lerobot-record \
     --dataset.num_episodes=2 \
     --dataset.episode_time_s=5 \
     --dataset.reset_time_s=5 \
-    --dataset.push_to_hub=true \
+    --dataset.push_to_hub=true
-    --dataset.streaming_encoding=true \
-    # --dataset.vcodec=auto \
-    --dataset.encoder_threads=2
 ```
 
 Example simulation dataset: [nepyope/teleop_test_sim](https://huggingface.co/datasets/nepyope/teleop_test_sim)
@@ -269,7 +266,7 @@ lerobot-teleoperate \
 ### Record Dataset on Real Robot
 
 ```bash
-lerobot-record \
+python -m lerobot.scripts.lerobot_record \
     --robot.type=unitree_g1 \
     --robot.is_simulation=false \
     --robot.cameras='{"global_view": {"type": "zmq", "server_address": "172.18.129.215", "port": 5555, "camera_name": "head_camera", "width": 640, "height": 480, "fps": 30}}' \
@@ -282,10 +279,7 @@ lerobot-record \
     --dataset.num_episodes=2 \
     --dataset.episode_time_s=5 \
     --dataset.reset_time_s=5 \
-    --dataset.push_to_hub=true \
+    --dataset.push_to_hub=true
-    --dataset.streaming_encoding=true \
-    # --dataset.vcodec=auto \
-    --dataset.encoder_threads=2
 ```
 
 **Note**: Update `server_address` to match your robot's camera server IP.

docs/source/using_dataset_tools.mdx

@@ -12,7 +12,6 @@ LeRobot provides several utilities for manipulating datasets:
 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.
 
 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`.
@@ -157,30 +156,6 @@ lerobot-edit-dataset \
 
 **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.
 
-### Show the information of datasets
-
-Show the information of datasets such as number of episode, number of frame, File size and so on.
-No change will be made to the dataset
-
-```bash
-
-# Show dataset information without feature details
-lerobot-edit-dataset \
-    --repo_id lerobot/pusht_image \
-    --operation.type info \
-
-# Show dataset information with feature details
-lerobot-edit-dataset \
-    --repo_id lerobot/pusht_image \
-    --operation.type info \
-    --operation.show_features true
-
-```
-
-**Parameters:**
-
-- `parameters`: The flag to control show or no show dataset information with feature details.(default=false)
-
 ### Push to Hub
 
 Add the `--push_to_hub true` flag to any command to automatically upload the resulting dataset to the Hugging Face Hub:

docs/source/walloss.mdx

@@ -45,7 +45,7 @@ policy.type=wall_x
 For training WallX, you can use the standard LeRobot training script with the appropriate configuration:
 
 ```bash
-lerobot-train \
+python src/lerobot/scripts/lerobot_train.py \
     --dataset.repo_id=your_dataset \
     --policy.type=wall_x \
     --output_dir=./outputs/wallx_training \

docs/source/xvla.mdx

@@ -154,7 +154,7 @@ lerobot-train \
 
 ```bash
 lerobot-train \
-  --dataset.repo_id=<USER>/bimanual-so100-handover-cube \
+  --dataset.repo_id=pepijn223/bimanual-so100-handover-cube \
   --output_dir=./outputs/xvla_bimanual \
   --job_name=xvla_so101_training \
   --policy.path="lerobot/xvla-base" \

examples/backward_compatibility/replay.py

@@ -22,7 +22,7 @@ lerobot-replay \
     --robot.type=so100_follower \
     --robot.port=/dev/tty.usbmodem58760431541 \
     --robot.id=black \
-    --dataset.repo_id=<USER>/record-test \
+    --dataset.repo_id=aliberts/record-test \
     --dataset.episode=2
 ```
 """
@@ -57,7 +57,7 @@ class DatasetReplayConfig:
     repo_id: str
     # Episode to replay.
     episode: int
-    # Root directory where the dataset will be stored (e.g. 'dataset/path'). If None, defaults to $HF_LEROBOT_HOME/repo_id.
+    # Root directory where the dataset will be stored (e.g. 'dataset/path').
     root: str | Path | None = None
     # Limit the frames per second. By default, uses the policy fps.
     fps: int = 30

examples/rtc/eval_dataset.py

@@ -27,8 +27,8 @@ measuring consistency and ground truth alignment.
 Usage:
     # Basic usage with smolvla policy
     uv run python examples/rtc/eval_dataset.py \
-        --policy.path=<USER>/smolvla_check_rtc_last3 \
+        --policy.path=helper2424/smolvla_check_rtc_last3 \
-        --dataset.repo_id=<USER>/check_rtc \
+        --dataset.repo_id=helper2424/check_rtc \
         --rtc.execution_horizon=8 \
         --device=mps \
         --rtc.max_guidance_weight=10.0 \
@@ -58,16 +58,16 @@ Usage:
         --device=cuda
 
     uv run python examples/rtc/eval_dataset.py \
-        --policy.path=<USER>/reuben_pi0 \
+        --policy.path=lipsop/reuben_pi0 \
-        --dataset.repo_id=<USER>/so101_cube_in_cup \
+        --dataset.repo_id=ReubenLim/so101_cube_in_cup \
         --rtc.execution_horizon=8 \
         --device=cuda
 
     # With torch.compile for faster inference (PyTorch 2.0+)
     # Note: CUDA graphs disabled by default due to in-place ops in denoising loop
     uv run python examples/rtc/eval_dataset.py \
-        --policy.path=<USER>/smolvla_check_rtc_last3 \
+        --policy.path=helper2424/smolvla_check_rtc_last3 \
-        --dataset.repo_id=<USER>/check_rtc \
+        --dataset.repo_id=helper2424/check_rtc \
         --rtc.execution_horizon=8 \
         --device=mps \
         --use_torch_compile=true \
@@ -75,8 +75,8 @@ Usage:
 
     # With torch.compile on CUDA (CUDA graphs disabled by default)
     uv run python examples/rtc/eval_dataset.py \
-        --policy.path=<USER>/smolvla_check_rtc_last3 \
+        --policy.path=helper2424/smolvla_check_rtc_last3 \
-        --dataset.repo_id=<USER>/check_rtc \
+        --dataset.repo_id=helper2424/check_rtc \
         --rtc.execution_horizon=8 \
         --device=cuda \
         --use_torch_compile=true \
@@ -84,8 +84,8 @@ Usage:
 
     # Enable CUDA graphs (advanced - may cause tensor aliasing errors)
     uv run python examples/rtc/eval_dataset.py \
-        --policy.path=<USER>/smolvla_check_rtc_last3 \
+        --policy.path=helper2424/smolvla_check_rtc_last3 \
-        --dataset.repo_id=<USER>/check_rtc \
+        --dataset.repo_id=helper2424/check_rtc \
         --use_torch_compile=true \
         --torch_compile_backend=inductor \
         --torch_compile_mode=max-autotune \

examples/rtc/eval_with_real_robot.py

@@ -28,7 +28,7 @@ For simulation environments, see eval_with_simulation.py
 Usage:
     # Run RTC with Real robot with RTC
     uv run examples/rtc/eval_with_real_robot.py \
-        --policy.path=<USER>/smolvla_check_rtc_last3 \
+        --policy.path=helper2424/smolvla_check_rtc_last3 \
         --policy.device=mps \
         --rtc.enabled=true \
         --rtc.execution_horizon=20 \
@@ -41,7 +41,7 @@ Usage:
 
     # Run RTC with Real robot without RTC
     uv run examples/rtc/eval_with_real_robot.py \
-        --policy.path=<USER>/smolvla_check_rtc_last3 \
+        --policy.path=helper2424/smolvla_check_rtc_last3 \
         --policy.device=mps \
         --rtc.enabled=false \
         --robot.type=so100_follower \
@@ -53,7 +53,7 @@ Usage:
 
     # Run RTC with Real robot with pi0.5 policy
     uv run examples/rtc/eval_with_real_robot.py \
-        --policy.path=<USER>/pi05_check_rtc \
+        --policy.path=helper2424/pi05_check_rtc \
         --policy.device=mps \
         --rtc.enabled=true \
         --rtc.execution_horizon=20 \

pyproject.toml

@@ -25,7 +25,7 @@ discord = "https://discord.gg/s3KuuzsPFb"
 
 [project]
 name = "lerobot"
-version = "0.4.5"
+version = "0.4.4"
 description = "🤗 LeRobot: State-of-the-art Machine Learning for Real-World Robotics in Pytorch"
 dynamic = ["readme"]
 license = { text = "Apache-2.0" }
@@ -59,7 +59,7 @@ keywords = ["lerobot", "huggingface", "robotics",  "machine learning", "artifici
 dependencies = [
 
     # Hugging Face dependencies
-    "datasets>=4.0.0,<5.0.0",
+    "datasets>=4.0.0,<4.2.0",
     "diffusers>=0.27.2,<0.36.0",
     "huggingface-hub[hf-transfer,cli]>=0.34.2,<0.36.0",
     "accelerate>=1.10.0,<2.0.0",
@@ -76,9 +76,9 @@ dependencies = [
     "pyserial>=3.5,<4.0",
     "wandb>=0.24.0,<0.25.0",
 
-    "torch>=2.2.1,<2.11.0", # TODO: Bump dependency
+    "torch>=2.2.1,<2.8.0", # TODO: Bumb dependency
-    "torchcodec>=0.2.1,<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')", # TODO: Bump dependency
+    "torchcodec>=0.2.1,<0.6.0; 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')", # TODO: Bumb dependency
-    "torchvision>=0.21.0,<0.26.0", # TODO: Bump dependency
+    "torchvision>=0.21.0,<0.23.0", # TODO: Bumb dependency
 
     "draccus==0.10.0", # TODO: Remove ==
     "gymnasium>=1.1.1,<2.0.0",
@@ -98,13 +98,11 @@ pygame-dep = ["pygame>=2.5.1,<2.7.0"]
 placo-dep = ["placo>=0.9.6,<0.10.0"]
 transformers-dep = ["transformers>=4.57.1,<5.0.0"]
 grpcio-dep = ["grpcio==1.73.1", "protobuf>=6.31.1,<6.32.0"]
-can-dep = ["python-can>=4.2.0,<5.0.0"]
 
 # Motors
 feetech = ["feetech-servo-sdk>=1.0.0,<2.0.0"]
 dynamixel = ["dynamixel-sdk>=3.7.31,<3.9.0"]
-damiao = ["lerobot[can-dep]"]
+damiao = ["python-can>=4.2.0,<5.0.0"]
-robstride = ["lerobot[can-dep]"]
 
 # Robots
 openarms = ["lerobot[damiao]"]
@@ -214,9 +212,6 @@ lerobot-edit-dataset="lerobot.scripts.lerobot_edit_dataset:main"
 lerobot-setup-can="lerobot.scripts.lerobot_setup_can:main"
 
 # ---------------- Tool Configurations ----------------
-[tool.setuptools.package-data]
-lerobot = ["envs/*.json"]
-
 [tool.setuptools.packages.find]
 where = ["src"]
 

src/lerobot/async_inference/robot_client.py

@@ -49,18 +49,23 @@ import torch
 
 from lerobot.cameras.opencv.configuration_opencv import OpenCVCameraConfig  # noqa: F401
 from lerobot.cameras.realsense.configuration_realsense import RealSenseCameraConfig  # noqa: F401
-from lerobot.robots import (
+from lerobot.robots import (  # noqa: F401
-    RobotConfig,  # noqa: F401
+    Robot,
+    RobotConfig,
+    bi_so_follower,
+    koch_follower,
     make_robot_from_config,
+    omx_follower,
+    so_follower,
 )
 from lerobot.transport import (
     services_pb2,  # type: ignore
     services_pb2_grpc,  # type: ignore
 )
 from lerobot.transport.utils import grpc_channel_options, send_bytes_in_chunks
-from lerobot.utils.import_utils import register_third_party_plugins
 
 from .configs import RobotClientConfig
+from .constants import SUPPORTED_ROBOTS
 from .helpers import (
     Action,
     FPSTracker,
@@ -480,9 +485,8 @@ class RobotClient:
 def async_client(cfg: RobotClientConfig):
     logging.info(pformat(asdict(cfg)))
 
-    # TODO: Assert if checking robot support is still needed with the plugin system
+    if cfg.robot.type not in SUPPORTED_ROBOTS:
-    # if cfg.robot.type not in SUPPORTED_ROBOTS:
+        raise ValueError(f"Robot {cfg.robot.type} not yet supported!")
-    #     raise ValueError(f"Robot {cfg.robot.type} not yet supported!")
 
     client = RobotClient(cfg)
 
@@ -508,5 +512,4 @@ def async_client(cfg: RobotClientConfig):
 
 
 if __name__ == "__main__":
-    register_third_party_plugins()
     async_client()  # run the client

src/lerobot/cameras/camera.py

@@ -150,7 +150,7 @@ class Camera(abc.ABC):
         """
         pass
 
-    def read_latest(self, max_age_ms: int = 500) -> NDArray[Any]:
+    def read_latest(self, max_age_ms: int = 1000) -> NDArray[Any]:
         """Return the most recent frame captured immediately (Peeking).
 
         This method is non-blocking and returns whatever is currently in the

src/lerobot/cameras/opencv/camera_opencv.py

@@ -530,7 +530,7 @@ class OpenCVCamera(Camera):
         return frame
 
     @check_if_not_connected
-    def read_latest(self, max_age_ms: int = 500) -> NDArray[Any]:
+    def read_latest(self, max_age_ms: int = 1000) -> NDArray[Any]:
         """Return the most recent frame captured immediately (Peeking).
 
         This method is non-blocking and returns whatever is currently in the

src/lerobot/cameras/reachy2_camera/reachy2_camera.py

@@ -201,7 +201,7 @@ class Reachy2Camera(Camera):
         return self.read()
 
     @check_if_not_connected
-    def read_latest(self, max_age_ms: int = 500) -> NDArray[Any]:
+    def read_latest(self, max_age_ms: int = 1000) -> NDArray[Any]:
         """Return the most recent frame captured immediately (Peeking).
 
         This method is non-blocking and returns whatever is currently in the

src/lerobot/cameras/realsense/camera_realsense.py

@@ -573,7 +573,7 @@ class RealSenseCamera(Camera):
 
     # NOTE(Steven): Missing implementation for depth for now
     @check_if_not_connected
-    def read_latest(self, max_age_ms: int = 500) -> NDArray[Any]:
+    def read_latest(self, max_age_ms: int = 1000) -> NDArray[Any]:
         """Return the most recent (color) frame captured immediately (Peeking).
 
         This method is non-blocking and returns whatever is currently in the

src/lerobot/configs/default.py

@@ -27,7 +27,7 @@ class DatasetConfig:
     # "dataset_index" into the returned item. The index mapping is made according to the order in which the
     # datasets are provided.
     repo_id: str
-    # Root directory where the dataset will be stored (e.g. 'dataset/path'). If None, defaults to $HF_LEROBOT_HOME/repo_id.
+    # Root directory where the dataset will be stored (e.g. 'dataset/path').
     root: str | None = None
     episodes: list[int] | None = None
     image_transforms: ImageTransformsConfig = field(default_factory=ImageTransformsConfig)

src/lerobot/datasets/aggregate.py

@@ -21,8 +21,6 @@ from pathlib import Path
 
 import datasets
 import pandas as pd
-import pyarrow as pa
-import pyarrow.parquet as pq
 import tqdm
 
 from lerobot.datasets.compute_stats import aggregate_stats
@@ -37,6 +35,7 @@ from lerobot.datasets.utils import (
     get_file_size_in_mb,
     get_hf_features_from_features,
     get_parquet_file_size_in_mb,
+    to_parquet_with_hf_images,
     update_chunk_file_indices,
     write_info,
     write_stats,
@@ -81,41 +80,28 @@ def validate_all_metadata(all_metadata: list[LeRobotDatasetMetadata]):
     return fps, robot_type, features
 
 
-def update_data_table(table: pa.Table, src_meta, dst_meta) -> pa.Table:
+def update_data_df(df, src_meta, dst_meta):
-    """Updates a pyarrow Table with new indices and task mappings for aggregation.
+    """Updates a data DataFrame with new indices and task mappings for aggregation.
 
     Adjusts episode indices, frame indices, and task indices to account for
     previously aggregated data in the destination dataset.
 
     Args:
-        table: pyarrow Table containing the data to be updated.
+        df: DataFrame containing the data to be updated.
         src_meta: Source dataset metadata.
         dst_meta: Destination dataset metadata.
 
     Returns:
-        pa.Table: Updated Table with adjusted indices.
+        pd.DataFrame: Updated DataFrame with adjusted indices.
     """
-    ep_offset = dst_meta.info["total_episodes"]
-    idx_offset = dst_meta.info["total_frames"]
 
-    ep_col = table.column("episode_index")
+    df["episode_index"] = df["episode_index"] + dst_meta.info["total_episodes"]
-    new_ep = pa.array([v + ep_offset for v in ep_col.to_pylist()], type=ep_col.type)
+    df["index"] = df["index"] + dst_meta.info["total_frames"]
-    table = table.set_column(table.column_names.index("episode_index"), "episode_index", new_ep)
 
-    idx_col = table.column("index")
+    src_task_names = src_meta.tasks.index.take(df["task_index"].to_numpy())
-    new_idx = pa.array([v + idx_offset for v in idx_col.to_pylist()], type=idx_col.type)
+    df["task_index"] = dst_meta.tasks.loc[src_task_names, "task_index"].to_numpy()
-    table = table.set_column(table.column_names.index("index"), "index", new_idx)
 
-    old_task_indices = table.column("task_index").to_pylist()
+    return df
-    src_task_names = src_meta.tasks.index.take(old_task_indices)
-    new_task_indices = dst_meta.tasks.loc[src_task_names, "task_index"].to_numpy().tolist()
-    table = table.set_column(
-        table.column_names.index("task_index"),
-        "task_index",
-        pa.array(new_task_indices, type=table.schema.field("task_index").type),
-    )
-
-    return table
 
 
 def update_meta_data(
@@ -303,9 +289,7 @@ def aggregate_datasets(
 
     logging.info("Find all tasks")
     unique_tasks = pd.concat([m.tasks for m in all_metadata]).index.unique()
-    dst_meta.tasks = pd.DataFrame(
+    dst_meta.tasks = pd.DataFrame({"task_index": range(len(unique_tasks))}, index=unique_tasks)
-        {"task_index": range(len(unique_tasks))}, index=pd.Index(unique_tasks, name="task")
-    )
 
     meta_idx = {"chunk": 0, "file": 0}
     data_idx = {"chunk": 0, "file": 0}
@@ -482,13 +466,18 @@ def aggregate_data(src_meta, dst_meta, data_idx, data_files_size_in_mb, chunk_si
         src_path = src_meta.root / DEFAULT_DATA_PATH.format(
             chunk_index=src_chunk_idx, file_index=src_file_idx
         )
-        table = pq.read_table(src_path)
+        if contains_images:
-        table = update_data_table(table, src_meta, dst_meta)
+            # Use HuggingFace datasets to read source data to preserve image format
+            src_ds = datasets.Dataset.from_parquet(str(src_path))
+            df = src_ds.to_pandas()
+        else:
+            df = pd.read_parquet(src_path)
+        df = update_data_df(df, src_meta, dst_meta)
 
         # Write data and get the actual destination file it was written to
         # This avoids duplicating the rotation logic here
         data_idx, (dst_chunk, dst_file) = append_or_create_parquet_file(
-            table,
+            df,
             src_path,
             data_idx,
             data_files_size_in_mb,
@@ -563,16 +552,8 @@ def aggregate_metadata(src_meta, dst_meta, meta_idx, data_idx, videos_idx):
     return meta_idx
 
 
-def _write_table_with_hf_images(
-    table: pa.Table, path: Path, features: datasets.Features | None = None
-) -> None:
-    """Write a pyarrow Table to parquet with proper HF image encoding."""
-    ds = datasets.Dataset.from_dict(table.to_pydict(), features=features)
-    ds.to_parquet(path)
-
-
 def append_or_create_parquet_file(
-    data: pd.DataFrame | pa.Table,
+    df: pd.DataFrame,
     src_path: Path,
     idx: dict[str, int],
     max_mb: float,
@@ -588,7 +569,7 @@ def append_or_create_parquet_file(
     from becoming too large. Handles both regular parquet files and those containing images.
 
     Args:
-        data: Data to write, as a pandas DataFrame or pyarrow Table.
+        df: DataFrame to write to the parquet file.
         src_path: Path to the source file (used for size estimation).
         idx: Dictionary containing current 'chunk' and 'file' indices.
         max_mb: Maximum allowed file size in MB before rotation.
@@ -602,17 +583,15 @@ def append_or_create_parquet_file(
         tuple: (updated_idx, (dst_chunk, dst_file)) where updated_idx is the index dict
                and (dst_chunk, dst_file) is the actual destination file the data was written to.
     """
-    table = data if isinstance(data, pa.Table) else pa.Table.from_pandas(data)
-
     dst_chunk, dst_file = idx["chunk"], idx["file"]
     dst_path = aggr_root / default_path.format(chunk_index=dst_chunk, file_index=dst_file)
 
     if not dst_path.exists():
         dst_path.parent.mkdir(parents=True, exist_ok=True)
         if contains_images:
-            _write_table_with_hf_images(table, dst_path, features=hf_features)
+            to_parquet_with_hf_images(df, dst_path, features=hf_features)
         else:
-            pq.write_table(table, dst_path)
+            df.to_parquet(dst_path)
         return idx, (dst_chunk, dst_file)
 
     src_size = get_parquet_file_size_in_mb(src_path)
@@ -623,17 +602,22 @@ def append_or_create_parquet_file(
         dst_chunk, dst_file = idx["chunk"], idx["file"]
         new_path = aggr_root / default_path.format(chunk_index=dst_chunk, file_index=dst_file)
         new_path.parent.mkdir(parents=True, exist_ok=True)
-        final_table = table
+        final_df = df
         target_path = new_path
     else:
-        existing_table = pq.read_table(dst_path)
+        if contains_images:
-        final_table = pa.concat_tables([existing_table, table], promote_options="permissive")
+            # Use HuggingFace datasets to read existing data to preserve image format
+            existing_ds = datasets.Dataset.from_parquet(str(dst_path))
+            existing_df = existing_ds.to_pandas()
+        else:
+            existing_df = pd.read_parquet(dst_path)
+        final_df = pd.concat([existing_df, df], ignore_index=True)
         target_path = dst_path
 
     if contains_images:
-        _write_table_with_hf_images(final_table, target_path, features=hf_features)
+        to_parquet_with_hf_images(final_df, target_path, features=hf_features)
     else:
-        pq.write_table(final_table, target_path)
+        final_df.to_parquet(target_path)
 
     return idx, (dst_chunk, dst_file)
 

src/lerobot/datasets/card_template.md

@@ -7,13 +7,6 @@
 
 This dataset was created using [LeRobot](https://github.com/huggingface/lerobot).
 
-{% if repo_id is defined and repo_id %}
-<a class="flex" href="https://huggingface.co/spaces/lerobot/visualize_dataset?path={{ repo_id }}">
-<img class="block dark:hidden" src="https://huggingface.co/datasets/huggingface/badges/resolve/main/visualize-this-dataset-xl.svg"/>
-<img class="hidden dark:block" src="https://huggingface.co/datasets/huggingface/badges/resolve/main/visualize-this-dataset-xl-dark.svg"/>
-</a>
-{% endif %}
-
 ## Dataset Description
 
 {{ dataset_description | default("", true) }}

src/lerobot/datasets/dataset_tools.py

@@ -25,6 +25,7 @@ This module provides utilities for:
 
 import logging
 import shutil
+from collections import defaultdict
 from collections.abc import Callable
 from concurrent.futures import ThreadPoolExecutor, as_completed
 from pathlib import Path
@@ -32,15 +33,12 @@ from pathlib import Path
 import datasets
 import numpy as np
 import pandas as pd
-import pyarrow as pa
-import pyarrow.compute as pc
-import pyarrow.dataset as pa_ds
 import pyarrow.parquet as pq
 import torch
 from tqdm import tqdm
 
 from lerobot.datasets.aggregate import aggregate_datasets
-from lerobot.datasets.compute_stats import aggregate_stats
+from lerobot.datasets.compute_stats import aggregate_stats, compute_episode_stats, get_feature_stats
 from lerobot.datasets.lerobot_dataset import LeRobotDataset, LeRobotDatasetMetadata
 from lerobot.datasets.utils import (
     DATA_DIR,
@@ -48,6 +46,8 @@ from lerobot.datasets.utils import (
     DEFAULT_DATA_FILE_SIZE_IN_MB,
     DEFAULT_DATA_PATH,
     DEFAULT_EPISODES_PATH,
+    DEFAULT_SUBTASKS_PATH,
+    flatten_dict,
     get_parquet_file_size_in_mb,
     load_episodes,
     update_chunk_file_indices,
@@ -92,8 +92,8 @@ def delete_episodes(
     Args:
         dataset: The source LeRobotDataset.
         episode_indices: List of episode indices to delete.
-        output_dir: Root directory where the edited dataset will be stored. If not specified, defaults to $HF_LEROBOT_HOME/repo_id. Equivalent to new_root in EditDatasetConfig.
+        output_dir: Directory to save the new dataset. If None, uses default location.
-        repo_id: Edited dataset identifier. Equivalent to new_repo_id in EditDatasetConfig.
+        repo_id: Repository ID for the new dataset. If None, appends "_modified" to original.
     """
     if not episode_indices:
         raise ValueError("No episodes to delete")
@@ -144,6 +144,315 @@ def delete_episodes(
     return new_dataset
 
 
+def trim_episode_start(
+    dataset: LeRobotDataset,
+    seconds: float,
+    episode_indices: list[int] | None = None,
+    output_dir: str | Path | None = None,
+    repo_id: str | None = None,
+) -> LeRobotDataset:
+    """Trim the first N seconds from selected episodes and create a new dataset.
+
+    The operation rewrites data parquet files and updates episode metadata so that:
+    - frame_index starts at 0 for each trimmed episode
+    - timestamp starts at 0 for each trimmed episode
+    - global index remains contiguous across the full dataset
+    - dataset_from_index / dataset_to_index reflect new frame ranges
+
+    Video files are copied as-is and per-episode video timestamps are shifted forward
+    for trimmed episodes.
+
+    Episodes selected for trimming that are too short (length <= trim_frames) are skipped
+    from the output dataset.
+
+    Args:
+        dataset: The source LeRobotDataset.
+        seconds: Number of seconds to remove from episode starts.
+        episode_indices: Optional list of episode indices to trim. If None, trims all episodes.
+        output_dir: Directory to save the new dataset. If None, uses default location.
+        repo_id: Repository ID for the new dataset. If None, appends "_trimmed" to original.
+    """
+    if seconds <= 0:
+        raise ValueError(f"seconds must be strictly positive, got {seconds}")
+
+    if dataset.meta.episodes is None:
+        dataset.meta.episodes = load_episodes(dataset.meta.root)
+
+    trim_frames = int(seconds * dataset.meta.fps)
+    if trim_frames <= 0:
+        raise ValueError(
+            f"seconds={seconds} corresponds to 0 frames at fps={dataset.meta.fps}. "
+            "Increase seconds so at least one frame is trimmed."
+        )
+
+    if episode_indices is None:
+        episode_indices = list(range(dataset.meta.total_episodes))
+
+    if len(episode_indices) == 0:
+        raise ValueError("No episodes specified to trim")
+
+    episode_indices = sorted(set(episode_indices))
+    valid_indices = set(range(dataset.meta.total_episodes))
+    invalid = set(episode_indices) - valid_indices
+    if invalid:
+        raise ValueError(f"Invalid episode indices: {invalid}")
+
+    too_short = sorted(
+        ep_idx for ep_idx in episode_indices if int(dataset.meta.episodes[ep_idx]["length"]) <= trim_frames
+    )
+    trim_set = set(episode_indices)
+    skipped_set = set(too_short)
+    trim_set -= skipped_set
+
+    if too_short:
+        logging.warning(
+            f"Skipping {len(too_short)} episode(s) that are too short to trim "
+            f"({trim_frames} frames): {too_short}"
+        )
+
+    episodes_to_keep = [ep_idx for ep_idx in range(dataset.meta.total_episodes) if ep_idx not in skipped_set]
+    if not episodes_to_keep:
+        raise ValueError(
+            "All episodes selected for trimming are too short and would be skipped. "
+            "Try a smaller trim duration."
+        )
+
+    logging.info(
+        f"Trimming {len(trim_set)} episode(s) by {seconds}s and keeping {len(episodes_to_keep)} "
+        f"episode(s) in output"
+    )
+
+    if repo_id is None:
+        repo_id = f"{dataset.repo_id}_trimmed"
+    output_dir = Path(output_dir) if output_dir is not None else HF_LEROBOT_HOME / repo_id
+
+    new_meta = LeRobotDatasetMetadata.create(
+        repo_id=repo_id,
+        fps=dataset.meta.fps,
+        features=dataset.meta.features,
+        robot_type=dataset.meta.robot_type,
+        root=output_dir,
+        use_videos=len(dataset.meta.video_keys) > 0,
+        chunks_size=dataset.meta.chunks_size,
+        data_files_size_in_mb=dataset.meta.data_files_size_in_mb,
+        video_files_size_in_mb=dataset.meta.video_files_size_in_mb,
+    )
+
+    if dataset.meta.tasks is not None:
+        write_tasks(dataset.meta.tasks, new_meta.root)
+        new_meta.tasks = dataset.meta.tasks.copy()
+
+    subtasks_path = dataset.root / DEFAULT_SUBTASKS_PATH
+    if subtasks_path.exists():
+        dst_subtasks_path = new_meta.root / DEFAULT_SUBTASKS_PATH
+        dst_subtasks_path.parent.mkdir(parents=True, exist_ok=True)
+        shutil.copy(subtasks_path, dst_subtasks_path)
+
+    episode_mapping = {old_idx: new_idx for new_idx, old_idx in enumerate(episodes_to_keep)}
+    trim_duration_s = trim_frames / dataset.meta.fps
+
+    episode_lengths: dict[int, int] = {}
+    episode_ranges: dict[int, tuple[int, int]] = {}
+    total_frames = 0
+    for old_ep_idx in episodes_to_keep:
+        new_ep_idx = episode_mapping[old_ep_idx]
+        src_length = int(dataset.meta.episodes[old_ep_idx]["length"])
+        new_length = src_length - trim_frames if old_ep_idx in trim_set else src_length
+        episode_lengths[new_ep_idx] = new_length
+        episode_ranges[new_ep_idx] = (total_frames, total_frames + new_length)
+        total_frames += new_length
+
+    numeric_features = {
+        k: v
+        for k, v in dataset.meta.features.items()
+        if v["dtype"] not in ["image", "video", "string"]
+    }
+    episode_stats_parts: dict[int, list[dict[str, dict]]] = defaultdict(list)
+    episode_file_metadata: dict[int, dict[str, int]] = {}
+
+    data_dir = dataset.root / DATA_DIR
+    parquet_files = sorted(data_dir.glob("*/*.parquet"))
+    if not parquet_files:
+        raise ValueError(f"No parquet files found in {data_dir}")
+
+    for src_path in tqdm(parquet_files, desc="Trimming data files"):
+        df = pd.read_parquet(src_path).reset_index(drop=True)
+
+        if len(df) == 0:
+            continue
+
+        if skipped_set:
+            keep_mask = ~df["episode_index"].isin(skipped_set)
+            if not keep_mask.all():
+                df = df.loc[keep_mask].copy().reset_index(drop=True)
+
+        if len(df) == 0:
+            continue
+
+        if trim_set:
+            trim_mask = df["episode_index"].isin(trim_set) & (df["frame_index"] < trim_frames)
+            if trim_mask.any():
+                df = df.loc[~trim_mask].copy().reset_index(drop=True)
+
+        if len(df) == 0:
+            continue
+
+        relative_path = src_path.relative_to(dataset.root)
+        chunk_idx = int(relative_path.parts[1].split("-")[1])
+        file_idx = int(relative_path.parts[2].split("-")[1].split(".")[0])
+
+        for old_ep_idx in sorted(df["episode_index"].unique().tolist()):
+            ep_mask = df["episode_index"] == old_ep_idx
+            new_ep_idx = episode_mapping[old_ep_idx]
+
+            if old_ep_idx in trim_set:
+                df.loc[ep_mask, "frame_index"] = df.loc[ep_mask, "frame_index"] - trim_frames
+                shifted_timestamps = df.loc[ep_mask, "timestamp"].to_numpy(dtype=np.float64) - trim_duration_s
+                df.loc[ep_mask, "timestamp"] = np.clip(shifted_timestamps, a_min=0.0, a_max=None)
+
+            df.loc[ep_mask, "episode_index"] = new_ep_idx
+
+            ep_start, _ = episode_ranges[new_ep_idx]
+            new_indices = ep_start + df.loc[ep_mask, "frame_index"].to_numpy(dtype=np.int64)
+            df.loc[ep_mask, "index"] = new_indices
+
+            if new_ep_idx in episode_file_metadata:
+                existing = episode_file_metadata[new_ep_idx]
+                if (
+                    existing["data/chunk_index"] != chunk_idx
+                    or existing["data/file_index"] != file_idx
+                ):
+                    raise ValueError(
+                        f"Episode {old_ep_idx} spans multiple data files. "
+                        "trim_episode_start currently expects one data file per episode."
+                    )
+            else:
+                episode_file_metadata[new_ep_idx] = {
+                    "data/chunk_index": chunk_idx,
+                    "data/file_index": file_idx,
+                }
+
+            if numeric_features:
+                ep_df = df.loc[ep_mask]
+                episode_data: dict[str, np.ndarray] = {}
+                episode_feature_spec: dict[str, dict] = {}
+
+                for key, feature in numeric_features.items():
+                    if key not in ep_df.columns:
+                        continue
+
+                    values = ep_df[key].to_numpy()
+                    if len(values) == 0:
+                        continue
+
+                    first_value = values[0]
+                    if isinstance(first_value, np.ndarray):
+                        episode_data[key] = np.stack(values)
+                    elif isinstance(first_value, (list, tuple)):
+                        episode_data[key] = np.stack(values)
+                    else:
+                        episode_data[key] = np.asarray(values)
+
+                    episode_feature_spec[key] = feature
+
+                if episode_data:
+                    episode_stats_parts[new_ep_idx].append(
+                        compute_episode_stats(episode_data, episode_feature_spec)
+                    )
+
+        df["index"] = df["index"].astype(np.int64)
+        if "frame_index" in df.columns:
+            df["frame_index"] = df["frame_index"].astype(np.int64)
+
+        dst_path = new_meta.root / DEFAULT_DATA_PATH.format(chunk_index=chunk_idx, file_index=file_idx)
+        dst_path.parent.mkdir(parents=True, exist_ok=True)
+        _write_parquet(df, dst_path, new_meta)
+
+    all_episode_stats = []
+    for old_ep_idx in tqdm(episodes_to_keep, desc="Writing episode metadata"):
+        new_ep_idx = episode_mapping[old_ep_idx]
+
+        if new_ep_idx not in episode_file_metadata:
+            raise ValueError(f"Missing data file metadata for episode {old_ep_idx}")
+
+        from_idx, to_idx = episode_ranges[new_ep_idx]
+        src_episode = dataset.meta.episodes[old_ep_idx]
+        ep_data_meta = episode_file_metadata[new_ep_idx]
+
+        stats_parts = episode_stats_parts.get(new_ep_idx, [])
+        ep_stats = aggregate_stats(stats_parts) if len(stats_parts) > 1 else (stats_parts[0] if stats_parts else {})
+        if ep_stats:
+            all_episode_stats.append(ep_stats)
+
+        episode_meta = {
+            "data/chunk_index": ep_data_meta["data/chunk_index"],
+            "data/file_index": ep_data_meta["data/file_index"],
+            "dataset_from_index": from_idx,
+            "dataset_to_index": to_idx,
+        }
+
+        for video_key in dataset.meta.video_keys:
+            from_ts = src_episode[f"videos/{video_key}/from_timestamp"]
+            if old_ep_idx in trim_set:
+                from_ts += trim_duration_s
+            episode_meta.update(
+                {
+                    f"videos/{video_key}/chunk_index": src_episode[f"videos/{video_key}/chunk_index"],
+                    f"videos/{video_key}/file_index": src_episode[f"videos/{video_key}/file_index"],
+                    f"videos/{video_key}/from_timestamp": from_ts,
+                    f"videos/{video_key}/to_timestamp": src_episode[f"videos/{video_key}/to_timestamp"],
+                }
+            )
+
+        episode_dict = {
+            "episode_index": new_ep_idx,
+            "tasks": src_episode["tasks"],
+            "length": episode_lengths[new_ep_idx],
+        }
+        episode_dict.update(episode_meta)
+        if ep_stats:
+            episode_dict.update(flatten_dict({"stats": ep_stats}))
+
+        new_meta._save_episode_metadata(episode_dict)
+
+    new_meta._close_writer()
+
+    if new_meta.video_keys:
+        _copy_videos(dataset, new_meta)
+
+    new_meta.info.update(
+        {
+            "total_episodes": len(episodes_to_keep),
+            "total_frames": total_frames,
+            "total_tasks": len(new_meta.tasks) if new_meta.tasks is not None else 0,
+            "splits": {"train": f"0:{len(episodes_to_keep)}"},
+        }
+    )
+
+    if new_meta.video_keys and dataset.meta.video_keys:
+        for key in new_meta.video_keys:
+            if key in dataset.meta.features:
+                new_meta.info["features"][key]["info"] = dataset.meta.info["features"][key].get("info", {})
+
+    write_info(new_meta.info, new_meta.root)
+
+    merged_stats = aggregate_stats(all_episode_stats) if all_episode_stats else {}
+    if dataset.meta.stats:
+        for key, value in dataset.meta.stats.items():
+            if key not in merged_stats:
+                merged_stats[key] = value
+    if merged_stats:
+        write_stats(merged_stats, new_meta.root)
+
+    return LeRobotDataset(
+        repo_id=repo_id,
+        root=output_dir,
+        image_transforms=dataset.image_transforms,
+        delta_timestamps=dataset.delta_timestamps,
+        tolerance_s=dataset.tolerance_s,
+    )
+
+
 def split_dataset(
     dataset: LeRobotDataset,
     splits: dict[str, float | list[int]],
@@ -155,7 +464,7 @@ def split_dataset(
         dataset: The source LeRobotDataset to split.
         splits: Either a dict mapping split names to episode indices, or a dict mapping
                 split names to fractions (must sum to <= 1.0).
-        output_dir: Root directory where the split datasets will be stored. If not specified, defaults to $HF_LEROBOT_HOME/repo_id.
+        output_dir: Base directory for output datasets. If None, uses default location.
 
     Examples:
       Split by specific episodes
@@ -246,8 +555,8 @@ def merge_datasets(
 
     Args:
         datasets: List of LeRobotDatasets to merge.
-        output_repo_id: Merged dataset identifier.
+        output_repo_id: Repository ID for the merged dataset.
-        output_dir: Root directory where the merged dataset will be stored. If not specified, defaults to $HF_LEROBOT_HOME/output_repo_id.
+        output_dir: Directory to save the merged dataset. If None, uses default location.
     """
     if not datasets:
         raise ValueError("No datasets to merge")
@@ -291,8 +600,8 @@ def modify_features(
         dataset: The source LeRobotDataset.
         add_features: Optional dict mapping feature names to (feature_values, feature_info) tuples.
         remove_features: Optional feature name(s) to remove. Can be a single string or list.
-        output_dir: Root directory where the edited dataset will be stored. If not specified, defaults to $HF_LEROBOT_HOME/repo_id. Equivalent to new_root in EditDatasetConfig.
+        output_dir: Directory to save the new dataset. If None, uses default location.
-        repo_id: Edited dataset identifier. Equivalent to new_repo_id in EditDatasetConfig.
+        repo_id: Repository ID for the new dataset. If None, appends "_modified" to original.
 
     Returns:
         New dataset with features modified.
@@ -393,8 +702,8 @@ def add_features(
     Args:
         dataset: The source LeRobotDataset.
         features: Dictionary mapping feature names to (feature_values, feature_info) tuples.
-        output_dir: Root directory where the edited dataset will be stored. If not specified, defaults to $HF_LEROBOT_HOME/repo_id. Equivalent to new_root in EditDatasetConfig.
+        output_dir: Directory to save the new dataset. If None, uses default location.
-        repo_id: Edited dataset identifier. Equivalent to new_repo_id in EditDatasetConfig.
+        repo_id: Repository ID for the new dataset. If None, appends "_modified" to original.
 
     Returns:
         New dataset with all features added.
@@ -430,8 +739,8 @@ def remove_feature(
     Args:
         dataset: The source LeRobotDataset.
         feature_names: Name(s) of features to remove. Can be a single string or list.
-        output_dir: Root directory where the edited dataset will be stored. If not specified, defaults to $HF_LEROBOT_HOME/repo_id. Equivalent to new_root in EditDatasetConfig.
+        output_dir: Directory to save the new dataset. If None, uses default location.
-        repo_id: Edited dataset identifier. Equivalent to new_repo_id in EditDatasetConfig.
+        repo_id: Repository ID for the new dataset. If None, appends "_modified" to original.
 
     Returns:
         New dataset with features removed.
@@ -499,16 +808,13 @@ def _copy_and_reindex_data(
     global_index = 0
     episode_data_metadata: dict[int, dict] = {}
 
-    episode_keys = list(episode_mapping.keys())
-    ep_filter = pa_ds.field("episode_index").isin(episode_keys)
-
     if dst_meta.tasks is None:
-        all_task_indices: set[int] = set()
+        all_task_indices = set()
         for src_path in file_to_episodes:
-            table = pq.read_table(
+            df = pd.read_parquet(src_dataset.root / src_path)
-                src_dataset.root / src_path, columns=["episode_index", "task_index"], filters=ep_filter
+            mask = df["episode_index"].isin(list(episode_mapping.keys()))
-            )
+            task_series: pd.Series = df[mask]["task_index"]
-            all_task_indices.update(pc.unique(table.column("task_index")).to_pylist())
+            all_task_indices.update(task_series.unique().tolist())
         tasks = [src_dataset.meta.tasks.iloc[idx].name for idx in all_task_indices]
         dst_meta.save_episode_tasks(list(set(tasks)))
 
@@ -520,41 +826,52 @@ def _copy_and_reindex_data(
             task_mapping[old_task_idx] = new_task_idx
 
     for src_path in tqdm(sorted(file_to_episodes.keys()), desc="Processing data files"):
-        table = pq.read_table(src_dataset.root / src_path, filters=ep_filter)
+        df = pd.read_parquet(src_dataset.root / src_path)
+
+        all_episodes_in_file = set(df["episode_index"].unique())
         episodes_to_keep = file_to_episodes[src_path]
 
-        if table.num_rows == 0:
+        if all_episodes_in_file == episodes_to_keep:
-            continue
+            df["episode_index"] = df["episode_index"].replace(episode_mapping)
+            df["index"] = range(global_index, global_index + len(df))
+            df["task_index"] = df["task_index"].replace(task_mapping)
 
-        table = _replace_column_values(table, "episode_index", episode_mapping)
+            first_ep_old_idx = min(episodes_to_keep)
-        col_pos = table.column_names.index("index")
+            src_ep = src_dataset.meta.episodes[first_ep_old_idx]
-        new_indices = pa.array(range(global_index, global_index + table.num_rows), type=pa.int64())
+            chunk_idx = src_ep["data/chunk_index"]
-        table = table.set_column(col_pos, "index", new_indices)
+            file_idx = src_ep["data/file_index"]
-        table = _replace_column_values(table, "task_index", task_mapping)
+        else:
+            mask = df["episode_index"].isin(list(episode_mapping.keys()))
+            df = df[mask].copy().reset_index(drop=True)
 
-        first_ep_old_idx = min(episodes_to_keep)
+            if len(df) == 0:
-        src_ep = src_dataset.meta.episodes[first_ep_old_idx]
+                continue
-        chunk_idx = src_ep["data/chunk_index"]
+
-        file_idx = src_ep["data/file_index"]
+            df["episode_index"] = df["episode_index"].replace(episode_mapping)
+            df["index"] = range(global_index, global_index + len(df))
+            df["task_index"] = df["task_index"].replace(task_mapping)
+
+            first_ep_old_idx = min(episodes_to_keep)
+            src_ep = src_dataset.meta.episodes[first_ep_old_idx]
+            chunk_idx = src_ep["data/chunk_index"]
+            file_idx = src_ep["data/file_index"]
 
         dst_path = dst_meta.root / DEFAULT_DATA_PATH.format(chunk_index=chunk_idx, file_index=file_idx)
         dst_path.parent.mkdir(parents=True, exist_ok=True)
 
-        _write_parquet(table, dst_path, dst_meta)
+        _write_parquet(df, dst_path, dst_meta)
 
-        ep_col = table.column("episode_index").to_pylist()
-        idx_col = table.column("index").to_pylist()
         for ep_old_idx in episodes_to_keep:
             ep_new_idx = episode_mapping[ep_old_idx]
-            ep_indices = [idx_col[i] for i, e in enumerate(ep_col) if e == ep_new_idx]
+            ep_df = df[df["episode_index"] == ep_new_idx]
             episode_data_metadata[ep_new_idx] = {
                 "data/chunk_index": chunk_idx,
                 "data/file_index": file_idx,
-                "dataset_from_index": min(ep_indices),
+                "dataset_from_index": int(ep_df["index"].min()),
-                "dataset_to_index": max(ep_indices) + 1,
+                "dataset_to_index": int(ep_df["index"].max() + 1),
             }
 
-        global_index += table.num_rows
+        global_index += len(df)
 
     return episode_data_metadata
 
@@ -562,22 +879,20 @@ def _copy_and_reindex_data(
 def _keep_episodes_from_video_with_av(
     input_path: Path,
     output_path: Path,
-    episodes_to_keep: list[tuple[int, int]],
+    episodes_to_keep: list[tuple[float, float]],
     fps: float,
     vcodec: str = "libsvtav1",
     pix_fmt: str = "yuv420p",
 ) -> None:
     """Keep only specified episodes from a video file using PyAV.
 
-    This function decodes frames from specified frame ranges and re-encodes them with
+    This function decodes frames from specified time ranges and re-encodes them with
     properly reset timestamps to ensure monotonic progression.
 
     Args:
         input_path: Source video file path.
         output_path: Destination video file path.
-        episodes_to_keep: List of (start_frame, end_frame) tuples for episodes to keep.
+        episodes_to_keep: List of (start_time, end_time) tuples for episodes to keep.
-            Ranges are half-open intervals: [start_frame, end_frame), where start_frame
-            is inclusive and end_frame is exclusive.
         fps: Frame rate of the video.
         vcodec: Video codec to use for encoding.
         pix_fmt: Pixel format for output video.
@@ -619,10 +934,9 @@ def _keep_episodes_from_video_with_av(
 
     # Create set of (start, end) ranges for fast lookup.
     # Convert to a sorted list for efficient checking.
-    frame_ranges = sorted(episodes_to_keep)
+    time_ranges = sorted(episodes_to_keep)
 
     # Track frame index for setting PTS and current range being processed.
-    src_frame_count = 0
     frame_count = 0
     range_idx = 0
 
@@ -632,20 +946,21 @@ def _keep_episodes_from_video_with_av(
             if frame is None:
                 continue
 
-            # Check if frame is in any of our desired frame ranges.
+            # Get frame timestamp.
+            frame_time = float(frame.pts * frame.time_base) if frame.pts is not None else 0.0
+
+            # Check if frame is in any of our desired time ranges.
             # Skip ranges that have already passed.
-            while range_idx < len(frame_ranges) and src_frame_count >= frame_ranges[range_idx][1]:
+            while range_idx < len(time_ranges) and frame_time >= time_ranges[range_idx][1]:
                 range_idx += 1
 
             # If we've passed all ranges, stop processing.
-            if range_idx >= len(frame_ranges):
+            if range_idx >= len(time_ranges):
                 break
 
             # Check if frame is in current range.
-            start_frame = frame_ranges[range_idx][0]
+            start_ts, end_ts = time_ranges[range_idx]
-
+            if frame_time < start_ts:
-            if src_frame_count < start_frame:
-                src_frame_count += 1
                 continue
 
             # Frame is in range - create a new frame with reset timestamps.
@@ -658,7 +973,6 @@ def _keep_episodes_from_video_with_av(
             for pkt in v_out.encode(new_frame):
                 out.mux(pkt)
 
-            src_frame_count += 1
             frame_count += 1
 
     # Flush encoder.
@@ -747,17 +1061,15 @@ def _copy_and_reindex_videos(
                         f"videos/{video_key}/to_timestamp"
                     ]
             else:
-                # Build list of frame ranges to keep, in sorted order.
+                # Build list of time ranges to keep, in sorted order.
                 sorted_keep_episodes = sorted(episodes_in_file, key=lambda x: episode_mapping[x])
-                episodes_to_keep_ranges: list[tuple[int, int]] = []
+                episodes_to_keep_ranges: list[tuple[float, float]] = []
+
                 for old_idx in sorted_keep_episodes:
                     src_ep = src_dataset.meta.episodes[old_idx]
-                    from_frame = round(src_ep[f"videos/{video_key}/from_timestamp"] * src_dataset.meta.fps)
+                    from_ts = src_ep[f"videos/{video_key}/from_timestamp"]
-                    to_frame = round(src_ep[f"videos/{video_key}/to_timestamp"] * src_dataset.meta.fps)
+                    to_ts = src_ep[f"videos/{video_key}/to_timestamp"]
-                    assert src_ep["length"] == to_frame - from_frame, (
+                    episodes_to_keep_ranges.append((from_ts, to_ts))
-                        f"Episode length mismatch: {src_ep['length']} vs {to_frame - from_frame}"
-                    )
-                    episodes_to_keep_ranges.append((from_frame, to_frame))
 
                 # Use PyAV filters to efficiently re-encode only the desired segments.
                 assert src_dataset.meta.video_path is not None
@@ -905,39 +1217,15 @@ def _copy_and_reindex_episodes_metadata(
     write_stats(filtered_stats, dst_meta.root)
 
 
-def _replace_column_values(table: pa.Table, column: str, mapping: dict) -> pa.Table:
+def _write_parquet(df: pd.DataFrame, path: Path, meta: LeRobotDatasetMetadata) -> None:
-    """Replace values in a pyarrow Table column using a mapping dict."""
+    """Write DataFrame to parquet
-    old_values = table.column(column).to_pylist()
-    new_values = [mapping.get(v, v) for v in old_values]
-    col_pos = table.column_names.index(column)
-    return table.set_column(col_pos, column, pa.array(new_values, type=table.schema.field(column).type))
-
-
-def _write_parquet(
-    data: pd.DataFrame | pa.Table | dict, path: Path, meta: LeRobotDatasetMetadata
-) -> None:
-    """Write data to parquet.
 
     This ensures images are properly embedded and the file can be loaded correctly by HF datasets.
-
-    Args:
-        data: Input data as a pandas DataFrame, pyarrow Table, or dict of lists.
-        path: Destination parquet file path.
-        meta: Dataset metadata for feature schema.
     """
     from lerobot.datasets.utils import embed_images, get_hf_features_from_features
 
-    if isinstance(data, pd.DataFrame):
-        data_dict = data.to_dict(orient="list")
-    elif isinstance(data, pa.Table):
-        data_dict = data.to_pydict()
-    elif isinstance(data, dict):
-        data_dict = data
-    else:
-        raise TypeError(f"Unsupported data type: {type(data)}")
-
     hf_features = get_hf_features_from_features(meta.features)
-    ep_dataset = datasets.Dataset.from_dict(data_dict, features=hf_features, split="train")
+    ep_dataset = datasets.Dataset.from_dict(df.to_dict(orient="list"), features=hf_features, split="train")
 
     if len(meta.image_keys) > 0:
         ep_dataset = embed_images(ep_dataset)
@@ -1494,9 +1782,7 @@ def modify_tasks(
 
     # Collect all unique tasks and create new task mapping
     unique_tasks = sorted(set(episode_to_task.values()))
-    new_task_df = pd.DataFrame(
+    new_task_df = pd.DataFrame({"task_index": list(range(len(unique_tasks)))}, index=unique_tasks)
-        {"task_index": list(range(len(unique_tasks)))}, index=pd.Index(unique_tasks, name="task")
-    )
     task_to_index = {task: idx for idx, task in enumerate(unique_tasks)}
 
     logging.info(f"Modifying tasks in {dataset.repo_id}")
@@ -1548,9 +1834,125 @@ def modify_tasks(
     return dataset
 
 
+def recompute_stats(
+    dataset: LeRobotDataset,
+    skip_image_video: bool = True,
+    delta_action: bool = False,
+    delta_exclude_joints: list[str] | None = None,
+) -> LeRobotDataset:
+    """Recompute stats.json from scratch by iterating all episodes.
+
+    Args:
+        dataset: The LeRobotDataset to recompute stats for.
+        skip_image_video: If True (default), only recompute stats for numeric features
+            (action, state, etc.) and keep existing image/video stats unchanged.
+        delta_action: If True, compute action stats as delta (action - state).
+            Useful when training with use_delta_actions=True so normalization matches.
+        delta_exclude_joints: Joint names to exclude from delta conversion when
+            delta_action=True. These dims keep absolute stats. Uses dataset's
+            action feature names to build the mask. Default: ["gripper"].
+
+    Returns:
+        The same dataset with updated stats.
+    """
+    features = dataset.meta.features
+    numeric_features = {
+        k: v for k, v in features.items()
+        if v["dtype"] not in ["image", "video", "string"]
+        and k not in ["index", "episode_index", "task_index", "frame_index", "timestamp"]
+    }
+
+    if skip_image_video:
+        features_to_compute = numeric_features
+    else:
+        features_to_compute = {
+            k: v for k, v in features.items()
+            if v["dtype"] != "string"
+            and k not in ["index", "episode_index", "task_index", "frame_index", "timestamp"]
+        }
+
+    # Build delta mask if delta_action is enabled
+    delta_mask = None
+    if delta_action and "action" in features and "observation.state" in features:
+        if delta_exclude_joints is None:
+            delta_exclude_joints = ["gripper"]
+        action_names = features["action"].get("names")
+        if action_names is not None:
+            exclude = set(delta_exclude_joints)
+            delta_mask = [n not in exclude for n in action_names]
+        else:
+            action_dim = features["action"]["shape"][0]
+            delta_mask = [True] * action_dim
+        # Only recompute action stats when delta is enabled — state stays unchanged
+        features_to_compute = {"action": features["action"]}
+        logging.info(f"Recomputing action stats as delta (exclude: {delta_exclude_joints})")
+    else:
+        logging.info(f"Recomputing stats for features: {list(features_to_compute.keys())}")
+
+    data_dir = dataset.root / DATA_DIR
+    parquet_files = sorted(data_dir.glob("*/*.parquet"))
+    if not parquet_files:
+        raise ValueError(f"No parquet files found in {data_dir}")
+
+    all_episode_stats = []
+    numeric_keys = [k for k, v in features_to_compute.items() if v["dtype"] not in ["image", "video"]]
+    # Also need state for delta computation even though we don't recompute state stats
+    needs_state = delta_mask is not None
+
+    for parquet_path in tqdm(parquet_files, desc="Computing stats from data files"):
+        df = pd.read_parquet(parquet_path)
+
+        for ep_idx in sorted(df["episode_index"].unique()):
+            ep_df = df[df["episode_index"] == ep_idx]
+            episode_data = {}
+            for key in numeric_keys:
+                if key in ep_df.columns:
+                    values = ep_df[key].values
+                    if hasattr(values[0], "__len__"):
+                        episode_data[key] = np.stack(values)
+                    else:
+                        episode_data[key] = np.array(values)
+
+            # Apply delta conversion to actions before computing stats
+            if delta_mask is not None and "action" in episode_data:
+                from lerobot.processor.delta_action_processor import to_delta_actions
+
+                # Load state for delta even if we're not computing state stats
+                if needs_state and "observation.state" in ep_df.columns:
+                    state_values = ep_df["observation.state"].values
+                    if hasattr(state_values[0], "__len__"):
+                        states = np.stack(state_values)
+                    else:
+                        states = np.array(state_values)
+                    actions_t = torch.from_numpy(episode_data["action"]).float()
+                    states_t = torch.from_numpy(states).float()
+                    episode_data["action"] = to_delta_actions(actions_t, states_t, delta_mask).numpy()
+
+            ep_stats = compute_episode_stats(episode_data, features_to_compute)
+            all_episode_stats.append(ep_stats)
+
+    if not all_episode_stats:
+        logging.warning("No episode stats computed")
+        return dataset
+
+    new_stats = aggregate_stats(all_episode_stats)
+
+    # Merge: keep existing stats for features we didn't recompute
+    if dataset.meta.stats:
+        for key, value in dataset.meta.stats.items():
+            if key not in new_stats:
+                new_stats[key] = value
+
+    write_stats(new_stats, dataset.root)
+    dataset.meta.stats = new_stats
+
+    logging.info(f"Stats recomputed for {len(all_episode_stats)} episodes")
+    return dataset
+
+
 def convert_image_to_video_dataset(
     dataset: LeRobotDataset,
-    output_dir: Path | None = None,
+    output_dir: Path,
     repo_id: str | None = None,
     vcodec: str = "libsvtav1",
     pix_fmt: str = "yuv420p",
@@ -1569,8 +1971,8 @@ def convert_image_to_video_dataset(
 
     Args:
         dataset: The source LeRobot dataset with images
-        output_dir: Root directory where the edited dataset will be stored. If not specified, defaults to $HF_LEROBOT_HOME/repo_id. Equivalent to new_root in EditDatasetConfig.
+        output_dir: Directory to save the new video dataset
-        repo_id: Edited dataset identifier. Equivalent to new_repo_id in EditDatasetConfig.
+        repo_id: Repository ID for the new dataset (default: original_id + "_video")
         vcodec: Video codec (default: libsvtav1)
         pix_fmt: Pixel format (default: yuv420p)
         g: Group of pictures size (default: 2)
@@ -1621,7 +2023,6 @@ def convert_image_to_video_dataset(
             # Video info will be updated after episodes are encoded
 
     # Create new metadata for video dataset
-    output_dir = Path(output_dir) if output_dir is not None else HF_LEROBOT_HOME / repo_id
     new_meta = LeRobotDatasetMetadata.create(
         repo_id=repo_id,
         fps=dataset.meta.fps,

src/lerobot/datasets/lerobot_dataset.py

@@ -68,7 +68,6 @@ from lerobot.datasets.utils import (
     write_tasks,
 )
 from lerobot.datasets.video_utils import (
-    StreamingVideoEncoder,
     VideoFrame,
     concatenate_video_files,
     decode_video_frames,
@@ -76,11 +75,11 @@ from lerobot.datasets.video_utils import (
     get_safe_default_codec,
     get_video_duration_in_s,
     get_video_info,
-    resolve_vcodec,
 )
 from lerobot.utils.constants import HF_LEROBOT_HOME
 
 CODEBASE_VERSION = "v3.0"
+VALID_VIDEO_CODECS = {"h264", "hevc", "libsvtav1"}
 
 
 class LeRobotDatasetMetadata:
@@ -314,7 +313,7 @@ class LeRobotDatasetMetadata:
         if self.tasks is None:
             new_tasks = tasks
             task_indices = range(len(tasks))
-            self.tasks = pd.DataFrame({"task_index": task_indices}, index=pd.Index(tasks, name="task"))
+            self.tasks = pd.DataFrame({"task_index": task_indices}, index=tasks)
         else:
             new_tasks = [task for task in tasks if task not in self.tasks.index]
             new_task_indices = range(len(self.tasks), len(self.tasks) + len(new_tasks))
@@ -546,19 +545,12 @@ class LeRobotDatasetMetadata:
 
 
 def _encode_video_worker(
-    video_key: str,
+    video_key: str, episode_index: int, root: Path, fps: int, vcodec: str = "libsvtav1"
-    episode_index: int,
-    root: Path,
-    fps: int,
-    vcodec: str = "libsvtav1",
-    encoder_threads: int | None = None,
 ) -> Path:
     temp_path = Path(tempfile.mkdtemp(dir=root)) / f"{video_key}_{episode_index:03d}.mp4"
     fpath = DEFAULT_IMAGE_PATH.format(image_key=video_key, episode_index=episode_index, frame_index=0)
     img_dir = (root / fpath).parent
-    encode_video_frames(
+    encode_video_frames(img_dir, temp_path, fps, vcodec=vcodec, overwrite=True)
-        img_dir, temp_path, fps, vcodec=vcodec, overwrite=True, encoder_threads=encoder_threads
-    )
     shutil.rmtree(img_dir)
     return temp_path
 
@@ -578,9 +570,6 @@ class LeRobotDataset(torch.utils.data.Dataset):
         video_backend: str | None = None,
         batch_encoding_size: int = 1,
         vcodec: str = "libsvtav1",
-        streaming_encoding: bool = False,
-        encoder_queue_maxsize: int = 30,
-        encoder_threads: int | None = None,
     ):
         """
         2 modes are available for instantiating this class, depending on 2 different use cases:
@@ -664,11 +653,11 @@ class LeRobotDataset(torch.utils.data.Dataset):
         for the README).
 
         Args:
-            repo_id (str): This is the repo id that will be used to fetch the dataset.
+            repo_id (str): This is the repo id that will be used to fetch the dataset. Locally, the dataset
-            root (Path | None, optional): Local directory where the dataset will be downloaded and
+                will be stored under root/repo_id.
-                stored. If set, all dataset files will be stored directly under this path. If not set, the
+            root (Path | None, optional): Local directory to use for downloading/writing files. You can also
-                dataset files will be stored under $HF_LEROBOT_HOME/repo_id (configurable via the
+                set the LEROBOT_HOME environment variable to point to a different location. Defaults to
-                HF_LEROBOT_HOME environment variable).
+                '~/.cache/huggingface/lerobot'.
             episodes (list[int] | None, optional): If specified, this will only load episodes specified by
                 their episode_index in this list. Defaults to None.
             image_transforms (Callable | None, optional): You can pass standard v2 image transforms from
@@ -694,17 +683,12 @@ class LeRobotDataset(torch.utils.data.Dataset):
             batch_encoding_size (int, optional): Number of episodes to accumulate before batch encoding videos.
                 Set to 1 for immediate encoding (default), or higher for batched encoding. Defaults to 1.
             vcodec (str, optional): Video codec for encoding videos during recording. Options: 'h264', 'hevc',
-                'libsvtav1', 'auto', or hardware-specific codecs like 'h264_videotoolbox', 'h264_nvenc'.
+                'libsvtav1'. Defaults to 'libsvtav1'. Use 'h264' for faster encoding on systems where AV1
-                Defaults to 'libsvtav1'. Use 'auto' to auto-detect the best available hardware encoder.
+                encoding is CPU-heavy.
-            streaming_encoding (bool, optional): If True, encode video frames in real-time during capture
-                instead of writing PNG images first. This makes save_episode() near-instant. Defaults to False.
-            encoder_queue_maxsize (int, optional): Maximum number of frames to buffer per camera when using
-                streaming encoding. Defaults to 30 (~1s at 30fps).
-            encoder_threads (int | None, optional): Number of threads per encoder instance. None lets the
-                codec auto-detect (default). Lower values reduce CPU usage per encoder. Maps to 'lp' (via svtav1-params) for
-                libsvtav1 and 'threads' for h264/hevc.
         """
         super().__init__()
+        if vcodec not in VALID_VIDEO_CODECS:
+            raise ValueError(f"Invalid vcodec '{vcodec}'. Must be one of: {sorted(VALID_VIDEO_CODECS)}")
         self.repo_id = repo_id
         self.root = Path(root) if root else HF_LEROBOT_HOME / repo_id
         self.image_transforms = image_transforms
@@ -716,8 +700,7 @@ class LeRobotDataset(torch.utils.data.Dataset):
         self.delta_indices = None
         self.batch_encoding_size = batch_encoding_size
         self.episodes_since_last_encoding = 0
-        self.vcodec = resolve_vcodec(vcodec)
+        self.vcodec = vcodec
-        self._encoder_threads = encoder_threads
 
         # Unused attributes
         self.image_writer = None
@@ -725,7 +708,6 @@ class LeRobotDataset(torch.utils.data.Dataset):
         self.writer = None
         self.latest_episode = None
         self._current_file_start_frame = None  # Track the starting frame index of the current parquet file
-        self._streaming_encoder = None
 
         self.root.mkdir(exist_ok=True, parents=True)
 
@@ -747,7 +729,7 @@ class LeRobotDataset(torch.utils.data.Dataset):
             # Check if cached dataset contains all requested episodes
             if not self._check_cached_episodes_sufficient():
                 raise FileNotFoundError("Cached dataset doesn't contain all requested episodes")
-        except (FileNotFoundError, NotADirectoryError):
+        except (AssertionError, FileNotFoundError, NotADirectoryError):
             if is_valid_version(self.revision):
                 self.revision = get_safe_version(self.repo_id, self.revision)
             self.download(download_videos)
@@ -767,19 +749,6 @@ class LeRobotDataset(torch.utils.data.Dataset):
             check_delta_timestamps(self.delta_timestamps, self.fps, self.tolerance_s)
             self.delta_indices = get_delta_indices(self.delta_timestamps, self.fps)
 
-        # Initialize streaming encoder for resumed recording
-        if streaming_encoding and len(self.meta.video_keys) > 0:
-            self._streaming_encoder = StreamingVideoEncoder(
-                fps=self.meta.fps,
-                vcodec=self.vcodec,
-                pix_fmt="yuv420p",
-                g=2,
-                crf=30,
-                preset=None,
-                queue_maxsize=encoder_queue_maxsize,
-                encoder_threads=encoder_threads,
-            )
-
     def _close_writer(self) -> None:
         """Close and cleanup the parquet writer if it exists."""
         writer = getattr(self, "writer", None)
@@ -839,7 +808,7 @@ class LeRobotDataset(torch.utils.data.Dataset):
             hub_api.upload_folder(**upload_kwargs)
 
         card = create_lerobot_dataset_card(
-            tags=tags, dataset_info=self.meta.info, license=license, repo_id=self.repo_id, **card_kwargs
+            tags=tags, dataset_info=self.meta.info, license=license, **card_kwargs
         )
         card.push_to_hub(repo_id=self.repo_id, repo_type="dataset", revision=branch)
 
@@ -1135,8 +1104,6 @@ class LeRobotDataset(torch.utils.data.Dataset):
         """
         self._close_writer()
         self.meta._close_writer()
-        if self._streaming_encoder is not None:
-            self._streaming_encoder.close()
 
     def create_episode_buffer(self, episode_index: int | None = None) -> dict:
         current_ep_idx = self.meta.total_episodes if episode_index is None else episode_index
@@ -1191,13 +1158,6 @@ class LeRobotDataset(torch.utils.data.Dataset):
         self.episode_buffer["timestamp"].append(timestamp)
         self.episode_buffer["task"].append(frame.pop("task"))  # Remove task from frame after processing
 
-        # Start streaming encoder on first frame of episode (once, before iterating keys)
-        if frame_index == 0 and self._streaming_encoder is not None:
-            self._streaming_encoder.start_episode(
-                video_keys=list(self.meta.video_keys),
-                temp_dir=self.root,
-            )
-
         # Add frame features to episode_buffer
         for key in frame:
             if key not in self.features:
@@ -1205,10 +1165,7 @@ class LeRobotDataset(torch.utils.data.Dataset):
                     f"An element of the frame is not in the features. '{key}' not in '{self.features.keys()}'."
                 )
 
-            if self.features[key]["dtype"] == "video" and self._streaming_encoder is not None:
+            if self.features[key]["dtype"] in ["image", "video"]:
-                self._streaming_encoder.feed_frame(key, frame[key])
-                self.episode_buffer[key].append(None)  # Placeholder (video keys are skipped in parquet)
-            elif self.features[key]["dtype"] in ["image", "video"]:
                 img_path = self._get_image_file_path(
                     episode_index=self.episode_buffer["episode_index"], image_key=key, frame_index=frame_index
                 )
@@ -1269,38 +1226,13 @@ class LeRobotDataset(torch.utils.data.Dataset):
 
         # Wait for image writer to end, so that episode stats over images can be computed
         self._wait_image_writer()
-
+        ep_stats = compute_episode_stats(episode_buffer, self.features)
-        has_video_keys = len(self.meta.video_keys) > 0
-        use_streaming = self._streaming_encoder is not None and has_video_keys
-        use_batched_encoding = self.batch_encoding_size > 1
-
-        if use_streaming:
-            # Compute stats for non-video features only (video stats come from encoder)
-            non_video_buffer = {
-                k: v
-                for k, v in episode_buffer.items()
-                if self.features.get(k, {}).get("dtype") not in ("video",)
-            }
-            non_video_features = {k: v for k, v in self.features.items() if v["dtype"] != "video"}
-            ep_stats = compute_episode_stats(non_video_buffer, non_video_features)
-        else:
-            ep_stats = compute_episode_stats(episode_buffer, self.features)
 
         ep_metadata = self._save_episode_data(episode_buffer)
+        has_video_keys = len(self.meta.video_keys) > 0
+        use_batched_encoding = self.batch_encoding_size > 1
 
-        if use_streaming:
+        if has_video_keys and not use_batched_encoding:
-            # Finish streaming encoding and collect results
-            streaming_results = self._streaming_encoder.finish_episode()
-            for video_key in self.meta.video_keys:
-                temp_path, video_stats = streaming_results[video_key]
-                if video_stats is not None:
-                    # Format stats same as compute_episode_stats: normalize to [0,1], reshape to (C,1,1)
-                    ep_stats[video_key] = {
-                        k: v if k == "count" else np.squeeze(v.reshape(1, -1, 1, 1) / 255.0, axis=0)
-                        for k, v in video_stats.items()
-                    }
-                ep_metadata.update(self._save_episode_video(video_key, episode_index, temp_path=temp_path))
-        elif has_video_keys and not use_batched_encoding:
             num_cameras = len(self.meta.video_keys)
             if parallel_encoding and num_cameras > 1:
                 # TODO(Steven): Ideally we would like to control the number of threads per encoding such that:
@@ -1314,7 +1246,6 @@ class LeRobotDataset(torch.utils.data.Dataset):
                             self.root,
                             self.fps,
                             self.vcodec,
-                            self._encoder_threads,
                         ): video_key
                         for video_key in self.meta.video_keys
                     }
@@ -1583,10 +1514,6 @@ class LeRobotDataset(torch.utils.data.Dataset):
         return metadata
 
     def clear_episode_buffer(self, delete_images: bool = True) -> None:
-        # Cancel streaming encoder if active
-        if self._streaming_encoder is not None:
-            self._streaming_encoder.cancel_episode()
-
         # Clean up image files for the current episode buffer
         if delete_images:
             # Wait for the async image writer to finish
@@ -1634,9 +1561,7 @@ class LeRobotDataset(torch.utils.data.Dataset):
         Note: `encode_video_frames` is a blocking call. Making it asynchronous shouldn't speedup encoding,
         since video encoding with ffmpeg is already using multithreading.
         """
-        return _encode_video_worker(
+        return _encode_video_worker(video_key, episode_index, self.root, self.fps, self.vcodec)
-            video_key, episode_index, self.root, self.fps, self.vcodec, self._encoder_threads
-        )
 
     @classmethod
     def create(
@@ -1653,13 +1578,10 @@ class LeRobotDataset(torch.utils.data.Dataset):
         video_backend: str | None = None,
         batch_encoding_size: int = 1,
         vcodec: str = "libsvtav1",
-        metadata_buffer_size: int = 10,
-        streaming_encoding: bool = False,
-        encoder_queue_maxsize: int = 30,
-        encoder_threads: int | None = None,
     ) -> "LeRobotDataset":
         """Create a LeRobot Dataset from scratch in order to record data."""
-        vcodec = resolve_vcodec(vcodec)
+        if vcodec not in VALID_VIDEO_CODECS:
+            raise ValueError(f"Invalid vcodec '{vcodec}'. Must be one of: {sorted(VALID_VIDEO_CODECS)}")
         obj = cls.__new__(cls)
         obj.meta = LeRobotDatasetMetadata.create(
             repo_id=repo_id,
@@ -1668,7 +1590,6 @@ class LeRobotDataset(torch.utils.data.Dataset):
             features=features,
             root=root,
             use_videos=use_videos,
-            metadata_buffer_size=metadata_buffer_size,
         )
         obj.repo_id = obj.meta.repo_id
         obj.root = obj.meta.root
@@ -1678,7 +1599,6 @@ class LeRobotDataset(torch.utils.data.Dataset):
         obj.batch_encoding_size = batch_encoding_size
         obj.episodes_since_last_encoding = 0
         obj.vcodec = vcodec
-        obj._encoder_threads = encoder_threads
 
         if image_writer_processes or image_writer_threads:
             obj.start_image_writer(image_writer_processes, image_writer_threads)
@@ -1700,22 +1620,6 @@ class LeRobotDataset(torch.utils.data.Dataset):
         obj._lazy_loading = False
         obj._recorded_frames = 0
         obj._writer_closed_for_reading = False
-
-        # Initialize streaming encoder
-        if streaming_encoding and len(obj.meta.video_keys) > 0:
-            obj._streaming_encoder = StreamingVideoEncoder(
-                fps=fps,
-                vcodec=vcodec,
-                pix_fmt="yuv420p",
-                g=2,
-                crf=30,
-                preset=None,
-                queue_maxsize=encoder_queue_maxsize,
-                encoder_threads=encoder_threads,
-            )
-        else:
-            obj._streaming_encoder = None
-
         return obj
 
 
@@ -1771,12 +1675,11 @@ class MultiLeRobotDataset(torch.utils.data.Dataset):
             )
         for repo_id, ds in zip(self.repo_ids, self._datasets, strict=True):
             extra_keys = set(ds.features).difference(intersection_features)
-            if extra_keys:
+            logging.warning(
-                logging.warning(
+                f"keys {extra_keys} of {repo_id} were disabled as they are not contained in all the "
-                    f"keys {extra_keys} of {repo_id} were disabled as they are not contained in all the "
+                "other datasets."
-                    "other datasets."
+            )
-                )
+            self.disabled_features.update(extra_keys)
-                self.disabled_features.update(extra_keys)
 
         self.image_transforms = image_transforms
         self.delta_timestamps = delta_timestamps

src/lerobot/datasets/utils.py

@@ -122,9 +122,19 @@ def load_nested_dataset(
         raise FileNotFoundError(f"Provided directory does not contain any parquet file: {pq_dir}")
 
     with SuppressProgressBars():
-        # We use .from_parquet() memory-mapped loading for efficiency
+        # When no filtering needed, Dataset uses memory-mapped loading for efficiency
-        filters = pa_ds.field("episode_index").isin(episodes) if episodes is not None else None
+        # PyArrow loads the entire dataset into memory
-        return Dataset.from_parquet([str(path) for path in paths], filters=filters, features=features)
+        if episodes is None:
+            return Dataset.from_parquet([str(path) for path in paths], features=features)
+
+        arrow_dataset = pa_ds.dataset(paths, format="parquet")
+        filter_expr = pa_ds.field("episode_index").isin(episodes)
+        table = arrow_dataset.to_table(filter=filter_expr)
+
+        if features is not None:
+            table = table.cast(features.arrow_schema)
+
+        return Dataset(table)
 
 
 def get_parquet_num_frames(parquet_path: str | Path) -> int:
@@ -341,7 +351,6 @@ def write_tasks(tasks: pandas.DataFrame, local_dir: Path) -> None:
 
 def load_tasks(local_dir: Path) -> pandas.DataFrame:
     tasks = pd.read_parquet(local_dir / DEFAULT_TASKS_PATH)
-    tasks.index.name = "task"
     return tasks
 
 

src/lerobot/datasets/v30/convert_dataset_v21_to_v30.py

@@ -36,11 +36,8 @@ Convert a local dataset (works in place):
 ```bash
 python src/lerobot/datasets/v30/convert_dataset_v21_to_v30.py \
     --repo-id=lerobot/pusht \
-    --root=/path/to/local/dataset/directory \
+    --root=/path/to/local/dataset/directory
     --push-to-hub=false
-
-N.B. Path semantics (v2): --root is the exact dataset folder containing
-meta/, data/, videos/. When omitted, defaults to $HF_LEROBOT_HOME/{repo_id}.
 ```
 
 """
@@ -108,7 +105,7 @@ episodes.jsonl
 {"episode_index": 1, "tasks": ["Put the blue block in the green bowl"], "length": 266}
 
 NEW
-meta/episodes/chunk-000/file_000.parquet
+meta/episodes/chunk-000/episodes_000.parquet
 episode_index | video_chunk_index | video_file_index | data_chunk_index | data_file_index | tasks | length
 -------------------------
 OLD
@@ -116,16 +113,15 @@ tasks.jsonl
 {"task_index": 1, "task": "Put the blue block in the green bowl"}
 
 NEW
-meta/tasks.parquet
+meta/tasks/chunk-000/file_000.parquet
 task_index | task
 -------------------------
 OLD
 episodes_stats.jsonl
-{"episode_index": 1, "stats": {"feature_name": {"min": ..., "max": ..., "mean": ..., "std": ..., "count": ...}}}
 
 NEW
-meta/episodes/chunk-000/file_000.parquet
+meta/episodes_stats/chunk-000/file_000.parquet
-episode_index | feature_name/min | feature_name/max | feature_name/mean | feature_name/std | feature_name/count
+episode_index | mean | std | min | max
 -------------------------
 UPDATE
 meta/info.json
@@ -174,7 +170,7 @@ def convert_tasks(root, new_root):
     tasks, _ = legacy_load_tasks(root)
     task_indices = tasks.keys()
     task_strings = tasks.values()
-    df_tasks = pd.DataFrame({"task_index": task_indices}, index=pd.Index(task_strings, name="task"))
+    df_tasks = pd.DataFrame({"task_index": task_indices}, index=task_strings)
     write_tasks(df_tasks, new_root)
 
 
@@ -205,6 +201,7 @@ def convert_data(root: Path, new_root: Path, data_file_size_in_mb: int):
 
     image_keys = get_image_keys(root)
 
+    ep_idx = 0
     chunk_idx = 0
     file_idx = 0
     size_in_mb = 0
@@ -214,24 +211,9 @@ def convert_data(root: Path, new_root: Path, data_file_size_in_mb: int):
 
     logging.info(f"Converting data files from {len(ep_paths)} episodes")
 
-    for ep_idx, ep_path in enumerate(tqdm.tqdm(ep_paths, desc="convert data files")):
+    for ep_path in tqdm.tqdm(ep_paths, desc="convert data files"):
         ep_size_in_mb = get_parquet_file_size_in_mb(ep_path)
         ep_num_frames = get_parquet_num_frames(ep_path)
-
-        # Check if we need to start a new file BEFORE creating metadata
-        if size_in_mb + ep_size_in_mb >= data_file_size_in_mb and len(paths_to_cat) > 0:
-            # Write the accumulated data files
-            concat_data_files(paths_to_cat, new_root, chunk_idx, file_idx, image_keys)
-
-            # Move to next file
-            chunk_idx, file_idx = update_chunk_file_indices(chunk_idx, file_idx, DEFAULT_CHUNK_SIZE)
-
-            # Reset for the next file
-            size_in_mb = 0
-            num_frames += ep_num_frames  # Still need to accumulate total frames
-            paths_to_cat = []
-
-        # Now create metadata with correct chunk/file indices
         ep_metadata = {
             "episode_index": ep_idx,
             "data/chunk_index": chunk_idx,
@@ -242,7 +224,20 @@ def convert_data(root: Path, new_root: Path, data_file_size_in_mb: int):
         size_in_mb += ep_size_in_mb
         num_frames += ep_num_frames
         episodes_metadata.append(ep_metadata)
-        paths_to_cat.append(ep_path)
+        ep_idx += 1
+
+        if size_in_mb < data_file_size_in_mb:
+            paths_to_cat.append(ep_path)
+            continue
+
+        if paths_to_cat:
+            concat_data_files(paths_to_cat, new_root, chunk_idx, file_idx, image_keys)
+
+        # Reset for the next file
+        size_in_mb = ep_size_in_mb
+        paths_to_cat = [ep_path]
+
+        chunk_idx, file_idx = update_chunk_file_indices(chunk_idx, file_idx, DEFAULT_CHUNK_SIZE)
 
     # Write remaining data if any
     if paths_to_cat:
@@ -474,7 +469,7 @@ def convert_dataset(
 
     # Set root based on whether local dataset path is provided
     use_local_dataset = False
-    root = HF_LEROBOT_HOME / repo_id if root is None else Path(root)
+    root = HF_LEROBOT_HOME / repo_id if root is None else Path(root) / repo_id
     if root.exists():
         validate_local_dataset_version(root)
         use_local_dataset = True
@@ -534,7 +529,7 @@ if __name__ == "__main__":
         type=str,
         required=True,
         help="Repository identifier on Hugging Face: a community or a user name `/` the name of the dataset "
-        "(e.g. `lerobot/pusht`, `<USER>/aloha_sim_insertion_human`).",
+        "(e.g. `lerobot/pusht`, `cadene/aloha_sim_insertion_human`).",
     )
     parser.add_argument(
         "--branch",
@@ -558,7 +553,7 @@ if __name__ == "__main__":
         "--root",
         type=str,
         default=None,
-        help="Local directory to use for downloading/writing the dataset. Defaults to $HF_LEROBOT_HOME/repo_id.",
+        help="Local directory to use for downloading/writing the dataset.",
     )
     parser.add_argument(
         "--push-to-hub",

src/lerobot/datasets/video_utils.py

@@ -13,106 +13,25 @@
 # 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.
-import contextlib
 import glob
 import importlib
 import logging
-import queue
 import shutil
 import tempfile
-import threading
 import warnings
 from dataclasses import dataclass, field
-from fractions import Fraction
 from pathlib import Path
 from threading import Lock
 from typing import Any, ClassVar
 
 import av
 import fsspec
-import numpy as np
 import pyarrow as pa
 import torch
 import torchvision
 from datasets.features.features import register_feature
 from PIL import Image
 
-# List of hardware encoders to probe for auto-selection. Availability depends on the platform and FFmpeg build.
-# Determines the order of preference for auto-selection when vcodec="auto" is used.
-HW_ENCODERS = [
-    "h264_videotoolbox",  # macOS
-    "hevc_videotoolbox",  # macOS
-    "h264_nvenc",  # NVIDIA GPU
-    "hevc_nvenc",  # NVIDIA GPU
-    "h264_vaapi",  # Linux Intel/AMD
-    "h264_qsv",  # Intel Quick Sync
-]
-
-VALID_VIDEO_CODECS = {"h264", "hevc", "libsvtav1", "auto"} | set(HW_ENCODERS)
-
-
-def _get_codec_options(
-    vcodec: str,
-    g: int | None = 2,
-    crf: int | None = 30,
-    preset: int | None = None,
-) -> dict:
-    """Build codec-specific options dict for video encoding."""
-    options = {}
-
-    # GOP size (keyframe interval) - supported by VideoToolbox and software encoders
-    if g is not None and (vcodec in ("h264_videotoolbox", "hevc_videotoolbox") or vcodec not in HW_ENCODERS):
-        options["g"] = str(g)
-
-    # Quality control (codec-specific parameter names)
-    if crf is not None:
-        if vcodec in ("h264", "hevc", "libsvtav1"):
-            options["crf"] = str(crf)
-        elif vcodec in ("h264_videotoolbox", "hevc_videotoolbox"):
-            quality = max(1, min(100, int(100 - crf * 2)))
-            options["q:v"] = str(quality)
-        elif vcodec in ("h264_nvenc", "hevc_nvenc"):
-            options["rc"] = "constqp"
-            options["qp"] = str(crf)
-        elif vcodec in ("h264_vaapi",):
-            options["qp"] = str(crf)
-        elif vcodec in ("h264_qsv",):
-            options["global_quality"] = str(crf)
-
-    # Preset (only for libsvtav1)
-    if vcodec == "libsvtav1":
-        options["preset"] = str(preset) if preset is not None else "12"
-
-    return options
-
-
-def detect_available_hw_encoders() -> list[str]:
-    """Probe PyAV/FFmpeg for available hardware video encoders."""
-    available = []
-    for codec_name in HW_ENCODERS:
-        try:
-            av.codec.Codec(codec_name, "w")
-            available.append(codec_name)
-        except Exception:  # nosec B110
-            pass  # nosec B110
-    return available
-
-
-def resolve_vcodec(vcodec: str) -> str:
-    """Validate vcodec and resolve 'auto' to best available HW encoder, fallback to libsvtav1."""
-    if vcodec not in VALID_VIDEO_CODECS:
-        raise ValueError(f"Invalid vcodec '{vcodec}'. Must be one of: {sorted(VALID_VIDEO_CODECS)}")
-    if vcodec != "auto":
-        logging.info(f"Using video codec: {vcodec}")
-        return vcodec
-    available = detect_available_hw_encoders()
-    for encoder in HW_ENCODERS:
-        if encoder in available:
-            logging.info(f"Auto-selected video codec: {encoder}")
-            return encoder
-    logging.info("No hardware encoder available, falling back to software encoder 'libsvtav1'")
-    return "libsvtav1"
-
 
 def get_safe_default_codec():
     if importlib.util.find_spec("torchcodec"):
@@ -227,17 +146,16 @@ def decode_video_frames_torchvision(
     min_, argmin_ = dist.min(1)
 
     is_within_tol = min_ < tolerance_s
-    if not is_within_tol.all():
+    assert is_within_tol.all(), (
-        raise FrameTimestampError(
+        f"One or several query timestamps unexpectedly violate the tolerance ({min_[~is_within_tol]} > {tolerance_s=})."
-            f"One or several query timestamps unexpectedly violate the tolerance ({min_[~is_within_tol]} > {tolerance_s=})."
+        "It means that the closest frame that can be loaded from the video is too far away in time."
-            " It means that the closest frame that can be loaded from the video is too far away in time."
+        "This might be due to synchronization issues with timestamps during data collection."
-            " This might be due to synchronization issues with timestamps during data collection."
+        "To be safe, we advise to ignore this item during training."
-            " To be safe, we advise to ignore this item during training."
+        f"\nqueried timestamps: {query_ts}"
-            f"\nqueried timestamps: {query_ts}"
+        f"\nloaded timestamps: {loaded_ts}"
-            f"\nloaded timestamps: {loaded_ts}"
+        f"\nvideo: {video_path}"
-            f"\nvideo: {video_path}"
+        f"\nbackend: {backend}"
-            f"\nbackend: {backend}"
+    )
-        )
 
     # get closest frames to the query timestamps
     closest_frames = torch.stack([loaded_frames[idx] for idx in argmin_])
@@ -249,11 +167,7 @@ def decode_video_frames_torchvision(
     # convert to the pytorch format which is float32 in [0,1] range (and channel first)
     closest_frames = closest_frames.type(torch.float32) / 255
 
-    if len(timestamps) != len(closest_frames):
+    assert len(timestamps) == len(closest_frames)
-        raise FrameTimestampError(
-            f"Number of retrieved frames ({len(closest_frames)}) does not match "
-            f"number of queried timestamps ({len(timestamps)})"
-        )
     return closest_frames
 
 
@@ -358,16 +272,15 @@ def decode_video_frames_torchcodec(
     min_, argmin_ = dist.min(1)
 
     is_within_tol = min_ < tolerance_s
-    if not is_within_tol.all():
+    assert is_within_tol.all(), (
-        raise FrameTimestampError(
+        f"One or several query timestamps unexpectedly violate the tolerance ({min_[~is_within_tol]} > {tolerance_s=})."
-            f"One or several query timestamps unexpectedly violate the tolerance ({min_[~is_within_tol]} > {tolerance_s=})."
+        "It means that the closest frame that can be loaded from the video is too far away in time."
-            " It means that the closest frame that can be loaded from the video is too far away in time."
+        "This might be due to synchronization issues with timestamps during data collection."
-            " This might be due to synchronization issues with timestamps during data collection."
+        "To be safe, we advise to ignore this item during training."
-            " To be safe, we advise to ignore this item during training."
+        f"\nqueried timestamps: {query_ts}"
-            f"\nqueried timestamps: {query_ts}"
+        f"\nloaded timestamps: {loaded_ts}"
-            f"\nloaded timestamps: {loaded_ts}"
+        f"\nvideo: {video_path}"
-            f"\nvideo: {video_path}"
+    )
-        )
 
     # get closest frames to the query timestamps
     closest_frames = torch.stack([loaded_frames[idx] for idx in argmin_])
@@ -396,13 +309,14 @@ def encode_video_frames(
     g: int | None = 2,
     crf: int | None = 30,
     fast_decode: int = 0,
-    log_level: int | None = av.logging.WARNING,
+    log_level: int | None = av.logging.ERROR,
     overwrite: bool = False,
     preset: int | None = None,
-    encoder_threads: int | None = None,
 ) -> None:
     """More info on ffmpeg arguments tuning on `benchmark/video/README.md`"""
-    vcodec = resolve_vcodec(vcodec)
+    # Check encoder availability
+    if vcodec not in ["h264", "hevc", "libsvtav1"]:
+        raise ValueError(f"Unsupported video codec: {vcodec}. Supported codecs are: h264, hevc, libsvtav1.")
 
     video_path = Path(video_path)
     imgs_dir = Path(imgs_dir)
@@ -433,22 +347,21 @@ def encode_video_frames(
         width, height = dummy_image.size
 
     # Define video codec options
-    video_options = _get_codec_options(vcodec, g, crf, preset)
+    video_options = {}
+
+    if g is not None:
+        video_options["g"] = str(g)
+
+    if crf is not None:
+        video_options["crf"] = str(crf)
 
     if fast_decode:
         key = "svtav1-params" if vcodec == "libsvtav1" else "tune"
         value = f"fast-decode={fast_decode}" if vcodec == "libsvtav1" else "fastdecode"
         video_options[key] = value
 
-    if encoder_threads is not None:
+    if vcodec == "libsvtav1":
-        if vcodec == "libsvtav1":
+        video_options["preset"] = str(preset) if preset is not None else "12"
-            lp_param = f"lp={encoder_threads}"
-            if "svtav1-params" in video_options:
-                video_options["svtav1-params"] += f":{lp_param}"
-            else:
-                video_options["svtav1-params"] = lp_param
-        else:
-            video_options["threads"] = str(encoder_threads)
 
     # Set logging level
     if log_level is not None:
@@ -567,348 +480,6 @@ def concatenate_video_files(
     Path(tmp_concatenate_path).unlink()
 
 
-class _CameraEncoderThread(threading.Thread):
-    """A thread that encodes video frames streamed via a queue into an MP4 file.
-
-    One instance is created per camera per episode. Frames are received as numpy arrays
-    from the main thread, encoded in real-time using PyAV (which releases the GIL during
-    encoding), and written to disk. Stats are computed incrementally using
-    RunningQuantileStats and returned via result_queue.
-    """
-
-    def __init__(
-        self,
-        video_path: Path,
-        fps: int,
-        vcodec: str,
-        pix_fmt: str,
-        g: int | None,
-        crf: int | None,
-        preset: int | None,
-        frame_queue: queue.Queue,
-        result_queue: queue.Queue,
-        stop_event: threading.Event,
-        encoder_threads: int | None = None,
-    ):
-        super().__init__(daemon=True)
-        self.video_path = video_path
-        self.fps = fps
-        self.vcodec = vcodec
-        self.pix_fmt = pix_fmt
-        self.g = g
-        self.crf = crf
-        self.preset = preset
-        self.frame_queue = frame_queue
-        self.result_queue = result_queue
-        self.stop_event = stop_event
-        self.encoder_threads = encoder_threads
-
-    def run(self) -> None:
-        from lerobot.datasets.compute_stats import RunningQuantileStats, auto_downsample_height_width
-
-        container = None
-        output_stream = None
-        stats_tracker = RunningQuantileStats()
-        frame_count = 0
-
-        try:
-            logging.getLogger("libav").setLevel(av.logging.WARNING)
-
-            while True:
-                try:
-                    frame_data = self.frame_queue.get(timeout=1)
-                except queue.Empty:
-                    if self.stop_event.is_set():
-                        break
-                    continue
-
-                if frame_data is None:
-                    # Sentinel: flush and close
-                    break
-
-                # Ensure HWC uint8 numpy array
-                if isinstance(frame_data, np.ndarray):
-                    if frame_data.ndim == 3 and frame_data.shape[0] == 3:
-                        # CHW -> HWC
-                        frame_data = frame_data.transpose(1, 2, 0)
-                    if frame_data.dtype != np.uint8:
-                        frame_data = (frame_data * 255).astype(np.uint8)
-
-                # Open container on first frame (to get width/height)
-                if container is None:
-                    height, width = frame_data.shape[:2]
-                    video_options = _get_codec_options(self.vcodec, self.g, self.crf, self.preset)
-                    if self.encoder_threads is not None:
-                        if self.vcodec == "libsvtav1":
-                            lp_param = f"lp={self.encoder_threads}"
-                            if "svtav1-params" in video_options:
-                                video_options["svtav1-params"] += f":{lp_param}"
-                            else:
-                                video_options["svtav1-params"] = lp_param
-                        else:
-                            video_options["threads"] = str(self.encoder_threads)
-                    Path(self.video_path).parent.mkdir(parents=True, exist_ok=True)
-                    container = av.open(str(self.video_path), "w")
-                    output_stream = container.add_stream(self.vcodec, self.fps, options=video_options)
-                    output_stream.pix_fmt = self.pix_fmt
-                    output_stream.width = width
-                    output_stream.height = height
-                    output_stream.time_base = Fraction(1, self.fps)
-
-                # Encode frame with explicit timestamps
-                pil_img = Image.fromarray(frame_data)
-                video_frame = av.VideoFrame.from_image(pil_img)
-                video_frame.pts = frame_count
-                video_frame.time_base = Fraction(1, self.fps)
-                packet = output_stream.encode(video_frame)
-                if packet:
-                    container.mux(packet)
-
-                # Update stats with downsampled frame (per-channel stats like compute_episode_stats)
-                img_chw = frame_data.transpose(2, 0, 1)  # HWC -> CHW
-                img_downsampled = auto_downsample_height_width(img_chw)
-                # Reshape CHW to (H*W, C) for per-channel stats
-                channels = img_downsampled.shape[0]
-                img_for_stats = img_downsampled.transpose(1, 2, 0).reshape(-1, channels)
-                stats_tracker.update(img_for_stats)
-
-                frame_count += 1
-
-            # Flush encoder
-            if output_stream is not None:
-                packet = output_stream.encode()
-                if packet:
-                    container.mux(packet)
-
-            if container is not None:
-                container.close()
-
-            av.logging.restore_default_callback()
-
-            # Get stats and put on result queue
-            if frame_count >= 2:
-                stats = stats_tracker.get_statistics()
-                self.result_queue.put(("ok", stats))
-            else:
-                self.result_queue.put(("ok", None))
-
-        except Exception as e:
-            logging.error(f"Encoder thread error: {e}")
-            if container is not None:
-                with contextlib.suppress(Exception):
-                    container.close()
-            self.result_queue.put(("error", str(e)))
-
-
-class StreamingVideoEncoder:
-    """Manages per-camera encoder threads for real-time video encoding during recording.
-
-    Instead of writing frames as PNG images and then encoding to MP4 at episode end,
-    this class streams frames directly to encoder threads, eliminating the
-    PNG round-trip and making save_episode() near-instant.
-
-    Uses threading instead of multiprocessing to avoid the overhead of pickling large
-    numpy arrays through multiprocessing.Queue. PyAV's encode() releases the GIL,
-    so encoding runs in parallel with the main recording loop.
-    """
-
-    def __init__(
-        self,
-        fps: int,
-        vcodec: str = "libsvtav1",
-        pix_fmt: str = "yuv420p",
-        g: int | None = 2,
-        crf: int | None = 30,
-        preset: int | None = None,
-        queue_maxsize: int = 30,
-        encoder_threads: int | None = None,
-    ):
-        self.fps = fps
-        self.vcodec = resolve_vcodec(vcodec)
-        self.pix_fmt = pix_fmt
-        self.g = g
-        self.crf = crf
-        self.preset = preset
-        self.queue_maxsize = queue_maxsize
-        self.encoder_threads = encoder_threads
-
-        self._frame_queues: dict[str, queue.Queue] = {}
-        self._result_queues: dict[str, queue.Queue] = {}
-        self._threads: dict[str, _CameraEncoderThread] = {}
-        self._stop_events: dict[str, threading.Event] = {}
-        self._video_paths: dict[str, Path] = {}
-        self._dropped_frames: dict[str, int] = {}
-        self._episode_active = False
-
-    def start_episode(self, video_keys: list[str], temp_dir: Path) -> None:
-        """Start encoder threads for a new episode.
-
-        Args:
-            video_keys: List of video feature keys (e.g. ["observation.images.laptop"])
-            temp_dir: Base directory for temporary MP4 files
-        """
-        if self._episode_active:
-            self.cancel_episode()
-
-        self._dropped_frames.clear()
-
-        for video_key in video_keys:
-            frame_queue: queue.Queue = queue.Queue(maxsize=self.queue_maxsize)
-            result_queue: queue.Queue = queue.Queue(maxsize=1)
-            stop_event = threading.Event()
-
-            temp_video_dir = Path(tempfile.mkdtemp(dir=temp_dir))
-            video_path = temp_video_dir / f"{video_key.replace('/', '_')}_streaming.mp4"
-
-            encoder_thread = _CameraEncoderThread(
-                video_path=video_path,
-                fps=self.fps,
-                vcodec=self.vcodec,
-                pix_fmt=self.pix_fmt,
-                g=self.g,
-                crf=self.crf,
-                preset=self.preset,
-                frame_queue=frame_queue,
-                result_queue=result_queue,
-                stop_event=stop_event,
-                encoder_threads=self.encoder_threads,
-            )
-            encoder_thread.start()
-
-            self._frame_queues[video_key] = frame_queue
-            self._result_queues[video_key] = result_queue
-            self._threads[video_key] = encoder_thread
-            self._stop_events[video_key] = stop_event
-            self._video_paths[video_key] = video_path
-
-        self._episode_active = True
-
-    def feed_frame(self, video_key: str, image: np.ndarray) -> None:
-        """Feed a frame to the encoder for a specific camera.
-
-        A copy of the image is made before enqueueing to prevent race conditions
-        with camera drivers that may reuse buffers. If the encoder queue is full
-        (encoder can't keep up), the frame is dropped with a warning instead of
-        crashing the recording session.
-
-        Args:
-            video_key: The video feature key
-            image: numpy array in (H,W,C) or (C,H,W) format, uint8 or float
-
-        Raises:
-            RuntimeError: If the encoder thread has crashed
-        """
-        if not self._episode_active:
-            raise RuntimeError("No active episode. Call start_episode() first.")
-
-        thread = self._threads[video_key]
-        if not thread.is_alive():
-            # Check for error
-            try:
-                status, msg = self._result_queues[video_key].get_nowait()
-                if status == "error":
-                    raise RuntimeError(f"Encoder thread for {video_key} crashed: {msg}")
-            except queue.Empty:
-                pass
-            raise RuntimeError(f"Encoder thread for {video_key} is not alive")
-
-        try:
-            self._frame_queues[video_key].put(image.copy(), timeout=0.1)
-        except queue.Full:
-            self._dropped_frames[video_key] = self._dropped_frames.get(video_key, 0) + 1
-            count = self._dropped_frames[video_key]
-            # Log periodically to avoid spam (1st, then every 10th)
-            if count == 1 or count % 10 == 0:
-                logging.warning(
-                    f"Encoder queue full for {video_key}, dropped {count} frame(s). "
-                    f"Consider using vcodec='auto' for hardware encoding or increasing encoder_queue_maxsize."
-                )
-
-    def finish_episode(self) -> dict[str, tuple[Path, dict | None]]:
-        """Finish encoding the current episode.
-
-        Sends sentinel values, waits for encoder threads to complete,
-        and collects results.
-
-        Returns:
-            Dict mapping video_key to (mp4_path, stats_dict_or_None)
-        """
-        if not self._episode_active:
-            raise RuntimeError("No active episode to finish.")
-
-        results = {}
-
-        # Report dropped frames
-        for video_key, count in self._dropped_frames.items():
-            if count > 0:
-                logging.warning(f"Episode finished with {count} dropped frame(s) for {video_key}.")
-
-        # Send sentinel to all queues
-        for video_key in self._frame_queues:
-            self._frame_queues[video_key].put(None)
-
-        # Wait for all threads and collect results
-        for video_key in self._threads:
-            self._threads[video_key].join(timeout=120)
-            if self._threads[video_key].is_alive():
-                logging.error(f"Encoder thread for {video_key} did not finish in time")
-                self._stop_events[video_key].set()
-                self._threads[video_key].join(timeout=5)
-                results[video_key] = (self._video_paths[video_key], None)
-                continue
-
-            try:
-                status, data = self._result_queues[video_key].get(timeout=5)
-                if status == "error":
-                    raise RuntimeError(f"Encoder thread for {video_key} failed: {data}")
-                results[video_key] = (self._video_paths[video_key], data)
-            except queue.Empty:
-                logging.error(f"No result from encoder thread for {video_key}")
-                results[video_key] = (self._video_paths[video_key], None)
-
-        self._cleanup()
-        self._episode_active = False
-        return results
-
-    def cancel_episode(self) -> None:
-        """Cancel the current episode, stopping encoder threads and cleaning up."""
-        if not self._episode_active:
-            return
-
-        # Signal all threads to stop
-        for video_key in self._stop_events:
-            self._stop_events[video_key].set()
-
-        # Wait for threads to finish
-        for video_key in self._threads:
-            self._threads[video_key].join(timeout=5)
-
-            # Clean up temp MP4 files
-            video_path = self._video_paths.get(video_key)
-            if video_path is not None and video_path.exists():
-                shutil.rmtree(str(video_path.parent), ignore_errors=True)
-
-        self._cleanup()
-        self._episode_active = False
-
-    def close(self) -> None:
-        """Close the encoder, canceling any in-progress episode."""
-        if self._episode_active:
-            self.cancel_episode()
-
-    def _cleanup(self) -> None:
-        """Clean up queues and thread tracking dicts."""
-        for q in self._frame_queues.values():
-            with contextlib.suppress(Exception):
-                while not q.empty():
-                    q.get_nowait()
-        self._frame_queues.clear()
-        self._result_queues.clear()
-        self._threads.clear()
-        self._stop_events.clear()
-        self._video_paths.clear()
-
-
 @dataclass
 class VideoFrame:
     # TODO(rcadene, lhoestq): move to Hugging Face `datasets` repo
@@ -943,7 +514,7 @@ with warnings.catch_warnings():
 
 def get_audio_info(video_path: Path | str) -> dict:
     # Set logging level
-    logging.getLogger("libav").setLevel(av.logging.WARNING)
+    logging.getLogger("libav").setLevel(av.logging.ERROR)
 
     # Getting audio stream information
     audio_info = {}
@@ -975,7 +546,7 @@ def get_audio_info(video_path: Path | str) -> dict:
 
 def get_video_info(video_path: Path | str) -> dict:
     # Set logging level
-    logging.getLogger("libav").setLevel(av.logging.WARNING)
+    logging.getLogger("libav").setLevel(av.logging.ERROR)
 
     # Getting video stream information
     video_info = {}
@@ -1061,15 +632,8 @@ class VideoEncodingManager:
         return self
 
     def __exit__(self, exc_type, exc_val, exc_tb):
-        streaming_encoder = getattr(self.dataset, "_streaming_encoder", None)
+        # Handle any remaining episodes that haven't been batch encoded
-
+        if self.dataset.episodes_since_last_encoding > 0:
-        if streaming_encoder is not None:
-            # Handle streaming encoder cleanup
-            if exc_type is not None:
-                streaming_encoder.cancel_episode()
-            streaming_encoder.close()
-        elif self.dataset.episodes_since_last_encoding > 0:
-            # Handle any remaining episodes that haven't been batch encoded
             if exc_type is not None:
                 logging.info("Exception occurred. Encoding remaining episodes before exit...")
             else:
@@ -1086,8 +650,8 @@ class VideoEncodingManager:
         # Finalize the dataset to properly close all writers
         self.dataset.finalize()
 
-        # Clean up episode images if recording was interrupted (only for non-streaming mode)
+        # Clean up episode images if recording was interrupted
-        if exc_type is not None and streaming_encoder is None:
+        if exc_type is not None:
             interrupted_episode_index = self.dataset.num_episodes
             for key in self.dataset.meta.video_keys:
                 img_dir = self.dataset._get_image_file_path(
@@ -1101,12 +665,14 @@ class VideoEncodingManager:
 
         # Clean up any remaining images directory if it's empty
         img_dir = self.dataset.root / "images"
-        if img_dir.exists():
+        # Check for any remaining PNG files
-            png_files = list(img_dir.rglob("*.png"))
+        png_files = list(img_dir.rglob("*.png"))
-            if len(png_files) == 0:
+        if len(png_files) == 0:
+            # Only remove the images directory if no PNG files remain
+            if img_dir.exists():
                 shutil.rmtree(img_dir)
                 logging.debug("Cleaned up empty images directory")
-            else:
+        else:
-                logging.debug(f"Images directory is not empty, containing {len(png_files)} PNG files")
+            logging.debug(f"Images directory is not empty, containing {len(png_files)} PNG files")
 
         return False  # Don't suppress the original exception

src/lerobot/motors/robstride/__init__.py

@@ -1,18 +0,0 @@
-#!/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 .robstride import RobstrideMotorsBus
-from .tables import *

src/lerobot/motors/robstride/tables.py

@@ -1,120 +0,0 @@
-# Copyright 2025 The HuggingFace Inc. team. All rights reserved.
-#
-# Licensed under the Apache License, Version 2.0 (the "License");
-# you may not use this file except in compliance with the License.
-# You may obtain a copy of the License at
-#
-#     http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing, software
-# distributed under the License is distributed on an "AS IS" BASIS,
-# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-# See the License for the specific language governing permissions and
-# limitations under the License.
-
-"""Configuration tables for Damiao motors."""
-
-from enum import IntEnum
-
-
-# Motor type definitions
-class MotorType(IntEnum):
-    O0 = 0
-    O1 = 1
-    O2 = 2
-    O3 = 3
-    O4 = 4
-    O5 = 5
-    ELO5 = 6
-    O6 = 7
-
-
-class CommMode(IntEnum):
-    PrivateProtocole = 0
-    CANopen = 1
-    MIT = 2
-
-
-# Control modes
-class ControlMode(IntEnum):
-    MIT = 0
-    POS_VEL = 1
-    VEL = 2
-
-
-# Motor limit parameters [PMAX, VMAX, TMAX]
-# PMAX: Maximum position (rad)
-# VMAX: Maximum velocity (rad/s)
-# TMAX: Maximum torque (N·m)
-MOTOR_LIMIT_PARAMS: dict[MotorType, tuple[float, float, float]] = {
-    MotorType.O0: (12.57, 33, 14),
-    MotorType.O1: (12.57, 44, 17),
-    MotorType.O2: (12.57, 33, 20),
-    MotorType.O3: (12.57, 33, 60),
-    MotorType.O4: (12.57, 33, 120),
-    MotorType.O5: (12.57, 50, 5.5),
-    MotorType.ELO5: (12.57, 50, 6),
-    MotorType.O6: (112.5, 50, 36),
-}
-
-# Motor model names
-MODEL_NAMES = {
-    MotorType.O0: "O0",
-    MotorType.O1: "O1",
-    MotorType.O2: "O2",
-    MotorType.O3: "O3",
-    MotorType.O4: "O4",
-    MotorType.O5: "O5",
-    MotorType.ELO5: "ELO5",
-    MotorType.O6: "O6",
-}
-
-# Motor resolution table (encoder counts per revolution)
-MODEL_RESOLUTION = {
-    "O0": 65536,
-    "O1": 65536,
-    "O2": 65536,
-    "O3": 65536,
-    "O4": 65536,
-    "O5": 65536,
-    "ELO5": 65536,
-    "O6": 65536,
-}
-
-# CAN baudrates supported by Robstride motors
-AVAILABLE_BAUDRATES = [
-    1000000,  # 4: 1 mbps (default)
-]
-DEFAULT_BAUDRATE = 1000000
-
-# Default timeout in milliseconds
-DEFAULT_TIMEOUT_MS = 0  # disabled by default, otherwise 20000 is 1s
-
-
-# Data that should be normalized
-NORMALIZED_DATA = ["Present_Position", "Goal_Position"]
-
-
-# MIT control parameter ranges
-MIT_KP_RANGE = (0.0, 500.0)
-MIT_KD_RANGE = (0.0, 5.0)
-
-# CAN frame command IDs
-CAN_CMD_ENABLE = 0xFC
-CAN_CMD_DISABLE = 0xFD
-CAN_CMD_SET_ZERO = 0xFE
-CAN_CMD_CLEAR_FAULT = 0xFB
-
-
-CAN_CMD_QUERY_PARAM = 0x33
-CAN_CMD_WRITE_PARAM = 0x55
-CAN_CMD_SAVE_PARAM = 0xAA
-
-# CAN ID for parameter operations
-CAN_PARAM_ID = 0x7FF
-
-
-RUNNING_TIMEOUT = 0.001
-PARAM_TIMEOUT = 0.01
-
-STATE_CACHE_TTL_S = 0.02

src/lerobot/policies/diffusion/configuration_diffusion.py

@@ -55,16 +55,10 @@ class DiffusionConfig(PreTrainedConfig):
         normalization_mapping: A dictionary that maps from a str value of FeatureType (e.g., "STATE", "VISUAL") to
             a corresponding NormalizationMode (e.g., NormalizationMode.MIN_MAX)
         vision_backbone: Name of the torchvision resnet backbone to use for encoding images.
-        resize_shape: (H, W) shape to resize images to as a preprocessing step for the vision
+        crop_shape: (H, W) shape to crop images to as a preprocessing step for the vision backbone. Must fit
-            backbone. If None, no resizing is done and the original image resolution is used.
+            within the image size. If None, no cropping is done.
-        crop_ratio: Ratio in (0, 1] used to derive the crop size from resize_shape
+        crop_is_random: Whether the crop should be random at training time (it's always a center crop in eval
-            (crop_h = int(resize_shape[0] * crop_ratio), likewise for width).
+            mode).
-            Set to 1.0 to disable cropping. Only takes effect when resize_shape is not None.
-        crop_shape: (H, W) shape to crop images to. When resize_shape is set and crop_ratio < 1.0,
-            this is computed automatically. Can also be set directly for legacy configs that use
-            crop-only (without resize). If None and no derivation applies, no cropping is done.
-        crop_is_random: Whether the crop should be random at training time (it's always a center
-            crop in eval mode).
         pretrained_backbone_weights: Pretrained weights from torchvision to initialize the backbone.
             `None` means no pretrained weights.
         use_group_norm: Whether to replace batch normalization with group normalization in the backbone.
@@ -120,9 +114,7 @@ class DiffusionConfig(PreTrainedConfig):
     # Architecture / modeling.
     # Vision backbone.
     vision_backbone: str = "resnet18"
-    resize_shape: tuple[int, int] | None = None
+    crop_shape: tuple[int, int] | None = (84, 84)
-    crop_ratio: float = 1.0
-    crop_shape: tuple[int, int] | None = None
     crop_is_random: bool = True
     pretrained_backbone_weights: str | None = None
     use_group_norm: bool = True
@@ -147,10 +139,6 @@ class DiffusionConfig(PreTrainedConfig):
     # Inference
     num_inference_steps: int | None = None
 
-    # Optimization
-    compile_model: bool = False
-    compile_mode: str = "reduce-overhead"
-
     # Loss computation
     do_mask_loss_for_padding: bool = False
 
@@ -183,25 +171,6 @@ class DiffusionConfig(PreTrainedConfig):
                 f"Got {self.noise_scheduler_type}."
             )
 
-        if self.resize_shape is not None and (
-            len(self.resize_shape) != 2 or any(d <= 0 for d in self.resize_shape)
-        ):
-            raise ValueError(f"`resize_shape` must be a pair of positive integers. Got {self.resize_shape}.")
-        if not (0 < self.crop_ratio <= 1.0):
-            raise ValueError(f"`crop_ratio` must be in (0, 1]. Got {self.crop_ratio}.")
-
-        if self.resize_shape is not None:
-            if self.crop_ratio < 1.0:
-                self.crop_shape = (
-                    int(self.resize_shape[0] * self.crop_ratio),
-                    int(self.resize_shape[1] * self.crop_ratio),
-                )
-            else:
-                # Explicitly disable cropping for resize+ratio path when crop_ratio == 1.0.
-                self.crop_shape = None
-        if self.crop_shape is not None and (self.crop_shape[0] <= 0 or self.crop_shape[1] <= 0):
-            raise ValueError(f"`crop_shape` must have positive dimensions. Got {self.crop_shape}.")
-
         # Check that the horizon size and U-Net downsampling is compatible.
         # U-Net downsamples by 2 with each stage.
         downsampling_factor = 2 ** len(self.down_dims)
@@ -229,12 +198,13 @@ class DiffusionConfig(PreTrainedConfig):
         if len(self.image_features) == 0 and self.env_state_feature is None:
             raise ValueError("You must provide at least one image or the environment state among the inputs.")
 
-        if self.resize_shape is None and self.crop_shape is not None:
+        if self.crop_shape is not None:
             for key, image_ft in self.image_features.items():
                 if self.crop_shape[0] > image_ft.shape[1] or self.crop_shape[1] > image_ft.shape[2]:
                     raise ValueError(
-                        f"`crop_shape` should fit within the image shapes. Got {self.crop_shape} "
+                        f"`crop_shape` should fit within the images shapes. Got {self.crop_shape} "
-                        f"for `crop_shape` and {image_ft.shape} for `{key}`."
+                        f"for `crop_shape` and {image_ft.shape} for "
+                        f"`{key}`."
                     )
 
         # Check that all input images have the same shape.

src/lerobot/policies/diffusion/modeling_diffusion.py

@@ -142,9 +142,6 @@ class DiffusionPolicy(PreTrainedPolicy):
         """Run the batch through the model and compute the loss for training or validation."""
         if self.config.image_features:
             batch = dict(batch)  # shallow copy so that adding a key doesn't modify the original
-            for key in self.config.image_features:
-                if self.config.n_obs_steps == 1 and batch[key].ndim == 4:
-                    batch[key] = batch[key].unsqueeze(1)
             batch[OBS_IMAGES] = torch.stack([batch[key] for key in self.config.image_features], dim=-4)
         loss = self.diffusion.compute_loss(batch)
         # no output_dict so returning None
@@ -185,11 +182,6 @@ class DiffusionModel(nn.Module):
 
         self.unet = DiffusionConditionalUnet1d(config, global_cond_dim=global_cond_dim * config.n_obs_steps)
 
-        if config.compile_model:
-            # Compile the U-Net. "reduce-overhead" is preferred for the small-batch repetitive loops
-            # common in diffusion inference.
-            self.unet = torch.compile(self.unet, mode=config.compile_mode)
-
         self.noise_scheduler = _make_noise_scheduler(
             config.noise_scheduler_type,
             num_train_timesteps=config.num_train_timesteps,
@@ -454,18 +446,12 @@ class DiffusionRgbEncoder(nn.Module):
     def __init__(self, config: DiffusionConfig):
         super().__init__()
         # Set up optional preprocessing.
-        if config.resize_shape is not None:
+        if config.crop_shape is not None:
-            self.resize = torchvision.transforms.Resize(config.resize_shape)
-        else:
-            self.resize = None
-
-        crop_shape = config.crop_shape
-        if crop_shape is not None:
             self.do_crop = True
             # Always use center crop for eval
-            self.center_crop = torchvision.transforms.CenterCrop(crop_shape)
+            self.center_crop = torchvision.transforms.CenterCrop(config.crop_shape)
             if config.crop_is_random:
-                self.maybe_random_crop = torchvision.transforms.RandomCrop(crop_shape)
+                self.maybe_random_crop = torchvision.transforms.RandomCrop(config.crop_shape)
             else:
                 self.maybe_random_crop = self.center_crop
         else:
@@ -491,16 +477,13 @@ class DiffusionRgbEncoder(nn.Module):
 
         # Set up pooling and final layers.
         # Use a dry run to get the feature map shape.
-        # The dummy shape mirrors the runtime preprocessing order: resize -> crop.
+        # The dummy input should take the number of image channels from `config.image_features` and it should
+        # use the height and width from `config.crop_shape` if it is provided, otherwise it should use the
+        # height and width from `config.image_features`.
 
         # Note: we have a check in the config class to make sure all images have the same shape.
         images_shape = next(iter(config.image_features.values())).shape
-        if config.crop_shape is not None:
+        dummy_shape_h_w = config.crop_shape if config.crop_shape is not None else images_shape[1:]
-            dummy_shape_h_w = config.crop_shape
-        elif config.resize_shape is not None:
-            dummy_shape_h_w = config.resize_shape
-        else:
-            dummy_shape_h_w = images_shape[1:]
         dummy_shape = (1, images_shape[0], *dummy_shape_h_w)
         feature_map_shape = get_output_shape(self.backbone, dummy_shape)[1:]
 
@@ -516,10 +499,7 @@ class DiffusionRgbEncoder(nn.Module):
         Returns:
             (B, D) image feature.
         """
-        # Preprocess: resize if configured, then crop if configured.
+        # Preprocess: maybe crop (if it was set up in the __init__).
-
-        if self.resize is not None:
-            x = self.resize(x)
         if self.do_crop:
             if self.training:  # noqa: SIM108
                 x = self.maybe_random_crop(x)

src/lerobot/policies/factory.py

@@ -470,6 +470,13 @@ def make_policy(
     cfg.output_features = {key: ft for key, ft in features.items() if ft.type is FeatureType.ACTION}
     if not cfg.input_features:
         cfg.input_features = {key: ft for key, ft in features.items() if key not in cfg.output_features}
+
+    # Store action feature names for delta_exclude_joints support
+    if ds_meta is not None and hasattr(cfg, "action_feature_names"):
+        action_names = ds_meta.features.get(ACTION, {}).get("names")
+        if action_names is not None:
+            cfg.action_feature_names = list(action_names)
+
     kwargs["config"] = cfg
 
     # Pass dataset_stats to the policy if available (needed for some policies like SARM)

src/lerobot/policies/pi0/configuration_pi0.py

@@ -50,6 +50,13 @@ class PI0Config(PreTrainedConfig):
     min_period: float = 4e-3
     max_period: float = 4.0
 
+    # Delta actions: converts absolute actions to delta (relative to state).
+    use_delta_actions: bool = False
+    # Joint names to exclude from delta (kept absolute). Empty list = all dims delta.
+    delta_exclude_joints: list[str] = field(default_factory=lambda: ["gripper"])
+    # Populated at runtime from dataset metadata by make_policy.
+    action_feature_names: list[str] | None = None
+
     # Real-Time Chunking (RTC) configuration
     rtc_config: RTCConfig | None = None
 

src/lerobot/policies/pi0/processor_pi0.py

@@ -21,8 +21,10 @@ import torch
 from lerobot.configs.types import PipelineFeatureType, PolicyFeature
 from lerobot.policies.pi0.configuration_pi0 import PI0Config
 from lerobot.processor import (
+    AbsoluteActionsProcessorStep,
     AddBatchDimensionProcessorStep,
     ComplementaryDataProcessorStep,
+    DeltaActionsProcessorStep,
     DeviceProcessorStep,
     NormalizerProcessorStep,
     PolicyAction,
@@ -126,7 +128,13 @@ def make_pi0_pre_post_processors(
         A tuple containing the configured pre-processor and post-processor pipelines.
     """
 
-    # Add remaining processors
+    delta_step = DeltaActionsProcessorStep(
+        enabled=config.use_delta_actions,
+        exclude_joints=getattr(config, "delta_exclude_joints", []),
+        action_names=getattr(config, "action_feature_names", None),
+    )
+
+    # OpenPI order: raw → delta → normalize → model → unnormalize → absolute
     input_steps: list[ProcessorStep] = [
         RenameObservationsProcessorStep(rename_map={}),  # To mimic the same processor as pretrained one
         AddBatchDimensionProcessorStep(),
@@ -138,6 +146,7 @@ def make_pi0_pre_post_processors(
             padding="max_length",
         ),
         DeviceProcessorStep(device=config.device),
+        delta_step,
         NormalizerProcessorStep(
             features={**config.input_features, **config.output_features},
             norm_map=config.normalization_mapping,
@@ -149,6 +158,7 @@ def make_pi0_pre_post_processors(
         UnnormalizerProcessorStep(
             features=config.output_features, norm_map=config.normalization_mapping, stats=dataset_stats
         ),
+        AbsoluteActionsProcessorStep(enabled=config.use_delta_actions, delta_step=delta_step),
         DeviceProcessorStep(device="cpu"),
     ]
 

src/lerobot/policies/pi05/configuration_pi05.py

@@ -50,6 +50,13 @@ class PI05Config(PreTrainedConfig):
     min_period: float = 4e-3
     max_period: float = 4.0
 
+    # Delta actions: converts absolute actions to delta (relative to state).
+    use_delta_actions: bool = False
+    # Joint names to exclude from delta (kept absolute). Empty list = all dims delta.
+    delta_exclude_joints: list[str] = field(default_factory=lambda: ["gripper"])
+    # Populated at runtime from dataset metadata by make_policy.
+    action_feature_names: list[str] | None = None
+
     # Real-Time Chunking (RTC) configuration
     rtc_config: RTCConfig | None = None
 

src/lerobot/policies/pi05/processor_pi05.py

@@ -25,7 +25,9 @@ from lerobot.configs.types import PipelineFeatureType, PolicyFeature
 from lerobot.policies.pi05.configuration_pi05 import PI05Config
 from lerobot.policies.pi05.modeling_pi05 import pad_vector
 from lerobot.processor import (
+    AbsoluteActionsProcessorStep,
     AddBatchDimensionProcessorStep,
+    DeltaActionsProcessorStep,
     DeviceProcessorStep,
     NormalizerProcessorStep,
     PolicyAction,
@@ -129,10 +131,19 @@ def make_pi05_pre_post_processors(
         A tuple containing the configured pre-processor and post-processor pipelines.
     """
 
-    # Add remaining processors
+    delta_step = DeltaActionsProcessorStep(
+        enabled=config.use_delta_actions,
+        exclude_joints=getattr(config, "delta_exclude_joints", []),
+        action_names=getattr(config, "action_feature_names", None),
+    )
+
+    # OpenPI order: raw → delta → normalize → model → unnormalize → absolute
+    # NOTE: NormalizerProcessorStep MUST come before Pi05PrepareStateTokenizerProcessorStep
+    # because the tokenizer step expects normalized state in [-1, 1] range for discretization
     input_steps: list[ProcessorStep] = [
         RenameObservationsProcessorStep(rename_map={}),  # To mimic the same processor as pretrained one
         AddBatchDimensionProcessorStep(),
+        delta_step,
         # NOTE: NormalizerProcessorStep MUST come before Pi05PrepareStateTokenizerProcessorStep
         # because the tokenizer step expects normalized state in [-1, 1] range for discretization
         NormalizerProcessorStep(
@@ -154,6 +165,7 @@ def make_pi05_pre_post_processors(
         UnnormalizerProcessorStep(
             features=config.output_features, norm_map=config.normalization_mapping, stats=dataset_stats
         ),
+        AbsoluteActionsProcessorStep(enabled=config.use_delta_actions, delta_step=delta_step),
         DeviceProcessorStep(device="cpu"),
     ]
 

src/lerobot/policies/pi0_fast/configuration_pi0_fast.py

@@ -41,6 +41,9 @@ class PI0FastConfig(PreTrainedConfig):
     max_action_dim: int = 32
     max_action_tokens: int = 256
 
+    # Delta actions: converts absolute actions to delta (relative to state).
+    use_delta_actions: bool = False
+
     # Real-Time Chunking (RTC) configuration
     rtc_config: RTCConfig | None = None
 

src/lerobot/policies/pi0_fast/modeling_pi0_fast.py

@@ -48,12 +48,14 @@ from lerobot.configs.policies import PreTrainedConfig
 from lerobot.policies.pi0_fast.configuration_pi0_fast import PI0FastConfig
 from lerobot.policies.pretrained import PreTrainedPolicy, T
 from lerobot.policies.rtc.modeling_rtc import RTCProcessor
+from lerobot.processor.delta_action_processor import to_absolute_actions
 from lerobot.utils.constants import (
     ACTION,
     ACTION_TOKEN_MASK,
     ACTION_TOKENS,
     OBS_LANGUAGE_ATTENTION_MASK,
     OBS_LANGUAGE_TOKENS,
+    OBS_STATE,
     OPENPI_ATTENTION_MASK_VALUE,
 )
 
@@ -1315,6 +1317,12 @@ class PI0FastPolicy(PreTrainedPolicy):
             action_tokens, action_horizon=action_horizon, action_dim=action_dim
         )
 
+        if self.config.use_delta_actions and OBS_STATE in batch:
+            state = pad_vector(batch[OBS_STATE], self.config.max_state_dim)
+            continuous_actions = to_absolute_actions(
+                continuous_actions, state, [True] * continuous_actions.shape[-1]
+        )
+
         return continuous_actions
 
     def forward(self, batch: dict[str, Tensor]) -> tuple[Tensor, dict]:

src/lerobot/policies/pi0_fast/processor_pi0_fast.py

@@ -27,6 +27,7 @@ from lerobot.policies.pi0_fast.modeling_pi0_fast import pad_vector
 from lerobot.processor import (
     ActionTokenizerProcessorStep,
     AddBatchDimensionProcessorStep,
+    DeltaActionsProcessorStep,
     DeviceProcessorStep,
     NormalizerProcessorStep,
     PolicyAction,
@@ -147,6 +148,7 @@ def make_pi0_fast_pre_post_processors(
             padding_side="right",
             padding="max_length",
         ),
+        DeltaActionsProcessorStep(enabled=config.use_delta_actions),
         ActionTokenizerProcessorStep(
             action_tokenizer_name=config.action_tokenizer_name,
             max_action_tokens=config.max_action_tokens,

src/lerobot/policies/sarm/compute_rabc_weights.py

@@ -27,18 +27,18 @@ Usage:
     # Full RA-BC computation with visualizations
     python src/lerobot/policies/sarm/compute_rabc_weights.py \\
         --dataset-repo-id lerobot/aloha_sim_insertion_human \\
-        --reward-model-path <USER>/sarm_single_uni4
+        --reward-model-path pepijn223/sarm_single_uni4
 
     # Faster computation with stride (compute every 5 frames, interpolate the rest)
     python src/lerobot/policies/sarm/compute_rabc_weights.py \\
         --dataset-repo-id lerobot/aloha_sim_insertion_human \\
-        --reward-model-path <USER>/sarm_single_uni4 \\
+        --reward-model-path pepijn223/sarm_single_uni4 \\
         --stride 5
 
     # Visualize predictions only (no RA-BC computation)
     python src/lerobot/policies/sarm/compute_rabc_weights.py \\
         --dataset-repo-id lerobot/aloha_sim_insertion_human \\
-        --reward-model-path <USER>/sarm_single_uni4 \\
+        --reward-model-path pepijn223/sarm_single_uni4 \\
         --visualize-only \\
         --num-visualizations 5
 
@@ -714,12 +714,12 @@ Examples:
     # Full RA-BC computation with visualizations
     python src/lerobot/policies/sarm/compute_rabc_weights.py \\
         --dataset-repo-id lerobot/aloha_sim_insertion_human \\
-        --reward-model-path <USER>/sarm_single_uni4
+        --reward-model-path pepijn223/sarm_single_uni4
 
     # Visualize predictions only (no RA-BC computation)
     python src/lerobot/policies/sarm/compute_rabc_weights.py \\
         --dataset-repo-id lerobot/aloha_sim_insertion_human \\
-        --reward-model-path <USER>/sarm_single_uni4 \\
+        --reward-model-path pepijn223/sarm_single_uni4 \\
         --visualize-only \\
         --num-visualizations 10
         """,

src/lerobot/policies/sarm/processor_sarm.py

@@ -277,7 +277,9 @@ class SARMEncodingProcessorStep(ProcessorStep):
 
         # When language is perturbed, targets are zero so perturbed samples don't contribute to progress loss
         if self.dataset_meta is not None:
-            episodes_df = self.dataset_meta.episodes.to_pandas()
+            episodes_df = None
+            if self.sparse_subtask_names != ["task"]:
+                episodes_df = self.dataset_meta.episodes.to_pandas()
 
             # Generate sparse targets
             if self.sparse_temporal_proportions is not None:

src/lerobot/policies/smolvla/configuration_smolvla.py

@@ -85,7 +85,7 @@ class SmolVLAConfig(PreTrainedConfig):
     scheduler_decay_lr: float = 2.5e-6
 
     vlm_model_name: str = "HuggingFaceTB/SmolVLM2-500M-Video-Instruct"  # Select the VLM backbone.
-    load_vlm_weights: bool = False  # Set to False in case of training the expert from scratch. True when init from pretrained SmolVLA weights
+    load_vlm_weights: bool = False  # Set to True in case of training the expert from scratch. True when init from pretrained SmolVLA weights
 
     add_image_special_tokens: bool = False  # Whether to use special image tokens around image features.
 
@@ -106,9 +106,6 @@ class SmolVLAConfig(PreTrainedConfig):
     # Real-Time Chunking (RTC) configuration
     rtc_config: RTCConfig | None = None
 
-    compile_model: bool = False  # Whether to use torch.compile for model optimization
-    compile_mode: str = "max-autotune"  # Torch compile mode
-
     def __post_init__(self):
         super().__post_init__()
 

src/lerobot/policies/smolvla/modeling_smolvla.py

@@ -30,7 +30,7 @@ Example of finetuning the smolvla pretrained model (`smolvla_base`):
 ```bash
 lerobot-train \
 --policy.path=lerobot/smolvla_base \
---dataset.repo_id=<USER>/svla_so100_task1_v3 \
+--dataset.repo_id=danaaubakirova/svla_so100_task1_v3 \
 --batch_size=64 \
 --steps=200000
 ```
@@ -40,7 +40,7 @@ and an action expert.
 ```bash
 lerobot-train \
 --policy.type=smolvla \
---dataset.repo_id=<USER>/svla_so100_task1_v3 \
+--dataset.repo_id=danaaubakirova/svla_so100_task1_v3 \
 --batch_size=64 \
 --steps=200000
 ```
@@ -593,12 +593,6 @@ class VLAFlowMatching(nn.Module):
         self.prefix_length = self.config.prefix_length
         self.rtc_processor = rtc_processor
 
-        # Compile model if requested
-        if config.compile_model:
-            torch.set_float32_matmul_precision("high")
-            self.sample_actions = torch.compile(self.sample_actions, mode=config.compile_mode)
-            self.forward = torch.compile(self.forward, mode=config.compile_mode)
-
     def _rtc_enabled(self):
         return self.config.rtc_config is not None and self.config.rtc_config.enabled
 

src/lerobot/policies/smolvla/smolvlm_with_expert.py

@@ -77,6 +77,7 @@ class SmolVLMWithExpertModel(nn.Module):
             print(f"Loading  {model_id} weights ...")
             self.vlm = AutoModelForImageTextToText.from_pretrained(
                 model_id,
+                device_map=device,
                 torch_dtype="bfloat16",
                 low_cpu_mem_usage=True,
             )

src/lerobot/processor/__init__.py

@@ -28,7 +28,14 @@ from .core import (
     RobotObservation,
     TransitionKey,
 )
-from .delta_action_processor import MapDeltaActionToRobotActionStep, MapTensorToDeltaActionDictStep
+from .delta_action_processor import (
+    AbsoluteActionsProcessorStep,
+    DeltaActionsProcessorStep,
+    MapDeltaActionToRobotActionStep,
+    MapTensorToDeltaActionDictStep,
+    to_absolute_actions,
+    to_delta_actions,
+)
 from .device_processor import DeviceProcessorStep
 from .factory import (
     make_default_processors,
@@ -44,7 +51,6 @@ from .hil_processor import (
     AddTeleopActionAsComplimentaryDataStep,
     AddTeleopEventsAsInfoStep,
     GripperPenaltyProcessorStep,
-    GymHILAdapterProcessorStep,
     ImageCropResizeProcessorStep,
     InterventionActionProcessorStep,
     RewardClassifierProcessorStep,
@@ -88,7 +94,6 @@ __all__ = [
     "DoneProcessorStep",
     "EnvAction",
     "EnvTransition",
-    "GymHILAdapterProcessorStep",
     "GripperPenaltyProcessorStep",
     "hotswap_stats",
     "IdentityProcessorStep",
@@ -99,6 +104,8 @@ __all__ = [
     "make_default_teleop_action_processor",
     "make_default_robot_action_processor",
     "make_default_robot_observation_processor",
+    "AbsoluteActionsProcessorStep",
+    "DeltaActionsProcessorStep",
     "MapDeltaActionToRobotActionStep",
     "MapTensorToDeltaActionDictStep",
     "NormalizerProcessorStep",
@@ -128,6 +135,8 @@ __all__ = [
     "transition_to_batch",
     "TransitionKey",
     "TruncatedProcessorStep",
+    "to_absolute_actions",
+    "to_delta_actions",
     "UnnormalizerProcessorStep",
     "VanillaObservationProcessorStep",
 ]

src/lerobot/processor/delta_action_processor.py

@@ -14,12 +14,54 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-from dataclasses import dataclass
+from collections.abc import Sequence
+from dataclasses import dataclass, field
+from typing import Any
+
+import torch
+from torch import Tensor
 
 from lerobot.configs.types import FeatureType, PipelineFeatureType, PolicyFeature
+from lerobot.utils.constants import OBS_STATE
 
-from .core import PolicyAction, RobotAction
+from .core import EnvTransition, PolicyAction, RobotAction, TransitionKey
-from .pipeline import ActionProcessorStep, ProcessorStepRegistry, RobotActionProcessorStep
+from .pipeline import ActionProcessorStep, ProcessorStep, ProcessorStepRegistry, RobotActionProcessorStep
+
+
+def to_delta_actions(actions: Tensor, state: Tensor, mask: Sequence[bool]) -> Tensor:
+    """Convert absolute actions to delta: delta = action - state (for masked dims).
+
+    Args:
+        actions: (B, T, action_dim) or (B, action_dim).
+        state: (B, state_dim). Broadcast across time dimension.
+        mask: Which dims to convert. Can be shorter than action_dim.
+    """
+    mask_t = torch.tensor(mask, dtype=actions.dtype, device=actions.device)
+    dims = mask_t.shape[0]
+    state_offset = state[..., :dims] * mask_t
+    if actions.ndim == 3:
+        state_offset = state_offset.unsqueeze(-2)
+    actions = actions.clone()
+    actions[..., :dims] -= state_offset
+    return actions
+
+
+def to_absolute_actions(actions: Tensor, state: Tensor, mask: Sequence[bool]) -> Tensor:
+    """Convert delta actions back to absolute: absolute = delta + state (for masked dims).
+
+    Args:
+        actions: (B, T, action_dim) or (B, action_dim).
+        state: (B, state_dim). Broadcast across time dimension.
+        mask: Which dims to convert. Can be shorter than action_dim.
+    """
+    mask_t = torch.tensor(mask, dtype=actions.dtype, device=actions.device)
+    dims = mask_t.shape[0]
+    state_offset = state[..., :dims] * mask_t
+    if actions.ndim == 3:
+        state_offset = state_offset.unsqueeze(-2)
+    actions = actions.clone()
+    actions[..., :dims] += state_offset
+    return actions
 
 
 @ProcessorStepRegistry.register("map_tensor_to_delta_action_dict")
@@ -141,3 +183,126 @@ class MapDeltaActionToRobotActionStep(RobotActionProcessorStep):
             )
 
         return features
+
+
+@ProcessorStepRegistry.register("delta_actions_processor")
+@dataclass
+class DeltaActionsProcessorStep(ProcessorStep):
+    """Converts absolute actions to delta actions (action -= state) for masked dimensions.
+
+    Mirrors OpenPI's DeltaActions transform. Applied during preprocessing so the model
+    trains on relative offsets instead of absolute positions.
+    Caches the last seen state so a paired AbsoluteActionsProcessorStep can reverse
+    the conversion during postprocessing.
+
+    Attributes:
+        enabled: Whether to apply the delta conversion.
+        exclude_joints: Joint names to keep absolute (not converted to delta).
+        action_names: Action dimension names from dataset metadata, used to build
+            the mask from exclude_joints. If None, all dims are converted.
+    """
+
+    enabled: bool = False
+    exclude_joints: list[str] = field(default_factory=list)
+    action_names: list[str] | None = None
+    _last_state: torch.Tensor | None = field(default=None, init=False, repr=False)
+
+    def _build_mask(self, action_dim: int) -> list[bool]:
+        if not self.exclude_joints or self.action_names is None:
+            return [True] * action_dim
+
+        exclude_tokens = [str(name).lower() for name in self.exclude_joints if name]
+        if not exclude_tokens:
+            return [True] * action_dim
+
+        mask = []
+        for name in self.action_names[:action_dim]:
+            action_name = str(name).lower()
+            is_excluded = any(token == action_name or token in action_name for token in exclude_tokens)
+            mask.append(not is_excluded)
+
+        if len(mask) < action_dim:
+            mask.extend([True] * (action_dim - len(mask)))
+
+        return mask
+
+    def __call__(self, transition: EnvTransition) -> EnvTransition:
+        observation = transition.get(TransitionKey.OBSERVATION, {})
+        state = observation.get(OBS_STATE) if observation else None
+
+        # Always cache state for the paired AbsoluteActionsProcessorStep
+        if state is not None:
+            self._last_state = state
+
+        if not self.enabled:
+            return transition
+
+        new_transition = transition.copy()
+        action = new_transition.get(TransitionKey.ACTION)
+        if action is None or state is None:
+            return new_transition
+
+        mask = self._build_mask(action.shape[-1])
+        new_transition[TransitionKey.ACTION] = to_delta_actions(action, state, mask)
+        return new_transition
+
+    def get_config(self) -> dict[str, Any]:
+        return {"enabled": self.enabled, "exclude_joints": self.exclude_joints}
+
+    def transform_features(
+        self, features: dict[PipelineFeatureType, dict[str, PolicyFeature]]
+    ) -> dict[PipelineFeatureType, dict[str, PolicyFeature]]:
+        return features
+
+
+@ProcessorStepRegistry.register("absolute_actions_processor")
+@dataclass
+class AbsoluteActionsProcessorStep(ProcessorStep):
+    """Converts delta actions back to absolute actions (action += state) for all dimensions.
+
+    Mirrors OpenPI's AbsoluteActions transform. Applied during postprocessing so
+    predicted deltas are converted back to absolute positions for execution.
+    Reads the cached state from its paired DeltaActionsProcessorStep.
+
+    Attributes:
+        enabled: Whether to apply the absolute conversion.
+        delta_step: Reference to the paired DeltaActionsProcessorStep that caches state.
+    """
+
+    enabled: bool = False
+    delta_step: DeltaActionsProcessorStep | None = field(default=None, repr=False)
+
+    def __call__(self, transition: EnvTransition) -> EnvTransition:
+        if not self.enabled:
+            return transition
+
+        if self.delta_step is None:
+            raise RuntimeError(
+                "AbsoluteActionsProcessorStep requires a paired DeltaActionsProcessorStep "
+                "but delta_step is None. Ensure delta_step is set when constructing the postprocessor."
+            )
+
+        if self.delta_step._last_state is None:
+            raise RuntimeError(
+                "AbsoluteActionsProcessorStep requires state from DeltaActionsProcessorStep "
+                "but no state has been cached. Ensure the preprocessor runs before the postprocessor."
+            )
+
+        new_transition = transition.copy()
+        action = new_transition.get(TransitionKey.ACTION)
+        if action is None:
+            return new_transition
+
+        mask = self.delta_step._build_mask(action.shape[-1])
+        new_transition[TransitionKey.ACTION] = to_absolute_actions(
+            action, self.delta_step._last_state, mask
+        )
+        return new_transition
+
+    def get_config(self) -> dict[str, Any]:
+        return {"enabled": self.enabled}
+
+    def transform_features(
+        self, features: dict[PipelineFeatureType, dict[str, PolicyFeature]]
+    ) -> dict[PipelineFeatureType, dict[str, PolicyFeature]]:
+        return features

src/lerobot/processor/gym_action_processor.py

@@ -20,7 +20,6 @@ from lerobot.configs.types import PipelineFeatureType, PolicyFeature
 
 from .converters import to_tensor
 from .core import EnvAction, EnvTransition, PolicyAction
-from .hil_processor import TELEOP_ACTION_KEY
 from .pipeline import ActionProcessorStep, ProcessorStep, ProcessorStepRegistry
 
 
@@ -90,13 +89,6 @@ class Numpy2TorchActionProcessorStep(ProcessorStep):
             torch_action = to_tensor(action, dtype=None)  # Preserve original dtype
             new_transition[TransitionKey.ACTION] = torch_action
 
-        complementary_data = new_transition.get(TransitionKey.COMPLEMENTARY_DATA, {})
-        if TELEOP_ACTION_KEY in complementary_data:
-            teleop_action = complementary_data[TELEOP_ACTION_KEY]
-            if isinstance(teleop_action, EnvAction):
-                complementary_data[TELEOP_ACTION_KEY] = to_tensor(teleop_action)
-            new_transition[TransitionKey.COMPLEMENTARY_DATA] = complementary_data
-
         return new_transition
 
     def transform_features(

src/lerobot/processor/hil_processor.py

@@ -312,37 +312,6 @@ class TimeLimitProcessorStep(TruncatedProcessorStep):
         return features
 
 
-@ProcessorStepRegistry.register("gym_hil_adapter_processor")
-class GymHILAdapterProcessorStep(ProcessorStep):
-    """
-    Adapts the output of the `gym-hil` environment to the format expected by `lerobot` processors.
-
-    This step normalizes the `transition` object by:
-    1. Copying `teleop_action` from `info` to `complementary_data`.
-    2. Copying `is_intervention` from `info` (using the string key) to `info` (using the enum key).
-    """
-
-    def __call__(self, transition: EnvTransition) -> EnvTransition:
-        info = transition.get(TransitionKey.INFO, {})
-        complementary_data = transition.get(TransitionKey.COMPLEMENTARY_DATA, {})
-
-        if TELEOP_ACTION_KEY in info:
-            complementary_data[TELEOP_ACTION_KEY] = info[TELEOP_ACTION_KEY]
-
-        if "is_intervention" in info:
-            info[TeleopEvents.IS_INTERVENTION] = info["is_intervention"]
-
-        transition[TransitionKey.INFO] = info
-        transition[TransitionKey.COMPLEMENTARY_DATA] = complementary_data
-
-        return transition
-
-    def transform_features(
-        self, features: dict[PipelineFeatureType, dict[str, PolicyFeature]]
-    ) -> dict[PipelineFeatureType, dict[str, PolicyFeature]]:
-        return features
-
-
 @dataclass
 @ProcessorStepRegistry.register("gripper_penalty_processor")
 class GripperPenaltyProcessorStep(ProcessorStep):

src/lerobot/processor/normalize_processor.py

@@ -331,11 +331,9 @@ class _NormalizationMixin:
                 )
 
             mean, std = stats["mean"], stats["std"]
-            # Avoid division by zero by adding a small epsilon.
-            denom = std + self.eps
             if inverse:
-                return tensor * std + mean
+                return tensor * (std + 1e-6) + mean
-            return (tensor - mean) / denom
+            return (tensor - mean) / (std + 1e-6)
 
         if norm_mode == NormalizationMode.MIN_MAX:
             min_val = stats.get("min", None)
@@ -367,11 +365,7 @@ class _NormalizationMixin:
                     "QUANTILES normalization mode requires q01 and q99 stats, please update the dataset with the correct stats using the `augment_dataset_quantile_stats.py` script"
                 )
 
-            denom = q99 - q01
+            denom = q99 - q01 + 1e-6
-            # Avoid division by zero by adding epsilon when quantiles are identical
-            denom = torch.where(
-                denom == 0, torch.tensor(self.eps, device=tensor.device, dtype=tensor.dtype), denom
-            )
             if inverse:
                 return (tensor + 1.0) * denom / 2.0 + q01
             return 2.0 * (tensor - q01) / denom - 1.0

src/lerobot/rl/gym_manipulator.py

@@ -36,7 +36,6 @@ from lerobot.processor import (
     DeviceProcessorStep,
     EnvTransition,
     GripperPenaltyProcessorStep,
-    GymHILAdapterProcessorStep,
     ImageCropResizeProcessorStep,
     InterventionActionProcessorStep,
     MapDeltaActionToRobotActionStep,
@@ -380,7 +379,6 @@ def make_processors(
         ]
 
         env_pipeline_steps = [
-            GymHILAdapterProcessorStep(),
             Numpy2TorchActionProcessorStep(),
             VanillaObservationProcessorStep(),
             AddBatchDimensionProcessorStep(),
@@ -610,14 +608,7 @@ def control_loop(
 
     dataset = None
     if cfg.mode == "record":
-        if teleop_device:
+        action_features = teleop_device.action_features
-            action_features = teleop_device.action_features
-        else:
-            action_features = {
-                "dtype": "float32",
-                "shape": (4,),
-                "names": ["delta_x", "delta_y", "delta_z", "gripper"],
-            }
         features = {
             ACTION: action_features,
             REWARD: {"dtype": "float32", "shape": (1,), "names": None},
@@ -665,7 +656,7 @@ def control_loop(
         # Create a neutral action (no movement)
         neutral_action = torch.tensor([0.0, 0.0, 0.0], dtype=torch.float32)
         if use_gripper:
-            neutral_action = torch.cat([neutral_action, torch.tensor([0.0])])  # Gripper stay
+            neutral_action = torch.cat([neutral_action, torch.tensor([1.0])])  # Gripper stay
 
         # Use the new step function
         transition = step_env_and_process_transition(
@@ -734,8 +725,6 @@ def control_loop(
         precise_sleep(max(dt - (time.perf_counter() - step_start_time), 0.0))
 
     if dataset is not None and cfg.dataset.push_to_hub:
-        logging.info("Finalizing dataset before pushing to hub")
-        dataset.finalize()
         logging.info("Pushing dataset to hub")
         dataset.push_to_hub()
 

src/lerobot/robots/hope_jr/hope_jr_arm.py

@@ -140,7 +140,7 @@ class HopeJrArm(Robot):
         # Capture images from cameras
         for cam_key, cam in self.cameras.items():
             start = time.perf_counter()
-            obs_dict[cam_key] = cam.read_latest()
+            obs_dict[cam_key] = cam.async_read()
             dt_ms = (time.perf_counter() - start) * 1e3
             logger.debug(f"{self} read {cam_key}: {dt_ms:.1f}ms")
 

src/lerobot/robots/hope_jr/hope_jr_hand.py

@@ -171,7 +171,7 @@ class HopeJrHand(Robot):
         # Capture images from cameras
         for cam_key, cam in self.cameras.items():
             start = time.perf_counter()
-            obs_dict[cam_key] = cam.read_latest()
+            obs_dict[cam_key] = cam.async_read()
             dt_ms = (time.perf_counter() - start) * 1e3
             logger.debug(f"{self} read {cam_key}: {dt_ms:.1f}ms")
 

src/lerobot/robots/koch_follower/koch_follower.py

@@ -193,7 +193,7 @@ class KochFollower(Robot):
         # Capture images from cameras
         for cam_key, cam in self.cameras.items():
             start = time.perf_counter()
-            obs_dict[cam_key] = cam.read_latest()
+            obs_dict[cam_key] = cam.async_read()
             dt_ms = (time.perf_counter() - start) * 1e3
             logger.debug(f"{self} read {cam_key}: {dt_ms:.1f}ms")
 

src/lerobot/robots/lekiwi/lekiwi.py

@@ -360,7 +360,7 @@ class LeKiwi(Robot):
         # Capture images from cameras
         for cam_key, cam in self.cameras.items():
             start = time.perf_counter()
-            obs_dict[cam_key] = cam.read_latest()
+            obs_dict[cam_key] = cam.async_read()
             dt_ms = (time.perf_counter() - start) * 1e3
             logger.debug(f"{self} read {cam_key}: {dt_ms:.1f}ms")
 

src/lerobot/robots/omx_follower/omx_follower.py

@@ -176,7 +176,7 @@ class OmxFollower(Robot):
         # Capture images from cameras
         for cam_key, cam in self.cameras.items():
             start = time.perf_counter()
-            obs_dict[cam_key] = cam.read_latest()
+            obs_dict[cam_key] = cam.async_read()
             dt_ms = (time.perf_counter() - start) * 1e3
             logger.debug(f"{self} read {cam_key}: {dt_ms:.1f}ms")
 

src/lerobot/robots/openarm_follower/openarm_follower.py

@@ -241,7 +241,7 @@ class OpenArmFollower(Robot):
         # Capture images from cameras
         for cam_key, cam in self.cameras.items():
             start = time.perf_counter()
-            obs_dict[cam_key] = cam.read_latest()
+            obs_dict[cam_key] = cam.async_read()
             dt_ms = (time.perf_counter() - start) * 1e3
             logger.debug(f"{self} read {cam_key}: {dt_ms:.1f}ms")
 

src/lerobot/robots/reachy2/robot_reachy2.py

@@ -180,7 +180,7 @@ class Reachy2Robot(Robot):
 
         # Capture images from cameras
         for cam_key, cam in self.cameras.items():
-            obs_dict[cam_key] = cam.read_latest()
+            obs_dict[cam_key] = cam.async_read()
 
         return obs_dict
 

src/lerobot/robots/so_follower/so_follower.py

@@ -187,7 +187,7 @@ class SOFollower(Robot):
         # Capture images from cameras
         for cam_key, cam in self.cameras.items():
             start = time.perf_counter()
-            obs_dict[cam_key] = cam.read_latest()
+            obs_dict[cam_key] = cam.async_read()
             dt_ms = (time.perf_counter() - start) * 1e3
             logger.debug(f"{self} read {cam_key}: {dt_ms:.1f}ms")
 

src/lerobot/robots/unitree_g1/unitree_g1.py

@@ -324,7 +324,7 @@ class UnitreeG1(Robot):
 
         # Cameras - read images from ZMQ cameras
         for cam_name, cam in self._cameras.items():
-            obs[cam_name] = cam.read_latest()
+            obs[cam_name] = cam.async_read()
 
         return obs
 

src/lerobot/scripts/lerobot_calibrate.py

@@ -56,7 +56,6 @@ from lerobot.teleoperators import (  # noqa: F401
     make_teleoperator_from_config,
     omx_leader,
     openarm_leader,
-    openarm_mini,
     so_leader,
     unitree_g1,
 )

src/lerobot/scripts/lerobot_dataset_viz.py

@@ -47,14 +47,16 @@ local$ rerun lerobot_pusht_episode_0.rrd
 ```
 
 - Visualize data stored on a distant machine through streaming:
+(You need to forward the websocket port to the distant machine, with
+`ssh -L 9087:localhost:9087 username@remote-host`)
 ```
 distant$ lerobot-dataset-viz \
     --repo-id lerobot/pusht \
     --episode-index 0 \
     --mode distant \
-    --grpc-port 9876
+    --ws-port 9087
 
-local$ rerun rerun+http://IP:GRPC_PORT/proxy
+local$ rerun ws://localhost:9087
 ```
 
 """
@@ -73,7 +75,6 @@ import tqdm
 
 from lerobot.datasets.lerobot_dataset import LeRobotDataset
 from lerobot.utils.constants import ACTION, DONE, OBS_STATE, REWARD
-from lerobot.utils.utils import init_logging
 
 
 def to_hwc_uint8_numpy(chw_float32_torch: torch.Tensor) -> np.ndarray:
@@ -92,11 +93,10 @@ def visualize_dataset(
     num_workers: int = 0,
     mode: str = "local",
     web_port: int = 9090,
-    grpc_port: int = 9876,
+    ws_port: int = 9087,
     save: bool = False,
     output_dir: Path | None = None,
     display_compressed_images: bool = False,
-    **kwargs,
 ) -> Path | None:
     if save:
         assert output_dir is not None, (
@@ -126,19 +126,14 @@ def visualize_dataset(
     gc.collect()
 
     if mode == "distant":
-        server_uri = rr.serve_grpc(grpc_port=grpc_port)
+        rr.serve_web_viewer(open_browser=False, web_port=web_port)
-        logging.info(f"Connect to a Rerun Server: rerun rerun+http://IP:{grpc_port}/proxy")
-        rr.serve_web_viewer(open_browser=False, web_port=web_port, connect_to=server_uri)
 
     logging.info("Logging to Rerun")
 
-    first_index = None
     for batch in tqdm.tqdm(dataloader, total=len(dataloader)):
-        if first_index is None:
-            first_index = batch["index"][0].item()
         # iterate over the batch
         for i in range(len(batch["index"])):
-            rr.set_time("frame_index", sequence=batch["index"][i].item() - first_index)
+            rr.set_time("frame_index", sequence=batch["frame_index"][i].item())
             rr.set_time("timestamp", timestamp=batch["timestamp"][i].item())
 
             # display each camera image
@@ -231,7 +226,7 @@ def main():
             "Mode of viewing between 'local' or 'distant'. "
             "'local' requires data to be on a local machine. It spawns a viewer to visualize the data locally. "
             "'distant' creates a server on the distant machine where the data is stored. "
-            "Visualize the data by connecting to the server with `rerun rerun+http://IP:GRPC_PORT/proxy` on the local machine."
+            "Visualize the data by connecting to the server with `rerun ws://localhost:PORT` on the local machine."
         ),
     )
     parser.add_argument(
@@ -243,13 +238,8 @@ def main():
     parser.add_argument(
         "--ws-port",
         type=int,
-        help="deprecated, please use --grpc-port instead.",
+        default=9087,
-    )
+        help="Web socket port for rerun.io when `--mode distant` is set.",
-    parser.add_argument(
-        "--grpc-port",
-        type=int,
-        default=9876,
-        help="gRPC port for rerun.io when `--mode distant` is set.",
     )
     parser.add_argument(
         "--save",
@@ -275,7 +265,9 @@ def main():
 
     parser.add_argument(
         "--display-compressed-images",
-        action="store_true",
+        type=bool,
+        required=True,
+        default=False,
         help="If set, display compressed images in Rerun instead of uncompressed ones.",
     )
 
@@ -285,14 +277,6 @@ def main():
     root = kwargs.pop("root")
     tolerance_s = kwargs.pop("tolerance_s")
 
-    if kwargs["ws_port"] is not None:
-        logging.warning(
-            "--ws-port is deprecated and will be removed in future versions. Please use --grpc-port instead."
-        )
-        logging.warning("Setting grpc_port to ws_port value.")
-        kwargs["grpc_port"] = kwargs.pop("ws_port")
-
-    init_logging()
     logging.info("Loading dataset")
     dataset = LeRobotDataset(repo_id, episodes=[args.episode_index], root=root, tolerance_s=tolerance_s)
 

src/lerobot/scripts/lerobot_edit_dataset.py

@@ -21,142 +21,104 @@ This script allows you to delete episodes, split datasets, merge datasets,
 remove features, modify tasks, and convert image datasets to video format.
 When new_repo_id is specified, creates a new dataset.
 
-Path semantics (v2): --root and --new_root are exact dataset folders containing
-meta/, data/, videos/. When omitted, defaults to $HF_LEROBOT_HOME/{repo_id}.
-
 Usage Examples:
 
 Delete episodes 0, 2, and 5 from a dataset:
-    lerobot-edit-dataset \
+    python -m lerobot.scripts.lerobot_edit_dataset \
         --repo_id lerobot/pusht \
         --operation.type delete_episodes \
         --operation.episode_indices "[0, 2, 5]"
 
-Delete episodes from a local dataset at a specific path:
+Delete episodes and save to a new dataset:
-    lerobot-edit-dataset \
+    python -m lerobot.scripts.lerobot_edit_dataset \
-        --repo_id lerobot/pusht \
-        --root /path/to/pusht \
-        --operation.type delete_episodes \
-        --operation.episode_indices "[0, 2, 5]"
-
-Delete episodes and save to a new dataset at a specific path and with a new repo_id:
-    lerobot-edit-dataset \
         --repo_id lerobot/pusht \
         --new_repo_id lerobot/pusht_filtered \
-        --new_root /path/to/pusht_filtered \
         --operation.type delete_episodes \
         --operation.episode_indices "[0, 2, 5]"
 
-Split dataset by fractions (pusht_train, pusht_val):
+Split dataset by fractions:
-    lerobot-edit-dataset \
+    python -m lerobot.scripts.lerobot_edit_dataset \
         --repo_id lerobot/pusht \
         --operation.type split \
         --operation.splits '{"train": 0.8, "val": 0.2}'
 
-Split dataset by fractions and save split datasets to a specific folder (base_folder/train, base_folder/val):
-    lerobot-edit-dataset \
-        --repo_id lerobot/pusht \
-        --new_root /path/to/base_folder \
-        --operation.type split \
-        --operation.splits '{"train": 0.8, "val": 0.2}'
-
 Split dataset by episode indices:
-    lerobot-edit-dataset \
+    python -m lerobot.scripts.lerobot_edit_dataset \
         --repo_id lerobot/pusht \
         --operation.type split \
         --operation.splits '{"train": [0, 1, 2, 3], "val": [4, 5]}'
 
 Split into more than two splits:
-    lerobot-edit-dataset \
+    python -m lerobot.scripts.lerobot_edit_dataset \
         --repo_id lerobot/pusht \
         --operation.type split \
         --operation.splits '{"train": 0.6, "val": 0.2, "test": 0.2}'
 
 Merge multiple datasets:
-    lerobot-edit-dataset \
+    python -m lerobot.scripts.lerobot_edit_dataset \
-        --new_repo_id lerobot/pusht_merged \
+        --repo_id lerobot/pusht_merged \
         --operation.type merge \
         --operation.repo_ids "['lerobot/pusht_train', 'lerobot/pusht_val']"
 
-Merge multiple datasets to a specific output path:
-    lerobot-edit-dataset \
-        --new_repo_id lerobot/pusht_merged \
-        --new_root /path/to/pusht_merged \
-        --operation.type merge \
-        --operation.repo_ids "['lerobot/pusht_train', 'lerobot/pusht_val']"
-
-Merge multiple datasets from a list of local dataset paths:
-    lerobot-edit-dataset \
-        --new_repo_id lerobot/pusht_merged \
-        --operation.type merge \
-        --operation.repo_ids "['pusht_train', 'pusht_val']" \
-        --operation.roots "['/path/to/pusht_train', '/path/to/pusht_val']"
-
 Remove camera feature:
-    lerobot-edit-dataset \
+    python -m lerobot.scripts.lerobot_edit_dataset \
         --repo_id lerobot/pusht \
         --operation.type remove_feature \
-        --operation.feature_names "['observation.image']"
+        --operation.feature_names "['observation.images.top']"
 
 Modify tasks - set a single task for all episodes (WARNING: modifies in-place):
-    lerobot-edit-dataset \
+    python -m lerobot.scripts.lerobot_edit_dataset \
         --repo_id lerobot/pusht \
         --operation.type modify_tasks \
         --operation.new_task "Pick up the cube and place it"
 
 Modify tasks - set different tasks for specific episodes (WARNING: modifies in-place):
-    lerobot-edit-dataset \
+    python -m lerobot.scripts.lerobot_edit_dataset \
         --repo_id lerobot/pusht \
         --operation.type modify_tasks \
         --operation.episode_tasks '{"0": "Task A", "1": "Task B", "2": "Task A"}'
 
 Modify tasks - set default task with overrides for specific episodes (WARNING: modifies in-place):
-    lerobot-edit-dataset \
+    python -m lerobot.scripts.lerobot_edit_dataset \
         --repo_id lerobot/pusht \
         --operation.type modify_tasks \
         --operation.new_task "Default task" \
         --operation.episode_tasks '{"5": "Special task for episode 5"}'
 
+Trim first 3 seconds from all episodes:
+    python -m lerobot.scripts.lerobot_edit_dataset \
+        --repo_id lerobot/pusht \
+        --new_repo_id lerobot/pusht_trim3s \
+        --operation.type trim_episode_start \
+        --operation.seconds 3.0
+
 Convert image dataset to video format and save locally:
-    lerobot-edit-dataset \
+    python -m lerobot.scripts.lerobot_edit_dataset \
         --repo_id lerobot/pusht_image \
-        --new_root /path/to/output/pusht_video \
+        --operation.type convert_image_to_video \
-        --operation.type convert_image_to_video
+        --operation.output_dir /path/to/output/pusht_video
 
 Convert image dataset to video format and save with new repo_id:
-    lerobot-edit-dataset \
+    python -m lerobot.scripts.lerobot_edit_dataset \
         --repo_id lerobot/pusht_image \
         --new_repo_id lerobot/pusht_video \
         --operation.type convert_image_to_video
 
 Convert image dataset to video format and push to hub:
-    lerobot-edit-dataset \
+    python -m lerobot.scripts.lerobot_edit_dataset \
         --repo_id lerobot/pusht_image \
         --new_repo_id lerobot/pusht_video \
         --operation.type convert_image_to_video \
         --push_to_hub true
 
-Show dataset information:
-    lerobot-edit-dataset \
-        --repo_id lerobot/pusht_image \
-        --operation.type info \
-        --operation.show_features true
-
-Show dataset information without feature details:
-    lerobot-edit-dataset \
-        --repo_id lerobot/pusht_image \
-        --operation.type info \
-        --operation.show_features false
-
 Using JSON config file:
-    lerobot-edit-dataset \
+    python -m lerobot.scripts.lerobot_edit_dataset \
         --config_path path/to/edit_config.json
 """
 
 import abc
 import logging
 import shutil
-import sys
 from dataclasses import dataclass
 from pathlib import Path
 
@@ -170,6 +132,7 @@ from lerobot.datasets.dataset_tools import (
     modify_tasks,
     remove_feature,
     split_dataset,
+    trim_episode_start,
 )
 from lerobot.datasets.lerobot_dataset import LeRobotDataset
 from lerobot.utils.constants import HF_LEROBOT_HOME
@@ -199,7 +162,6 @@ class SplitConfig(OperationConfig):
 @dataclass
 class MergeConfig(OperationConfig):
     repo_ids: list[str] | None = None
-    roots: list[str] | None = None
 
 
 @OperationConfig.register_subclass("remove_feature")
@@ -215,6 +177,13 @@ class ModifyTasksConfig(OperationConfig):
     episode_tasks: dict[str, str] | None = None
 
 
+@OperationConfig.register_subclass("trim_episode_start")
+@dataclass
+class TrimEpisodeStartConfig(OperationConfig):
+    seconds: float | None = None
+    episode_indices: list[int] | None = None
+
+
 @OperationConfig.register_subclass("convert_image_to_video")
 @dataclass
 class ConvertImageToVideoConfig(OperationConfig):
@@ -230,49 +199,32 @@ class ConvertImageToVideoConfig(OperationConfig):
     max_frames_per_batch: int | None = None
 
 
-@OperationConfig.register_subclass("info")
-@dataclass
-class InfoConfig(OperationConfig):
-    show_features: bool = False
-
-
 @dataclass
 class EditDatasetConfig:
-    # Operation configuration.
+    repo_id: str
     operation: OperationConfig
-    # Input dataset identifier. Always required unless for Merge operation.
-    repo_id: str | None = None
-    # Root directory where the input dataset is stored. If not specified, defaults to $HF_LEROBOT_HOME/repo_id.
     root: str | None = None
-    # Edited dataset identifier. When both new_repo_id (resp. new_root) and repo_id (resp. root) are identical, modifications are applied in-place and a backup of the original dataset is created. Required for Merge operation.
     new_repo_id: str | None = None
-    # Root directory where the edited dataset will be stored. If not specified, defaults to $HF_LEROBOT_HOME/new_repo_id. For Split operation, this is the base directory for the split datasets.
-    new_root: str | None = None
-    # Upload dataset to Hugging Face hub.
     push_to_hub: bool = False
 
 
-def get_output_path(
+def get_output_path(repo_id: str, new_repo_id: str | None, root: Path | None) -> tuple[str, Path]:
-    repo_id: str,
+    if new_repo_id:
-    new_repo_id: str | None,
+        output_repo_id = new_repo_id
-    root: Path | str | None,
+        output_dir = root / new_repo_id if root else HF_LEROBOT_HOME / new_repo_id
-    new_root: Path | str | None,
+    else:
-) -> tuple[str, Path]:
+        output_repo_id = repo_id
-    input_path = Path(root) if root else HF_LEROBOT_HOME / repo_id
+        dataset_path = root / repo_id if root else HF_LEROBOT_HOME / repo_id
+        old_path = Path(str(dataset_path) + "_old")
 
-    output_repo_id = new_repo_id if new_repo_id else repo_id
+        if dataset_path.exists():
-    output_path = Path(new_root) if new_root else HF_LEROBOT_HOME / output_repo_id
+            if old_path.exists():
+                shutil.rmtree(old_path)
+            shutil.move(str(dataset_path), str(old_path))
 
-    # In case of in-place modification, create a backup of the original dataset (if it exists)
+        output_dir = dataset_path
-    if output_path == input_path:
-        backup_path = input_path.with_name(input_path.name + "_old")
 
-        if input_path.exists():
+    return output_repo_id, output_dir
-            if backup_path.exists():
-                shutil.rmtree(backup_path)
-            shutil.move(input_path, backup_path)
-
-    return output_repo_id, output_path
 
 
 def handle_delete_episodes(cfg: EditDatasetConfig) -> None:
@@ -284,15 +236,11 @@ def handle_delete_episodes(cfg: EditDatasetConfig) -> None:
 
     dataset = LeRobotDataset(cfg.repo_id, root=cfg.root)
     output_repo_id, output_dir = get_output_path(
-        cfg.repo_id,
+        cfg.repo_id, cfg.new_repo_id, Path(cfg.root) if cfg.root else None
-        new_repo_id=cfg.new_repo_id,
-        root=cfg.root,
-        new_root=cfg.new_root,
     )
 
-    # In case of in-place modification, make the dataset point to the backup directory
+    if cfg.new_repo_id is None:
-    if output_dir == dataset.root:
+        dataset.root = Path(str(dataset.root) + "_old")
-        dataset.root = dataset.root.with_name(dataset.root.name + "_old")
 
     logging.info(f"Deleting episodes {cfg.operation.episode_indices} from {cfg.repo_id}")
     new_dataset = delete_episodes(
@@ -319,27 +267,19 @@ def handle_split(cfg: EditDatasetConfig) -> None:
             "splits dict must be specified with split names as keys and fractions/episode lists as values"
         )
 
-    if cfg.new_repo_id is not None:
-        logging.warning(
-            "split uses the original dataset identifier --repo_id to generate split names. The --new_repo_id parameter is ignored."
-        )
-
     dataset = LeRobotDataset(cfg.repo_id, root=cfg.root)
 
     logging.info(f"Splitting dataset {cfg.repo_id} with splits: {cfg.operation.splits}")
-    split_datasets = split_dataset(
+    split_datasets = split_dataset(dataset, splits=cfg.operation.splits)
-        dataset,
-        splits=cfg.operation.splits,
-        output_dir=cfg.new_root,
-    )
 
     for split_name, split_ds in split_datasets.items():
+        split_repo_id = f"{cfg.repo_id}_{split_name}"
         logging.info(
             f"{split_name}: {split_ds.meta.total_episodes} episodes, {split_ds.meta.total_frames} frames"
         )
 
         if cfg.push_to_hub:
-            logging.info(f"Pushing {split_name} split to hub as {split_ds.repo_id}")
+            logging.info(f"Pushing {split_name} split to hub as {split_repo_id}")
             LeRobotDataset(split_ds.repo_id, root=split_ds.root).push_to_hub()
 
 
@@ -350,29 +290,18 @@ def handle_merge(cfg: EditDatasetConfig) -> None:
     if not cfg.operation.repo_ids:
         raise ValueError("repo_ids must be specified for merge operation")
 
-    if cfg.repo_id is not None or cfg.root is not None:
+    if not cfg.repo_id:
-        logging.warning(
+        raise ValueError("repo_id must be specified as the output repository for merged dataset")
-            "merge uses --new_repo_id and --new_root for the merged dataset. The --repo_id and --root parameters are ignored."
-        )
 
-    if cfg.operation.roots:
+    logging.info(f"Loading {len(cfg.operation.repo_ids)} datasets to merge")
-        if len(cfg.operation.roots) != len(cfg.operation.repo_ids):
+    datasets = [LeRobotDataset(repo_id, root=cfg.root) for repo_id in cfg.operation.repo_ids]
-            raise ValueError("repo_ids and roots must have the same length for merge operation")
-        logging.info(f"Loading {len(cfg.operation.roots)} datasets to merge")
-        datasets = [
-            LeRobotDataset(repo_id=repo_id, root=root)
-            for repo_id, root in zip(cfg.operation.repo_ids, cfg.operation.roots, strict=True)
-        ]
-    else:
-        logging.info(f"Loading {len(cfg.operation.repo_ids)} datasets to merge")
-        datasets = [LeRobotDataset(repo_id) for repo_id in cfg.operation.repo_ids]
 
-    output_dir = Path(cfg.new_root) if cfg.new_root else HF_LEROBOT_HOME / cfg.new_repo_id
+    output_dir = Path(cfg.root) / cfg.repo_id if cfg.root else HF_LEROBOT_HOME / cfg.repo_id
 
-    logging.info(f"Merging datasets into {cfg.new_repo_id}")
+    logging.info(f"Merging datasets into {cfg.repo_id}")
     merged_dataset = merge_datasets(
         datasets,
-        output_repo_id=cfg.new_repo_id,
+        output_repo_id=cfg.repo_id,
         output_dir=output_dir,
     )
 
@@ -382,7 +311,7 @@ def handle_merge(cfg: EditDatasetConfig) -> None:
     )
 
     if cfg.push_to_hub:
-        logging.info(f"Pushing to hub as {cfg.new_repo_id}")
+        logging.info(f"Pushing to hub as {cfg.repo_id}")
         LeRobotDataset(merged_dataset.repo_id, root=output_dir).push_to_hub()
 
 
@@ -395,15 +324,11 @@ def handle_remove_feature(cfg: EditDatasetConfig) -> None:
 
     dataset = LeRobotDataset(cfg.repo_id, root=cfg.root)
     output_repo_id, output_dir = get_output_path(
-        cfg.repo_id,
+        cfg.repo_id, cfg.new_repo_id, Path(cfg.root) if cfg.root else None
-        new_repo_id=cfg.new_repo_id,
-        root=cfg.root,
-        new_root=cfg.new_root,
     )
 
-    # In case of in-place modification, make the dataset point to the backup directory
+    if cfg.new_repo_id is None:
-    if output_dir == dataset.root:
+        dataset.root = Path(str(dataset.root) + "_old")
-        dataset.root = dataset.root.with_name(dataset.root.name + "_old")
 
     logging.info(f"Removing features {cfg.operation.feature_names} from {cfg.repo_id}")
     new_dataset = remove_feature(
@@ -431,10 +356,9 @@ def handle_modify_tasks(cfg: EditDatasetConfig) -> None:
     if new_task is None and episode_tasks_raw is None:
         raise ValueError("Must specify at least one of new_task or episode_tasks for modify_tasks operation")
 
-    if cfg.new_repo_id is not None or cfg.new_root is not None:
+    # Warn about in-place modification behavior
-        logging.warning(
+    if cfg.new_repo_id is not None:
-            "modify_tasks modifies datasets in-place. The --new_repo_id and --new_root parameters are ignored."
+        logging.warning("modify_tasks modifies datasets in-place. The --new_repo_id parameter is ignored.")
-        )
 
     dataset = LeRobotDataset(cfg.repo_id, root=cfg.root)
     logging.warning(f"Modifying dataset in-place at {dataset.root}. Original data will be overwritten.")
@@ -464,36 +388,73 @@ def handle_modify_tasks(cfg: EditDatasetConfig) -> None:
         modified_dataset.push_to_hub()
 
 
+def handle_trim_episode_start(cfg: EditDatasetConfig) -> None:
+    if not isinstance(cfg.operation, TrimEpisodeStartConfig):
+        raise ValueError("Operation config must be TrimEpisodeStartConfig")
+
+    if cfg.operation.seconds is None:
+        raise ValueError("seconds must be specified for trim_episode_start operation")
+
+    dataset = LeRobotDataset(cfg.repo_id, root=cfg.root)
+    output_repo_id, output_dir = get_output_path(
+        cfg.repo_id, cfg.new_repo_id, Path(cfg.root) if cfg.root else None
+    )
+
+    if cfg.new_repo_id is None:
+        dataset.root = Path(str(dataset.root) + "_old")
+
+    logging.info(
+        f"Trimming first {cfg.operation.seconds}s from episodes "
+        f"{cfg.operation.episode_indices if cfg.operation.episode_indices else 'ALL'} in {cfg.repo_id}"
+    )
+    new_dataset = trim_episode_start(
+        dataset=dataset,
+        seconds=cfg.operation.seconds,
+        episode_indices=cfg.operation.episode_indices,
+        output_dir=output_dir,
+        repo_id=output_repo_id,
+    )
+
+    logging.info(f"Dataset saved to {output_dir}")
+    logging.info(f"Episodes: {new_dataset.meta.total_episodes}, Frames: {new_dataset.meta.total_frames}")
+
+    if cfg.push_to_hub:
+        logging.info(f"Pushing to hub as {output_repo_id}")
+        LeRobotDataset(output_repo_id, root=output_dir).push_to_hub()
+
+
 def handle_convert_image_to_video(cfg: EditDatasetConfig) -> None:
     # Note: Parser may create any config type with the right fields, so we access fields directly
     # instead of checking isinstance()
     dataset = LeRobotDataset(cfg.repo_id, root=cfg.root)
 
     # Determine output directory and repo_id
-    # Priority: 1) new_root, 2) new_repo_id, 3) operation.output_dir, 4) auto-generated name
+    # Priority: 1) new_repo_id, 2) operation.output_dir, 3) auto-generated name
     output_dir_config = getattr(cfg.operation, "output_dir", None)
-    if output_dir_config:
-        logging.warning(
-            "--operation.output_dir is deprecated and will be removed in future versions. "
-            "Please use --new_root instead."
-        )
 
-    if cfg.new_root:
+    if cfg.new_repo_id:
-        output_dir = Path(cfg.new_root)
+        # Use new_repo_id for both local storage and hub push
-        output_repo_id = cfg.new_repo_id or f"{cfg.repo_id}_video"
-        logging.info(f"Saving to new_root: {output_dir} as {output_repo_id}")
-    elif cfg.new_repo_id:
         output_repo_id = cfg.new_repo_id
-        output_dir = HF_LEROBOT_HOME / cfg.new_repo_id
+        # Place new dataset as a sibling to the original dataset
+        # Get the parent of the actual dataset root (not cfg.root which might be the lerobot cache dir)
+        # Extract just the dataset name (after last slash) for the local directory
+        local_dir_name = cfg.new_repo_id.split("/")[-1]
+        output_dir = dataset.root.parent / local_dir_name
         logging.info(f"Saving to new dataset: {cfg.new_repo_id} at {output_dir}")
     elif output_dir_config:
+        # Use custom output directory for local-only storage
         output_dir = Path(output_dir_config)
+        # Extract repo name from output_dir for the dataset
         output_repo_id = output_dir.name
-        logging.info(f"Saving to local directory: {output_dir} as {output_repo_id}")
+        logging.info(f"Saving to local directory: {output_dir}")
     else:
+        # Auto-generate name: append "_video" to original repo_id
         output_repo_id = f"{cfg.repo_id}_video"
-        output_dir = HF_LEROBOT_HOME / output_repo_id
+        # Place new dataset as a sibling to the original dataset
-        logging.info(f"Saving to auto-generated location: {output_dir} as {output_repo_id}")
+        # Extract just the dataset name (after last slash) for the local directory
+        local_dir_name = output_repo_id.split("/")[-1]
+        output_dir = dataset.root.parent / local_dir_name
+        logging.info(f"Saving to auto-generated location: {output_dir}")
 
     logging.info(f"Converting dataset {cfg.repo_id} to video format")
 
@@ -525,63 +486,8 @@ def handle_convert_image_to_video(cfg: EditDatasetConfig) -> None:
         logging.info("Dataset saved locally (not pushed to hub)")
 
 
-def _get_dataset_size(repo_path):
-    import os
-
-    total = 0
-    with os.scandir(repo_path) as it:
-        for entry in it:
-            if entry.is_file():
-                total += entry.stat().st_size
-            elif entry.is_dir():
-                total += _get_dataset_size(entry.path)
-    return total
-
-
-def handle_info(cfg: EditDatasetConfig):
-    if not isinstance(cfg.operation, InfoConfig):
-        raise ValueError("Operation config must be InfoConfig")
-
-    dataset = LeRobotDataset(cfg.repo_id, root=cfg.root)
-    sys.stdout.write(f"======Info {dataset.meta.repo_id}\n")
-    sys.stdout.write(f"Repository ID: {dataset.meta.repo_id} \n")
-    sys.stdout.write(f"Total episode: {dataset.meta.total_episodes} \n")
-    sys.stdout.write(f"Total task: {dataset.meta.total_tasks} \n")
-    sys.stdout.write(f"Total frame(Actual Count): {dataset.meta.total_frames}({len(dataset)}) \n")
-    sys.stdout.write(
-        f"Average frame per episode: {dataset.meta.total_frames / dataset.meta.total_episodes:.1f}\n"
-    )
-    sys.stdout.write(
-        f"Average episode time(sec): {(dataset.meta.total_frames / dataset.meta.total_episodes) / dataset.meta.fps:.1f}\n"
-    )
-    sys.stdout.write(f"FPS: {dataset.meta.fps}\n")
-
-    total_file_size = _get_dataset_size(dataset.root)
-    sys.stdout.write(f"Size: {total_file_size / (1024 * 1024):.1f} MB\n")
-    if cfg.operation.show_features:
-        import json
-
-        feature_dump_str = json.dumps(
-            dataset.meta.features, ensure_ascii=False, indent=4, sort_keys=True, separators=(",", ": ")
-        )
-        sys.stdout.write("Features:\n")
-        sys.stdout.write(f"{feature_dump_str}\n")
-
-
-def _validate_config(cfg: EditDatasetConfig) -> None:
-    if isinstance(cfg.operation, MergeConfig):
-        if not cfg.new_repo_id:
-            raise ValueError("--new_repo_id is required for merge operation (the merged dataset identifier)")
-    else:
-        if not cfg.repo_id:
-            raise ValueError(
-                f"--repo_id is required for {cfg.operation.type} operation (the input dataset identifier)"
-            )
-
-
 @parser.wrap()
 def edit_dataset(cfg: EditDatasetConfig) -> None:
-    _validate_config(cfg)
     operation_type = cfg.operation.type
 
     if operation_type == "delete_episodes":
@@ -594,10 +500,10 @@ def edit_dataset(cfg: EditDatasetConfig) -> None:
         handle_remove_feature(cfg)
     elif operation_type == "modify_tasks":
         handle_modify_tasks(cfg)
+    elif operation_type == "trim_episode_start":
+        handle_trim_episode_start(cfg)
     elif operation_type == "convert_image_to_video":
         handle_convert_image_to_video(cfg)
-    elif operation_type == "info":
-        handle_info(cfg)
     else:
         available = ", ".join(OperationConfig.get_known_choices())
         raise ValueError(f"Unknown operation: {operation_type}\nAvailable operations: {available}")

src/lerobot/scripts/lerobot_find_joint_limits.py

@@ -61,7 +61,6 @@ from lerobot.teleoperators import (  # noqa: F401
     make_teleoperator_from_config,
     omx_leader,
     openarm_leader,
-    openarm_mini,
     so_leader,
 )
 from lerobot.utils.robot_utils import precise_sleep

src/lerobot/scripts/lerobot_mirror_dataset.py

@@ -0,0 +1,366 @@
+# Copyright 2025 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""
+Mirror a bimanual robot dataset by swapping left/right arms and inverting joint values.
+
+This script creates a mirrored version of a dataset where:
+1. Left and right arm observations/actions are swapped
+2. Joint values are inverted according to a mirroring mask
+3. Video frames are horizontally flipped
+
+Example usage:
+```shell
+python -m lerobot.scripts.lerobot_mirror_dataset \
+    --repo_id=pepijn/openarm_bimanual \
+    --output_repo_id=pepijn/openarm_bimanual_mirrored
+```
+"""
+
+import argparse
+import logging
+import os
+import subprocess
+from concurrent.futures import ThreadPoolExecutor, as_completed
+from pathlib import Path
+
+import numpy as np
+import pandas as pd
+from tqdm import tqdm
+
+from lerobot.datasets.lerobot_dataset import LeRobotDataset, LeRobotDatasetMetadata
+from lerobot.datasets.utils import (
+    DATA_DIR,
+    DEFAULT_DATA_PATH,
+    write_info,
+    write_stats,
+    write_tasks,
+)
+from lerobot.utils.constants import HF_LEROBOT_HOME
+
+logger = logging.getLogger(__name__)
+
+OPENARM_MIRRORING_MASK = {
+    "joint_1": -1,  # Pan - invert
+    "joint_2": -1,  # Lift - invert
+    "joint_3": -1,  # Roll - invert
+    "joint_4": 1,   # Elbow - no invert
+    "joint_5": -1,  # W-Roll - invert
+    "joint_6": -1,  # W-Pitch - invert
+    "joint_7": -1,  # W-Yaw - invert
+    "gripper": 1,   # Gripper - no invert
+}
+
+
+def get_mirroring_mask(robot_type: str) -> dict[str, int]:
+    """Get the mirroring mask for a given robot type."""
+    if robot_type in ["bi_openarm_follower", "openarm_follower", "bi_openarms_follower", "openarms_follower"]:
+        return OPENARM_MIRRORING_MASK
+    raise ValueError(f"Unknown robot type: {robot_type}. Add a mirroring mask for this robot.")
+
+
+def swap_left_right_name(name: str) -> str:
+    """Swap 'left' and 'right' in a feature name."""
+    # Use placeholder to avoid double-swap
+    result = name.replace("left_", "LEFT_PLACEHOLDER_")
+    result = result.replace("right_", "left_")
+    result = result.replace("LEFT_PLACEHOLDER_", "right_")
+    return result
+
+
+def mirror_feature_names(names: list[str]) -> tuple[list[str], dict[int, int]]:
+    """Mirror feature names by swapping left/right and return the new names and index mapping."""
+    mirrored_names = [swap_left_right_name(n) for n in names]
+    old_to_new_idx = {}
+    for old_idx, old_name in enumerate(names):
+        new_name = swap_left_right_name(old_name)
+        new_idx = mirrored_names.index(new_name)
+        old_to_new_idx[old_idx] = new_idx
+    return mirrored_names, old_to_new_idx
+
+
+def apply_mirroring_mask(
+    value: float,
+    feature_name: str,
+    mirroring_mask: dict[str, int],
+) -> float:
+    """Apply mirroring mask to a joint value."""
+    name_without_prefix = feature_name.split("_", 1)[1] if "_" in feature_name else feature_name
+    joint_name = name_without_prefix.split(".")[0]
+    if joint_name in mirroring_mask:
+        return value * mirroring_mask[joint_name]
+    return value
+
+
+def mirror_array(
+    array: np.ndarray,
+    names: list[str],
+    mirroring_mask: dict[str, int],
+) -> np.ndarray:
+    """Mirror an array of values (action or state) by swapping left/right and applying mask."""
+    mirrored_names, idx_mapping = mirror_feature_names(names)
+    result = np.zeros_like(array)
+    for old_idx, new_idx in idx_mapping.items():
+        old_name = names[old_idx]
+        new_name = mirrored_names[new_idx]
+        value = array[old_idx]
+        mirrored_value = apply_mirroring_mask(value, new_name, mirroring_mask)
+        result[new_idx] = mirrored_value
+    return result
+
+
+def flip_video_frames(
+    input_path: Path,
+    output_path: Path,
+    fps: float,
+    vcodec: str = "libsvtav1",
+):
+    """Flip video frames horizontally using FFmpeg with same settings as encode_video_frames."""
+    output_path.parent.mkdir(parents=True, exist_ok=True)
+    cmd = [
+        "ffmpeg", "-y", "-i", str(input_path),
+        "-vf", "hflip",
+        "-c:v", vcodec,
+        "-g", "2",
+        "-crf", "30",
+        "-r", str(int(fps)),
+        "-pix_fmt", "yuv420p",
+        "-loglevel", "error",
+    ]
+    if vcodec == "libsvtav1":
+        cmd.extend(["-preset", "12"])
+    cmd.append(str(output_path))
+    result = subprocess.run(cmd, capture_output=True, text=True)
+    if result.returncode != 0:
+        raise RuntimeError(f"FFmpeg failed: {result.stderr}")
+
+
+def mirror_dataset(
+    repo_id: str,
+    output_repo_id: str,
+    root: str | Path | None = None,
+    output_root: str | Path | None = None,
+    mirroring_mask: dict[str, int] | None = None,
+    vcodec: str = "libsvtav1",
+    num_workers: int | None = None,
+) -> LeRobotDataset:
+    """Mirror a bimanual robot dataset."""
+    logger.info(f"Loading dataset: {repo_id}")
+    dataset = LeRobotDataset(repo_id, root=root)
+
+    if mirroring_mask is None:
+        robot_type = dataset.meta.robot_type or "bi_openarms_follower"
+        mirroring_mask = get_mirroring_mask(robot_type)
+        logger.info(f"Using mirroring mask for robot type: {robot_type}")
+
+    output_root = Path(output_root) if output_root else HF_LEROBOT_HOME / output_repo_id
+
+    mirrored_features = {}
+    for key, feat in dataset.meta.features.items():
+        new_key = swap_left_right_name(key)
+        new_feat = feat.copy()
+        if "names" in new_feat and new_feat["names"]:
+            new_feat["names"] = [swap_left_right_name(n) for n in new_feat["names"]]
+        mirrored_features[new_key] = new_feat
+
+    logger.info("Creating mirrored dataset metadata...")
+    new_meta = LeRobotDatasetMetadata.create(
+        repo_id=output_repo_id,
+        fps=dataset.meta.fps,
+        features=mirrored_features,
+        robot_type=dataset.meta.robot_type,
+        root=output_root,
+        use_videos=len(dataset.meta.video_keys) > 0,
+    )
+
+    if dataset.meta.tasks is not None:
+        write_tasks(dataset.meta.tasks, new_meta.root)
+        new_meta.tasks = dataset.meta.tasks.copy()
+
+    _mirror_data(dataset, new_meta, mirroring_mask)
+    _mirror_videos(dataset, new_meta, vcodec, num_workers)
+    _copy_episodes_metadata(dataset, new_meta)
+
+    logger.info(f"Mirrored dataset saved to: {output_root}")
+    return LeRobotDataset(output_repo_id, root=output_root)
+
+
+def _mirror_data(
+    src_dataset: LeRobotDataset,
+    dst_meta: LeRobotDatasetMetadata,
+    mirroring_mask: dict[str, int],
+) -> None:
+    """Mirror parquet data files."""
+    data_dir = src_dataset.root / DATA_DIR
+    parquet_files = sorted(data_dir.glob("*/*.parquet"))
+
+    if not parquet_files:
+        raise ValueError(f"No parquet files found in {data_dir}")
+
+    action_names = src_dataset.meta.features.get("action", {}).get("names", [])
+    state_names = src_dataset.meta.features.get("observation.state", {}).get("names", [])
+
+    for src_path in tqdm(parquet_files, desc="Mirroring data files"):
+        df = pd.read_parquet(src_path).reset_index(drop=True)
+        relative_path = src_path.relative_to(src_dataset.root)
+        chunk_dir = relative_path.parts[1]
+        file_name = relative_path.parts[2]
+        chunk_idx = int(chunk_dir.split("-")[1])
+        file_idx = int(file_name.split("-")[1].split(".")[0])
+
+        if "action" in df.columns and action_names:
+            actions = np.stack(df["action"].values)
+            mirrored_actions = np.array([
+                mirror_array(row, action_names, mirroring_mask) for row in actions
+            ])
+            df["action"] = list(mirrored_actions)
+
+        if "observation.state" in df.columns and state_names:
+            states = np.stack(df["observation.state"].values)
+            mirrored_states = np.array([
+                mirror_array(row, state_names, mirroring_mask) for row in states
+            ])
+            df["observation.state"] = list(mirrored_states)
+
+        dst_path = dst_meta.root / DEFAULT_DATA_PATH.format(chunk_index=chunk_idx, file_index=file_idx)
+        dst_path.parent.mkdir(parents=True, exist_ok=True)
+        df.to_parquet(dst_path, index=False)
+
+
+def _mirror_videos(
+    src_dataset: LeRobotDataset,
+    dst_meta: LeRobotDatasetMetadata,
+    vcodec: str,
+    num_workers: int | None = None,
+) -> None:
+    """Mirror video files by flipping horizontally and swapping left/right names."""
+    if not src_dataset.meta.video_keys:
+        return
+
+    video_tasks = []
+    for old_video_key in src_dataset.meta.video_keys:
+        new_video_key = swap_left_right_name(old_video_key)
+        for ep_idx in range(src_dataset.meta.total_episodes):
+            try:
+                src_path = src_dataset.root / src_dataset.meta.get_video_file_path(ep_idx, old_video_key)
+                dst_relative = src_dataset.meta.get_video_file_path(ep_idx, old_video_key)
+                dst_relative_str = str(dst_relative).replace(old_video_key, new_video_key)
+                dst_path = dst_meta.root / dst_relative_str
+                if src_path.exists():
+                    video_tasks.append((src_path, dst_path))
+            except KeyError:
+                continue
+
+    def process_video(task, pbar):
+        src_path, dst_path = task
+        pbar.set_postfix_str(src_path.name)
+        flip_video_frames(src_path, dst_path, src_dataset.meta.fps, vcodec)
+        return src_path
+
+    if num_workers is None:
+        num_workers = os.cpu_count() or 16
+    num_workers = min(len(video_tasks), num_workers)
+    logger.info(f"Processing {len(video_tasks)} videos with {num_workers} workers")
+    with tqdm(total=len(video_tasks), desc="Mirroring videos") as pbar:
+        with ThreadPoolExecutor(max_workers=num_workers) as executor:
+            futures = {executor.submit(process_video, t, pbar): t for t in video_tasks}
+            for future in as_completed(futures):
+                task = futures[future]
+                future.result()
+                pbar.set_postfix_str(f"done: {task[0].name}")
+                pbar.update(1)
+
+
+def _copy_episodes_metadata(
+    src_dataset: LeRobotDataset,
+    dst_meta: LeRobotDatasetMetadata,
+) -> None:
+    """Copy episodes metadata with swapped video keys."""
+    episodes_dir = src_dataset.root / "meta/episodes"
+    dst_episodes_dir = dst_meta.root / "meta/episodes"
+
+    if episodes_dir.exists():
+        dst_episodes_dir.mkdir(parents=True, exist_ok=True)
+        for src_parquet in episodes_dir.glob("*/*.parquet"):
+            df = pd.read_parquet(src_parquet)
+            columns_to_rename = {}
+            for col in df.columns:
+                if col.startswith("videos/"):
+                    parts = col.split("/")
+                    if len(parts) >= 2:
+                        video_key = parts[1]
+                        new_video_key = swap_left_right_name(video_key)
+                        new_col = col.replace(f"videos/{video_key}/", f"videos/{new_video_key}/")
+                        columns_to_rename[col] = new_col
+            if columns_to_rename:
+                df = df.rename(columns=columns_to_rename)
+            dst_parquet = dst_episodes_dir / src_parquet.relative_to(episodes_dir)
+            dst_parquet.parent.mkdir(parents=True, exist_ok=True)
+            df.to_parquet(dst_parquet, index=False)
+
+    dst_meta.info.update({
+        "total_episodes": src_dataset.meta.total_episodes,
+        "total_frames": src_dataset.meta.total_frames,
+        "total_tasks": src_dataset.meta.total_tasks,
+        "total_videos": src_dataset.meta.total_videos,
+        "total_chunks": src_dataset.meta.total_chunks,
+    })
+    write_info(dst_meta.info, dst_meta.root)
+
+    if src_dataset.meta.stats is not None:
+        mirrored_stats = _mirror_stats(src_dataset.meta.stats)
+        write_stats(mirrored_stats, dst_meta.root)
+
+
+def _mirror_stats(stats: dict) -> dict:
+    """Mirror stats by swapping left/right in feature names."""
+    mirrored = {}
+    for key, value in stats.items():
+        new_key = swap_left_right_name(key)
+        if isinstance(value, dict):
+            mirrored[new_key] = _mirror_stats(value)
+        else:
+            mirrored[new_key] = value
+    return mirrored
+
+
+def main():
+    logging.basicConfig(level=logging.INFO)
+    parser = argparse.ArgumentParser(description="Mirror a bimanual robot dataset")
+    parser.add_argument("--repo_id", type=str, required=True, help="Source dataset repo_id")
+    parser.add_argument("--output_repo_id", type=str, required=True, help="Output dataset repo_id")
+    parser.add_argument("--root", type=str, default=None, help="Source dataset root directory")
+    parser.add_argument("--output_root", type=str, default=None, help="Output dataset root directory")
+    parser.add_argument("--vcodec", type=str, default="libsvtav1", help="Video codec (libsvtav1, h264, hevc)")
+    parser.add_argument("--num_workers", type=int, default=None, help="Number of parallel workers for video processing")
+    parser.add_argument("--push-to-hub", action="store_true", help="Push mirrored dataset to HuggingFace Hub")
+    args = parser.parse_args()
+
+    dataset = mirror_dataset(
+        repo_id=args.repo_id,
+        output_repo_id=args.output_repo_id,
+        root=args.root,
+        output_root=args.output_root,
+        vcodec=args.vcodec,
+        num_workers=args.num_workers,
+    )
+
+    if getattr(args, "push_to_hub", False):
+        logger.info(f"Pushing dataset to HuggingFace Hub: {args.output_repo_id}")
+        dataset.push_to_hub()
+
+
+if __name__ == "__main__":
+    main()
+

src/lerobot/scripts/lerobot_record.py

@@ -26,10 +26,8 @@ lerobot-record \
     --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 \
     --display_data=true
-    # <- Optional: specify video codec (auto, h264, hevc, libsvtav1). Default is libsvtav1. \
+    # <- Optional: specify video codec (h264, hevc, libsvtav1). Default is libsvtav1. \
     # --dataset.vcodec=h264 \
     # <- Teleop optional if you want to teleoperate to record or in between episodes with a policy \
     # --teleop.type=so100_leader \
@@ -60,10 +58,7 @@ lerobot-record \
   --display_data=true \
   --dataset.repo_id=${HF_USER}/bimanual-so-handover-cube \
   --dataset.num_episodes=25 \
-  --dataset.single_task="Grab and handover the red cube to the other arm" \
+  --dataset.single_task="Grab and handover the red cube to the other arm"
-  --dataset.streaming_encoding=true \
-  # --dataset.vcodec=auto \
-  --dataset.encoder_threads=2
 ```
 """
 
@@ -125,7 +120,6 @@ from lerobot.teleoperators import (  # noqa: F401
     make_teleoperator_from_config,
     omx_leader,
     openarm_leader,
-    openarm_mini,
     reachy2_teleoperator,
     so_leader,
     unitree_g1,
@@ -155,7 +149,7 @@ class DatasetRecordConfig:
     repo_id: str
     # A short but accurate description of the task performed during the recording (e.g. "Pick the Lego block and drop it in the box on the right.")
     single_task: str
-    # Root directory where the dataset will be stored (e.g. 'dataset/path'). If None, defaults to $HF_LEROBOT_HOME/repo_id.
+    # Root directory where the dataset will be stored (e.g. 'dataset/path').
     root: str | Path | None = None
     # Limit the frames per second.
     fps: int = 30
@@ -185,19 +179,9 @@ class DatasetRecordConfig:
     # Number of episodes to record before batch encoding videos
     # Set to 1 for immediate encoding (default behavior), or higher for batched encoding
     video_encoding_batch_size: int = 1
-    # Video codec for encoding videos. Options: 'h264', 'hevc', 'libsvtav1', 'auto',
+    # Video codec for encoding videos. Options: 'h264', 'hevc', 'libsvtav1'.
-    # or hardware-specific: 'h264_videotoolbox', 'h264_nvenc', 'h264_vaapi', 'h264_qsv'.
+    # Use 'h264' for faster encoding on systems where AV1 encoding is CPU-heavy.
-    # Use 'auto' to auto-detect the best available hardware encoder.
     vcodec: str = "libsvtav1"
-    # Enable streaming video encoding: encode frames in real-time during capture instead
-    # of writing PNG images first. Makes save_episode() near-instant. More info in the documentation: https://huggingface.co/docs/lerobot/streaming_video_encoding
-    streaming_encoding: bool = False
-    # Maximum number of frames to buffer per camera when using streaming encoding.
-    # ~1s buffer at 30fps. Provides backpressure if the encoder can't keep up.
-    encoder_queue_maxsize: int = 30
-    # Number of threads per encoder instance. None = auto (codec default).
-    # Lower values reduce CPU usage, maps to 'lp' (via svtav1-params) for libsvtav1 and 'threads' for h264/hevc..
-    encoder_threads: int | None = None
     # Rename map for the observation to override the image and state keys
     rename_map: dict[str, str] = field(default_factory=dict)
 
@@ -334,7 +318,6 @@ def record_loop(
         preprocessor.reset()
         postprocessor.reset()
 
-    no_action_count = 0
     timestamp = 0
     start_episode_t = time.perf_counter()
     while timestamp < control_time_s:
@@ -382,13 +365,11 @@ def record_loop(
             act = {**arm_action, **base_action} if len(base_action) > 0 else arm_action
             act_processed_teleop = teleop_action_processor((act, obs))
         else:
-            no_action_count += 1
+            logging.info(
-            if no_action_count == 1 or no_action_count % 10 == 0:
+                "No policy or teleoperator provided, skipping action generation."
-                logging.warning(
+                "This is likely to happen when resetting the environment without a teleop device."
-                    "No policy or teleoperator provided, skipping action generation. "
+                "The robot won't be at its rest position at the start of the next episode."
-                    "This is likely to happen when resetting the environment without a teleop device. "
+            )
-                    "The robot won't be at its rest position at the start of the next episode."
-                )
             continue
 
         # Applies a pipeline to the action, default is IdentityProcessor
@@ -417,14 +398,7 @@ def record_loop(
             )
 
         dt_s = time.perf_counter() - start_loop_t
-
+        precise_sleep(max(1 / fps - dt_s, 0.0))
-        sleep_time_s: float = 1 / fps - dt_s
-        if sleep_time_s < 0:
-            logging.warning(
-                f"Record loop is running slower ({1 / dt_s:.1f} Hz) than the target FPS ({fps} Hz). Dataset frames might be dropped and robot control might be unstable. Common causes are: 1) Camera FPS not keeping up 2) Policy inference taking too long 3) CPU starvation"
-            )
-
-        precise_sleep(max(sleep_time_s, 0.0))
 
         timestamp = time.perf_counter() - start_episode_t
 
@@ -471,9 +445,6 @@ def record(cfg: RecordConfig) -> LeRobotDataset:
                 root=cfg.dataset.root,
                 batch_encoding_size=cfg.dataset.video_encoding_batch_size,
                 vcodec=cfg.dataset.vcodec,
-                streaming_encoding=cfg.dataset.streaming_encoding,
-                encoder_queue_maxsize=cfg.dataset.encoder_queue_maxsize,
-                encoder_threads=cfg.dataset.encoder_threads,
             )
 
             if hasattr(robot, "cameras") and len(robot.cameras) > 0:
@@ -496,9 +467,6 @@ def record(cfg: RecordConfig) -> LeRobotDataset:
                 image_writer_threads=cfg.dataset.num_image_writer_threads_per_camera * len(robot.cameras),
                 batch_encoding_size=cfg.dataset.video_encoding_batch_size,
                 vcodec=cfg.dataset.vcodec,
-                streaming_encoding=cfg.dataset.streaming_encoding,
-                encoder_queue_maxsize=cfg.dataset.encoder_queue_maxsize,
-                encoder_threads=cfg.dataset.encoder_threads,
             )
 
         # Load pretrained policy
@@ -522,11 +490,6 @@ def record(cfg: RecordConfig) -> LeRobotDataset:
 
         listener, events = init_keyboard_listener()
 
-        if not cfg.dataset.streaming_encoding:
-            logging.info(
-                "Streaming encoding is disabled. If you have capable hardware, consider enabling it for way faster episode saving. --dataset.streaming_encoding=true --dataset.encoder_threads=2 # --dataset.vcodec=auto. More info in the documentation: https://huggingface.co/docs/lerobot/streaming_video_encoding"
-            )
-
         with VideoEncodingManager(dataset):
             recorded_episodes = 0
             while recorded_episodes < cfg.dataset.num_episodes and not events["stop_recording"]:

src/lerobot/scripts/lerobot_replay.py

@@ -22,7 +22,7 @@ lerobot-replay \
     --robot.type=so100_follower \
     --robot.port=/dev/tty.usbmodem58760431541 \
     --robot.id=black \
-    --dataset.repo_id=<USER>/record-test \
+    --dataset.repo_id=aliberts/record-test \
     --dataset.episode=0
 ```
 
@@ -80,7 +80,7 @@ class DatasetReplayConfig:
     repo_id: str
     # Episode to replay.
     episode: int
-    # Root directory where the dataset will be stored (e.g. 'dataset/path'). If None, defaults to $HF_LEROBOT_HOME/repo_id.
+    # Root directory where the dataset will be stored (e.g. 'dataset/path').
     root: str | Path | None = None
     # Limit the frames per second. By default, uses the policy fps.
     fps: int = 30

src/lerobot/scripts/lerobot_setup_can.py

@@ -152,7 +152,6 @@ def test_motor(bus, motor_id: int, timeout: float, use_fd: bool):
     )
     try:
         bus.send(disable_msg)
-        bus.recv(timeout=0.1)  # Clear any pending responses
     except Exception:
         print(f"Error sending message to motor 0x{motor_id:02X}")
 

src/lerobot/scripts/lerobot_setup_motors.py

@@ -43,7 +43,6 @@ from lerobot.teleoperators import (  # noqa: F401
     koch_leader,
     make_teleoperator_from_config,
     omx_leader,
-    openarm_mini,
     so_leader,
 )
 
@@ -52,7 +51,6 @@ COMPATIBLE_DEVICES = [
     "koch_leader",
     "omx_follower",
     "omx_leader",
-    "openarm_mini",
     "so100_follower",
     "so100_leader",
     "so101_follower",

src/lerobot/scripts/lerobot_teleoperate.py

@@ -94,7 +94,6 @@ from lerobot.teleoperators import (  # noqa: F401
     make_teleoperator_from_config,
     omx_leader,
     openarm_leader,
-    openarm_mini,
     reachy2_teleoperator,
     so_leader,
     unitree_g1,

src/lerobot/scripts/lerobot_train.py

@@ -24,7 +24,6 @@ import torch
 from accelerate import Accelerator
 from termcolor import colored
 from torch.optim import Optimizer
-from tqdm import tqdm
 
 from lerobot.configs import parser
 from lerobot.configs.train import TrainPipelineConfig
@@ -52,7 +51,6 @@ from lerobot.utils.utils import (
     format_big_number,
     has_method,
     init_logging,
-    inside_slurm,
 )
 
 
@@ -177,8 +175,6 @@ def train(cfg: TrainPipelineConfig, accelerator: Accelerator | None = None):
         from accelerate.utils import DistributedDataParallelKwargs
 
         ddp_kwargs = DistributedDataParallelKwargs(find_unused_parameters=True)
-        # Accelerate auto-detects the device based on the available hardware and ignores the policy.device setting.
-        # Force the device to be CPU when policy.device is set to CPU.
         force_cpu = cfg.policy.device == "cpu"
         accelerator = Accelerator(
             step_scheduler_with_optimizer=False,
@@ -213,16 +209,98 @@ def train(cfg: TrainPipelineConfig, accelerator: Accelerator | None = None):
     torch.backends.cuda.matmul.allow_tf32 = True
 
     # Dataset loading synchronization: main process downloads first to avoid race conditions
+    delta_action_stats = None
     if is_main_process:
         logging.info("Creating dataset")
         dataset = make_dataset(cfg)
 
+        # Compute delta action stats BEFORE distributed sync to avoid NCCL timeout
+        if getattr(cfg.policy, "use_delta_actions", False):
+            import numpy as np
+
+            from lerobot.datasets.compute_stats import get_feature_stats
+            from lerobot.processor.delta_action_processor import DeltaActionsProcessorStep, to_delta_actions
+
+            chunk_size = cfg.policy.chunk_size
+            hf = dataset.hf_dataset
+            total_frames = len(hf)
+            sample_upper_bound = total_frames - chunk_size
+            if sample_upper_bound <= 0:
+                raise ValueError(
+                    f"Cannot compute delta action stats: total_frames={total_frames}, chunk_size={chunk_size}"
+                )
+
+            max_samples = min(100_000, sample_upper_bound)
+            indices = np.random.choice(sample_upper_bound, max_samples, replace=False)
+
+            action_names = dataset.meta.features.get("action", {}).get("names")
+            delta_mask_step = DeltaActionsProcessorStep(
+                enabled=True,
+                exclude_joints=getattr(cfg.policy, "delta_exclude_joints", []),
+                action_names=action_names,
+            )
+            delta_mask = delta_mask_step._build_mask(dataset.meta.features["action"]["shape"][0])
+            logging.info(
+                f"use_delta_actions is enabled — computing delta action stats "
+                f"from {max_samples} chunk samples (chunk_size={chunk_size})"
+            )
+
+            all_delta_actions = []
+            episode_indices = np.array(hf["episode_index"])
+            for idx in indices:
+                idx = int(idx)
+                ep_idx = episode_indices[idx]
+                end_idx = min(idx + chunk_size, total_frames)
+                if end_idx > idx and episode_indices[end_idx - 1] != ep_idx:
+                    continue
+
+                chunk_data = hf[idx:end_idx]
+                actions = torch.tensor(np.stack([np.asarray(a) for a in chunk_data["action"]])).float()
+                state = torch.tensor(np.asarray(chunk_data["observation.state"][0])).float()
+
+                delta = to_delta_actions(actions.unsqueeze(0), state.unsqueeze(0), delta_mask).squeeze(0)
+                all_delta_actions.append(delta.numpy())
+
+            if not all_delta_actions:
+                raise RuntimeError("Failed to compute delta action stats: no valid chunks found.")
+
+            all_delta = np.concatenate(all_delta_actions, axis=0)
+            delta_stats = get_feature_stats(all_delta, axis=0, keepdims=all_delta.ndim == 1)
+            delta_action_stats = delta_stats
+            dataset.meta.stats["action"] = delta_action_stats
+
+            norm_type = "UNKNOWN"
+            if hasattr(cfg.policy, "normalization_mapping"):
+                from lerobot.configs.types import NormalizationMode
+                action_norm = cfg.policy.normalization_mapping.get("ACTION", None)
+                norm_type = action_norm.value if action_norm else "UNKNOWN"
+
+            excluded_dims = len(delta_mask) - sum(delta_mask)
+            logging.info(
+                f"Delta action stats ({len(all_delta_actions)} chunks, {len(all_delta)} values, norm={norm_type}): "
+                f"delta_dims={sum(delta_mask)}/{len(delta_mask)} (excluded={excluded_dims}), "
+                f"mean={np.abs(delta_stats['mean']).mean():.4f}, std={delta_stats['std'].mean():.4f}, "
+                f"q01={delta_stats['q01'].mean():.4f}, q99={delta_stats['q99'].mean():.4f}"
+            )
+            if norm_type == "QUANTILES":
+                q_range = (delta_stats['q99'] - delta_stats['q01']).mean()
+                logging.info(f"  Quantile range (q99-q01): {q_range:.4f}")
+
     accelerator.wait_for_everyone()
 
     # Now all other processes can safely load the dataset
     if not is_main_process:
         dataset = make_dataset(cfg)
 
+    # Ensure all ranks use the exact same delta action stats.
+    if getattr(cfg.policy, "use_delta_actions", False):
+        if accelerator.num_processes > 1 and torch.distributed.is_initialized():
+            stats_list = [delta_action_stats]
+            torch.distributed.broadcast_object_list(stats_list, src=0)
+            delta_action_stats = stats_list[0]
+        if delta_action_stats is not None:
+            dataset.meta.stats["action"] = delta_action_stats
+
     # Create environment used for evaluating checkpoints during training on simulation data.
     # On real-world data, no need to create an environment as evaluations are done outside train.py,
     # using the eval.py instead, with gym_dora environment and dora-rs.
@@ -248,10 +326,22 @@ def train(cfg: TrainPipelineConfig, accelerator: Accelerator | None = None):
     # Wait for all processes to finish policy creation before continuing
     accelerator.wait_for_everyone()
 
+    processor_pretrained_path = cfg.policy.pretrained_path
+    if (
+        getattr(cfg.policy, "use_delta_actions", False)
+        and processor_pretrained_path is not None
+        and not cfg.resume
+    ):
+        logging.warning(
+            "use_delta_actions=true with pretrained processors can skip delta transforms if "
+            "the checkpoint processors do not define them. Building processors from current policy config."
+        )
+        processor_pretrained_path = None
+
     # Create processors - only provide dataset_stats if not resuming from saved processors
     processor_kwargs = {}
     postprocessor_kwargs = {}
-    if (cfg.policy.pretrained_path and not cfg.resume) or not cfg.policy.pretrained_path:
+    if (processor_pretrained_path and not cfg.resume) or not processor_pretrained_path:
         # Only provide dataset_stats when not resuming from saved processor state
         processor_kwargs["dataset_stats"] = dataset.meta.stats
 
@@ -259,7 +349,7 @@ def train(cfg: TrainPipelineConfig, accelerator: Accelerator | None = None):
     if cfg.policy.type == "sarm":
         processor_kwargs["dataset_meta"] = dataset.meta
 
-    if cfg.policy.pretrained_path is not None:
+    if processor_pretrained_path is not None:
         processor_kwargs["preprocessor_overrides"] = {
             "device_processor": {"device": device.type},
             "normalizer_processor": {
@@ -281,7 +371,7 @@ def train(cfg: TrainPipelineConfig, accelerator: Accelerator | None = None):
 
     preprocessor, postprocessor = make_pre_post_processors(
         policy_cfg=cfg.policy,
-        pretrained_path=cfg.policy.pretrained_path,
+        pretrained_path=processor_pretrained_path,
         **processor_kwargs,
         **postprocessor_kwargs,
     )
@@ -380,10 +470,10 @@ def train(cfg: TrainPipelineConfig, accelerator: Accelerator | None = None):
         "dataloading_s": AverageMeter("data_s", ":.3f"),
     }
 
-    # Keep global batch size for logging; MetricsTracker handles world size internally.
+    # Use effective batch size for proper epoch calculation in distributed training
     effective_batch_size = cfg.batch_size * accelerator.num_processes
     train_tracker = MetricsTracker(
-        cfg.batch_size,
+        effective_batch_size,
         dataset.num_frames,
         dataset.num_episodes,
         train_metrics,
@@ -392,14 +482,6 @@ def train(cfg: TrainPipelineConfig, accelerator: Accelerator | None = None):
     )
 
     if is_main_process:
-        progbar = tqdm(
-            total=cfg.steps - step,
-            desc="Training",
-            unit="step",
-            disable=inside_slurm(),
-            position=0,
-            leave=True,
-        )
         logging.info(
             f"Start offline training on a fixed dataset, with effective batch size: {effective_batch_size}"
         )
@@ -407,7 +489,36 @@ def train(cfg: TrainPipelineConfig, accelerator: Accelerator | None = None):
     for _ in range(step, cfg.steps):
         start_time = time.perf_counter()
         batch = next(dl_iter)
+
+        # Debug logging for first few steps and periodically
+        if is_main_process and (step < 3 or (cfg.log_freq > 0 and step % (cfg.log_freq * 10) == 0)):
+            action = batch.get("action")
+            state = batch.get("observation.state")
+            if action is not None and state is not None:
+                logging.info(
+                    f"[DEBUG step={step}] PRE-PROCESSOR — "
+                    f"action: shape={tuple(action.shape)}, mean={action.mean():.4f}, std={action.std():.4f}, "
+                    f"min={action.min():.4f}, max={action.max():.4f} | "
+                    f"state: shape={tuple(state.shape)}, mean={state.mean():.4f}"
+                )
+
         batch = preprocessor(batch)
+
+        if is_main_process and (step < 3 or (cfg.log_freq > 0 and step % (cfg.log_freq * 10) == 0)):
+            action = batch.get("action")
+            state = batch.get("observation.state")
+            if action is not None:
+                logging.info(
+                    f"[DEBUG step={step}] POST-PROCESSOR — "
+                    f"action: shape={tuple(action.shape)}, mean={action.mean():.4f}, std={action.std():.4f}, "
+                    f"min={action.min():.4f}, max={action.max():.4f}"
+                )
+                if state is not None:
+                    logging.info(
+                        f"[DEBUG step={step}] POST-PROCESSOR — "
+                        f"state: shape={tuple(state.shape)}, mean={state.mean():.4f}, std={state.std():.4f}"
+                    )
+
         train_tracker.dataloading_s = time.perf_counter() - start_time
 
         train_tracker, output_dict = update_policy(
@@ -424,8 +535,6 @@ def train(cfg: TrainPipelineConfig, accelerator: Accelerator | None = None):
         # Note: eval and checkpoint happens *after* the `step`th training update has completed, so we
         # increment `step` here.
         step += 1
-        if is_main_process:
-            progbar.update(1)
         train_tracker.step()
         is_log_step = cfg.log_freq > 0 and step % cfg.log_freq == 0 and is_main_process
         is_saving_step = step % cfg.save_freq == 0 or step == cfg.steps
@@ -519,9 +628,6 @@ def train(cfg: TrainPipelineConfig, accelerator: Accelerator | None = None):
 
             accelerator.wait_for_everyone()
 
-    if is_main_process:
-        progbar.close()
-
     if eval_env:
         close_envs(eval_env)
 

src/lerobot/teleoperators/openarm_mini/__init__.py

@@ -1,20 +0,0 @@
-#!/usr/bin/env python
-
-# Copyright 2025 The HuggingFace Inc. team. All rights reserved.
-#
-# Licensed under the Apache License, Version 2.0 (the "License");
-# you may not use this file except in compliance with the License.
-# You may obtain a copy of the License at
-#
-#     http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing, software
-# distributed under the License is distributed on an "AS IS" BASIS,
-# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-# See the License for the specific language governing permissions and
-# limitations under the License.
-
-from .config_openarm_mini import OpenArmMiniConfig
-from .openarm_mini import OpenArmMini
-
-__all__ = ["OpenArmMini", "OpenArmMiniConfig"]

src/lerobot/teleoperators/openarm_mini/config_openarm_mini.py

@@ -1,30 +0,0 @@
-#!/usr/bin/env python
-
-# Copyright 2025 The HuggingFace Inc. team. All rights reserved.
-#
-# Licensed under the Apache License, Version 2.0 (the "License");
-# you may not use this file except in compliance with the License.
-# You may obtain a copy of the License at
-#
-#     http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing, software
-# distributed under the License is distributed on an "AS IS" BASIS,
-# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-# See the License for the specific language governing permissions and
-# limitations under the License.
-
-from dataclasses import dataclass
-
-from ..config import TeleoperatorConfig
-
-
-@TeleoperatorConfig.register_subclass("openarm_mini")
-@dataclass
-class OpenArmMiniConfig(TeleoperatorConfig):
-    """Configuration for OpenArm Mini teleoperator with Feetech motors (dual arms)."""
-
-    port_right: str = "/dev/ttyUSB0"
-    port_left: str = "/dev/ttyUSB1"
-
-    use_degrees: bool = True

src/lerobot/teleoperators/openarm_mini/openarm_mini.py

@@ -1,296 +0,0 @@
-#!/usr/bin/env python
-
-# Copyright 2025 The HuggingFace Inc. team. All rights reserved.
-#
-# Licensed under the Apache License, Version 2.0 (the "License");
-# you may not use this file except in compliance with the License.
-# You may obtain a copy of the License at
-#
-#     http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing, software
-# distributed under the License is distributed on an "AS IS" BASIS,
-# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-# See the License for the specific language governing permissions and
-# limitations under the License.
-
-import logging
-import time
-from typing import Any
-
-from lerobot.motors import Motor, MotorCalibration, MotorNormMode
-from lerobot.motors.feetech import (
-    FeetechMotorsBus,
-    OperatingMode,
-)
-from lerobot.processor import RobotAction
-from lerobot.utils.decorators import check_if_already_connected, check_if_not_connected
-
-from ..teleoperator import Teleoperator
-from .config_openarm_mini import OpenArmMiniConfig
-
-logger = logging.getLogger(__name__)
-
-# Motors whose direction is inverted during readout
-RIGHT_MOTORS_TO_FLIP = ["joint_1", "joint_2", "joint_3", "joint_4", "joint_5"]
-LEFT_MOTORS_TO_FLIP = ["joint_1", "joint_3", "joint_4", "joint_5", "joint_6", "joint_7"]
-
-
-class OpenArmMini(Teleoperator):
-    """
-    OpenArm Mini Teleoperator with dual Feetech-based arms (8 motors per arm).
-
-    Each arm has 7 joints plus a gripper, using Feetech STS3215 servos.
-    """
-
-    config_class = OpenArmMiniConfig
-    name = "openarm_mini"
-
-    def __init__(self, config: OpenArmMiniConfig):
-        super().__init__(config)
-        self.config = config
-
-        norm_mode_body = MotorNormMode.DEGREES
-
-        motors_right = {
-            "joint_1": Motor(1, "sts3215", norm_mode_body),
-            "joint_2": Motor(2, "sts3215", norm_mode_body),
-            "joint_3": Motor(3, "sts3215", norm_mode_body),
-            "joint_4": Motor(4, "sts3215", norm_mode_body),
-            "joint_5": Motor(5, "sts3215", norm_mode_body),
-            "joint_6": Motor(6, "sts3215", norm_mode_body),
-            "joint_7": Motor(7, "sts3215", norm_mode_body),
-            "gripper": Motor(8, "sts3215", MotorNormMode.RANGE_0_100),
-        }
-
-        motors_left = {
-            "joint_1": Motor(1, "sts3215", norm_mode_body),
-            "joint_2": Motor(2, "sts3215", norm_mode_body),
-            "joint_3": Motor(3, "sts3215", norm_mode_body),
-            "joint_4": Motor(4, "sts3215", norm_mode_body),
-            "joint_5": Motor(5, "sts3215", norm_mode_body),
-            "joint_6": Motor(6, "sts3215", norm_mode_body),
-            "joint_7": Motor(7, "sts3215", norm_mode_body),
-            "gripper": Motor(8, "sts3215", MotorNormMode.RANGE_0_100),
-        }
-
-        cal_right = {
-            k.replace("right_", ""): v for k, v in (self.calibration or {}).items() if k.startswith("right_")
-        }
-        cal_left = {
-            k.replace("left_", ""): v for k, v in (self.calibration or {}).items() if k.startswith("left_")
-        }
-
-        self.bus_right = FeetechMotorsBus(
-            port=self.config.port_right,
-            motors=motors_right,
-            calibration=cal_right,
-        )
-
-        self.bus_left = FeetechMotorsBus(
-            port=self.config.port_left,
-            motors=motors_left,
-            calibration=cal_left,
-        )
-
-    @property
-    def action_features(self) -> dict[str, type]:
-        features: dict[str, type] = {}
-        for motor in self.bus_right.motors:
-            features[f"right_{motor}.pos"] = float
-        for motor in self.bus_left.motors:
-            features[f"left_{motor}.pos"] = float
-        return features
-
-    @property
-    def feedback_features(self) -> dict[str, type]:
-        return {}
-
-    @property
-    def is_connected(self) -> bool:
-        return self.bus_right.is_connected and self.bus_left.is_connected
-
-    @check_if_already_connected
-    def connect(self, calibrate: bool = True) -> None:
-        logger.info(f"Connecting right arm on {self.config.port_right}...")
-        self.bus_right.connect()
-        logger.info(f"Connecting left arm on {self.config.port_left}...")
-        self.bus_left.connect()
-
-        if calibrate:
-            self.calibrate()
-
-        self.configure()
-        logger.info(f"{self} connected.")
-
-    @property
-    def is_calibrated(self) -> bool:
-        return self.bus_right.is_calibrated and self.bus_left.is_calibrated
-
-    def calibrate(self) -> None:
-        """
-        Run calibration procedure for OpenArm Mini.
-
-        1. Disable torque
-        2. Ask user to position arms in hanging position with grippers closed
-        3. Set this as zero position via half-turn homing
-        4. Interactive gripper calibration (open/close positions)
-        5. Save calibration
-        """
-        if self.calibration:
-            user_input = input(
-                f"Press ENTER to use existing calibration for {self.id}, "
-                f"or type 'c' and press ENTER to run new calibration: "
-            )
-            if user_input.strip().lower() != "c":
-                logger.info(f"Using existing calibration for {self.id}")
-                cal_right = {
-                    k.replace("right_", ""): v for k, v in self.calibration.items() if k.startswith("right_")
-                }
-                cal_left = {
-                    k.replace("left_", ""): v for k, v in self.calibration.items() if k.startswith("left_")
-                }
-                self.bus_right.write_calibration(cal_right)
-                self.bus_left.write_calibration(cal_left)
-                return
-
-        logger.info(f"\nRunning calibration for {self}")
-
-        self._calibrate_arm("right", self.bus_right)
-        self._calibrate_arm("left", self.bus_left)
-
-        self._save_calibration()
-        print(f"\nCalibration complete and saved to {self.calibration_fpath}")
-
-    def _calibrate_arm(self, arm_name: str, bus: FeetechMotorsBus) -> None:
-        """Calibrate a single arm with Feetech motors."""
-        logger.info(f"\n=== Calibrating {arm_name.upper()} arm ===")
-
-        bus.disable_torque()
-
-        logger.info(f"Setting Phase to 12 for all motors in {arm_name.upper()} arm...")
-        for motor in bus.motors:
-            bus.write("Phase", motor, 12)
-
-        for motor in bus.motors:
-            bus.write("Operating_Mode", motor, OperatingMode.POSITION.value)
-
-        input(
-            f"\nCalibration: Zero Position ({arm_name.upper()} arm)\n"
-            "Position the arm in the following configuration:\n"
-            "  - Arm hanging straight down\n"
-            "  - Gripper closed\n"
-            "Press ENTER when ready..."
-        )
-
-        homing_offsets = bus.set_half_turn_homings()
-        logger.info(f"{arm_name.capitalize()} arm zero position set.")
-
-        print(f"\nSetting motor ranges for {arm_name.upper()} arm\n")
-
-        if self.calibration is None:
-            self.calibration = {}
-
-        motor_resolution = bus.model_resolution_table[list(bus.motors.values())[0].model]
-        max_res = motor_resolution - 1
-
-        for motor_name, motor in bus.motors.items():
-            prefixed_name = f"{arm_name}_{motor_name}"
-
-            if motor_name == "gripper":
-                input(
-                    f"\nGripper Calibration ({arm_name.upper()} arm)\n"
-                    f"Step 1: CLOSE the gripper fully\n"
-                    f"Press ENTER when gripper is closed..."
-                )
-                closed_pos = bus.read("Present_Position", motor_name, normalize=False)
-                logger.info(f"  Gripper closed position recorded: {closed_pos}")
-
-                input("\nStep 2: OPEN the gripper fully\nPress ENTER when gripper is fully open...")
-                open_pos = bus.read("Present_Position", motor_name, normalize=False)
-                logger.info(f"  Gripper open position recorded: {open_pos}")
-
-                if closed_pos < open_pos:
-                    range_min = int(closed_pos)
-                    range_max = int(open_pos)
-                    drive_mode = 0
-                else:
-                    range_min = int(open_pos)
-                    range_max = int(closed_pos)
-                    drive_mode = 1
-
-                logger.info(
-                    f"  {prefixed_name}: range set to [{range_min}, {range_max}] "
-                    f"(0=closed, 100=open, drive_mode={drive_mode})"
-                )
-            else:
-                range_min = 0
-                range_max = max_res
-                drive_mode = 0
-                logger.info(f"  {prefixed_name}: range set to [0, {max_res}] (full motor range)")
-
-            self.calibration[prefixed_name] = MotorCalibration(
-                id=motor.id,
-                drive_mode=drive_mode,
-                homing_offset=homing_offsets[motor_name],
-                range_min=range_min,
-                range_max=range_max,
-            )
-
-        cal_for_bus = {
-            k.replace(f"{arm_name}_", ""): v
-            for k, v in self.calibration.items()
-            if k.startswith(f"{arm_name}_")
-        }
-        bus.write_calibration(cal_for_bus)
-
-    def configure(self) -> None:
-        self.bus_right.disable_torque()
-        self.bus_right.configure_motors()
-        for motor in self.bus_right.motors:
-            self.bus_right.write("Operating_Mode", motor, OperatingMode.POSITION.value)
-
-        self.bus_left.disable_torque()
-        self.bus_left.configure_motors()
-        for motor in self.bus_left.motors:
-            self.bus_left.write("Operating_Mode", motor, OperatingMode.POSITION.value)
-
-    def setup_motors(self) -> None:
-        print("\nSetting up RIGHT arm motors...")
-        for motor in reversed(self.bus_right.motors):
-            input(f"Connect the controller board to the RIGHT '{motor}' motor only and press enter.")
-            self.bus_right.setup_motor(motor)
-            print(f"RIGHT '{motor}' motor id set to {self.bus_right.motors[motor].id}")
-
-        print("\nSetting up LEFT arm motors...")
-        for motor in reversed(self.bus_left.motors):
-            input(f"Connect the controller board to the LEFT '{motor}' motor only and press enter.")
-            self.bus_left.setup_motor(motor)
-            print(f"LEFT '{motor}' motor id set to {self.bus_left.motors[motor].id}")
-
-    @check_if_not_connected
-    def get_action(self) -> RobotAction:
-        """Get current action from both arms (read positions from all motors)."""
-        start = time.perf_counter()
-
-        right_positions = self.bus_right.sync_read("Present_Position")
-        left_positions = self.bus_left.sync_read("Present_Position")
-
-        action: dict[str, Any] = {}
-        for motor, val in right_positions.items():
-            action[f"right_{motor}.pos"] = -val if motor in RIGHT_MOTORS_TO_FLIP else val
-        for motor, val in left_positions.items():
-            action[f"left_{motor}.pos"] = -val if motor in LEFT_MOTORS_TO_FLIP else val
-
-        dt_ms = (time.perf_counter() - start) * 1e3
-        logger.debug(f"{self} read action: {dt_ms:.1f}ms")
-        return action
-
-    def send_feedback(self, feedback: dict[str, float]) -> None:
-        raise NotImplementedError("Feedback is not yet implemented for OpenArm Mini.")
-
-    @check_if_not_connected
-    def disconnect(self) -> None:
-        self.bus_right.disconnect()
-        self.bus_left.disconnect()
-        logger.info(f"{self} disconnected.")

src/lerobot/teleoperators/utils.py

@@ -95,10 +95,6 @@ def make_teleoperator_from_config(config: TeleoperatorConfig) -> "Teleoperator":
         from .bi_openarm_leader import BiOpenArmLeader
 
         return BiOpenArmLeader(config)
-    elif config.type == "openarm_mini":
-        from .openarm_mini import OpenArmMini
-
-        return OpenArmMini(config)
     else:
         try:
             return cast("Teleoperator", make_device_from_device_class(config))

src/lerobot/utils/control_utils.py

@@ -189,7 +189,7 @@ def sanity_check_dataset_name(repo_id, policy_cfg):
     # Check if dataset_name starts with "eval_" but policy is missing
     if dataset_name.startswith("eval_") and policy_cfg is None:
         raise ValueError(
-            f"Your dataset name begins with 'eval_' ({dataset_name}), but no policy is provided."
+            f"Your dataset name begins with 'eval_' ({dataset_name}), but no policy is provided ({policy_cfg.type})."
         )
 
     # Check if dataset_name does not start with "eval_" but policy is provided

src/lerobot/utils/logging_utils.py

@@ -104,10 +104,9 @@ class MetricsTracker:
         self.metrics = metrics
 
         self.steps = initial_step
-        world_size = accelerator.num_processes if accelerator else 1
         # A sample is an (observation,action) pair, where observation and action
         # can be on multiple timestamps. In a batch, we have `batch_size` number of samples.
-        self.samples = self.steps * self._batch_size * world_size
+        self.samples = self.steps * self._batch_size
         self.episodes = self.samples / self._avg_samples_per_ep
         self.epochs = self.samples / self._num_frames
         self.accelerator = accelerator
@@ -133,8 +132,7 @@ class MetricsTracker:
         Updates metrics that depend on 'step' for one step.
         """
         self.steps += 1
-        world_size = self.accelerator.num_processes if self.accelerator else 1
+        self.samples += self._batch_size * (self.accelerator.num_processes if self.accelerator else 1)
-        self.samples += self._batch_size * world_size
         self.episodes = self.samples / self._avg_samples_per_ep
         self.epochs = self.samples / self._num_frames
 

src/lerobot/utils/robot_utils.py

@@ -16,14 +16,14 @@ import platform
 import time
 
 
-def precise_sleep(seconds: float, spin_threshold: float = 0.010, sleep_margin: float = 0.005):
+def precise_sleep(seconds: float, spin_threshold: float = 0.010, sleep_margin: float = 0.003):
     """
     Wait for `seconds` with better precision than time.sleep alone at the expense of more CPU usage.
 
     Parameters:
       - seconds: duration to wait
       - spin_threshold: if remaining <= spin_threshold -> spin; otherwise sleep (seconds). Default 10ms
-      - sleep_margin: when sleeping leave this much time before deadline to avoid oversleep. Default 5ms
+      - sleep_margin: when sleeping leave this much time before deadline to avoid oversleep. Default 3ms
 
     Note:
         The default parameters are chosen to prioritize timing accuracy over CPU usage for the common 30 FPS use case.

tests/artifacts/policies/pusht_diffusion_/actions.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:54aecbc1af72a4cd5e9261492f5e7601890517516257aacdf2a0ffb3ce281f1b
+oid sha256:19eaaa85f66ba4aa6388dbb83819ffad6ea4363247208f871a8dc385689f6fc8
 size 992

tests/artifacts/policies/pusht_diffusion_/grad_stats.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:88a9c3775a2aa1e90a08850521970070a4fcf0f6b82aab43cd8ccc5cf77e0013
+oid sha256:227296eaeeb54acdc3dae2eb8af3d4d08fb87e245337624447140b1e91cfd002
 size 47424

tests/artifacts/policies/pusht_diffusion_/output_dict.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:91a2635e05a75fe187a5081504c5f35ce3417378813fa2deaf9ca4e8200e1819
+oid sha256:271b00cb2f0cd5fd26b1d53463638e3d1a6e92692ec625fcffb420ca190869e5
 size 68

tests/artifacts/policies/pusht_diffusion_/param_stats.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:645bff922ac7bea63ad018ebf77c303c0e4cd2c1c0dc5ef3192865281bef3dc6
+oid sha256:778fddbbaa64248cee35cb377c02cc2b6076f7ce5855146de677128900617ddf
 size 47424