Refactor SACObservationEncoder to improve modularity and readability. Split initialization into dedicated methods for image and state layers, and enhance caching logic for image features. Update forward method to streamline feature encoding and ensure proper normalization handling.

Co-authored-by: s1lent4gnt <kmeftah.khalil@gmail.com>

.github/workflows/build-docker-images.yml

@@ -40,24 +40,24 @@ jobs:
           git lfs install
 
       - name: Set up Docker Buildx
-        uses: docker/setup-buildx-action@b5ca514318bd6ebac0fb2aedd5d36ec1b5c232a2 # v3.10.0
+        uses: docker/setup-buildx-action@v3
         with:
           cache-binary: false
 
       - name: Check out code
-        uses: actions/checkout@11bd71901bbe5b1630ceea73d27597364c9af683 # v4.2.2
+        uses: actions/checkout@v4
         with:
           lfs: true
           persist-credentials: false
 
       - name: Login to DockerHub
-        uses: docker/login-action@74a5d142397b4f367a81961eba4e8cd7edddf772 # v3.4.0
+        uses: docker/login-action@v3
         with:
           username: ${{ secrets.DOCKERHUB_USERNAME }}
           password: ${{ secrets.DOCKERHUB_PASSWORD }}
 
       - name: Build and Push CPU
-        uses: docker/build-push-action@ca052bb54ab0790a636c9b5f226502c73d547a25 # v5.4.0
+        uses: docker/build-push-action@v5
         with:
           context: .
           file: ./docker/lerobot-cpu/Dockerfile
@@ -78,24 +78,24 @@ jobs:
           git lfs install
 
       - name: Set up Docker Buildx
-        uses: docker/setup-buildx-action@b5ca514318bd6ebac0fb2aedd5d36ec1b5c232a2 # v3.10.0
+        uses: docker/setup-buildx-action@v3
         with:
           cache-binary: false
 
       - name: Check out code
-        uses: actions/checkout@11bd71901bbe5b1630ceea73d27597364c9af683 # v4.2.2
+        uses: actions/checkout@v4
         with:
           lfs: true
           persist-credentials: false
 
       - name: Login to DockerHub
-        uses: docker/login-action@74a5d142397b4f367a81961eba4e8cd7edddf772 # v3.4.0
+        uses: docker/login-action@v3
         with:
           username: ${{ secrets.DOCKERHUB_USERNAME }}
           password: ${{ secrets.DOCKERHUB_PASSWORD }}
 
       - name: Build and Push GPU
-        uses: docker/build-push-action@ca052bb54ab0790a636c9b5f226502c73d547a25 # v5.4.0
+        uses: docker/build-push-action@v5
         with:
           context: .
           file: ./docker/lerobot-gpu/Dockerfile
@@ -110,23 +110,23 @@ jobs:
       group: aws-general-8-plus
     steps:
       - name: Set up Docker Buildx
-        uses: docker/setup-buildx-action@b5ca514318bd6ebac0fb2aedd5d36ec1b5c232a2 # v3.10.0
+        uses: docker/setup-buildx-action@v3
         with:
           cache-binary: false
 
       - name: Check out code
-        uses: actions/checkout@11bd71901bbe5b1630ceea73d27597364c9af683 # v4.2.2
+        uses: actions/checkout@v4
         with:
           persist-credentials: false
 
       - name: Login to DockerHub
-        uses: docker/login-action@74a5d142397b4f367a81961eba4e8cd7edddf772 # v3.4.0
+        uses: docker/login-action@v3
         with:
           username: ${{ secrets.DOCKERHUB_USERNAME }}
           password: ${{ secrets.DOCKERHUB_PASSWORD }}
 
       - name: Build and Push GPU dev
-        uses: docker/build-push-action@ca052bb54ab0790a636c9b5f226502c73d547a25 # v5.4.0
+        uses: docker/build-push-action@v5
         with:
           context: .
           file: ./docker/lerobot-gpu-dev/Dockerfile

.github/workflows/build_documentation.yml

@@ -1,23 +0,0 @@
-name: Build documentation
-
-on:
-  workflow_dispatch:
-  push:
-    paths:
-      - "docs/**"
-    branches:
-    - main
-    - doc-builder*
-    - v*-release
-
-
-jobs:
-  build:  # zizmor: ignore[excessive-permissions] We follow the same pattern as in Transformers
-    uses: huggingface/doc-builder/.github/workflows/build_main_documentation.yml@main
-    with:
-      commit_sha: ${{ github.sha }}
-      package: lerobot
-      additional_args: --not_python_module
-    secrets:
-      token: ${{ secrets.HUGGINGFACE_PUSH }}
-      hf_token: ${{ secrets.HF_DOC_BUILD_PUSH }}

.github/workflows/build_pr_documentation.yml

@@ -1,19 +0,0 @@
-name: Build PR Documentation
-
-on:
-  pull_request:
-    paths:
-      - "docs/**"
-
-concurrency:
-  group: ${{ github.workflow }}-${{ github.head_ref || github.run_id }}
-  cancel-in-progress: true
-
-jobs:
-  build:  # zizmor: ignore[excessive-permissions] We follow the same pattern as in Transformers
-    uses: huggingface/doc-builder/.github/workflows/build_pr_documentation.yml@main
-    with:
-      commit_sha: ${{ github.event.pull_request.head.sha }}
-      pr_number: ${{ github.event.number }}
-      package: lerobot
-      additional_args: --not_python_module

.github/workflows/nightly-tests.yml

@@ -33,7 +33,7 @@ jobs:
     runs-on:
       group: aws-general-8-plus
     container:
-      image: huggingface/lerobot-cpu:latest  # zizmor: ignore[unpinned-images]
+      image: huggingface/lerobot-cpu:latest
       options: --shm-size "16gb"
       credentials:
         username: ${{ secrets.DOCKERHUB_USERNAME }}
@@ -60,7 +60,7 @@ jobs:
       CUDA_VISIBLE_DEVICES: "0"
       TEST_TYPE: "single_gpu"
     container:
-      image: huggingface/lerobot-gpu:latest  # zizmor: ignore[unpinned-images]
+      image: huggingface/lerobot-gpu:latest
       options: --gpus all --shm-size "16gb"
       credentials:
         username: ${{ secrets.DOCKERHUB_USERNAME }}

.github/workflows/quality.yml

@@ -33,12 +33,12 @@ jobs:
     runs-on: ubuntu-latest
     steps:
       - name: Checkout Repository
-        uses: actions/checkout@11bd71901bbe5b1630ceea73d27597364c9af683 # v4.2.2
+        uses: actions/checkout@v4
         with:
           persist-credentials: false
 
       - name: Set up Python
-        uses: actions/setup-python@7f4fc3e22c37d6ff65e88745f38bd3157c663f7c # v4.9.1
+        uses: actions/setup-python@v4
         with:
           python-version: ${{ env.PYTHON_VERSION }}
 
@@ -64,9 +64,9 @@ jobs:
     runs-on: ubuntu-latest
     steps:
       - name: Checkout Repository
-        uses: actions/checkout@11bd71901bbe5b1630ceea73d27597364c9af683 # v4.2.2
+        uses: actions/checkout@v4
         with:
           persist-credentials: false
 
       - name: typos-action
-        uses: crate-ci/typos@db35ee91e80fbb447f33b0e5fbddb24d2a1a884f # v1.29.10
+        uses: crate-ci/typos@v1.29.10

.github/workflows/test-docker-build.yml

@@ -35,7 +35,7 @@ jobs:
       matrix: ${{ steps.set-matrix.outputs.matrix }}
     steps:
       - name: Check out code
-        uses: actions/checkout@11bd71901bbe5b1630ceea73d27597364c9af683 # v4.2.2
+        uses: actions/checkout@v4
         with:
           persist-credentials: false
 
@@ -64,17 +64,17 @@ jobs:
         docker-file: ${{ fromJson(needs.get_changed_files.outputs.matrix) }}
     steps:
       - name: Set up Docker Buildx
-        uses: docker/setup-buildx-action@b5ca514318bd6ebac0fb2aedd5d36ec1b5c232a2 # v3.10.0
+        uses: docker/setup-buildx-action@v3
         with:
           cache-binary: false
 
       - name: Check out code
-        uses: actions/checkout@11bd71901bbe5b1630ceea73d27597364c9af683 # v4.2.2
+        uses: actions/checkout@v4
         with:
           persist-credentials: false
 
       - name: Build Docker image
-        uses: docker/build-push-action@ca052bb54ab0790a636c9b5f226502c73d547a25 # v5.4.0
+        uses: docker/build-push-action@v5
         with:
           file: ${{ matrix.docker-file }}
           context: .

.github/workflows/test.yml

@@ -50,7 +50,7 @@ jobs:
     env:
       MUJOCO_GL: egl
     steps:
-      - uses: actions/checkout@11bd71901bbe5b1630ceea73d27597364c9af683 # v4.2.2
+      - uses: actions/checkout@v4
         with:
           lfs: true  # Ensure LFS files are pulled
           persist-credentials: false
@@ -62,7 +62,7 @@ jobs:
           sudo apt-get install -y libegl1-mesa-dev ffmpeg portaudio19-dev
 
       - name: Install uv and python
-        uses: astral-sh/setup-uv@d4b2f3b6ecc6e67c4457f6d3e41ec42d3d0fcb86 # v5.4.2
+        uses: astral-sh/setup-uv@v5
         with:
           enable-cache: true
           version: ${{ env.UV_VERSION }}
@@ -85,7 +85,7 @@ jobs:
     env:
       MUJOCO_GL: egl
     steps:
-      - uses: actions/checkout@11bd71901bbe5b1630ceea73d27597364c9af683 # v4.2.2
+      - uses: actions/checkout@v4
         with:
           lfs: true  # Ensure LFS files are pulled
           persist-credentials: false
@@ -94,7 +94,7 @@ jobs:
         run: sudo apt-get update && sudo apt-get install -y ffmpeg
 
       - name: Install uv and python
-        uses: astral-sh/setup-uv@d4b2f3b6ecc6e67c4457f6d3e41ec42d3d0fcb86 # v5.4.2
+        uses: astral-sh/setup-uv@v5
         with:
           enable-cache: true
           version: ${{ env.UV_VERSION }}
@@ -117,7 +117,7 @@ jobs:
     env:
       MUJOCO_GL: egl
     steps:
-      - uses: actions/checkout@11bd71901bbe5b1630ceea73d27597364c9af683 # v4.2.2
+      - uses: actions/checkout@v4
         with:
           lfs: true  # Ensure LFS files are pulled
           persist-credentials: false
@@ -129,7 +129,7 @@ jobs:
           sudo apt-get install -y libegl1-mesa-dev ffmpeg portaudio19-dev
 
       - name: Install uv and python
-        uses: astral-sh/setup-uv@d4b2f3b6ecc6e67c4457f6d3e41ec42d3d0fcb86 # v5.4.2
+        uses: astral-sh/setup-uv@v5
         with:
           enable-cache: true
           version: ${{ env.UV_VERSION }}

.github/workflows/trufflehog.yml

@@ -24,12 +24,12 @@ jobs:
     runs-on: ubuntu-latest
     steps:
     - name: Checkout code
-      uses: actions/checkout@11bd71901bbe5b1630ceea73d27597364c9af683 # v4.2.2
+      uses: actions/checkout@v4
       with:
         fetch-depth: 0
         persist-credentials: false
 
     - name: Secret Scanning
-      uses: trufflesecurity/trufflehog@90694bf9af66e7536abc5824e7a87246dbf933cb # v3.88.35
+      uses: trufflesecurity/trufflehog@main
       with:
         extra_args: --only-verified

.github/workflows/upload_pr_documentation.yml

@@ -1,16 +0,0 @@
-name: Upload PR Documentation
-
-on: # zizmor: ignore[dangerous-triggers] We follow the same pattern as in Transformers
-  workflow_run:
-    workflows: [ "Build PR Documentation" ]
-    types:
-    - completed
-
-jobs:
-  build:  # zizmor: ignore[excessive-permissions] We follow the same pattern as in Transformers
-    uses: huggingface/doc-builder/.github/workflows/upload_pr_documentation.yml@main
-    with:
-      package_name: lerobot
-    secrets:
-      hf_token: ${{ secrets.HF_DOC_BUILD_PUSH }}
-      comment_bot_token: ${{ secrets.COMMENT_BOT_TOKEN }}

.gitignore

@@ -26,6 +26,7 @@ outputs
 
 # VS Code
 .vscode
+.devcontainer
 
 # HPC
 nautilus/*.yaml

.pre-commit-config.yaml

@@ -36,19 +36,19 @@ repos:
       - id: end-of-file-fixer
       - id: trailing-whitespace
 
-  - repo: https://github.com/adhtruong/mirrors-typos
-    rev: v1.32.0
+  - repo: https://github.com/crate-ci/typos
+    rev: v1.30.2
     hooks:
       - id: typos
         args: [--force-exclude]
 
   - repo: https://github.com/asottile/pyupgrade
-    rev: v3.20.0
+    rev: v3.19.1
     hooks:
     -   id: pyupgrade
-
+        exclude: '^(.*_pb2_grpc\.py|.*_pb2\.py$)'
   - repo: https://github.com/astral-sh/ruff-pre-commit
-    rev: v0.11.11
+    rev: v0.9.10
     hooks:
       - id: ruff
         args: [--fix]
@@ -57,12 +57,12 @@ repos:
 
   ##### Security #####
   - repo: https://github.com/gitleaks/gitleaks
-    rev: v8.26.0
+    rev: v8.24.0
     hooks:
       - id: gitleaks
 
   - repo: https://github.com/woodruffw/zizmor-pre-commit
-    rev: v1.8.0
+    rev: v1.4.1
     hooks:
       - id: zizmor
 

README.md

@@ -23,35 +23,21 @@
 </div>
 
 <h2 align="center">
-    <p><a href="https://github.com/huggingface/lerobot/blob/main/examples/12_use_so101.md">
-        Build Your Own SO-101 Robot!</a></p>
+    <p><a href="https://github.com/huggingface/lerobot/blob/main/examples/10_use_so100.md">
+        Build Your Own SO-100 Robot!</a></p>
 </h2>
 
 <div align="center">
-  <div style="display: flex; gap: 1rem; justify-content: center; align-items: center;" >
-    <img
-      src="media/so101/so101.webp?raw=true"
-      alt="SO-101 follower arm"
-      title="SO-101 follower arm"
-      style="width: 40%;"
-    />
-    <img
-      src="media/so101/so101-leader.webp?raw=true"
-      alt="SO-101 leader arm"
-      title="SO-101 leader arm"
-      style="width: 40%;"
-    />
-  </div>
+  <img src="media/so100/leader_follower.webp?raw=true" alt="SO-100 leader and follower arms" title="SO-100 leader and follower arms" width="50%">
 
-
-  <p><strong>Meet the updated SO100, the SO-101 – Just €114 per arm!</strong></p>
+  <p><strong>Meet the SO-100 – Just $110 per arm!</strong></p>
   <p>Train it in minutes with a few simple moves on your laptop.</p>
   <p>Then sit back and watch your creation act autonomously! 🤯</p>
 
-  <p><a href="https://github.com/huggingface/lerobot/blob/main/examples/12_use_so101.md">
-      See the full SO-101 tutorial here.</a></p>
+  <p><a href="https://github.com/huggingface/lerobot/blob/main/examples/10_use_so100.md">
+      Get the full SO-100 tutorial here.</a></p>
 
-  <p>Want to take it to the next level? Make your SO-101 mobile by building LeKiwi!</p>
+  <p>Want to take it to the next level? Make your SO-100 mobile by building LeKiwi!</p>
   <p>Check out the <a href="https://github.com/huggingface/lerobot/blob/main/examples/11_use_lekiwi.md">LeKiwi tutorial</a> and bring your robot to life on wheels.</p>
 
   <img src="media/lekiwi/kiwi.webp?raw=true" alt="LeKiwi mobile robot" title="LeKiwi mobile robot" width="50%">
@@ -65,6 +51,7 @@
 
 ---
 
+
 🤗 LeRobot aims to provide models, datasets, and tools for real-world robotics in PyTorch. The goal is to lower the barrier to entry to robotics so that everyone can contribute and benefit from sharing datasets and pretrained models.
 
 🤗 LeRobot contains state-of-the-art approaches that have been shown to transfer to the real-world with a focus on imitation learning and reinforcement learning.
@@ -111,25 +98,18 @@ conda create -y -n lerobot python=3.10
 conda activate lerobot
 ```
 
-When using `miniconda`, install `ffmpeg` in your environment:
+When using `miniconda`, if you don't have `ffmpeg` in your environment:
 ```bash
-conda install ffmpeg -c conda-forge
+conda install ffmpeg
 ```
 
-> **NOTE:** This usually installs `ffmpeg 7.X` for your platform compiled with the `libsvtav1` encoder. If `libsvtav1` is not supported (check supported encoders with `ffmpeg -encoders`), you can:
->  - _[On any platform]_ Explicitly install `ffmpeg 7.X` using:
->  ```bash
->  conda install ffmpeg=7.1.1 -c conda-forge
->  ```
->  - _[On Linux only]_ 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`.
-
 Install 🤗 LeRobot:
 ```bash
-pip install -e .
+pip install --no-binary=av -e .
 ```
 
 > **NOTE:** If you encounter build errors, you may need to install additional dependencies (`cmake`, `build-essential`, and `ffmpeg libs`). On Linux, run:
-`sudo apt-get install cmake build-essential python3-dev pkg-config libavformat-dev libavcodec-dev libavdevice-dev libavutil-dev libswscale-dev libswresample-dev libavfilter-dev pkg-config`. For other systems, see: [Compiling PyAV](https://pyav.org/docs/develop/overview/installation.html#bring-your-own-ffmpeg)
+`sudo apt-get install cmake build-essential python-dev pkg-config libavformat-dev libavcodec-dev libavdevice-dev libavutil-dev libswscale-dev libswresample-dev libavfilter-dev pkg-config`. For other systems, see: [Compiling PyAV](https://pyav.org/docs/develop/overview/installation.html#bring-your-own-ffmpeg)
 
 For simulations, 🤗 LeRobot comes with gymnasium environments that can be installed as extras:
 - [aloha](https://github.com/huggingface/gym-aloha)
@@ -138,7 +118,7 @@ For simulations, 🤗 LeRobot comes with gymnasium environments that can be inst
 
 For instance, to install 🤗 LeRobot with aloha and pusht, use:
 ```bash
-pip install -e ".[aloha, pusht]"
+pip install --no-binary=av -e ".[aloha, pusht]"
 ```
 
 To use [Weights and Biases](https://docs.wandb.ai/quickstart) for experiment tracking, log in with
@@ -221,7 +201,7 @@ dataset attributes:
   │  ├ episode_index (int64): index of the episode for this sample
   │  ├ frame_index (int64): index of the frame for this sample in the episode ; starts at 0 for each episode
   │  ├ timestamp (float32): timestamp in the episode
-  │  ├ next.done (bool): indicates the end of an episode ; True for the last frame in each episode
+  │  ├ next.done (bool): indicates the end of en episode ; True for the last frame in each episode
   │  └ index (int64): general index in the whole dataset
   ├ episode_data_index: contains 2 tensors with the start and end indices of each episode
   │  ├ from (1D int64 tensor): first frame index for each episode — shape (num episodes,) starts with 0
@@ -270,7 +250,7 @@ See `python lerobot/scripts/eval.py --help` for more instructions.
 
 ### Train your own policy
 
-Check out [example 3](./examples/3_train_policy.py) that illustrates how to train a model using our core library in python, and [example 4](./examples/4_train_policy_with_script.md) that shows how to use our training script from command line.
+Check out [example 3](./examples/3_train_policy.py) that illustrate how to train a model using our core library in python, and [example 4](./examples/4_train_policy_with_script.md) that shows how to use our training script from command line.
 
 To use wandb for logging training and evaluation curves, make sure you've run `wandb login` as a one-time setup step. Then, when running the training command above, enable WandB in the configuration by adding `--wandb.enable=true`.
 
@@ -321,7 +301,7 @@ Once you have trained a policy you may upload it to the Hugging Face hub using a
 You first need to find the checkpoint folder located inside your experiment directory (e.g. `outputs/train/2024-05-05/20-21-12_aloha_act_default/checkpoints/002500`). Within that there is a `pretrained_model` directory which should contain:
 - `config.json`: A serialized version of the policy configuration (following the policy's dataclass config).
 - `model.safetensors`: A set of `torch.nn.Module` parameters, saved in [Hugging Face Safetensors](https://huggingface.co/docs/safetensors/index) format.
-- `train_config.json`: A consolidated configuration containing all parameters used for training. The policy configuration should match `config.json` exactly. This is useful for anyone who wants to evaluate your policy or for reproducibility.
+- `train_config.json`: A consolidated configuration containing all parameter userd for training. The policy configuration should match `config.json` exactly. Thisis useful for anyone who wants to evaluate your policy or for reproducibility.
 
 To upload these to the hub, run the following:
 ```bash
@@ -360,7 +340,7 @@ with profile(
 If you want, you can cite this work with:
 ```bibtex
 @misc{cadene2024lerobot,
-    author = {Cadene, Remi and Alibert, Simon and Soare, Alexander and Gallouedec, Quentin and Zouitine, Adil and Palma, Steven and Kooijmans, Pepijn and Aractingi, Michel and Shukor, Mustafa and Aubakirova, Dana and Russi, Martino and Capuano, Francesco and Pascale, Caroline and Choghari, Jade and Moss, Jess and Wolf, Thomas},
+    author = {Cadene, Remi and Alibert, Simon and Soare, Alexander and Gallouedec, Quentin and Zouitine, Adil and Wolf, Thomas},
     title = {LeRobot: State-of-the-art Machine Learning for Real-World Robotics in Pytorch},
     howpublished = "\url{https://github.com/huggingface/lerobot}",
     year = {2024}

benchmarks/video/capture_camera_feed.py

@@ -17,21 +17,12 @@
 
 import argparse
 import datetime as dt
-import os
-import time
 from pathlib import Path
 
 import cv2
-import rerun as rr
-
-# see https://rerun.io/docs/howto/visualization/limit-ram
-RERUN_MEMORY_LIMIT = os.getenv("LEROBOT_RERUN_MEMORY_LIMIT", "5%")
 
 
-def display_and_save_video_stream(output_dir: Path, fps: int, width: int, height: int, duration: int):
-    rr.init("lerobot_capture_camera_feed")
-    rr.spawn(memory_limit=RERUN_MEMORY_LIMIT)
-
+def display_and_save_video_stream(output_dir: Path, fps: int, width: int, height: int):
     now = dt.datetime.now()
     capture_dir = output_dir / f"{now:%Y-%m-%d}" / f"{now:%H-%M-%S}"
     if not capture_dir.exists():
@@ -48,21 +39,24 @@ def display_and_save_video_stream(output_dir: Path, fps: int, width: int, height
     cap.set(cv2.CAP_PROP_FRAME_HEIGHT, height)
 
     frame_index = 0
-    start_time = time.time()
-    while time.time() - start_time < duration:
+    while True:
         ret, frame = cap.read()
 
         if not ret:
             print("Error: Could not read frame.")
             break
-        rr.log("video/stream", rr.Image(frame.numpy()), static=True)
+
+        cv2.imshow("Video Stream", frame)
         cv2.imwrite(str(capture_dir / f"frame_{frame_index:06d}.png"), frame)
         frame_index += 1
 
-    # Release the capture
-    cap.release()
+        # Break the loop on 'q' key press
+        if cv2.waitKey(1) & 0xFF == ord("q"):
+            break
 
-    # TODO(Steven): Add a graceful shutdown via a close() method for the Viewer context, though not currently supported in the Rerun API.
+    # Release the capture and destroy all windows
+    cap.release()
+    cv2.destroyAllWindows()
 
 
 if __name__ == "__main__":
@@ -92,11 +86,5 @@ if __name__ == "__main__":
         default=720,
         help="Height of the captured images.",
     )
-    parser.add_argument(
-        "--duration",
-        type=int,
-        default=20,
-        help="Duration in seconds for which the video stream should be captured.",
-    )
     args = parser.parse_args()
     display_and_save_video_stream(**vars(args))

benchmarks/video/run_video_benchmark.py

@@ -32,7 +32,11 @@ import numpy as np
 import pandas as pd
 import PIL
 import torch
-from skimage.metrics import mean_squared_error, peak_signal_noise_ratio, structural_similarity
+from skimage.metrics import (
+    mean_squared_error,
+    peak_signal_noise_ratio,
+    structural_similarity,
+)
 from tqdm import tqdm
 
 from lerobot.common.datasets.lerobot_dataset import LeRobotDataset
@@ -94,7 +98,11 @@ def load_original_frames(imgs_dir: Path, timestamps: list[float], fps: int) -> t
 
 
 def save_decoded_frames(
-    imgs_dir: Path, save_dir: Path, frames: torch.Tensor, timestamps: list[float], fps: int
+    imgs_dir: Path,
+    save_dir: Path,
+    frames: torch.Tensor,
+    timestamps: list[float],
+    fps: int,
 ) -> None:
     if save_dir.exists() and len(list(save_dir.glob("frame_*.png"))) == len(timestamps):
         return
@@ -104,7 +112,10 @@ def save_decoded_frames(
         idx = int(ts * fps)
         frame_hwc = (frames[i].permute((1, 2, 0)) * 255).type(torch.uint8).cpu().numpy()
         PIL.Image.fromarray(frame_hwc).save(save_dir / f"frame_{idx:06d}_decoded.png")
-        shutil.copyfile(imgs_dir / f"frame_{idx:06d}.png", save_dir / f"frame_{idx:06d}_original.png")
+        shutil.copyfile(
+            imgs_dir / f"frame_{idx:06d}.png",
+            save_dir / f"frame_{idx:06d}_original.png",
+        )
 
 
 def save_first_episode(imgs_dir: Path, dataset: LeRobotDataset) -> None:
@@ -120,7 +131,11 @@ def save_first_episode(imgs_dir: Path, dataset: LeRobotDataset) -> None:
     imgs_dataset = hf_dataset.select_columns(img_keys[0])
 
     for i, item in enumerate(
-        tqdm(imgs_dataset, desc=f"saving {dataset.repo_id} first episode images", leave=False)
+        tqdm(
+            imgs_dataset,
+            desc=f"saving {dataset.repo_id} first episode images",
+            leave=False,
+        )
     ):
         img = item[img_keys[0]]
         img.save(str(imgs_dir / f"frame_{i:06d}.png"), quality=100)
@@ -275,7 +290,9 @@ def benchmark_encoding_decoding(
     random.seed(seed)
     benchmark_table = []
     for timestamps_mode in tqdm(
-        decoding_cfg["timestamps_modes"], desc="decodings (timestamps_modes)", leave=False
+        decoding_cfg["timestamps_modes"],
+        desc="decodings (timestamps_modes)",
+        leave=False,
     ):
         for backend in tqdm(decoding_cfg["backends"], desc="decodings (backends)", leave=False):
             benchmark_row = benchmark_decoding(
@@ -416,7 +433,7 @@ if __name__ == "__main__":
         "--vcodec",
         type=str,
         nargs="*",
-        default=["libx264", "hevc", "libsvtav1"],
+        default=["libx264", "libx265", "libsvtav1"],
         help="Video codecs to be tested",
     )
     parser.add_argument(
@@ -446,7 +463,7 @@ if __name__ == "__main__":
     #     nargs="*",
     #     default=[0, 1],
     #     help="Use the fastdecode tuning option. 0 disables it. "
-    #         "For libx264 and libx265/hevc, only 1 is possible. "
+    #         "For libx264 and libx265, only 1 is possible. "
     #         "For libsvtav1, 1, 2 or 3 are possible values with a higher number meaning a faster decoding optimization",
     # )
     parser.add_argument(

docker/lerobot-gpu-dev/Dockerfile

@@ -14,7 +14,7 @@ RUN apt-get update && apt-get install -y --no-install-recommends \
     tcpdump sysstat screen tmux \
     libglib2.0-0 libgl1-mesa-glx libegl1-mesa \
     speech-dispatcher portaudio19-dev libgeos-dev \
-    python${PYTHON_VERSION} python${PYTHON_VERSION}-venv python${PYTHON_VERSION}-dev \
+    python${PYTHON_VERSION} python${PYTHON_VERSION}-venv \
     && apt-get clean && rm -rf /var/lib/apt/lists/*
 
 # Install ffmpeg build dependencies. See:

docs/README.md

@@ -1,137 +0,0 @@
-<!---
-Copyright 2020 The HuggingFace 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.
--->
-
-# Generating the documentation
-
-To generate the documentation, you first have to build it. Several packages are necessary to build the doc,
-you can install them with the following command, at the root of the code repository:
-
-```bash
-pip install -e ".[docs]"
-```
-
-You will also need `nodejs`. Please refer to their [installation page](https://nodejs.org/en/download)
-
----
-**NOTE**
-
-You only need to generate the documentation to inspect it locally (if you're planning changes and want to
-check how they look before committing for instance). You don't have to `git commit` the built documentation.
-
----
-
-## Building the documentation
-
-Once you have setup the `doc-builder` and additional packages, you can generate the documentation by
-typing the following command:
-
-```bash
-doc-builder build lerobot docs/source/ --build_dir ~/tmp/test-build
-```
-
-You can adapt the `--build_dir` to set any temporary folder that you prefer. This command will create it and generate
-the MDX files that will be rendered as the documentation on the main website. You can inspect them in your favorite
-Markdown editor.
-
-## Previewing the documentation
-
-To preview the docs, first install the `watchdog` module with:
-
-```bash
-pip install watchdog
-```
-
-Then run the following command:
-
-```bash
-doc-builder preview lerobot docs/source/
-```
-
-The docs will be viewable at [http://localhost:3000](http://localhost:3000). You can also preview the docs once you have opened a PR. You will see a bot add a comment to a link where the documentation with your changes lives.
-
----
-**NOTE**
-
-The `preview` command only works with existing doc files. When you add a completely new file, you need to update `_toctree.yml` & restart `preview` command (`ctrl-c` to stop it & call `doc-builder preview ...` again).
-
----
-
-## Adding a new element to the navigation bar
-
-Accepted files are Markdown (.md).
-
-Create a file with its extension and put it in the source directory. You can then link it to the toc-tree by putting
-the filename without the extension in the [`_toctree.yml`](https://github.com/huggingface/lerobot/blob/main/docs/source/_toctree.yml) file.
-
-## Renaming section headers and moving sections
-
-It helps to keep the old links working when renaming the section header and/or moving sections from one document to another. This is because the old links are likely to be used in Issues, Forums, and Social media and it'd make for a much more superior user experience if users reading those months later could still easily navigate to the originally intended information.
-
-Therefore, we simply keep a little map of moved sections at the end of the document where the original section was. The key is to preserve the original anchor.
-
-So if you renamed a section from: "Section A" to "Section B", then you can add at the end of the file:
-
-```
-Sections that were moved:
-
-[ <a href="#section-b">Section A</a><a id="section-a"></a> ]
-```
-and of course, if you moved it to another file, then:
-
-```
-Sections that were moved:
-
-[ <a href="../new-file#section-b">Section A</a><a id="section-a"></a> ]
-```
-
-Use the relative style to link to the new file so that the versioned docs continue to work.
-
-For an example of a rich moved sections set please see the very end of [the transformers Trainer doc](https://github.com/huggingface/transformers/blob/main/docs/source/en/main_classes/trainer.md).
-
-### Adding a new tutorial
-
-Adding a new tutorial or section is done in two steps:
-
-- Add a new file under `./source`. This file can either be ReStructuredText (.rst) or Markdown (.md).
-- Link that file in `./source/_toctree.yml` on the correct toc-tree.
-
-Make sure to put your new file under the proper section. If you have a doubt, feel free to ask in a Github Issue or PR.
-
-### Writing source documentation
-
-Values that should be put in `code` should either be surrounded by backticks: \`like so\`. Note that argument names
-and objects like True, None or any strings should usually be put in `code`.
-
-#### Writing a multi-line code block
-
-Multi-line code blocks can be useful for displaying examples. They are done between two lines of three backticks as usual in Markdown:
-
-
-````
-```
-# first line of code
-# second line
-# etc
-```
-````
-
-#### Adding an image
-
-Due to the rapidly growing repository, it is important to make sure that no files that would significantly weigh down the repository are added. This includes images, videos, and other non-text files. We prefer to leverage a hf.co hosted `dataset` like
-the ones hosted on [`hf-internal-testing`](https://huggingface.co/hf-internal-testing) in which to place these files and reference
-them by URL. We recommend putting them in the following dataset: [huggingface/documentation-images](https://huggingface.co/datasets/huggingface/documentation-images).
-If an external contribution, feel free to add the images to your PR and ask a Hugging Face member to migrate your images
-to this dataset.

docs/source/_toctree.yml

@@ -1,12 +0,0 @@
-- sections:
-  - local: index
-    title: LeRobot
-  - local: installation
-    title: Installation
-  title: Get started
-- sections:
-  - local: assemble_so101
-    title: Assemble SO-101
-  - local: getting_started_real_world_robot
-    title: Getting Started with Real-World Robots
-  title: "Tutorials"

docs/source/assemble_so101.mdx

@@ -1,348 +0,0 @@
-# Assemble SO-101
-
-In the steps below we explain how to assemble our flagship robot, the SO-101.
-
-## Source the parts
-
-Follow this [README](https://github.com/TheRobotStudio/SO-ARM100). It contains the bill of materials, with a link to source the parts, as well as the instructions to 3D print the parts,
-and advice if it's your first time printing or if you don't own a 3D printer.
-
-Before assembling, you will first need to configure your motors. To this end, we provide a nice script, so let's first install LeRobot. After configuration, we will also guide you through assembly.
-
-## Install LeRobot
-
-To install LeRobot follow our [Installation Guide](./installation)
-
-## Configure motors
-
-To configure the motors designate one bus servo adapter and 6 motors for your leader arm, and similarly the other bus servo adapter and 6 motors for the follower arm. It's convenient to label them and write on each motor if it's for the follower `F` or for the leader `L` and it's ID from 1 to 6.
-
-You now should plug the 5V or 12V power supply to the motor bus. 5V for the STS3215 7.4V motors and 12V for the STS3215 12V motors. Note that the leader arm always uses the 7.4V motors, so watch out that you plug in the right power supply if you have 12V and 7.4V motors, otherwise you might burn your motors! Now, connect the motor bus to your computer via USB. Note that the USB doesn't provide any power, and both the power supply and USB have to be plugged in.
-
-### Find the USB ports associated to each arm
-
-To find the port for each bus servo adapter, run this script:
-```bash
-python lerobot/scripts/find_motors_bus_port.py
-```
-##### Example outputs of script
-
-<hfoptions id="example">
-<hfoption id="Mac">
-
-Example output leader arm's port: `/dev/tty.usbmodem575E0031751`
-
-```bash
-Finding all available ports for the MotorBus.
-['/dev/tty.usbmodem575E0032081', '/dev/tty.usbmodem575E0031751']
-Remove the usb cable from your MotorsBus and press Enter when done.
-
-[...Disconnect leader arm and press Enter...]
-
-The port of this MotorsBus is /dev/tty.usbmodem575E0031751
-Reconnect the usb cable.
-```
-
-Example output follower arm port: `/dev/tty.usbmodem575E0032081`
-
-```
-Finding all available ports for the MotorBus.
-['/dev/tty.usbmodem575E0032081', '/dev/tty.usbmodem575E0031751']
-Remove the usb cable from your MotorsBus and press Enter when done.
-
-[...Disconnect follower arm and press Enter...]
-
-The port of this MotorsBus is /dev/tty.usbmodem575E0032081
-Reconnect the usb cable.
-```
-
-</hfoption>
-<hfoption id="Linux">
-
-On Linux, you might need to give access to the USB ports by running:
-```bash
-sudo chmod 666 /dev/ttyACM0
-sudo chmod 666 /dev/ttyACM1
-```
-
-Example output leader arm port: `/dev/ttyACM0`
-
-```bash
-Finding all available ports for the MotorBus.
-['/dev/ttyACM0', '/dev/ttyACM1']
-Remove the usb cable from your MotorsBus and press Enter when done.
-
-[...Disconnect leader arm and press Enter...]
-
-The port of this MotorsBus is /dev/ttyACM0
-Reconnect the usb cable.
-```
-
-Example output follower arm port: `/dev/ttyACM1`
-
-```
-Finding all available ports for the MotorBus.
-['/dev/ttyACM0', '/dev/ttyACM1']
-Remove the usb cable from your MotorsBus and press Enter when done.
-
-[...Disconnect follower arm and press Enter...]
-
-The port of this MotorsBus is /dev/ttyACM1
-Reconnect the usb cable.
-```
-</hfoption>
-</hfoptions>
-
-#### Update config file
-
-Now that you have your ports, update the **port** default values of [`SO101RobotConfig`](https://github.com/huggingface/lerobot/blob/main/lerobot/common/robot_devices/robots/configs.py).
-You will find a class called `so101` where you can update the `port` values with your actual motor ports:
-```diff
-@RobotConfig.register_subclass("so101")
-@dataclass
-class So101RobotConfig(ManipulatorRobotConfig):
-    calibration_dir: str = ".cache/calibration/so101"
-    # `max_relative_target` limits the magnitude of the relative positional target vector for safety purposes.
-    # Set this to a positive scalar to have the same value for all motors, or a list that is the same length as
-    # the number of motors in your follower arms.
-    max_relative_target: int | None = None
-
-    leader_arms: dict[str, MotorsBusConfig] = field(
-        default_factory=lambda: {
-            "main": FeetechMotorsBusConfig(
--               port="/dev/tty.usbmodem58760431091",
-+               port="{ADD YOUR LEADER PORT}",
-                motors={
-                    # name: (index, model)
-                    "shoulder_pan": [1, "sts3215"],
-                    "shoulder_lift": [2, "sts3215"],
-                    "elbow_flex": [3, "sts3215"],
-                    "wrist_flex": [4, "sts3215"],
-                    "wrist_roll": [5, "sts3215"],
-                    "gripper": [6, "sts3215"],
-                },
-            ),
-        }
-    )
-
-    follower_arms: dict[str, MotorsBusConfig] = field(
-        default_factory=lambda: {
-            "main": FeetechMotorsBusConfig(
--                port="/dev/tty.usbmodem585A0076891",
-+                port="{ADD YOUR FOLLOWER PORT}",
-                motors={
-                    # name: (index, model)
-                    "shoulder_pan": [1, "sts3215"],
-                    "shoulder_lift": [2, "sts3215"],
-                    "elbow_flex": [3, "sts3215"],
-                    "wrist_flex": [4, "sts3215"],
-                    "wrist_roll": [5, "sts3215"],
-                    "gripper": [6, "sts3215"],
-                },
-            ),
-        }
-    )
-```
-
-Here is a video of the process:
-<div class="video-container">
-  <video controls width="600">
-    <source src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/lerobot/lerobot-find-motorbus.mp4" type="video/mp4" />
-  </video>
- </div>
-
-## Step-by-Step Assembly Instructions
-
-The follower arm uses 6x STS3215 motors with 1/345 gearing. The leader however uses three differently geared motors to make sure it can both sustain its own weight and it can be moved without requiring much force. Which motor is needed for which joint is shown in table below.
-
-| Leader-Arm Axis | Motor | Gear Ratio |
-|-----------------|:-------:|:----------:|
-| Base / Shoulder Yaw | 1 | 1 / 191 |
-| Shoulder Pitch      | 2 | 1 / 345 |
-| Elbow               | 3 | 1 / 191 |
-| Wrist Roll          | 4 | 1 / 147 |
-| Wrist Pitch         | 5 | 1 / 147 |
-| Gripper             | 6 | 1 / 147 |
-
-### Set motor IDs
-
-Plug your motor in one of the two ports of the motor bus and run this script to set its ID to 1. Replace the text after --port to the corresponding control board port.
-```bash
-python lerobot/scripts/configure_motor.py \
-  --port /dev/tty.usbmodem58760432961 \
-  --brand feetech \
-  --model sts3215 \
-  --baudrate 1000000 \
-  --ID 1
-```
-
-Then unplug your motor and plug the second motor and set its ID to 2.
-```bash
-python lerobot/scripts/configure_motor.py \
-  --port /dev/tty.usbmodem58760432961 \
-  --brand feetech \
-  --model sts3215 \
-  --baudrate 1000000 \
-  --ID 2
-```
-
-Redo this process for all your motors until ID 6. Do the same for the 6 motors of the leader arm, but make sure to change the power supply if you use motors with different voltage and make sure you give the right ID to the right motor according to the table above.
-
-Here is a video of the process:
-<div class="video-container">
-  <video controls width="600">
-    <source src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/lerobot/lerobot-configure-motor.mp4" type="video/mp4" />
-  </video>
-</div>
-
-### Clean Parts
-Remove all support material from the 3D-printed parts, the easiest way to do this is using a small screwdriver to get underneath the support material.
-
-### Joint 1
-
-- Place the first motor into the base.
-- Fasten the motor with 4 M2x6mm screws (smallest screws). Two from the top and two from bottom.
-- Slide over the first motor holder and fasten it using two M2x6mm screws (one on each side).
-- Install both motor horns, securing the top horn with a M3x6mm screw.
-- Attach the shoulder part.
-- Tighten the shoulder part with 4 M3x6mm screws on top and 4 M3x6mm screws on the bottom
-- Add the shoulder motor holder.
-
-<div class="video-container">
-  <video controls width="600">
-    <source src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/lerobot/Joint1_v2.mp4" type="video/mp4" />
-  </video>
-</div>
-
-### Joint 2
-
-- Slide the second motor in from the top.
-- Fasten the second motor with 4 M2x6mm screws.
-- Attach both motor horns to motor 2, again use the M3x6mm horn screw.
-- Attach the upper arm with 4 M3x6mm screws on each side.
-
-<div class="video-container">
-  <video controls width="600">
-    <source src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/lerobot/Joint2_v2.mp4" type="video/mp4" />
-  </video>
-</div>
-
-### Joint 3
-
-- Insert motor 3 and fasten using 4 M2x6mm screws
-- Attach both motor horns to motor 3 and secure one again with a M3x6mm horn screw.
-- Connect the forearm to motor 3 using 4 M3x6mm screws on each side.
-
-<div class="video-container">
-  <video controls width="600">
-    <source src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/lerobot/Joint3_v2.mp4" type="video/mp4" />
-  </video>
-</div>
-
-### Joint 4
-
-- Slide over motor holder 4.
-- Slide in motor 4.
-- Fasten motor 4 with 4 M2x6mm screws and attach its motor horns, use a M3x6mm horn screw.
-
-<div class="video-container">
-  <video controls width="600">
-    <source src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/lerobot/Joint4_v2.mp4" type="video/mp4" />
-  </video>
-</div>
-
-### Joint 5
-
-- Insert motor 5 into the wrist holder and secure it with 2 M2x6mm front screws.
-- Install only one motor horn on the wrist motor and secure it with a M3x6mm horn screw.
-- Secure the wrist to motor 4 using 4 M3x6mm screws on both sides.
-
-<div class="video-container">
-  <video controls width="600">
-    <source src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/lerobot/Joint5_v2.mp4" type="video/mp4" />
-  </video>
-</div>
-
-### Gripper / Handle
-
-<hfoptions id="assembly">
-<hfoption id="Follower">
-
-- Attach the gripper to motor 5, attach it to the motor horn on the wrist using 4 M3x6mm screws.
-- Insert the gripper motor and secure it with 2 M2x6mm screws on each side.
-- Attach the motor horns and again use a M3x6mm horn screw.
-- Install the gripper claw and secure it with 4 M3x6mm screws on both sides.
-
-<div class="video-container">
-  <video controls width="600">
-    <source src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/lerobot/Gripper_v2.mp4" type="video/mp4" />
-  </video>
-</div>
-
-</hfoption>
-<hfoption id="Leader">
-
-- Mount the leader holder onto the wrist and secure it with 4 M3x6mm screws.
-- Attach the handle to motor 5 using 1 M2x6mm screw.
-- Insert the gripper motor, secure it with 2 M2x6mm screws on each side, attach a motor horn using a M3x6mm horn screw.
-- Attach the follower trigger with 4 M3x6mm screws.
-
-<div class="video-container">
-  <video controls width="600">
-    <source src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/lerobot/Leader_v2.mp4" type="video/mp4" />
-  </video>
-</div>
-
-</hfoption>
-</hfoptions>
-
-##### Wiring
-
-- Attach the motor controller on the back.
-- Then insert all wires, use the wire guides everywhere to make sure the wires don't unplug themselves and stay in place.
-
-<div class="video-container">
-  <video controls width="600">
-    <source src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/lerobot/Wiring_v2.mp4" type="video/mp4" />
-  </video>
-</div>
-
-## Calibrate
-
-Next, you'll need to calibrate your SO-101 robot to ensure that the leader and follower arms have the same position values when they are in the same physical position.
-The calibration process is very important because it allows a neural network trained on one SO-101 robot to work on another.
-
-#### Manual calibration of follower arm
-
-You will need to move the follower arm to these positions sequentially, note that the rotated position is on the right side of the robot and you have to open the gripper fully.
-
-| 1. Middle position | 2. Zero position                                                                                                                                       | 3. Rotated position                                                                                                                                             | 4. Rest position                                                                                                                                       |
-| ------------ |------------------------------------------------------------------------------------------------------------------------------------------------------ | --------------------------------------------------------------------------------------------------------------------------------------------------------------- | ------------------------------------------------------------------------------------------------------------------------------------------------------ |
-| <img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/lerobot/follower_middle.webp?raw=true" alt="SO-101 leader arm middle position" title="SO-101 leader arm middle position" style="width:100%;"> | <img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/lerobot/follower_zero.webp?raw=true" alt="SO-101 leader arm zero position" title="SO-101 leader arm zero position" style="width:100%;"> | <img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/lerobot/follower_rotated.webp?raw=true" alt="SO-101 leader arm rotated position" title="SO-101 leader arm rotated position" style="width:100%;"> | <img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/lerobot/follower_rest.webp?raw=true" alt="SO-101 leader arm rest position" title="SO-101 leader arm rest position" style="width:100%;"> |
-
-Make sure both arms are connected and run this script to launch manual calibration:
-```bash
-python lerobot/scripts/control_robot.py \
-  --robot.type=so101 \
-  --robot.cameras='{}' \
-  --control.type=calibrate \
-  --control.arms='["main_follower"]'
-```
-
-#### Manual calibration of leader arm
-You will also need to move the leader arm to these positions sequentially:
-
-| 1. Middle position | 2. Zero position                                                                                                                                       | 3. Rotated position                                                                                                                                             | 4. Rest position                                                                                                                                       |
-| ------------ |------------------------------------------------------------------------------------------------------------------------------------------------------ | --------------------------------------------------------------------------------------------------------------------------------------------------------------- | ------------------------------------------------------------------------------------------------------------------------------------------------------ |
-| <img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/lerobot/leader_middle.webp?raw=true" alt="SO-101 leader arm middle position" title="SO-101 leader arm middle position" style="width:100%;"> | <img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/lerobot/leader_zero.webp?raw=true" alt="SO-101 leader arm zero position" title="SO-101 leader arm zero position" style="width:100%;"> | <img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/lerobot/leader_rotated.webp?raw=true" alt="SO-101 leader arm rotated position" title="SO-101 leader arm rotated position" style="width:100%;"> | <img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/lerobot/leader_rest.webp?raw=true" alt="SO-101 leader arm rest position" title="SO-101 leader arm rest position" style="width:100%;"> |
-
-Run this script to launch manual calibration:
-```bash
-python lerobot/scripts/control_robot.py \
-  --robot.type=so101 \
-  --robot.cameras='{}' \
-  --control.type=calibrate \
-  --control.arms='["main_leader"]'
-```
-
-Congrats 🎉, your robot is all set to learn a task on its own. Start training it by following this tutorial: [Getting started with real-world robots](./getting_started_real_world_robot)

docs/source/getting_started_real_world_robot.mdx

@@ -1,370 +0,0 @@
-# Getting Started with Real-World Robots
-
-This tutorial will explain you how to train a neural network to autonomously control a real robot.
-
-**You'll learn:**
-1. How to record and visualize your dataset.
-2. How to train a policy using your data and prepare it for evaluation.
-3. How to evaluate your policy and visualize the results.
-
-By following these steps, you'll be able to replicate tasks like picking up a Lego block and placing it in a bin with a high success rate, as demonstrated in [this video](https://x.com/RemiCadene/status/1814680760592572934).
-
-This tutorial is specifically made for the affordable [SO-101](https://github.com/TheRobotStudio/SO-ARM100) robot, but it contains additional information to be easily adapted to various types of robots like [Aloha bimanual robot](https://aloha-2.github.io) by changing some configurations. The SO-101 consists of a leader arm and a follower arm, each with 6 motors. It can work with one or several cameras to record the scene, which serve as visual sensors for the robot.
-
-During the data collection phase, you will control the follower arm by moving the leader arm. This process is known as "teleoperation." This technique is used to collect robot trajectories. Afterward, you'll train a neural network to imitate these trajectories and deploy the network to enable your robot to operate autonomously.
-
-If you encounter any issues at any step of the tutorial, feel free to seek help on [Discord](https://discord.com/invite/s3KuuzsPFb) or don't hesitate to iterate with us on the tutorial by creating issues or pull requests.
-
-## Setup and Calibrate
-
-If you haven't yet setup and calibrate the SO-101 follow these steps:
-1. [Find ports and update config file](./assemble_so101#find-the-usb-ports-associated-to-each-arm)
-2. [Calibrate](./assemble_so101#calibrate)
-
-## Teleoperate
-
-Run this simple script to teleoperate your robot (it won't connect and display the cameras):
-```bash
-python lerobot/scripts/control_robot.py \
-  --robot.type=so101 \
-  --robot.cameras='{}' \
-  --control.type=teleoperate
-```
-
-The teleoperate command will automatically:
-1. Identify any missing calibrations and initiate the calibration procedure.
-2. Connect the robot and start teleoperation.
-
-## Setup Cameras
-
-To connect a camera you have three options:
-1. OpenCVCamera which allows us to use any camera: usb, realsense, laptop webcam
-2. iPhone camera with MacOS
-3. Phone camera on Linux
-
-### Use OpenCVCamera
-
-The [`OpenCVCamera`](../lerobot/common/robot_devices/cameras/opencv.py) class allows you to efficiently record frames from most cameras using the [`opencv2`](https://docs.opencv.org) library.  For more details on compatibility, see [Video I/O with OpenCV Overview](https://docs.opencv.org/4.x/d0/da7/videoio_overview.html).
-
-To instantiate an [`OpenCVCamera`](../lerobot/common/robot_devices/cameras/opencv.py), you need a camera index (e.g. `OpenCVCamera(camera_index=0)`). When you only have one camera like a webcam of a laptop, the camera index is usually `0` but it might differ, and the camera index might change if you reboot your computer or re-plug your camera. This behavior depends on your operating system.
-
-To find the camera indices, run the following utility script, which will save a few frames from each detected camera:
-```bash
-python lerobot/common/robot_devices/cameras/opencv.py \
-    --images-dir outputs/images_from_opencv_cameras
-```
-
-The output will look something like this if you have two cameras connected:
-```
-Mac or Windows detected. Finding available camera indices through scanning all indices from 0 to 60
-[...]
-Camera found at index 0
-Camera found at index 1
-[...]
-Connecting cameras
-OpenCVCamera(0, fps=30.0, width=1920.0, height=1080.0, color_mode=rgb)
-OpenCVCamera(1, fps=24.0, width=1920.0, height=1080.0, color_mode=rgb)
-Saving images to outputs/images_from_opencv_cameras
-Frame: 0000	Latency (ms): 39.52
-[...]
-Frame: 0046	Latency (ms): 40.07
-Images have been saved to outputs/images_from_opencv_cameras
-```
-
-Check the saved images in `outputs/images_from_opencv_cameras` to identify which camera index corresponds to which physical camera (e.g. `0` for `camera_00` or `1` for `camera_01`):
-```
-camera_00_frame_000000.png
-[...]
-camera_00_frame_000047.png
-camera_01_frame_000000.png
-[...]
-camera_01_frame_000047.png
-```
-
-Note: Some cameras may take a few seconds to warm up, and the first frame might be black or green.
-
-Now that you have the camera indexes, you should specify the camera's in the config.
-
-### Use your phone
-<hfoptions id="use phone">
-<hfoption id="Mac">
-
-To use your iPhone as a camera on macOS, enable the Continuity Camera feature:
-- Ensure your Mac is running macOS 13 or later, and your iPhone is on iOS 16 or later.
-- Sign in both devices with the same Apple ID.
-- Connect your devices with a USB cable or turn on Wi-Fi and Bluetooth for a wireless connection.
-
-For more details, visit [Apple support](https://support.apple.com/en-gb/guide/mac-help/mchl77879b8a/mac).
-
-Your iPhone should be detected automatically when running the camera setup script in the next section.
-
-</hfoption>
-<hfoption id="Linux">
-
-If you want to use your phone as a camera on Linux, follow these steps to set up a virtual camera
-
-1. *Install `v4l2loopback-dkms` and `v4l-utils`*. Those packages are required to create virtual camera devices (`v4l2loopback`) and verify their settings with the `v4l2-ctl` utility from `v4l-utils`. Install them using:
-```python
-sudo apt install v4l2loopback-dkms v4l-utils
-```
-2. *Install [DroidCam](https://droidcam.app) on your phone*. This app is available for both iOS and Android.
-3. *Install [OBS Studio](https://obsproject.com)*. This software will help you manage the camera feed. Install it using [Flatpak](https://flatpak.org):
-```python
-flatpak install flathub com.obsproject.Studio
-```
-4. *Install the DroidCam OBS plugin*. This plugin integrates DroidCam with OBS Studio. Install it with:
-```python
-flatpak install flathub com.obsproject.Studio.Plugin.DroidCam
-```
-5. *Start OBS Studio*. Launch with:
-```python
-flatpak run com.obsproject.Studio
-```
-6. *Add your phone as a source*. Follow the instructions [here](https://droidcam.app/obs/usage). Be sure to set the resolution to `640x480`.
-7. *Adjust resolution settings*. In OBS Studio, go to `File > Settings > Video`. Change the `Base(Canvas) Resolution` and the `Output(Scaled) Resolution` to `640x480` by manually typing it in.
-8. *Start virtual camera*. In OBS Studio, follow the instructions [here](https://obsproject.com/kb/virtual-camera-guide).
-9. *Verify the virtual camera setup*. Use `v4l2-ctl` to list the devices:
-```python
-v4l2-ctl --list-devices
-```
-You should see an entry like:
-```
-VirtualCam (platform:v4l2loopback-000):
-/dev/video1
-```
-10. *Check the camera resolution*. Use `v4l2-ctl` to ensure that the virtual camera output resolution is `640x480`. Change `/dev/video1` to the port of your virtual camera from the output of `v4l2-ctl --list-devices`.
-```python
-v4l2-ctl -d /dev/video1 --get-fmt-video
-```
-You should see an entry like:
-```
->>> Format Video Capture:
->>>	Width/Height      : 640/480
->>>	Pixel Format      : 'YUYV' (YUYV 4:2:2)
-```
-
-Troubleshooting: If the resolution is not correct you will have to delete the Virtual Camera port and try again as it cannot be changed.
-
-If everything is set up correctly, you can proceed with the rest of the tutorial.
-
-</hfoption>
-</hfoptions>
-
-## Teleoperate with cameras
-
-We can now teleoperate again while at the same time visualizing the cameras and joint positions with `rerun`.
-
-```bash
-python lerobot/scripts/control_robot.py \
-  --robot.type=so101 \
-  --control.type=teleoperate
-  --control.display_data=true
-```
-
-## Record a dataset
-
-Once you're familiar with teleoperation, you can record your first dataset with SO-101.
-
-We use the Hugging Face hub features for uploading your dataset. If you haven't previously used the Hub, make sure you can login via the cli using a write-access token, this token can be generated from the [Hugging Face settings](https://huggingface.co/settings/tokens).
-
-Add your token to the cli by running this command:
-```bash
-huggingface-cli login --token ${HUGGINGFACE_TOKEN} --add-to-git-credential
-```
-
-Then store your Hugging Face repository name in a variable:
-```bash
-HF_USER=$(huggingface-cli whoami | head -n 1)
-echo $HF_USER
-```
-
-Now you can record a dataset, to record 2 episodes and upload your dataset to the hub execute this command:
-```bash
-python lerobot/scripts/control_robot.py \
-  --robot.type=so101 \
-  --control.type=record \
-  --control.fps=30 \
-  --control.single_task="Grasp a lego block and put it in the bin." \
-  --control.repo_id=${HF_USER}/so101_test \
-  --control.tags='["so101","tutorial"]' \
-  --control.warmup_time_s=5 \
-  --control.episode_time_s=30 \
-  --control.reset_time_s=30 \
-  --control.num_episodes=2 \
-  --control.push_to_hub=true
-```
-
-You will see a lot of lines appearing like this one:
-```
-INFO 2024-08-10 15:02:58 ol_robot.py:219 dt:33.34 (30.0hz) dtRlead: 5.06 (197.5hz) dtWfoll: 0.25 (3963.7hz) dtRfoll: 6.22 (160.7hz) dtRlaptop: 32.57 (30.7hz) dtRphone: 33.84 (29.5hz)
-```
-
-| Field | Meaning |
-|:---|:---|
-| `2024-08-10 15:02:58` | Timestamp when `print` was called. |
-| `ol_robot.py:219` | Source file and line number of the `print` call (`lerobot/scripts/control_robot.py` at line `219`). |
-| `dt: 33.34 (30.0 Hz)` | Delta time (ms) between teleop steps (target: 30.0 Hz, `--fps 30`). Yellow if step is too slow. |
-| `dtRlead: 5.06 (197.5 Hz)` | Delta time (ms) for reading present position from the **leader arm**. |
-| `dtWfoll: 0.25 (3963.7 Hz)` | Delta time (ms) for writing goal position to the **follower arm** (asynchronous). |
-| `dtRfoll: 6.22 (160.7 Hz)` | Delta time (ms) for reading present position from the **follower arm**. |
-| `dtRlaptop: 32.57 (30.7 Hz)` | Delta time (ms) for capturing an image from the **laptop camera** (async thread). |
-| `dtRphone: 33.84 (29.5 Hz)` | Delta time (ms) for capturing an image from the **phone camera** (async thread). |
-
-
-#### Dataset upload
-Locally your dataset is stored in this folder: `~/.cache/huggingface/lerobot/{repo-id}` (e.g. `data/cadene/so101_test`). At the end of data recording, your dataset will be uploaded on your Hugging Face page (e.g. https://huggingface.co/datasets/cadene/so101_test) that you can obtain by running:
-```bash
-echo https://huggingface.co/datasets/${HF_USER}/so101_test
-```
-Your dataset will be automatically tagged with `LeRobot` for the community to find it easily, and you can also add custom tags (in this case `tutorial` for example).
-
-You can look for other LeRobot datasets on the hub by searching for `LeRobot` [tags](https://huggingface.co/datasets?other=LeRobot).
-
-#### Record function
-
-The `record` function provides a suite of tools for capturing and managing data during robot operation:
-
-##### 1. Frame Capture and Video Encoding
-- Frames from cameras are saved to disk during recording.
-- At the end of each episode, frames are encoded into video files.
-
-##### 2. Data Storage
-- Data is stored using the `LeRobotDataset` format.
-- By default, the dataset is pushed to your Hugging Face page.
-  - To disable uploading, use `--control.push_to_hub=false`.
-
-##### 3. Checkpointing and Resuming
-- Checkpoints are automatically created during recording.
-- If an issue occurs, you can resume by re-running the same command with `--control.resume=true`.
-- To start recording from scratch, **manually delete** the dataset directory.
-
-##### 4. Recording Parameters
-Set the flow of data recording using command-line arguments:
-- `--control.warmup_time_s=10`
-  Number of seconds before starting data collection (default: **10 seconds**).
-  Allows devices to warm up and synchronize.
-- `--control.episode_time_s=60`
-  Duration of each data recording episode (default: **60 seconds**).
-- `--control.reset_time_s=60`
-  Duration for resetting the environment after each episode (default: **60 seconds**).
-- `--control.num_episodes=50`
-  Total number of episodes to record (default: **50**).
-
-##### 5. Keyboard Controls During Recording
-Control the data recording flow using keyboard shortcuts:
-- Press **Right Arrow (`→`)**: Early stop the current episode or reset time and move to the next.
-- Press **Left Arrow (`←`)**: Cancel the current episode and re-record it.
-- Press **Escape (`ESC`)**: Immediately stop the session, encode videos, and upload the dataset.
-
-#### Tips for gathering data
-
-Once you're comfortable with data recording, you can create a larger dataset for training. A good starting task is grasping an object at different locations and placing it in a bin. We suggest recording at least 50 episodes, with 10 episodes per location. Keep the cameras fixed and maintain consistent grasping behavior throughout the recordings. Also make sure the object you are manipulating is visible on the camera's. A good rule of thumb is you should be able to do the task yourself by only looking at the camera images.
-
-In the following sections, you’ll train your neural network. After achieving reliable grasping performance, you can start introducing more variations during data collection, such as additional grasp locations, different grasping techniques, and altering camera positions.
-
-Avoid adding too much variation too quickly, as it may hinder your results.
-
-
-#### Troubleshooting:
-- On Linux, if the left and right arrow keys and escape key don't have any effect during data recording, make sure you've set the `$DISPLAY` environment variable. See [pynput limitations](https://pynput.readthedocs.io/en/latest/limitations.html#linux).
-
-## Visualize a dataset
-
-If you uploaded your dataset to the hub with `--control.push_to_hub=true`, you can [visualize your dataset online](https://huggingface.co/spaces/lerobot/visualize_dataset) by copy pasting your repo id given by:
-```bash
-echo ${HF_USER}/so101_test
-```
-
-If you didn't upload with `--control.push_to_hub=false`, you can visualize it locally with (via a window in the browser `http://127.0.0.1:9090` with the visualization tool):
-```bash
-python lerobot/scripts/visualize_dataset_html.py \
-  --repo-id ${HF_USER}/so101_test \
-  --local-files-only 1
-```
-
-This will launch a local web server that looks like this:
-<div style="text-align:center;">
-  <img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/lerobot/visualize_dataset_html.webp?raw=true" alt="Koch v1.1 leader and follower arms" title="Koch v1.1 leader and follower arms" width="100%"></img>
-</div>
-
-## Replay an episode
-
-A useful feature is the `replay` function, which allows to replay on your robot any episode that you've recorded or episodes from any dataset out there. This function helps you test the repeatability of your robot's actions and assess transferability across robots of the same model.
-
-You can replay the first episode on your robot with:
-```bash
-python lerobot/scripts/control_robot.py \
-  --robot.type=so101 \
-  --control.type=replay \
-  --control.fps=30 \
-  --control.repo_id=${HF_USER}/so101_test \
-  --control.episode=0
-```
-
-Your robot should replicate movements similar to those you recorded. For example, check out [this video](https://x.com/RemiCadene/status/1793654950905680090) where we use `replay` on a Aloha robot from [Trossen Robotics](https://www.trossenrobotics.com).
-
-## Train a policy
-
-To train a policy to control your robot, use the [`python lerobot/scripts/train.py`](../lerobot/scripts/train.py) script. A few arguments are required. Here is an example command:
-```bash
-python lerobot/scripts/train.py \
-  --dataset.repo_id=${HF_USER}/so101_test \
-  --policy.type=act \
-  --output_dir=outputs/train/act_so101_test \
-  --job_name=act_so101_test \
-  --policy.device=cuda \
-  --wandb.enable=true
-```
-
-Let's explain the command:
-1. We provided the dataset as argument with `--dataset.repo_id=${HF_USER}/so101_test`.
-2. We provided the policy with `policy.type=act`. This loads configurations from [`configuration_act.py`](../lerobot/common/policies/act/configuration_act.py). Importantly, this policy will automatically adapt to the number of motor states, motor actions and cameras of your robot (e.g. `laptop` and `phone`) which have been saved in your dataset.
-4. We provided `policy.device=cuda` since we are training on a Nvidia GPU, but you could use `policy.device=mps` to train on Apple silicon.
-5. We provided `wandb.enable=true` to use [Weights and Biases](https://docs.wandb.ai/quickstart) for visualizing training plots. This is optional but if you use it, make sure you are logged in by running `wandb login`.
-
-Training should take several hours. You will find checkpoints in `outputs/train/act_so101_test/checkpoints`.
-
-To resume training from a checkpoint, below is an example command to resume from `last` checkpoint of the `act_so101_test` policy:
-```bash
-python lerobot/scripts/train.py \
-  --config_path=outputs/train/act_so101_test/checkpoints/last/pretrained_model/train_config.json \
-  --resume=true
-```
-
-#### Upload policy checkpoints
-
-Once training is done, upload the latest checkpoint with:
-```bash
-huggingface-cli upload ${HF_USER}/act_so101_test \
-  outputs/train/act_so101_test/checkpoints/last/pretrained_model
-```
-
-You can also upload intermediate checkpoints with:
-```bash
-CKPT=010000
-huggingface-cli upload ${HF_USER}/act_so101_test${CKPT} \
-  outputs/train/act_so101_test/checkpoints/${CKPT}/pretrained_model
-```
-
-## Evaluate your policy
-
-You can use the `record` function from [`lerobot/scripts/control_robot.py`](../lerobot/scripts/control_robot.py) but with a policy checkpoint as input. For instance, run this command to record 10 evaluation episodes:
-```bash
-python lerobot/scripts/control_robot.py \
-  --robot.type=so101 \
-  --control.type=record \
-  --control.fps=30 \
-  --control.single_task="Grasp a lego block and put it in the bin." \
-  --control.repo_id=${HF_USER}/eval_act_so101_test \
-  --control.tags='["tutorial"]' \
-  --control.warmup_time_s=5 \
-  --control.episode_time_s=30 \
-  --control.reset_time_s=30 \
-  --control.num_episodes=10 \
-  --control.push_to_hub=true \
-  --control.policy.path=outputs/train/act_so101_test/checkpoints/last/pretrained_model
-```
-
-As you can see, it's almost the same command as previously used to record your training dataset. Two things changed:
-1. There is an additional `--control.policy.path` argument which indicates the path to your policy checkpoint with  (e.g. `outputs/train/eval_act_so101_test/checkpoints/last/pretrained_model`). You can also use the model repository if you uploaded a model checkpoint to the hub (e.g. `${HF_USER}/act_so101_test`).
-2. The name of dataset begins by `eval` to reflect that you are running inference (e.g. `${HF_USER}/eval_act_so101_test`).

docs/source/index.mdx

@@ -1,19 +0,0 @@
-<div class="flex justify-center">
-  <a target="_blank" href="https://huggingface.co/lerobot">
-      <img alt="HuggingFace Expert Acceleration Program" src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/lerobot/lerobot-logo-thumbnail.png" style="width: 100%"></img>
-  </a>
-</div>
-
-# LeRobot
-
-**State-of-the-art machine learning for real-world robotics**
-
-🤗 LeRobot aims to provide models, datasets, and tools for real-world robotics in PyTorch. The goal is to lower the barrier for entry to robotics so that everyone can contribute and benefit from sharing datasets and pretrained models.
-
-🤗 LeRobot contains state-of-the-art approaches that have been shown to transfer to the real-world with a focus on imitation learning and reinforcement learning.
-
-🤗 LeRobot already provides a set of pretrained models, datasets with human collected demonstrations, and simulated environments so that everyone can get started.
-
-🤗 LeRobot hosts pretrained models and datasets on the LeRobot HuggingFace page.
-
-Join the LeRobot community on [Discord](https://discord.gg/s3KuuzsPFb)

docs/source/installation.mdx

@@ -1,84 +0,0 @@
-# Installation
-
-## Install LeRobot
-
-Download our source code:
-```bash
-git clone https://github.com/huggingface/lerobot.git
-cd lerobot
-```
-
-Create a virtual environment with Python 3.10, using [`Miniconda`](https://docs.anaconda.com/miniconda/install/#quick-command-line-install)
-```bash
-conda create -y -n lerobot python=3.10
-```
-
-Now restart the shell by running:
-<hfoptions id="shell_restart">
-<hfoption id="Windows">
-
-```bash
-source ~/.bashrc
-```
-</hfoption>
-<hfoption id="Mac">
-
-```bash
-source ~/.bash_profile
-```
-</hfoption>
-<hfoption id="zshell">
-
-```bash
-source ~/.zshrc
-```
-</hfoption>
-</hfoptions>
-
-Then activate your conda environment, you have to do this each time you open a shell to use lerobot:
-```bash
-conda activate lerobot
-```
-
-When using `miniconda`, install `ffmpeg` in your environment:
-```bash
-conda install ffmpeg -c conda-forge
-```
-
-> [!TIP]
-> This usually installs `ffmpeg 7.X` for your platform compiled with the `libsvtav1` encoder. If `libsvtav1` is not supported (check supported encoders with `ffmpeg -encoders`), you can:
->  - _[On any platform]_ Explicitly install `ffmpeg 7.X` using:
->  ```bash
->  conda install ffmpeg=7.1.1 -c conda-forge
->  ```
->  - _[On Linux only]_ 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`.
-
-Install 🤗 LeRobot:
-```bash
-cd lerobot && pip install -e ".[feetech]"
-```
-
-## Troubleshooting
-If you encounter build errors, you may need to install additional dependencies: `cmake`, `build-essential`, and `ffmpeg libs`.
-To install these for linux run:
-```bash
-sudo apt-get install cmake build-essential python-dev pkg-config libavformat-dev libavcodec-dev libavdevice-dev libavutil-dev libswscale-dev libswresample-dev libavfilter-dev pkg-config
-```
-For other systems, see: [Compiling PyAV](https://pyav.org/docs/develop/overview/installation.html#bring-your-own-ffmpeg)
-
-## Sim
-For simulations, 🤗 LeRobot comes with gymnasium environments that can be installed as extras:
-- [aloha](https://github.com/huggingface/gym-aloha)
-- [xarm](https://github.com/huggingface/gym-xarm)
-- [pusht](https://github.com/huggingface/gym-pusht)
-
-For instance, to install 🤗 LeRobot with aloha and pusht, use:
-```bash
-pip install -e ".[aloha, pusht]"
-```
-
-## W&B
-To use [Weights and Biases](https://docs.wandb.ai/quickstart) for experiment tracking, log in with
-```bash
-wandb login
-```

examples/10_use_so100.md

@@ -57,15 +57,9 @@ conda activate lerobot
 git clone https://github.com/huggingface/lerobot.git ~/lerobot
 ```
 
-#### 5. Install ffmpeg in your environment:
-When using `miniconda`, install `ffmpeg` in your environment:
+#### 5. Install LeRobot with dependencies for the feetech motors:
 ```bash
-conda install ffmpeg -c conda-forge
-```
-
-#### 6. Install LeRobot with dependencies for the feetech motors:
-```bash
-cd ~/lerobot && pip install -e ".[feetech]"
+cd ~/lerobot && pip install --no-binary=av -e ".[feetech]"
 ```
 
 Great :hugs:! You are now done installing LeRobot and we can begin assembling the SO100 arms :robot:.
@@ -128,7 +122,7 @@ sudo chmod 666 /dev/ttyACM1
 #### d. Update config file
 
 IMPORTANTLY: Now that you have your ports, update the **port** default values of [`SO100RobotConfig`](../lerobot/common/robot_devices/robots/configs.py). You will find something like:
-```diff
+```python
 @RobotConfig.register_subclass("so100")
 @dataclass
 class So100RobotConfig(ManipulatorRobotConfig):
@@ -141,8 +135,7 @@ class So100RobotConfig(ManipulatorRobotConfig):
     leader_arms: dict[str, MotorsBusConfig] = field(
         default_factory=lambda: {
             "main": FeetechMotorsBusConfig(
--                port="/dev/tty.usbmodem58760431091",
-+                port="{ADD YOUR LEADER PORT}",
+                port="/dev/tty.usbmodem58760431091",  <-- UPDATE HERE
                 motors={
                     # name: (index, model)
                     "shoulder_pan": [1, "sts3215"],
@@ -159,8 +152,7 @@ class So100RobotConfig(ManipulatorRobotConfig):
     follower_arms: dict[str, MotorsBusConfig] = field(
         default_factory=lambda: {
             "main": FeetechMotorsBusConfig(
--                port="/dev/tty.usbmodem585A0076891",
-+                port="{ADD YOUR FOLLOWER PORT}",
+                port="/dev/tty.usbmodem585A0076891",  <-- UPDATE HERE
                 motors={
                     # name: (index, model)
                     "shoulder_pan": [1, "sts3215"],
@@ -447,16 +439,18 @@ For the leader configuration, perform **Steps 1–23**. Make sure that you remov
 
 ## E. Calibrate
 
-Next, you'll need to calibrate your SO-100 robot to ensure that the leader and follower arms have the same position values when they are in the same physical position.
-The calibration process is very important because it allows a neural network trained on one SO-100 robot to work on another.
+Next, you'll need to calibrate your SO-100 robot to ensure that the leader and follower arms have the same position values when they are in the same physical position. This calibration is essential because it allows a neural network trained on one SO-100 robot to work on another.
 
-#### Manual calibration of follower arm
+#### a. Manual calibration of follower arm
 
-You will need to move the follower arm to these positions sequentially, note that the rotated position is on the right side of the robot and you have to open the gripper fully.
+> [!IMPORTANT]
+> Contrarily to step 6 of the [assembly video](https://youtu.be/FioA2oeFZ5I?t=724) which illustrates the auto calibration, we will actually do manual calibration of follower for now.
 
-| 1. Middle position | 2. Zero position                                                                                                                                       | 3. Rotated position                                                                                                                                             | 4. Rest position                                                                                                                                       |
-| ------------ |------------------------------------------------------------------------------------------------------------------------------------------------------ | --------------------------------------------------------------------------------------------------------------------------------------------------------------- | ------------------------------------------------------------------------------------------------------------------------------------------------------ |
-| <img src="../media/so101/follower_middle.webp?raw=true" alt="SO-101 leader arm middle position" title="SO-101 leader arm middle position" style="width:100%;"> | <img src="../media/so101/follower_zero.webp?raw=true" alt="SO-101 leader arm zero position" title="SO-101 leader arm zero position" style="width:100%;"> | <img src="../media/so101/follower_rotated.webp?raw=true" alt="SO-101 leader arm rotated position" title="SO-101 leader arm rotated position" style="width:100%;"> | <img src="../media/so101/follower_rest.webp?raw=true" alt="SO-101 leader arm rest position" title="SO-101 leader arm rest position" style="width:100%;"> |
+You will need to move the follower arm to these positions sequentially:
+
+| 1. Zero position                                                                                                                                             | 2. Rotated position                                                                                                                                                   | 3. Rest position                                                                                                                                             |
+| ------------------------------------------------------------------------------------------------------------------------------------------------------------ | --------------------------------------------------------------------------------------------------------------------------------------------------------------------- | ------------------------------------------------------------------------------------------------------------------------------------------------------------ |
+| <img src="../media/so100/follower_zero.webp?raw=true" alt="SO-100 follower arm zero position" title="SO-100 follower arm zero position" style="width:100%;"> | <img src="../media/so100/follower_rotated.webp?raw=true" alt="SO-100 follower arm rotated position" title="SO-100 follower arm rotated position" style="width:100%;"> | <img src="../media/so100/follower_rest.webp?raw=true" alt="SO-100 follower arm rest position" title="SO-100 follower arm rest position" style="width:100%;"> |
 
 Make sure both arms are connected and run this script to launch manual calibration:
 ```bash
@@ -467,12 +461,12 @@ python lerobot/scripts/control_robot.py \
   --control.arms='["main_follower"]'
 ```
 
-#### Manual calibration of leader arm
-You will also need to move the leader arm to these positions sequentially:
+#### b. Manual calibration of leader arm
+Follow step 6 of the [assembly video](https://youtu.be/FioA2oeFZ5I?t=724) which illustrates the manual calibration. You will need to move the leader arm to these positions sequentially:
 
-| 1. Middle position | 2. Zero position                                                                                                                                       | 3. Rotated position                                                                                                                                             | 4. Rest position                                                                                                                                       |
-| ------------ |------------------------------------------------------------------------------------------------------------------------------------------------------ | --------------------------------------------------------------------------------------------------------------------------------------------------------------- | ------------------------------------------------------------------------------------------------------------------------------------------------------ |
-| <img src="../media/so101/leader_middle.webp?raw=true" alt="SO-100 leader arm middle position" title="SO-100 leader arm middle position" style="width:100%;"> | <img src="../media/so101/leader_zero.webp?raw=true" alt="SO-100 leader arm zero position" title="SO-100 leader arm zero position" style="width:100%;"> | <img src="../media/so101/leader_rotated.webp?raw=true" alt="SO-100 leader arm rotated position" title="SO-100 leader arm rotated position" style="width:100%;"> | <img src="../media/so101/leader_rest.webp?raw=true" alt="SO-100 leader arm rest position" title="SO-100 leader arm rest position" style="width:100%;"> |
+| 1. Zero position                                                                                                                                       | 2. Rotated position                                                                                                                                             | 3. Rest position                                                                                                                                       |
+| ------------------------------------------------------------------------------------------------------------------------------------------------------ | --------------------------------------------------------------------------------------------------------------------------------------------------------------- | ------------------------------------------------------------------------------------------------------------------------------------------------------ |
+| <img src="../media/so100/leader_zero.webp?raw=true" alt="SO-100 leader arm zero position" title="SO-100 leader arm zero position" style="width:100%;"> | <img src="../media/so100/leader_rotated.webp?raw=true" alt="SO-100 leader arm rotated position" title="SO-100 leader arm rotated position" style="width:100%;"> | <img src="../media/so100/leader_rest.webp?raw=true" alt="SO-100 leader arm rest position" title="SO-100 leader arm rest position" style="width:100%;"> |
 
 Run this script to launch manual calibration:
 ```bash
@@ -497,9 +491,6 @@ python lerobot/scripts/control_robot.py \
 
 #### a. Teleop with displaying cameras
 Follow [this guide to setup your cameras](https://github.com/huggingface/lerobot/blob/main/examples/7_get_started_with_real_robot.md#c-add-your-cameras-with-opencvcamera). Then you will be able to display the cameras on your computer while you are teleoperating by running the following code. This is useful to prepare your setup before recording your first dataset.
-
-> **NOTE:** To visualize the data, enable `--control.display_data=true`. This streams the data using `rerun`.
-
 ```bash
 python lerobot/scripts/control_robot.py \
   --robot.type=so100 \
@@ -580,7 +571,7 @@ python lerobot/scripts/train.py \
 
 Let's explain it:
 1. We provided the dataset as argument with `--dataset.repo_id=${HF_USER}/so100_test`.
-2. We provided the policy with `policy.type=act`. This loads configurations from [`configuration_act.py`](../lerobot/common/policies/act/configuration_act.py). Importantly, this policy will automatically adapt to the number of motor states, motor actions and cameras of your robot (e.g. `laptop` and `phone`) which have been saved in your dataset.
+2. We provided the policy with `policy.type=act`. This loads configurations from [`configuration_act.py`](../lerobot/common/policies/act/configuration_act.py). Importantly, this policy will automatically adapt to the number of motor sates, motor actions and cameras of your robot (e.g. `laptop` and `phone`) which have been saved in your dataset.
 4. We provided `policy.device=cuda` since we are training on a Nvidia GPU, but you could use `policy.device=mps` to train on Apple silicon.
 5. We provided `wandb.enable=true` to use [Weights and Biases](https://docs.wandb.ai/quickstart) for visualizing training plots. This is optional but if you use it, make sure you are logged in by running `wandb login`.
 

examples/11_use_lekiwi.md

@@ -67,15 +67,9 @@ conda activate lerobot
 git clone https://github.com/huggingface/lerobot.git ~/lerobot
 ```
 
-#### 5. Install ffmpeg in your environment:
-When using `miniconda`, install `ffmpeg` in your environment:
+#### 5. Install LeRobot with dependencies for the feetech motors:
 ```bash
-conda install ffmpeg -c conda-forge
-```
-
-#### 6. Install LeRobot with dependencies for the feetech motors:
-```bash
-cd ~/lerobot && pip install -e ".[feetech]"
+cd ~/lerobot && pip install --no-binary=av -e ".[feetech]"
 ```
 
 ## C. Install LeRobot on laptop
@@ -114,15 +108,9 @@ conda activate lerobot
 git clone https://github.com/huggingface/lerobot.git ~/lerobot
 ```
 
-#### 5. Install ffmpeg in your environment:
-When using `miniconda`, install `ffmpeg` in your environment:
+#### 5. Install LeRobot with dependencies for the feetech motors:
 ```bash
-conda install ffmpeg -c conda-forge
-```
-
-#### 6. Install LeRobot with dependencies for the feetech motors:
-```bash
-cd ~/lerobot && pip install -e ".[feetech]"
+cd ~/lerobot && pip install --no-binary=av -e ".[feetech]"
 ```
 
 Great :hugs:! You are now done installing LeRobot and we can begin assembling the SO100 arms and Mobile base :robot:.
@@ -134,7 +122,7 @@ First we will assemble the two SO100 arms. One to attach to the mobile base and
 
 ## SO100 Arms
 ### Configure motors
-The instructions for configuring the motors can be found [Here](https://github.com/huggingface/lerobot/blob/main/examples/10_use_so100.md#c-configure-the-motors) in step C of the SO100 tutorial. Besides the ID's for the arm motors we also need to set the motor ID's for the mobile base. These need to be in a specific order to work. Below an image of the motor ID's and motor mounting positions for the mobile base. Note that we only use one Motor Control board on LeKiwi. This means the motor ID's for the wheels are 7, 8 and 9.
+The instructions for configuring the motors can be found [Here](https://github.com/huggingface/lerobot/blob/main/examples/10_use_so100.md#c-configure-the-motors) in step C of the SO100 tutorial. Besides the ID's for the arm motors we also need to set the motor ID's for the mobile base. These needs to be in a specific order to work. Below an image of the motor ID's and motor mounting positions for the mobile base. Note that we only use one Motor Control board on LeKiwi. This means the motor ID's for the wheels are 7, 8 and 9.
 
 <img src="../media/lekiwi/motor_ids.webp?raw=true" alt="Motor ID's for mobile robot" title="Motor ID's for mobile robot" width="60%">
 
@@ -405,10 +393,6 @@ python lerobot/scripts/control_robot.py \
 ```
 
 # F. Teleoperate
-
-> [!TIP]
-> If you're using a Mac, you might need to give Terminal permission to access your keyboard. Go to System Preferences > Security & Privacy > Input Monitoring and check the box for Terminal.
-
 To teleoperate SSH into your Raspberry Pi, and run `conda activate lerobot` and this script:
 ```bash
 python lerobot/scripts/control_robot.py \
@@ -424,8 +408,6 @@ python lerobot/scripts/control_robot.py \
   --control.fps=30
 ```
 
-> **NOTE:** To visualize the data, enable `--control.display_data=true`. This streams the data using `rerun`. For the `--control.type=remote_robot` you will also need to set `--control.viewer_ip` and `--control.viewer_port`
-
 You should see on your laptop something like this: ```[INFO] Connected to remote robot at tcp://172.17.133.91:5555 and video stream at tcp://172.17.133.91:5556.``` Now you can move the leader arm and use the keyboard (w,a,s,d) to drive forward, left, backwards, right. And use (z,x) to turn left or turn right. You can use (r,f) to increase and decrease the speed of the mobile robot. There are three speed modes, see the table below:
 | Speed Mode | Linear Speed (m/s) | Rotation Speed (deg/s) |
 | ---------- | ------------------ | ---------------------- |
@@ -567,7 +549,7 @@ python lerobot/scripts/train.py \
 
 Let's explain it:
 1. We provided the dataset as argument with `--dataset.repo_id=${HF_USER}/lekiwi_test`.
-2. We provided the policy with `policy.type=act`. This loads configurations from [`configuration_act.py`](../lerobot/common/policies/act/configuration_act.py). Importantly, this policy will automatically adapt to the number of motor states, motor actions and cameras of your robot (e.g. `laptop` and `phone`) which have been saved in your dataset.
+2. We provided the policy with `policy.type=act`. This loads configurations from [`configuration_act.py`](../lerobot/common/policies/act/configuration_act.py). Importantly, this policy will automatically adapt to the number of motor sates, motor actions and cameras of your robot (e.g. `laptop` and `phone`) which have been saved in your dataset.
 4. We provided `policy.device=cuda` since we are training on a Nvidia GPU, but you could use `policy.device=mps` to train on Apple silicon.
 5. We provided `wandb.enable=true` to use [Weights and Biases](https://docs.wandb.ai/quickstart) for visualizing training plots. This is optional but if you use it, make sure you are logged in by running `wandb login`.
 

examples/11_use_moss.md

@@ -31,20 +31,14 @@ conda create -y -n lerobot python=3.10 && conda activate lerobot
 git clone https://github.com/huggingface/lerobot.git ~/lerobot
 ```
 
-5. Install ffmpeg in your environment:
-When using `miniconda`, install `ffmpeg` in your environment:
+5. Install LeRobot with dependencies for the feetech motors:
 ```bash
-conda install ffmpeg -c conda-forge
-```
-
-6. Install LeRobot with dependencies for the feetech motors:
-```bash
-cd ~/lerobot && pip install -e ".[feetech]"
+cd ~/lerobot && pip install --no-binary=av -e ".[feetech]"
 ```
 
 ## Configure the motors
 
-Follow step 1 of the [assembly video](https://www.youtube.com/watch?v=DA91NJOtMic) which illustrates the use of our scripts below.
+Follow steps 1 of the [assembly video](https://www.youtube.com/watch?v=DA91NJOtMic) which illustrates the use of our scripts below.
 
 **Find USB ports associated to your arms**
 To find the correct ports for each arm, run the utility script twice:
@@ -141,7 +135,7 @@ python lerobot/scripts/configure_motor.py \
   --ID 1
 ```
 
-Note: These motors are currently limited. They can take values between 0 and 4096 only, which corresponds to a full turn. They can't turn more than that. 2048 is at the middle of this range, so we can take -2048 steps (180 degrees anticlockwise) and reach the maximum range, or take +2048 steps (180 degrees clockwise) and reach the maximum range. The configuration step also sets the homing offset to 0, so that if you misassembled the arm, you can always update the homing offset to account for a shift up to ± 2048 steps (± 180 degrees).
+Note: These motors are currently limitated. They can take values between 0 and 4096 only, which corresponds to a full turn. They can't turn more than that. 2048 is at the middle of this range, so we can take -2048 steps (180 degrees anticlockwise) and reach the maximum range, or take +2048 steps (180 degrees clockwise) and reach the maximum range. The configuration step also sets the homing offset to 0, so that if you misassembled the arm, you can always update the homing offset to account for a shift up to ± 2048 steps (± 180 degrees).
 
 Then unplug your motor and plug the second motor and set its ID to 2.
 ```bash
@@ -164,7 +158,7 @@ Try to avoid rotating the motor while doing so to keep position 2048 set during
 
 ## Assemble the arms
 
-Follow step 4 of the [assembly video](https://www.youtube.com/watch?v=DA91NJOtMic). The first arm should take a bit more than 1 hour to assemble, but once you get used to it, you can do it under 1 hour for the second arm.
+Follow step 4 of the [assembly video](https://www.youtube.com/watch?v=DA91NJOtMic). The first arm should take a bit more than 1 hour to assemble, but once you get use to it, you can do it under 1 hour for the second arm.
 
 ## Calibrate
 
@@ -218,9 +212,6 @@ python lerobot/scripts/control_robot.py \
 
 **Teleop with displaying cameras**
 Follow [this guide to setup your cameras](https://github.com/huggingface/lerobot/blob/main/examples/7_get_started_with_real_robot.md#c-add-your-cameras-with-opencvcamera). Then you will be able to display the cameras on your computer while you are teleoperating by running the following code. This is useful to prepare your setup before recording your first dataset.
-
-> **NOTE:** To visualize the data, enable `--control.display_data=true`. This streams the data using `rerun`.
-
 ```bash
 python lerobot/scripts/control_robot.py \
   --robot.type=moss \
@@ -301,7 +292,7 @@ python lerobot/scripts/train.py \
 
 Let's explain it:
 1. We provided the dataset as argument with `--dataset.repo_id=${HF_USER}/moss_test`.
-2. We provided the policy with `policy.type=act`. This loads configurations from [`configuration_act.py`](../lerobot/common/policies/act/configuration_act.py). Importantly, this policy will automatically adapt to the number of motor states, motor actions and cameras of your robot (e.g. `laptop` and `phone`) which have been saved in your dataset.
+2. We provided the policy with `policy.type=act`. This loads configurations from [`configuration_act.py`](../lerobot/common/policies/act/configuration_act.py). Importantly, this policy will automatically adapt to the number of motor sates, motor actions and cameras of your robot (e.g. `laptop` and `phone`) which have been saved in your dataset.
 4. We provided `policy.device=cuda` since we are training on a Nvidia GPU, but you could use `policy.device=mps` to train on Apple silicon.
 5. We provided `wandb.enable=true` to use [Weights and Biases](https://docs.wandb.ai/quickstart) for visualizing training plots. This is optional but if you use it, make sure you are logged in by running `wandb login`.
 

examples/12_train_hilserl_classifier.md

@@ -0,0 +1,94 @@
+# Training a HIL-SERL Reward Classifier with LeRobot
+
+This tutorial provides step-by-step instructions for training a reward classifier using LeRobot.
+
+---
+
+## Training Script Overview
+
+LeRobot includes a ready-to-use training script located at [`lerobot/scripts/train_hilserl_classifier.py`](../../lerobot/scripts/train_hilserl_classifier.py). Here's an outline of its workflow:
+
+1. **Configuration Loading**
+   The script uses Hydra to load a configuration file for subsequent steps. (Details on Hydra follow below.)
+
+2. **Dataset Initialization**
+   It loads a `LeRobotDataset` containing images and rewards. To optimize performance, a weighted random sampler is used to balance class sampling.
+
+3. **Classifier Initialization**
+   A lightweight classification head is built on top of a frozen, pretrained image encoder from HuggingFace. The classifier outputs either:
+   - A single probability (binary classification), or
+   - Logits (multi-class classification).
+
+4. **Training Loop Execution**
+   The script performs:
+   - Forward and backward passes,
+   - Optimization steps,
+   - Periodic logging, evaluation, and checkpoint saving.
+
+---
+
+## Configuring with Hydra
+
+For detailed information about Hydra usage, refer to [`examples/4_train_policy_with_script.md`](../examples/4_train_policy_with_script.md). However, note that training the reward classifier differs slightly and requires a separate configuration file.
+
+### Config File Setup
+
+The default `default.yaml` cannot launch the reward classifier training directly. Instead, you need a configuration file like [`lerobot/configs/policy/hilserl_classifier.yaml`](../../lerobot/configs/policy/hilserl_classifier.yaml), with the following adjustment:
+
+Replace the `dataset_repo_id` field with the identifier for your dataset, which contains images and sparse rewards:
+
+```yaml
+# Example: lerobot/configs/policy/reward_classifier.yaml
+dataset_repo_id: "my_dataset_repo_id"
+## Typical logs and metrics
+```
+When you start the training process, you will first see your full configuration being printed in the terminal. You can check it to make sure that you config it correctly and your config is not overrided by other files. The final configuration will also be saved with the checkpoint.
+
+After that, you will see training log like this one:
+
+```
+[2024-11-29 18:26:36,999][root][INFO] -
+Epoch 5/5
+Training:  82%|██████████████████████████████████████████████████████████████████████████████▋                 | 91/111 [00:50<00:09,  2.04it/s, loss=0.2999, acc=69.99%]
+```
+
+or evaluation log like:
+
+```
+Validation: 100%|████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 28/28 [00:20<00:00,  1.37it/s]
+```
+
+### Metrics Tracking with Weights & Biases (WandB)
+
+If `wandb.enable` is set to `true`, the training and evaluation logs will also be saved in WandB. This allows you to track key metrics in real-time, including:
+
+- **Training Metrics**:
+  - `train/accuracy`
+  - `train/loss`
+  - `train/dataloading_s`
+- **Evaluation Metrics**:
+  - `eval/accuracy`
+  - `eval/loss`
+  - `eval/eval_s`
+
+#### Additional Features
+
+You can also log sample predictions during evaluation. Each logged sample will include:
+
+- The **input image**.
+- The **predicted label**.
+- The **true label**.
+- The **classifier's "confidence" (logits/probability)**.
+
+These logs can be useful for diagnosing and debugging performance issues.
+
+
+#### Generate protobuf files
+
+```bash
+python -m grpc_tools.protoc \
+    -I lerobot/scripts/server \
+    --python_out=lerobot/scripts/server \
+    --grpc_python_out=lerobot/scripts/server \
+    lerobot/scripts/server/hilserl.proto
+```

examples/12_use_so101.md

@@ -1,711 +0,0 @@
-# Assemble and use SO-101
-
-In the steps below we explain how to assemble and use our flagship robot, the SO-101 with LeRobot 🤗.
-
-## Source the parts
-
-Follow this [README](https://github.com/TheRobotStudio/SO-ARM100). It contains the bill of materials, with a link to source the parts, as well as the instructions to 3D print the parts,
-and advice if it's your first time printing or if you don't own a 3D printer.
-
-Before assembling, you will first need to configure your motors. To this end, we provide a nice script, so let's first install LeRobot. After configuration, we will also guide you through assembly.
-
-## Install LeRobot
-
-> [!TIP]
-> We use the Command Prompt (cmd) quite a lot. If you are not comfortable using the cmd or want to brush up using the command line you can have a look here: [Command line crash course](https://developer.mozilla.org/en-US/docs/Learn_web_development/Getting_started/Environment_setup/Command_line)
-
-Download our source code:
-```bash
-git clone https://github.com/huggingface/lerobot.git
-cd lerobot
-```
-
-Create a virtual environment with Python 3.10 and activate it, e.g. with [`miniconda`](https://docs.anaconda.com/miniconda/install/#quick-command-line-install):
-```bash
-conda create -y -n lerobot python=3.10
-```
-Now restart the shell by running:
-
-##### Windows:
-```bash
-`source ~/.bashrc`
-```
-
-##### Mac:
-```bash
-`source ~/.bash_profile`
-```
-
-##### zshell:
-```bash
-`source ~/.zshrc`
-```
-
-Then activate your conda environment, you have to do this each time you open a shell to use lerobot:
-```bash
-conda activate lerobot
-```
-
-When using `miniconda`, install `ffmpeg` in your environment:
-```bash
-conda install ffmpeg -c conda-forge
-```
-
-> [!NOTE]
-> This usually installs `ffmpeg 7.X` for your platform compiled with the `libsvtav1` encoder. If `libsvtav1` is not supported (check supported encoders with `ffmpeg -encoders`), you can:
->  - _[On any platform]_ Explicitly install `ffmpeg 7.X` using:
->  ```bash
->  conda install ffmpeg=7.1.1 -c conda-forge
->  ```
->  - _[On Linux only]_ 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`.
-
-Install 🤗 LeRobot:
-```bash
-cd lerobot && pip install -e ".[feetech]"
-```
-
-> [!NOTE]
-> If you encounter build errors, you may need to install additional dependencies (`cmake`, `build-essential`, and `ffmpeg libs`). On Linux, run: `sudo apt-get install cmake build-essential python3-dev pkg-config libavformat-dev libavcodec-dev libavdevice-dev libavutil-dev libswscale-dev libswresample-dev libavfilter-dev pkg-config`. For other systems, see: [Compiling PyAV](https://pyav.org/docs/develop/overview/installation.html#bring-your-own-ffmpeg)
-
-
-## Configure motors
-
-To configure the motors designate one bus servo adapter and 6 motors for your leader arm, and similarly the other bus servo adapter and 6 motors for the follower arm. It's convenient to label them and write on each motor if it's for the follower `F` or for the leader `L` and it's ID from 1 to 6.
-
-You now should plug the 5V or 12V power supply to the motor bus. 5V for the STS3215 7.4V motors and 12V for the STS3215 12V motors. Note that the leader arm always uses the 7.4V motors, so watch out that you plug in the right power supply if you have 12V and 7.4V motors, otherwise you might burn your motors! Now, connect the motor bus to your computer via USB. Note that the USB doesn't provide any power, and both the power supply and USB have to be plugged in.
-
-### Find the USB ports associated to each arm
-
-To find the port for each bus servo adapter, run this script:
-```bash
-python lerobot/scripts/find_motors_bus_port.py
-```
-#### Example outputs of script
-
-##### Mac:
-Example output leader arm's port: `/dev/tty.usbmodem575E0031751`
-
-```bash
-Finding all available ports for the MotorBus.
-['/dev/tty.usbmodem575E0032081', '/dev/tty.usbmodem575E0031751']
-Remove the usb cable from your MotorsBus and press Enter when done.
-
-[...Disconnect leader arm and press Enter...]
-
-The port of this MotorsBus is /dev/tty.usbmodem575E0031751
-Reconnect the usb cable.
-```
-
-Example output follower arm port: `/dev/tty.usbmodem575E0032081`
-
-```
-Finding all available ports for the MotorBus.
-['/dev/tty.usbmodem575E0032081', '/dev/tty.usbmodem575E0031751']
-Remove the usb cable from your MotorsBus and press Enter when done.
-
-[...Disconnect follower arm and press Enter...]
-
-The port of this MotorsBus is /dev/tty.usbmodem575E0032081
-Reconnect the usb cable.
-```
-
-##### Linux:
-On Linux, you might need to give access to the USB ports by running:
-```bash
-sudo chmod 666 /dev/ttyACM0
-sudo chmod 666 /dev/ttyACM1
-```
-
-Example output leader arm port: `/dev/ttyACM0`
-
-```bash
-Finding all available ports for the MotorBus.
-['/dev/ttyACM0', '/dev/ttyACM1']
-Remove the usb cable from your MotorsBus and press Enter when done.
-
-[...Disconnect leader arm and press Enter...]
-
-The port of this MotorsBus is /dev/ttyACM0
-Reconnect the usb cable.
-```
-
-Example output follower arm port: `/dev/ttyACM1`
-
-```
-Finding all available ports for the MotorBus.
-['/dev/ttyACM0', '/dev/ttyACM1']
-Remove the usb cable from your MotorsBus and press Enter when done.
-
-[...Disconnect follower arm and press Enter...]
-
-The port of this MotorsBus is /dev/ttyACM1
-Reconnect the usb cable.
-```
-
-#### Update config file
-
-Now that you have your ports, update the **port** default values of [`SO101RobotConfig`](https://github.com/huggingface/lerobot/blob/main/lerobot/common/robot_devices/robots/configs.py).
-You will find a class called `so101` where you can update the `port` values with your actual motor ports:
-```diff
-@RobotConfig.register_subclass("so101")
-@dataclass
-class So101RobotConfig(ManipulatorRobotConfig):
-    calibration_dir: str = ".cache/calibration/so101"
-    # `max_relative_target` limits the magnitude of the relative positional target vector for safety purposes.
-    # Set this to a positive scalar to have the same value for all motors, or a list that is the same length as
-    # the number of motors in your follower arms.
-    max_relative_target: int | None = None
-
-    leader_arms: dict[str, MotorsBusConfig] = field(
-        default_factory=lambda: {
-            "main": FeetechMotorsBusConfig(
--               port="/dev/tty.usbmodem58760431091",
-+               port="{ADD YOUR LEADER PORT}",
-                motors={
-                    # name: (index, model)
-                    "shoulder_pan": [1, "sts3215"],
-                    "shoulder_lift": [2, "sts3215"],
-                    "elbow_flex": [3, "sts3215"],
-                    "wrist_flex": [4, "sts3215"],
-                    "wrist_roll": [5, "sts3215"],
-                    "gripper": [6, "sts3215"],
-                },
-            ),
-        }
-    )
-
-    follower_arms: dict[str, MotorsBusConfig] = field(
-        default_factory=lambda: {
-            "main": FeetechMotorsBusConfig(
--                port="/dev/tty.usbmodem585A0076891",
-+                port="{ADD YOUR FOLLOWER PORT}",
-                motors={
-                    # name: (index, model)
-                    "shoulder_pan": [1, "sts3215"],
-                    "shoulder_lift": [2, "sts3215"],
-                    "elbow_flex": [3, "sts3215"],
-                    "wrist_flex": [4, "sts3215"],
-                    "wrist_roll": [5, "sts3215"],
-                    "gripper": [6, "sts3215"],
-                },
-            ),
-        }
-    )
-```
-
-Here is a video of the process:
-
-<video controls width="640" src="https://github.com/user-attachments/assets/fc45d756-31bb-4a61-b973-a87d633d08a7" type="video/mp4"></video>
-
-### Set motor IDs
-
-Now we need to set the motor ID for each motor. Plug your motor in only one of the two ports of the motor bus and run this script to set its ID to 1. Replace the text after --port to the corresponding control board port.
-```bash
-python lerobot/scripts/configure_motor.py \
-  --port /dev/tty.usbmodem58760432961 \
-  --brand feetech \
-  --model sts3215 \
-  --baudrate 1000000 \
-  --ID 1
-```
-
-Then unplug your motor and plug the second motor and set its ID to 2.
-```bash
-python lerobot/scripts/configure_motor.py \
-  --port /dev/tty.usbmodem58760432961 \
-  --brand feetech \
-  --model sts3215 \
-  --baudrate 1000000 \
-  --ID 2
-```
-
-Redo this process for all your motors until ID 6. Do the same for the 6 motors of the leader arm, but make sure to change the power supply if you use motors with different voltage.
-
-Here is a video of the process:
-
-<video controls width="640" src="https://github.com/user-attachments/assets/b31c115f-e706-4dcd-b7f1-4535da62416d" type="video/mp4"></video>
-
-## Step-by-Step Assembly Instructions
-
-The follower arm uses 6x STS3215 motors with 1/345 gearing. The leader however uses three differently geared motors to make sure it can both sustain its own weight and it can be moved without requiring much force. Which motor is needed for which joint is shown in table below.
-
-| Leader-Arm Axis | Motor | Gear Ratio |
-|-----------------|:-------:|:----------:|
-| Base / Shoulder Yaw | 1 | 1 / 191 |
-| Shoulder Pitch      | 2 | 1 / 345 |
-| Elbow               | 3 | 1 / 191 |
-| Wrist Roll          | 4 | 1 / 147 |
-| Wrist Pitch         | 5 | 1 / 147 |
-| Gripper             | 6 | 1 / 147 |
-
-
-### Clean Parts
-Remove all support material from the 3D-printed parts.
-
-### Joint 1
-
-- Place the first motor into the base.
-- Fasten the motor with 4 M2x6mm screws (smallest screws). Two from the top and two from bottom.
-- Slide over the first motor holder and fasten it using two M2x6mm screws (one on each side).
-- Install both motor horns, securing the top horn with a M3x6mm screw.
-- Attach the shoulder part.
-- Tighten the shoulder part with 4 M3x6mm screws on top and 4 M3x6mm screws on the bottom
-- Add the shoulder motor holder.
-
-<video controls width="640" src="https://github.com/user-attachments/assets/b0ee9dee-a2d0-445b-8489-02ebecb3d639" type="video/mp4"></video>
-
-### Joint 2
-
-- Slide the second motor in from the top.
-- Fasten the second motor with 4 M2x6mm screws.
-- Attach both motor horns to motor 2, again use the M3x6mm horn screw.
-- Attach the upper arm with 4 M3x6mm screws on each side.
-
-<video controls width="640" src="https://github.com/user-attachments/assets/32453dc2-5006-4140-9f56-f0d78eae5155" type="video/mp4"></video>
-
-### Joint 3
-
-- Insert motor 3 and fasten using 4 M2x6mm screws
-- Attach both motor horns to motor 3 and secure one again with a M3x6mm horn screw.
-- Connect the forearm to motor 3 using 4 M3x6mm screws on each side.
-
-<video controls width="640" src="https://github.com/user-attachments/assets/7384b9a7-a946-440c-b292-91391bcc4d6b" type="video/mp4"></video>
-
-### Joint 4
-
-- Slide over motor holder 4.
-- Slide in motor 4.
-- Fasten motor 4 with 4 M2x6mm screws and attach its motor horns, use a M3x6mm horn screw.
-
-<video controls width="640" src="https://github.com/user-attachments/assets/dca78ad0-7c36-4bdf-8162-c9ac42a1506f" type="video/mp4"></video>
-
-### Joint 5
-
-- Insert motor 5 into the wrist holder and secure it with 2 M2x6mm front screws.
-- Install only one motor horn on the wrist motor and secure it with a M3x6mm horn screw.
-- Secure the wrist to motor 4 using 4 M3x6mm screws on both sides.
-
-<video controls width="640" src="https://github.com/user-attachments/assets/55f5d245-976d-49ff-8b4a-59843c441b12" type="video/mp4"></video>
-
-### Gripper / Handle
-
-#### Follower:
-
-- Attach the gripper to motor 5, attach it to the motor horn on the wrist using 4 M3x6mm screws.
-- Insert the gripper motor and secure it with 2 M2x6mm screws on each side.
-- Attach the motor horns and again use a M3x6mm horn screw.
-- Install the gripper claw and secure it with 4 M3x6mm screws on both sides.
-
-<video controls width="640" src="https://github.com/user-attachments/assets/6f766aa9-cfae-4388-89e7-0247f198c086" type="video/mp4"></video>
-
-#### Leader:
-
-- Mount the leader holder onto the wrist and secure it with 4 M3x6mm screws.
-- Attach the handle to motor 5 using 1 M2x6mm screw.
-- Insert the gripper motor, secure it with 2 M2x6mm screws on each side, attach a motor horn using a M3x6mm horn screw.
-- Attach the follower trigger with 4 M3x6mm screws.
-
-<video controls width="640" src="https://github.com/user-attachments/assets/1308c93d-2ef1-4560-8e93-a3812568a202" type="video/mp4"></video>
-
-##### Wiring
-
-- Attach the motor controller on the back.
-- Then insert all wires, use the wire guides everywhere to make sure the wires don't unplug themselves and stay in place.
-
-<video controls width="640" src="https://github.com/user-attachments/assets/4c2cacfd-9276-4ee4-8bf2-ba2492667b78" type="video/mp4"></video>
-
-## Calibrate
-
-Next, you'll need to calibrate your SO-101 robot to ensure that the leader and follower arms have the same position values when they are in the same physical position.
-The calibration process is very important because it allows a neural network trained on one SO-101 robot to work on another.
-
-#### Manual calibration of follower arm
-
-You will need to move the follower arm to these positions sequentially, note that the rotated position is on the right side of the robot and you have to open the gripper fully.
-
-| 1. Middle position | 2. Zero position                                                                                                                                       | 3. Rotated position                                                                                                                                             | 4. Rest position                                                                                                                                       |
-| ------------ |------------------------------------------------------------------------------------------------------------------------------------------------------ | --------------------------------------------------------------------------------------------------------------------------------------------------------------- | ------------------------------------------------------------------------------------------------------------------------------------------------------ |
-| <img src="../media/so101/follower_middle.webp?raw=true" alt="SO-101 leader arm middle position" title="SO-101 leader arm middle position" style="width:100%;"> | <img src="../media/so101/follower_zero.webp?raw=true" alt="SO-101 leader arm zero position" title="SO-101 leader arm zero position" style="width:100%;"> | <img src="../media/so101/follower_rotated.webp?raw=true" alt="SO-101 leader arm rotated position" title="SO-101 leader arm rotated position" style="width:100%;"> | <img src="../media/so101/follower_rest.webp?raw=true" alt="SO-101 leader arm rest position" title="SO-101 leader arm rest position" style="width:100%;"> |
-
-Make sure both arms are connected and run this script to launch manual calibration:
-```bash
-python lerobot/scripts/control_robot.py \
-  --robot.type=so101 \
-  --robot.cameras='{}' \
-  --control.type=calibrate \
-  --control.arms='["main_follower"]'
-```
-
-#### Manual calibration of leader arm
-You will also need to move the leader arm to these positions sequentially:
-
-| 1. Middle position | 2. Zero position                                                                                                                                       | 3. Rotated position                                                                                                                                             | 4. Rest position                                                                                                                                       |
-| ------------ |------------------------------------------------------------------------------------------------------------------------------------------------------ | --------------------------------------------------------------------------------------------------------------------------------------------------------------- | ------------------------------------------------------------------------------------------------------------------------------------------------------ |
-| <img src="../media/so101/leader_middle.webp?raw=true" alt="SO-101 leader arm middle position" title="SO-101 leader arm middle position" style="width:100%;"> | <img src="../media/so101/leader_zero.webp?raw=true" alt="SO-101 leader arm zero position" title="SO-101 leader arm zero position" style="width:100%;"> | <img src="../media/so101/leader_rotated.webp?raw=true" alt="SO-101 leader arm rotated position" title="SO-101 leader arm rotated position" style="width:100%;"> | <img src="../media/so101/leader_rest.webp?raw=true" alt="SO-101 leader arm rest position" title="SO-101 leader arm rest position" style="width:100%;"> |
-
-Run this script to launch manual calibration:
-```bash
-python lerobot/scripts/control_robot.py \
-  --robot.type=so101 \
-  --robot.cameras='{}' \
-  --control.type=calibrate \
-  --control.arms='["main_leader"]'
-```
-## Control your robot
-
-Congrats 🎉, your robot is all set to learn a task on its own. Next we will explain to you how to train a neural network to autonomously control a real robot.
-
-**You'll learn to:**
-1. How to record and visualize your dataset.
-2. How to train a policy using your data and prepare it for evaluation.
-3. How to evaluate your policy and visualize the results.
-
-By following these steps, you'll be able to replicate tasks like picking up a Lego block and placing it in a bin with a high success rate, as demonstrated in [this video](https://x.com/RemiCadene/status/1814680760592572934).
-
-This tutorial is specifically made for the affordable [SO-101](https://github.com/TheRobotStudio/SO-ARM100) robot, but it contains additional information to be easily adapted to various types of robots like [Aloha bimanual robot](https://aloha-2.github.io) by changing some configurations. The SO-101 consists of a leader arm and a follower arm, each with 6 motors. It can work with one or several cameras to record the scene, which serve as visual sensors for the robot.
-
-During the data collection phase, you will control the follower arm by moving the leader arm. This process is known as "teleoperation." This technique is used to collect robot trajectories. Afterward, you'll train a neural network to imitate these trajectories and deploy the network to enable your robot to operate autonomously.
-
-If you encounter any issues at any step of the tutorial, feel free to seek help on [Discord](https://discord.com/invite/s3KuuzsPFb) or don't hesitate to iterate with us on the tutorial by creating issues or pull requests.
-
-## Teleoperate
-
-Run this simple script to teleoperate your robot (it won't connect and display the cameras):
-```bash
-python lerobot/scripts/control_robot.py \
-  --robot.type=so101 \
-  --robot.cameras='{}' \
-  --control.type=teleoperate
-```
-
-The teleoperate command will automatically:
-1. Identify any missing calibrations and initiate the calibration procedure.
-2. Connect the robot and start teleoperation.
-
-## Setup Cameras
-
-To connect a camera you have three options:
-1. OpenCVCamera which allows us to use any camera: usb, realsense, laptop webcam
-2. iPhone camera with MacOS
-3. Phone camera on Linux
-
-### Use OpenCVCamera
-
-The [`OpenCVCamera`](../lerobot/common/robot_devices/cameras/opencv.py) class allows you to efficiently record frames from most cameras using the [`opencv2`](https://docs.opencv.org) library.  For more details on compatibility, see [Video I/O with OpenCV Overview](https://docs.opencv.org/4.x/d0/da7/videoio_overview.html).
-
-To instantiate an [`OpenCVCamera`](../lerobot/common/robot_devices/cameras/opencv.py), you need a camera index (e.g. `OpenCVCamera(camera_index=0)`). When you only have one camera like a webcam of a laptop, the camera index is usually `0` but it might differ, and the camera index might change if you reboot your computer or re-plug your camera. This behavior depends on your operating system.
-
-To find the camera indices, run the following utility script, which will save a few frames from each detected camera:
-```bash
-python lerobot/common/robot_devices/cameras/opencv.py \
-    --images-dir outputs/images_from_opencv_cameras
-```
-
-The output will look something like this if you have two cameras connected:
-```
-Mac or Windows detected. Finding available camera indices through scanning all indices from 0 to 60
-[...]
-Camera found at index 0
-Camera found at index 1
-[...]
-Connecting cameras
-OpenCVCamera(0, fps=30.0, width=1920.0, height=1080.0, color_mode=rgb)
-OpenCVCamera(1, fps=24.0, width=1920.0, height=1080.0, color_mode=rgb)
-Saving images to outputs/images_from_opencv_cameras
-Frame: 0000 Latency (ms): 39.52
-[...]
-Frame: 0046 Latency (ms): 40.07
-Images have been saved to outputs/images_from_opencv_cameras
-```
-
-Check the saved images in `outputs/images_from_opencv_cameras` to identify which camera index corresponds to which physical camera (e.g. `0` for `camera_00` or `1` for `camera_01`):
-```
-camera_00_frame_000000.png
-[...]
-camera_00_frame_000047.png
-camera_01_frame_000000.png
-[...]
-camera_01_frame_000047.png
-```
-
-Note: Some cameras may take a few seconds to warm up, and the first frame might be black or green.
-
-Now that you have the camera indexes, you should change them in the config. You can also change the fps, width or height of the camera.
-
-The camera config is defined per robot, can be found here [`RobotConfig`](https://github.com/huggingface/lerobot/blob/main/lerobot/common/robot_devices/robots/configs.py) and looks like this:
-```python
-cameras: dict[str, CameraConfig] = field(
-        default_factory=lambda: {
-            "wrist": OpenCVCameraConfig(
-                camera_index=0,  <-- UPDATE HERE
-                fps=30,
-                width=640,
-                height=480,
-            ),
-            "base": OpenCVCameraConfig(
-                camera_index=1,   <-- UPDATE HERE
-                fps=30,
-                width=640,
-                height=480,
-            ),
-        }
-    )
-```
-
-### Use your phone
-#### Mac:
-
-To use your iPhone as a camera on macOS, enable the Continuity Camera feature:
-- Ensure your Mac is running macOS 13 or later, and your iPhone is on iOS 16 or later.
-- Sign in both devices with the same Apple ID.
-- Connect your devices with a USB cable or turn on Wi-Fi and Bluetooth for a wireless connection.
-
-For more details, visit [Apple support](https://support.apple.com/en-gb/guide/mac-help/mchl77879b8a/mac).
-
-Your iPhone should be detected automatically when running the camera setup script in the next section.
-
-#### Linux:
-
-If you want to use your phone as a camera on Linux, follow these steps to set up a virtual camera
-
-1. *Install `v4l2loopback-dkms` and `v4l-utils`*. Those packages are required to create virtual camera devices (`v4l2loopback`) and verify their settings with the `v4l2-ctl` utility from `v4l-utils`. Install them using:
-```python
-sudo apt install v4l2loopback-dkms v4l-utils
-```
-2. *Install [DroidCam](https://droidcam.app) on your phone*. This app is available for both iOS and Android.
-3. *Install [OBS Studio](https://obsproject.com)*. This software will help you manage the camera feed. Install it using [Flatpak](https://flatpak.org):
-```python
-flatpak install flathub com.obsproject.Studio
-```
-4. *Install the DroidCam OBS plugin*. This plugin integrates DroidCam with OBS Studio. Install it with:
-```python
-flatpak install flathub com.obsproject.Studio.Plugin.DroidCam
-```
-5. *Start OBS Studio*. Launch with:
-```python
-flatpak run com.obsproject.Studio
-```
-6. *Add your phone as a source*. Follow the instructions [here](https://droidcam.app/obs/usage). Be sure to set the resolution to `640x480`.
-7. *Adjust resolution settings*. In OBS Studio, go to `File > Settings > Video`. Change the `Base(Canvas) Resolution` and the `Output(Scaled) Resolution` to `640x480` by manually typing it in.
-8. *Start virtual camera*. In OBS Studio, follow the instructions [here](https://obsproject.com/kb/virtual-camera-guide).
-9. *Verify the virtual camera setup*. Use `v4l2-ctl` to list the devices:
-```python
-v4l2-ctl --list-devices
-```
-You should see an entry like:
-```
-VirtualCam (platform:v4l2loopback-000):
-/dev/video1
-```
-10. *Check the camera resolution*. Use `v4l2-ctl` to ensure that the virtual camera output resolution is `640x480`. Change `/dev/video1` to the port of your virtual camera from the output of `v4l2-ctl --list-devices`.
-```python
-v4l2-ctl -d /dev/video1 --get-fmt-video
-```
-You should see an entry like:
-```
->>> Format Video Capture:
->>> Width/Height      : 640/480
->>> Pixel Format      : 'YUYV' (YUYV 4:2:2)
-```
-
-Troubleshooting: If the resolution is not correct you will have to delete the Virtual Camera port and try again as it cannot be changed.
-
-If everything is set up correctly, you can proceed with the rest of the tutorial.
-
-### Add wrist camera
-If you have an additional camera you can add a wrist camera to the SO101. There are already many premade wrist camera holders that you can find in the SO101 repo: [Wrist camera's](https://github.com/TheRobotStudio/SO-ARM100#wrist-cameras)
-
-## Teleoperate with cameras
-
-We can now teleoperate again while at the same time visualizing the cameras and joint positions with `rerun`.
-
-```bash
-python lerobot/scripts/control_robot.py \
-  --robot.type=so101 \
-  --control.type=teleoperate \
-  --control.display_data=true
-```
-
-## Record a dataset
-
-Once you're familiar with teleoperation, you can record your first dataset with SO-101.
-
-We use the Hugging Face hub features for uploading your dataset. If you haven't previously used the Hub, make sure you can login via the cli using a write-access token, this token can be generated from the [Hugging Face settings](https://huggingface.co/settings/tokens).
-
-Add your token to the cli by running this command:
-```bash
-huggingface-cli login --token ${HUGGINGFACE_TOKEN} --add-to-git-credential
-```
-
-Then store your Hugging Face repository name in a variable:
-```bash
-HF_USER=$(huggingface-cli whoami | head -n 1)
-echo $HF_USER
-```
-
-Now you can record a dataset, to record 2 episodes and upload your dataset to the hub execute this command:
-```bash
-python lerobot/scripts/control_robot.py \
-  --robot.type=so101 \
-  --control.type=record \
-  --control.fps=30 \
-  --control.single_task="Grasp a lego block and put it in the bin." \
-  --control.repo_id=${HF_USER}/so101_test \
-  --control.tags='["so101","tutorial"]' \
-  --control.warmup_time_s=5 \
-  --control.episode_time_s=30 \
-  --control.reset_time_s=30 \
-  --control.num_episodes=2 \
-  --control.display_data=true \
-  --control.push_to_hub=true
-```
-
-You will see a lot of lines appearing like this one:
-```
-INFO 2024-08-10 15:02:58 ol_robot.py:219 dt:33.34 (30.0hz) dtRlead: 5.06 (197.5hz) dtWfoll: 0.25 (3963.7hz) dtRfoll: 6.22 (160.7hz) dtRlaptop: 32.57 (30.7hz) dtRphone: 33.84 (29.5hz)
-```
-It contains:
-- `2024-08-10 15:02:58` which is the date and time of the call to the print function,
-- `ol_robot.py:219` which is the end of the file name and the line number where the print function is called  (`lerobot/scripts/control_robot.py` line `219`).
-- `dt:33.34 (30.0hz)` which is the "delta time" or the number of milliseconds spent between the previous call to `robot.teleop_step(record_data=True)` and the current one, associated with the frequency (33.34 ms equals 30.0 Hz) ; note that we use `--fps 30` so we expect 30.0 Hz ; when a step takes more time, the line appears in yellow.
-- `dtRlead: 5.06 (197.5hz)` which is the delta time of reading the present position of the leader arm.
-- `dtWfoll: 0.25 (3963.7hz)` which is the delta time of writing the goal position on the follower arm ; writing is asynchronous so it takes less time than reading.
-- `dtRfoll: 6.22 (160.7hz)` which is the delta time of reading the present position on the follower arm.
-- `dtRlaptop:32.57 (30.7hz) ` which is the delta time of capturing an image from the laptop camera in the thread running asynchronously.
-- `dtRphone:33.84 (29.5hz)` which is the delta time of capturing an image from the phone camera in the thread running asynchronously.
-
-#### Dataset upload
-Locally your dataset is stored in this folder: `~/.cache/huggingface/lerobot/{repo-id}` (e.g. `data/cadene/so101_test`). At the end of data recording, your dataset will be uploaded on your Hugging Face page (e.g. https://huggingface.co/datasets/cadene/so101_test) that you can obtain by running:
-```bash
-echo https://huggingface.co/datasets/${HF_USER}/so101_test
-```
-Your dataset will be automatically tagged with `LeRobot` for the community to find it easily, and you can also add custom tags (in this case `tutorial` for example).
-
-You can look for other LeRobot datasets on the hub by searching for `LeRobot` [tags](https://huggingface.co/datasets?other=LeRobot).
-
-#### Record function
-
-The `record` function provides a suite of tools for capturing and managing data during robot operation:
-1. Set the flow of data recording using command line arguments:
-   - `--control.warmup_time_s=10` defines the number of seconds before starting data collection. It allows the robot devices to warmup and synchronize (10 seconds by default).
-   - `--control.episode_time_s=60` defines the number of seconds for data recording for each episode (60 seconds by default).
-   - `--control.reset_time_s=60` defines the number of seconds for resetting the environment after each episode (60 seconds by default).
-   - `--control.num_episodes=50` defines the number of episodes to record (50 by default).
-2. Control the flow during data recording using keyboard keys:
-   - Press right arrow `->` at any time during episode recording to early stop and go to resetting. Same during resetting, to early stop and to go to the next episode recording.
-   - Press left arrow `<-` at any time during episode recording or resetting to early stop, cancel the current episode, and re-record it.
-   - Press escape `ESC` at any time during episode recording to end the session early and go straight to video encoding and dataset uploading.
-3. Checkpoints are done set during recording, so if any issue occurs, you can resume recording by re-running the same command again with `--control.resume=true`. You will need to manually delete the dataset directory if you want to start recording from scratch.
-
-#### Tips for gathering data
-
-Once you're comfortable with data recording, you can create a larger dataset for training. A good starting task is grasping an object at different locations and placing it in a bin. We suggest recording at least 50 episodes, with 10 episodes per location. Keep the cameras fixed and maintain consistent grasping behavior throughout the recordings. Also make sure the object you are manipulating is visible on the camera's. A good rule of thumb is you should be able to do the task yourself by only looking at the camera images.
-
-In the following sections, you’ll train your neural network. After achieving reliable grasping performance, you can start introducing more variations during data collection, such as additional grasp locations, different grasping techniques, and altering camera positions.
-
-Avoid adding too much variation too quickly, as it may hinder your results.
-
-#### Troubleshooting:
-- On Linux, if the left and right arrow keys and escape key don't have any effect during data recording, make sure you've set the `$DISPLAY` environment variable. See [pynput limitations](https://pynput.readthedocs.io/en/latest/limitations.html#linux).
-
-## Visualize a dataset
-
-If you uploaded your dataset to the hub with `--control.push_to_hub=true`, you can [visualize your dataset online](https://huggingface.co/spaces/lerobot/visualize_dataset) by copy pasting your repo id given by:
-```bash
-echo ${HF_USER}/so101_test
-```
-
-If you didn't upload with `--control.push_to_hub=false`, you can visualize it locally with (via a window in the browser `http://127.0.0.1:9090` with the visualization tool):
-```bash
-python lerobot/scripts/visualize_dataset_html.py \
-  --repo-id ${HF_USER}/so101_test \
-  --local-files-only 1
-```
-
-This will launch a local web server that looks like this:
-
-<div style="text-align:center;">
-  <img src="../media/tutorial/visualize_dataset_html.webp?raw=true" alt="Koch v1.1 leader and follower arms" title="Koch v1.1 leader and follower arms" width="100%"></img>
-</div>
-
-## Replay an episode
-
-A useful feature is the `replay` function, which allows to replay on your robot any episode that you've recorded or episodes from any dataset out there. This function helps you test the repeatability of your robot's actions and assess transferability across robots of the same model.
-
-You can replay the first episode on your robot with:
-```bash
-python lerobot/scripts/control_robot.py \
-  --robot.type=so101 \
-  --control.type=replay \
-  --control.fps=30 \
-  --control.repo_id=${HF_USER}/so101_test \
-  --control.episode=0
-```
-
-Your robot should replicate movements similar to those you recorded. For example, check out [this video](https://x.com/RemiCadene/status/1793654950905680090) where we use `replay` on a Aloha robot from [Trossen Robotics](https://www.trossenrobotics.com).
-
-## Train a policy
-
-To train a policy to control your robot, use the [`python lerobot/scripts/train.py`](../lerobot/scripts/train.py) script. A few arguments are required. Here is an example command:
-```bash
-python lerobot/scripts/train.py \
-  --dataset.repo_id=${HF_USER}/so101_test \
-  --policy.type=act \
-  --output_dir=outputs/train/act_so101_test \
-  --job_name=act_so101_test \
-  --policy.device=cuda \
-  --wandb.enable=true
-```
-
-Let's explain the command:
-1. We provided the dataset as argument with `--dataset.repo_id=${HF_USER}/so101_test`.
-2. We provided the policy with `policy.type=act`. This loads configurations from [`configuration_act.py`](../lerobot/common/policies/act/configuration_act.py). Importantly, this policy will automatically adapt to the number of motor states, motor actions and cameras of your robot (e.g. `laptop` and `phone`) which have been saved in your dataset.
-4. We provided `policy.device=cuda` since we are training on a Nvidia GPU, but you could use `policy.device=mps` to train on Apple silicon.
-5. We provided `wandb.enable=true` to use [Weights and Biases](https://docs.wandb.ai/quickstart) for visualizing training plots. This is optional but if you use it, make sure you are logged in by running `wandb login`.
-
-Training should take several hours. You will find checkpoints in `outputs/train/act_so101_test/checkpoints`.
-
-To resume training from a checkpoint, below is an example command to resume from `last` checkpoint of the `act_so101_test` policy:
-```bash
-python lerobot/scripts/train.py \
-  --config_path=outputs/train/act_so101_test/checkpoints/last/pretrained_model/train_config.json \
-  --resume=true
-```
-
-#### Upload policy checkpoints
-
-Once training is done, upload the latest checkpoint with:
-```bash
-huggingface-cli upload ${HF_USER}/act_so101_test \
-  outputs/train/act_so101_test/checkpoints/last/pretrained_model
-```
-
-You can also upload intermediate checkpoints with:
-```bash
-CKPT=010000
-huggingface-cli upload ${HF_USER}/act_so101_test${CKPT} \
-  outputs/train/act_so101_test/checkpoints/${CKPT}/pretrained_model
-```
-
-## Evaluate your policy
-
-You can use the `record` function from [`lerobot/scripts/control_robot.py`](../lerobot/scripts/control_robot.py) but with a policy checkpoint as input. For instance, run this command to record 10 evaluation episodes:
-```bash
-python lerobot/scripts/control_robot.py \
-  --robot.type=so101 \
-  --control.type=record \
-  --control.fps=30 \
-  --control.single_task="Grasp a lego block and put it in the bin." \
-  --control.repo_id=${HF_USER}/eval_act_so101_test \
-  --control.tags='["tutorial"]' \
-  --control.warmup_time_s=5 \
-  --control.episode_time_s=30 \
-  --control.reset_time_s=30 \
-  --control.num_episodes=10 \
-  --control.push_to_hub=true \
-  --control.policy.path=outputs/train/act_so101_test/checkpoints/last/pretrained_model
-```
-
-As you can see, it's almost the same command as previously used to record your training dataset. Two things changed:
-1. There is an additional `--control.policy.path` argument which indicates the path to your policy checkpoint with  (e.g. `outputs/train/eval_act_so101_test/checkpoints/last/pretrained_model`). You can also use the model repository if you uploaded a model checkpoint to the hub (e.g. `${HF_USER}/act_so101_test`).
-2. The name of dataset begins by `eval` to reflect that you are running inference (e.g. `${HF_USER}/eval_act_so101_test`).

examples/1_load_lerobot_dataset.py

@@ -32,7 +32,10 @@ import torch
 from huggingface_hub import HfApi
 
 import lerobot
-from lerobot.common.datasets.lerobot_dataset import LeRobotDataset, LeRobotDatasetMetadata
+from lerobot.common.datasets.lerobot_dataset import (
+    LeRobotDataset,
+    LeRobotDatasetMetadata,
+)
 
 # We ported a number of existing datasets ourselves, use this to see the list:
 print("List of available datasets:")
@@ -119,7 +122,7 @@ print(dataset.features[camera_key]["shape"])
 delta_timestamps = {
     # loads 4 images: 1 second before current frame, 500 ms before, 200 ms before, and current frame
     camera_key: [-1, -0.5, -0.20, 0],
-    # loads 6 state vectors: 1.5 seconds before, 1 second before, ... 200 ms, 100 ms, and current frame
+    # loads 8 state vectors: 1.5 seconds before, 1 second before, ... 200 ms, 100 ms, and current frame
     "observation.state": [-1.5, -1, -0.5, -0.20, -0.10, 0],
     # loads 64 action vectors: current frame, 1 frame in the future, 2 frames, ... 63 frames in the future
     "action": [t / dataset.fps for t in range(64)],
@@ -143,6 +146,6 @@ dataloader = torch.utils.data.DataLoader(
 
 for batch in dataloader:
     print(f"{batch[camera_key].shape=}")  # (32, 4, c, h, w)
-    print(f"{batch['observation.state'].shape=}")  # (32, 6, c)
+    print(f"{batch['observation.state'].shape=}")  # (32, 5, c)
     print(f"{batch['action'].shape=}")  # (32, 64, c)
     break

examples/2_evaluate_pretrained_policy.py

@@ -13,12 +13,12 @@
 # limitations under the License.
 
 """
-This script demonstrates how to evaluate a pretrained policy from the HuggingFace Hub or from your local
+This scripts demonstrates how to evaluate a pretrained policy from the HuggingFace Hub or from your local
 training outputs directory. In the latter case, you might want to run examples/3_train_policy.py first.
 
 It requires the installation of the 'gym_pusht' simulation environment. Install it by running:
 ```bash
-pip install -e ".[pusht]"
+pip install --no-binary=av -e ".[pusht]"`
 ```
 """
 
@@ -119,7 +119,7 @@ while not done:
     rewards.append(reward)
     frames.append(env.render())
 
-    # The rollout is considered done when the success state is reached (i.e. terminated is True),
+    # The rollout is considered done when the success state is reach (i.e. terminated is True),
     # or the maximum number of iterations is reached (i.e. truncated is True)
     done = terminated | truncated | done
     step += 1

examples/3_train_policy.py

@@ -12,7 +12,7 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-"""This script demonstrates how to train Diffusion Policy on the PushT environment.
+"""This scripts demonstrates how to train Diffusion Policy on the PushT environment.
 
 Once you have trained a model with this script, you can try to evaluate it on
 examples/2_evaluate_pretrained_policy.py
@@ -22,7 +22,10 @@ from pathlib import Path
 
 import torch
 
-from lerobot.common.datasets.lerobot_dataset import LeRobotDataset, LeRobotDatasetMetadata
+from lerobot.common.datasets.lerobot_dataset import (
+    LeRobotDataset,
+    LeRobotDatasetMetadata,
+)
 from lerobot.common.datasets.utils import dataset_to_policy_features
 from lerobot.common.policies.diffusion.configuration_diffusion import DiffusionConfig
 from lerobot.common.policies.diffusion.modeling_diffusion import DiffusionPolicy
@@ -77,7 +80,24 @@ def main():
         # Load the previous action (-0.1), the next action to be executed (0.0),
         # and 14 future actions with a 0.1 seconds spacing. All these actions will be
         # used to supervise the policy.
-        "action": [-0.1, 0.0, 0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8, 0.9, 1.0, 1.1, 1.2, 1.3, 1.4],
+        "action": [
+            -0.1,
+            0.0,
+            0.1,
+            0.2,
+            0.3,
+            0.4,
+            0.5,
+            0.6,
+            0.7,
+            0.8,
+            0.9,
+            1.0,
+            1.1,
+            1.2,
+            1.3,
+            1.4,
+        ],
     }
 
     # We can then instantiate the dataset with these delta_timestamps configuration.

examples/4_train_policy_with_script.md

@@ -1,10 +1,10 @@
 This tutorial will explain the training script, how to use it, and particularly how to configure everything needed for the training run.
-> **Note:** The following assumes you're running these commands on a machine equipped with a cuda GPU. If you don't have one (or if you're using a Mac), you can add `--policy.device=cpu` (`--policy.device=mps` respectively). However, be advised that the code executes much slower on cpu.
+> **Note:** The following assume you're running these commands on a machine equipped with a cuda GPU. If you don't have one (or if you're using a Mac), you can add `--policy.device=cpu` (`--policy.device=mps` respectively). However, be advised that the code executes much slower on cpu.
 
 
 ## The training script
 
-LeRobot offers a training script at [`lerobot/scripts/train.py`](../lerobot/scripts/train.py). At a high level it does the following:
+LeRobot offers a training script at [`lerobot/scripts/train.py`](../../lerobot/scripts/train.py). At a high level it does the following:
 
 - Initialize/load a configuration for the following steps using.
 - Instantiates a dataset.
@@ -21,9 +21,9 @@ In the training script, the main function `train` expects a `TrainPipelineConfig
 def train(cfg: TrainPipelineConfig):
 ```
 
-You can inspect the `TrainPipelineConfig` defined in [`lerobot/configs/train.py`](../lerobot/configs/train.py) (which is heavily commented and meant to be a reference to understand any option)
+You can inspect the `TrainPipelineConfig` defined in [`lerobot/configs/train.py`](../../lerobot/configs/train.py) (which is heavily commented and meant to be a reference to understand any option)
 
-When running the script, inputs for the command line are parsed thanks to the `@parser.wrap()` decorator and an instance of this class is automatically generated. Under the hood, this is done with [Draccus](https://github.com/dlwh/draccus) which is a tool dedicated to this purpose. If you're familiar with Hydra, Draccus can similarly load configurations from config files (.json, .yaml) and also override their values through command line inputs. Unlike Hydra, these configurations are pre-defined in the code through dataclasses rather than being defined entirely in config files. This allows for more rigorous serialization/deserialization, typing, and to manipulate configuration as objects directly in the code and not as dictionaries or namespaces (which enables nice features in an IDE such as autocomplete, jump-to-def, etc.)
+When running the script, inputs for the command line are parsed thanks to the `@parser.wrap()` decorator and an instance of this class is automatically generated. Under the hood, this is done with [Draccus](https://github.com/dlwh/draccus) which is a tool dedicated for this purpose. If you're familiar with Hydra, Draccus can similarly load configurations from config files (.json, .yaml) and also override their values through command line inputs. Unlike Hydra, these configurations are pre-defined in the code through dataclasses rather than being defined entirely in config files. This allows for more rigorous serialization/deserialization, typing, and to manipulate configuration as objects directly in the code and not as dictionaries or namespaces (which enables nice features in an IDE such as autocomplete, jump-to-def, etc.)
 
 Let's have a look at a simplified example. Amongst other attributes, the training config has the following attributes:
 ```python
@@ -43,14 +43,14 @@ class DatasetConfig:
 ```
 
 This creates a hierarchical relationship where, for example assuming we have a `cfg` instance of `TrainPipelineConfig`, we can access the `repo_id` value with `cfg.dataset.repo_id`.
-From the command line, we can specify this value by using a very similar syntax `--dataset.repo_id=repo/id`.
+From the command line, we can specify this value with using a very similar syntax `--dataset.repo_id=repo/id`.
 
 By default, every field takes its default value specified in the dataclass. If a field doesn't have a default value, it needs to be specified either from the command line or from a config file – which path is also given in the command line (more in this below). In the example above, the `dataset` field doesn't have a default value which means it must be specified.
 
 
 ## Specifying values from the CLI
 
-Let's say that we want to train [Diffusion Policy](../lerobot/common/policies/diffusion) on the [pusht](https://huggingface.co/datasets/lerobot/pusht) dataset, using the [gym_pusht](https://github.com/huggingface/gym-pusht) environment for evaluation. The command to do so would look like this:
+Let's say that we want to train [Diffusion Policy](../../lerobot/common/policies/diffusion) on the [pusht](https://huggingface.co/datasets/lerobot/pusht) dataset, using the [gym_pusht](https://github.com/huggingface/gym-pusht) environment for evaluation. The command to do so would look like this:
 ```bash
 python lerobot/scripts/train.py \
     --dataset.repo_id=lerobot/pusht \
@@ -60,10 +60,10 @@ python lerobot/scripts/train.py \
 
 Let's break this down:
 - To specify the dataset, we just need to specify its `repo_id` on the hub which is the only required argument in the `DatasetConfig`. The rest of the fields have default values and in this case we are fine with those so we can just add the option `--dataset.repo_id=lerobot/pusht`.
-- To specify the policy, we can just select diffusion policy using `--policy` appended with `.type`. Here, `.type` is a special argument which allows us to select config classes inheriting from `draccus.ChoiceRegistry` and that have been decorated with the `register_subclass()` method. To have a better explanation of this feature, have a look at this [Draccus demo](https://github.com/dlwh/draccus?tab=readme-ov-file#more-flexible-configuration-with-choice-types). In our code, we use this mechanism mainly to select policies, environments, robots, and some other components like optimizers. The policies available to select are located in [lerobot/common/policies](../lerobot/common/policies)
-- Similarly, we select the environment with `--env.type=pusht`. The different environment configs are available in [`lerobot/common/envs/configs.py`](../lerobot/common/envs/configs.py)
+- To specify the policy, we can just select diffusion policy using `--policy` appended with `.type`. Here, `.type` is a special argument which allows us to select config classes inheriting from `draccus.ChoiceRegistry` and that have been decorated with the `register_subclass()` method. To have a better explanation of this feature, have a look at this [Draccus demo](https://github.com/dlwh/draccus?tab=readme-ov-file#more-flexible-configuration-with-choice-types). In our code, we use this mechanism mainly to select policies, environments, robots, and some other components like optimizers. The policies available to select are located in [lerobot/common/policies](../../lerobot/common/policies)
+- Similarly, we select the environment with `--env.type=pusht`. The different environment configs are available in [`lerobot/common/envs/configs.py`](../../lerobot/common/envs/configs.py)
 
-Let's see another example. Let's say you've been training [ACT](../lerobot/common/policies/act) on [lerobot/aloha_sim_insertion_human](https://huggingface.co/datasets/lerobot/aloha_sim_insertion_human) using the [gym-aloha](https://github.com/huggingface/gym-aloha) environment for evaluation with:
+Let's see another example. Let's say you've been training [ACT](../../lerobot/common/policies/act) on [lerobot/aloha_sim_insertion_human](https://huggingface.co/datasets/lerobot/aloha_sim_insertion_human) using the [gym-aloha](https://github.com/huggingface/gym-aloha) environment for evaluation with:
 ```bash
 python lerobot/scripts/train.py \
     --policy.type=act \
@@ -74,7 +74,7 @@ python lerobot/scripts/train.py \
 > Notice we added `--output_dir` to explicitly tell where to write outputs from this run (checkpoints, training state, configs etc.). This is not mandatory and if you don't specify it, a default directory will be created from the current date and time, env.type and policy.type. This will typically look like `outputs/train/2025-01-24/16-10-05_aloha_act`.
 
 We now want to train a different policy for aloha on another task. We'll change the dataset and use [lerobot/aloha_sim_transfer_cube_human](https://huggingface.co/datasets/lerobot/aloha_sim_transfer_cube_human) instead. Of course, we also need to change the task of the environment as well to match this other task.
-Looking at the [`AlohaEnv`](../lerobot/common/envs/configs.py) config, the task is `"AlohaInsertion-v0"` by default, which corresponds to the task we trained on in the command above. The [gym-aloha](https://github.com/huggingface/gym-aloha?tab=readme-ov-file#description) environment also has the `AlohaTransferCube-v0` task which corresponds to this other task we want to train on. Putting this together, we can train this new policy on this different task using:
+Looking at the [`AlohaEnv`](../../lerobot/common/envs/configs.py) config, the task is `"AlohaInsertion-v0"` by default, which corresponds to the task we trained on in the command above. The [gym-aloha](https://github.com/huggingface/gym-aloha?tab=readme-ov-file#description) environment also has the `AlohaTransferCube-v0` task which corresponds to this other task we want to train on. Putting this together, we can train this new policy on this different task using:
 ```bash
 python lerobot/scripts/train.py \
     --policy.type=act \
@@ -135,7 +135,7 @@ will start a training run with the same configuration used for training [lerobot
 
 ## Resume training
 
-Being able to resume a training run is important in case it crashed or aborted for any reason. We'll demonstrate how to do that here.
+Being able to resume a training run is important in case it crashed or aborted for any reason. We'll demonstrate how to that here.
 
 Let's reuse the command from the previous run and add a few more options:
 ```bash

examples/7_get_started_with_real_robot.md

@@ -33,7 +33,7 @@ First, install the additional dependencies required for robots built with dynami
 
 Using `pip`:
 ```bash
-pip install -e ".[dynamixel]"
+pip install --no-binary=av -e ".[dynamixel]"
 ```
 
 Using `poetry`:
@@ -55,9 +55,6 @@ Finally, connect both arms to your computer via USB. Note that the USB doesn't p
 Now you are ready to configure your motors for the first time, as detailed in the sections below. In the upcoming sections, you'll learn about our classes and functions by running some python code in an interactive session, or by copy-pasting it in a python file.
 
 If you have already configured your motors the first time, you can streamline the process by directly running the teleoperate script (which is detailed further in the tutorial):
-
-> **NOTE:** To visualize the data, enable `--control.display_data=true`. This streams the data using `rerun`.
-
 ```bash
 python lerobot/scripts/control_robot.py \
   --robot.type=koch \
@@ -377,7 +374,7 @@ robot = ManipulatorRobot(robot_config)
 
 The `KochRobotConfig` is used to set the associated settings and calibration process. For instance, we activate the torque of the gripper of the leader Koch v1.1 arm and position it at a 40 degree angle to use it as a trigger.
 
-For the [Aloha bimanual robot](https://aloha-2.github.io), we would use `AlohaRobotConfig` to set different settings such as a secondary ID for shadow joints (shoulder, elbow). Specific to Aloha, LeRobot comes with default calibration files stored in `.cache/calibration/aloha_default`. Assuming the motors have been properly assembled, no manual calibration step is expected for Aloha.
+For the [Aloha bimanual robot](https://aloha-2.github.io), we would use `AlohaRobotConfig` to set different settings such as a secondary ID for shadow joints (shoulder, elbow). Specific to Aloha, LeRobot comes with default calibration files stored in in `.cache/calibration/aloha_default`. Assuming the motors have been properly assembled, no manual calibration step is expected for Aloha.
 
 **Calibrate and Connect the ManipulatorRobot**
 
@@ -399,7 +396,7 @@ And here are the corresponding positions for the leader arm:
 
 You can watch a [video tutorial of the calibration procedure](https://youtu.be/8drnU9uRY24) for more details.
 
-During calibration, we count the number of full 360-degree rotations your motors have made since they were first used. That's why we ask you to move to this arbitrary "zero" position. We don't actually "set" the zero position, so you don't need to be accurate. After calculating these "offsets" to shift the motor values around 0, we need to assess the rotation direction of each motor, which might differ. That's why we ask you to rotate all motors to roughly 90 degrees, to measure if the values changed negatively or positively.
+During calibration, we count the number of full 360-degree rotations your motors have made since they were first used. That's why we ask yo to move to this arbitrary "zero" position. We don't actually "set" the zero position, so you don't need to be accurate. After calculating these "offsets" to shift the motor values around 0, we need to assess the rotation direction of each motor, which might differ. That's why we ask you to rotate all motors to roughly 90 degrees, to measure if the values changed negatively or positively.
 
 Finally, the rest position ensures that the follower and leader arms are roughly aligned after calibration, preventing sudden movements that could damage the motors when starting teleoperation.
 
@@ -622,7 +619,7 @@ camera_01_frame_000047.png
 
 Note: Some cameras may take a few seconds to warm up, and the first frame might be black or green.
 
-Finally, run this code to instantiate and connect your camera:
+Finally, run this code to instantiate and connectyour camera:
 ```python
 from lerobot.common.robot_devices.cameras.configs import OpenCVCameraConfig
 from lerobot.common.robot_devices.cameras.opencv import OpenCVCamera
@@ -830,6 +827,11 @@ It contains:
 - `dtRphone:33.84 (29.5hz)` which is the delta time of capturing an image from the phone camera in the thread running asynchronously.
 
 Troubleshooting:
+- On Linux, if you encounter any issue during video encoding with `ffmpeg: unknown encoder libsvtav1`, you can:
+  - install with conda-forge by running `conda install -c conda-forge ffmpeg` (it should be compiled with `libsvtav1`),
+  - or, install [Homebrew](https://brew.sh) and run `brew install ffmpeg` (it should be compiled with `libsvtav1`),
+  - or, 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`.
 - On Linux, if the left and right arrow keys and escape key don't have any effect during data recording, make sure you've set the `$DISPLAY` environment variable. See [pynput limitations](https://pynput.readthedocs.io/en/latest/limitations.html#linux).
 
 At the end of data recording, your dataset will be uploaded on your Hugging Face page (e.g. https://huggingface.co/datasets/cadene/koch_test) that you can obtain by running:

examples/8_use_stretch.md

@@ -43,19 +43,14 @@ conda create -y -n lerobot python=3.10 && conda activate lerobot
 git clone https://github.com/huggingface/lerobot.git ~/lerobot
 ```
 
-6. When using `miniconda`, install `ffmpeg` in your environment:
+6. Install LeRobot with stretch dependencies:
 ```bash
-conda install ffmpeg -c conda-forge
-```
-
-7. Install LeRobot with stretch dependencies:
-```bash
-cd ~/lerobot && pip install -e ".[stretch]"
+cd ~/lerobot && pip install --no-binary=av -e ".[stretch]"
 ```
 
 > **Note:** If you get this message, you can ignore it: `ERROR: pip's dependency resolver does not currently take into account all the packages that are installed.`
 
-8. Run a [system check](https://docs.hello-robot.com/0.3/getting_started/stretch_hardware_overview/#system-check) to make sure your robot is ready:
+7. Run a [system check](https://docs.hello-robot.com/0.3/getting_started/stretch_hardware_overview/#system-check) to make sure your robot is ready:
 ```bash
 stretch_system_check.py
 ```
@@ -99,11 +94,9 @@ This is equivalent to running `stretch_robot_home.py`
 > **Note:** If you run any of the LeRobot scripts below and Stretch is not properly homed, it will automatically home/calibrate first.
 
 **Teleoperate**
-Before trying teleoperation, you need to activate the gamepad controller by pressing the middle button. For more info, see Stretch's [doc](https://docs.hello-robot.com/0.3/getting_started/hello_robot/#gamepad-teleoperation).
+Before trying teleoperation, you need activate the gamepad controller by pressing the middle button. For more info, see Stretch's [doc](https://docs.hello-robot.com/0.3/getting_started/hello_robot/#gamepad-teleoperation).
 
 Now try out teleoperation (see above documentation to learn about the gamepad controls):
-
-> **NOTE:** To visualize the data, enable `--control.display_data=true`. This streams the data using `rerun`.
 ```bash
 python lerobot/scripts/control_robot.py \
     --robot.type=stretch \

examples/9_use_aloha.md

@@ -30,14 +30,9 @@ conda create -y -n lerobot python=3.10 && conda activate lerobot
 git clone https://github.com/huggingface/lerobot.git ~/lerobot
 ```
 
-5. When using `miniconda`, install `ffmpeg` in your environment:
+5. Install LeRobot with dependencies for the Aloha motors (dynamixel) and cameras (intelrealsense):
 ```bash
-conda install ffmpeg -c conda-forge
-```
-
-6. Install LeRobot with dependencies for the Aloha motors (dynamixel) and cameras (intelrealsense):
-```bash
-cd ~/lerobot && pip install -e ".[dynamixel, intelrealsense]"
+cd ~/lerobot && pip install --no-binary=av -e ".[dynamixel, intelrealsense]"
 ```
 
 ## Teleoperate
@@ -48,9 +43,6 @@ Teleoperation consists in manually operating the leader arms to move the followe
 2. Our code assumes that your robot has been assembled following Trossen Robotics instructions. This allows us to skip calibration, as we use the pre-defined calibration files in `.cache/calibration/aloha_default`. If you replace a motor, make sure you follow the exact instructions from Trossen Robotics.
 
 By running the following code, you can start your first **SAFE** teleoperation:
-
-> **NOTE:** To visualize the data, enable `--control.display_data=true`. This streams the data using `rerun`.
-
 ```bash
 python lerobot/scripts/control_robot.py \
   --robot.type=aloha \
@@ -142,7 +134,7 @@ python lerobot/scripts/train.py \
 
 Let's explain it:
 1. We provided the dataset as argument with `--dataset.repo_id=${HF_USER}/aloha_test`.
-2. We provided the policy with `policy.type=act`. This loads configurations from [`configuration_act.py`](../lerobot/common/policies/act/configuration_act.py). Importantly, this policy will automatically adapt to the number of motor states, motor actions and cameras of your robot (e.g. `laptop` and `phone`) which have been saved in your dataset.
+2. We provided the policy with `policy.type=act`. This loads configurations from [`configuration_act.py`](../lerobot/common/policies/act/configuration_act.py). Importantly, this policy will automatically adapt to the number of motor sates, motor actions and cameras of your robot (e.g. `laptop` and `phone`) which have been saved in your dataset.
 4. We provided `policy.device=cuda` since we are training on a Nvidia GPU, but you could use `policy.device=mps` to train on Apple silicon.
 5. We provided `wandb.enable=true` to use [Weights and Biases](https://docs.wandb.ai/quickstart) for visualizing training plots. This is optional but if you use it, make sure you are logged in by running `wandb login`.
 

examples/advanced/2_calculate_validation_loss.py

@@ -26,7 +26,10 @@ import math
 
 import torch
 
-from lerobot.common.datasets.lerobot_dataset import LeRobotDataset, LeRobotDatasetMetadata
+from lerobot.common.datasets.lerobot_dataset import (
+    LeRobotDataset,
+    LeRobotDatasetMetadata,
+)
 from lerobot.common.policies.diffusion.modeling_diffusion import DiffusionPolicy
 
 
@@ -51,7 +54,24 @@ def main():
         # Load the previous action (-0.1), the next action to be executed (0.0),
         # and 14 future actions with a 0.1 seconds spacing. All these actions will be
         # used to calculate the loss.
-        "action": [-0.1, 0.0, 0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8, 0.9, 1.0, 1.1, 1.2, 1.3, 1.4],
+        "action": [
+            -0.1,
+            0.0,
+            0.1,
+            0.2,
+            0.3,
+            0.4,
+            0.5,
+            0.6,
+            0.7,
+            0.8,
+            0.9,
+            1.0,
+            1.1,
+            1.2,
+            1.3,
+            1.4,
+        ],
     }
 
     # Load the last 10% of episodes of the dataset as a validation set.
@@ -66,7 +86,7 @@ def main():
     print(f"Number of episodes in full dataset: {total_episodes}")
     print(f"Number of episodes in training dataset (90% subset): {len(train_episodes)}")
     print(f"Number of episodes in validation dataset (10% subset): {len(val_episodes)}")
-    # - Load train and val datasets
+    # - Load train an val datasets
     train_dataset = LeRobotDataset(
         "lerobot/pusht", episodes=train_episodes, delta_timestamps=delta_timestamps
     )

lerobot/__init__.py

@@ -168,7 +168,12 @@ available_datasets = sorted(
 )
 
 # lists all available policies from `lerobot/common/policies`
-available_policies = ["act", "diffusion", "tdmpc", "vqbet"]
+available_policies = [
+    "act",
+    "diffusion",
+    "tdmpc",
+    "vqbet",
+]
 
 # lists all available robots from `lerobot/common/robot_devices/robots`
 available_robots = [
@@ -176,7 +181,6 @@ available_robots = [
     "koch_bimanual",
     "aloha",
     "so100",
-    "so101",
     "moss",
 ]
 

lerobot/common/datasets/compute_stats.py

@@ -19,7 +19,10 @@ from lerobot.common.datasets.utils import load_image_as_numpy
 
 
 def estimate_num_samples(
-    dataset_len: int, min_num_samples: int = 100, max_num_samples: int = 10_000, power: float = 0.75
+    dataset_len: int,
+    min_num_samples: int = 100,
+    max_num_samples: int = 10_000,
+    power: float = 0.75,
 ) -> int:
     """Heuristic to estimate the number of samples based on dataset size.
     The power controls the sample growth relative to dataset size.
@@ -123,7 +126,9 @@ def _assert_type_and_shape(stats_list: list[dict[str, dict]]):
                     raise ValueError(f"Shape of '{k}' must be (3,1,1), but is {v.shape} instead.")
 
 
-def aggregate_feature_stats(stats_ft_list: list[dict[str, dict]]) -> dict[str, dict[str, np.ndarray]]:
+def aggregate_feature_stats(
+    stats_ft_list: list[dict[str, dict]],
+) -> dict[str, dict[str, np.ndarray]]:
     """Aggregates stats for a single feature."""
     means = np.stack([s["mean"] for s in stats_ft_list])
     variances = np.stack([s["std"] ** 2 for s in stats_ft_list])
@@ -152,7 +157,9 @@ def aggregate_feature_stats(stats_ft_list: list[dict[str, dict]]) -> dict[str, d
     }
 
 
-def aggregate_stats(stats_list: list[dict[str, dict]]) -> dict[str, dict[str, np.ndarray]]:
+def aggregate_stats(
+    stats_list: list[dict[str, dict]],
+) -> dict[str, dict[str, np.ndarray]]:
     """Aggregate stats from multiple compute_stats outputs into a single set of stats.
 
     The final stats will have the union of all data keys from each of the stats dicts.

lerobot/common/datasets/factory.py

@@ -49,7 +49,7 @@ def resolve_delta_timestamps(
                 "observation.state": [-0.04, -0.02, 0]
                 "observation.action": [-0.02, 0, 0.02]
             }
-            returns `None` if the resulting dict is empty.
+            returns `None` if the the resulting dict is empty.
     """
     delta_timestamps = {}
     for key in ds_meta.features:

lerobot/common/datasets/image_writer.py

@@ -106,7 +106,7 @@ def worker_process(queue: queue.Queue, num_threads: int):
 class AsyncImageWriter:
     """
     This class abstract away the initialisation of processes or/and threads to
-    save images on disk asynchronously, which is critical to control a robot and record data
+    save images on disk asynchrounously, which is critical to control a robot and record data
     at a high frame rate.
 
     When `num_processes=0`, it creates a threads pool of size `num_threads`.

lerobot/common/datasets/lerobot_dataset.py

@@ -318,7 +318,7 @@ class LeRobotDatasetMetadata:
         obj.root.mkdir(parents=True, exist_ok=False)
 
         if robot is not None:
-            features = get_features_from_robot(robot, use_videos)
+            features = {**(features or {}), **get_features_from_robot(robot)}
             robot_type = robot.robot_type
             if not all(cam.fps == fps for cam in robot.cameras.values()):
                 logging.warning(
@@ -821,7 +821,9 @@ class LeRobotDataset(torch.utils.data.Dataset):
 
             if 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
+                    episode_index=self.episode_buffer["episode_index"],
+                    image_key=key,
+                    frame_index=frame_index,
                 )
                 if frame_index == 0:
                     img_path.parent.mkdir(parents=True, exist_ok=True)
@@ -867,7 +869,10 @@ class LeRobotDataset(torch.utils.data.Dataset):
         for key, ft in self.features.items():
             # index, episode_index, task_index are already processed above, and image and video
             # are processed separately by storing image path and frame info as meta data
-            if key in ["index", "episode_index", "task_index"] or ft["dtype"] in ["image", "video"]:
+            if key in ["index", "episode_index", "task_index"] or ft["dtype"] in [
+                "image",
+                "video",
+            ]:
                 continue
             episode_buffer[key] = np.stack(episode_buffer[key])
 
@@ -944,7 +949,7 @@ class LeRobotDataset(torch.utils.data.Dataset):
     def stop_image_writer(self) -> None:
         """
         Whenever wrapping this dataset inside a parallelized DataLoader, this needs to be called first to
-        remove the image_writer in order for the LeRobotDataset object to be picklable and parallelized.
+        remove the image_writer in order for the LeRobotDataset object to be pickleable and parallelized.
         """
         if self.image_writer is not None:
             self.image_writer.stop()
@@ -1053,7 +1058,7 @@ class MultiLeRobotDataset(torch.utils.data.Dataset):
         super().__init__()
         self.repo_ids = repo_ids
         self.root = Path(root) if root else HF_LEROBOT_HOME
-        self.tolerances_s = tolerances_s if tolerances_s else dict.fromkeys(repo_ids, 0.0001)
+        self.tolerances_s = tolerances_s if tolerances_s else {repo_id: 1e-4 for repo_id in repo_ids}
         # Construct the underlying datasets passing everything but `transform` and `delta_timestamps` which
         # are handled by this class.
         self._datasets = [

lerobot/common/datasets/online_buffer.py

@@ -154,14 +154,32 @@ class OnlineBuffer(torch.utils.data.Dataset):
             OnlineBuffer.NEXT_INDEX_KEY: {"dtype": np.dtype("int64"), "shape": ()},
             # Since the memmap is initialized with all-zeros, this keeps track of which indices are occupied
             # with real data rather than the dummy initialization.
-            OnlineBuffer.OCCUPANCY_MASK_KEY: {"dtype": np.dtype("?"), "shape": (buffer_capacity,)},
-            OnlineBuffer.INDEX_KEY: {"dtype": np.dtype("int64"), "shape": (buffer_capacity,)},
-            OnlineBuffer.FRAME_INDEX_KEY: {"dtype": np.dtype("int64"), "shape": (buffer_capacity,)},
-            OnlineBuffer.EPISODE_INDEX_KEY: {"dtype": np.dtype("int64"), "shape": (buffer_capacity,)},
-            OnlineBuffer.TIMESTAMP_KEY: {"dtype": np.dtype("float64"), "shape": (buffer_capacity,)},
+            OnlineBuffer.OCCUPANCY_MASK_KEY: {
+                "dtype": np.dtype("?"),
+                "shape": (buffer_capacity,),
+            },
+            OnlineBuffer.INDEX_KEY: {
+                "dtype": np.dtype("int64"),
+                "shape": (buffer_capacity,),
+            },
+            OnlineBuffer.FRAME_INDEX_KEY: {
+                "dtype": np.dtype("int64"),
+                "shape": (buffer_capacity,),
+            },
+            OnlineBuffer.EPISODE_INDEX_KEY: {
+                "dtype": np.dtype("int64"),
+                "shape": (buffer_capacity,),
+            },
+            OnlineBuffer.TIMESTAMP_KEY: {
+                "dtype": np.dtype("float64"),
+                "shape": (buffer_capacity,),
+            },
         }
         for k, v in data_spec.items():
-            complete_data_spec[k] = {"dtype": v["dtype"], "shape": (buffer_capacity, *v["shape"])}
+            complete_data_spec[k] = {
+                "dtype": v["dtype"],
+                "shape": (buffer_capacity, *v["shape"]),
+            }
         return complete_data_spec
 
     def add_data(self, data: dict[str, np.ndarray]):

lerobot/common/datasets/push_dataset_to_hub/utils.py

@@ -77,7 +77,9 @@ def check_repo_id(repo_id: str) -> None:
 
 
 # TODO(aliberts): remove
-def calculate_episode_data_index(hf_dataset: datasets.Dataset) -> Dict[str, torch.Tensor]:
+def calculate_episode_data_index(
+    hf_dataset: datasets.Dataset,
+) -> Dict[str, torch.Tensor]:
     """
     Calculate episode data index for the provided HuggingFace Dataset. Relies on episode_index column of hf_dataset.
 

lerobot/common/datasets/sampler.py

@@ -43,7 +43,10 @@ class EpisodeAwareSampler:
         ):
             if episode_indices_to_use is None or episode_idx in episode_indices_to_use:
                 indices.extend(
-                    range(start_index.item() + drop_n_first_frames, end_index.item() - drop_n_last_frames)
+                    range(
+                        start_index.item() + drop_n_first_frames,
+                        end_index.item() - drop_n_last_frames,
+                    )
                 )
 
         self.indices = indices

lerobot/common/datasets/transforms.py

@@ -128,7 +128,7 @@ class SharpnessJitter(Transform):
             raise TypeError(f"{sharpness=} should be a single number or a sequence with length 2.")
 
         if not 0.0 <= sharpness[0] <= sharpness[1]:
-            raise ValueError(f"sharpness values should be between (0., inf), but got {sharpness}.")
+            raise ValueError(f"sharpnesss values should be between (0., inf), but got {sharpness}.")
 
         return float(sharpness[0]), float(sharpness[1])
 

lerobot/common/datasets/utils.py

@@ -225,7 +225,10 @@ def load_episodes(local_dir: Path) -> dict:
 def write_episode_stats(episode_index: int, episode_stats: dict, local_dir: Path):
     # We wrap episode_stats in a dictionary since `episode_stats["episode_index"]`
     # is a dictionary of stats and not an integer.
-    episode_stats = {"episode_index": episode_index, "stats": serialize_dict(episode_stats)}
+    episode_stats = {
+        "episode_index": episode_index,
+        "stats": serialize_dict(episode_stats),
+    }
     append_jsonlines(episode_stats, local_dir / EPISODES_STATS_PATH)
 
 
@@ -240,7 +243,7 @@ def load_episodes_stats(local_dir: Path) -> dict:
 def backward_compatible_episodes_stats(
     stats: dict[str, dict[str, np.ndarray]], episodes: list[int]
 ) -> dict[str, dict[str, np.ndarray]]:
-    return dict.fromkeys(episodes, stats)
+    return {ep_idx: stats for ep_idx in episodes}
 
 
 def load_image_as_numpy(
@@ -409,7 +412,7 @@ def dataset_to_policy_features(features: dict[str, dict]) -> dict[str, PolicyFea
 
             names = ft["names"]
             # Backward compatibility for "channel" which is an error introduced in LeRobotDataset v2.0 for ported datasets.
-            if names[2] in ["channel", "channels"]:  # (h, w, c) -> (c, h, w)
+            if names is not None and names[2] in ["channel", "channels"]:  # (h, w, c) -> (c, h, w)
                 shape = (shape[2], shape[0], shape[1])
         elif key == "observation.environment_state":
             type = FeatureType.ENV
@@ -540,7 +543,10 @@ def check_timestamps_sync(
 
 
 def check_delta_timestamps(
-    delta_timestamps: dict[str, list[float]], fps: int, tolerance_s: float, raise_value_error: bool = True
+    delta_timestamps: dict[str, list[float]],
+    fps: int,
+    tolerance_s: float,
+    raise_value_error: bool = True,
 ) -> bool:
     """This will check if all the values in delta_timestamps are multiples of 1/fps +/- tolerance.
     This is to ensure that these delta_timestamps added to any timestamp from a dataset will themselves be

lerobot/common/datasets/v2/batch_convert_dataset_v1_to_v2.py

@@ -118,7 +118,10 @@ DATASETS = {
         "single_task": "Place the battery into the slot of the remote controller.",
         **ALOHA_STATIC_INFO,
     },
-    "aloha_static_candy": {"single_task": "Pick up the candy and unwrap it.", **ALOHA_STATIC_INFO},
+    "aloha_static_candy": {
+        "single_task": "Pick up the candy and unwrap it.",
+        **ALOHA_STATIC_INFO,
+    },
     "aloha_static_coffee": {
         "single_task": "Place the coffee capsule inside the capsule container, then place the cup onto the center of the cup tray, then push the 'Hot Water' and 'Travel Mug' buttons.",
         **ALOHA_STATIC_INFO,
@@ -167,13 +170,22 @@ DATASETS = {
         "single_task": "Pick up the plastic cup with the left arm, then pop its lid open with the right arm.",
         **ALOHA_STATIC_INFO,
     },
-    "aloha_static_ziploc_slide": {"single_task": "Slide open the ziploc bag.", **ALOHA_STATIC_INFO},
-    "aloha_sim_insertion_scripted": {"single_task": "Insert the peg into the socket.", **ALOHA_STATIC_INFO},
+    "aloha_static_ziploc_slide": {
+        "single_task": "Slide open the ziploc bag.",
+        **ALOHA_STATIC_INFO,
+    },
+    "aloha_sim_insertion_scripted": {
+        "single_task": "Insert the peg into the socket.",
+        **ALOHA_STATIC_INFO,
+    },
     "aloha_sim_insertion_scripted_image": {
         "single_task": "Insert the peg into the socket.",
         **ALOHA_STATIC_INFO,
     },
-    "aloha_sim_insertion_human": {"single_task": "Insert the peg into the socket.", **ALOHA_STATIC_INFO},
+    "aloha_sim_insertion_human": {
+        "single_task": "Insert the peg into the socket.",
+        **ALOHA_STATIC_INFO,
+    },
     "aloha_sim_insertion_human_image": {
         "single_task": "Insert the peg into the socket.",
         **ALOHA_STATIC_INFO,
@@ -194,10 +206,19 @@ DATASETS = {
         "single_task": "Pick up the cube with the right arm and transfer it to the left arm.",
         **ALOHA_STATIC_INFO,
     },
-    "pusht": {"single_task": "Push the T-shaped block onto the T-shaped target.", **PUSHT_INFO},
-    "pusht_image": {"single_task": "Push the T-shaped block onto the T-shaped target.", **PUSHT_INFO},
+    "pusht": {
+        "single_task": "Push the T-shaped block onto the T-shaped target.",
+        **PUSHT_INFO,
+    },
+    "pusht_image": {
+        "single_task": "Push the T-shaped block onto the T-shaped target.",
+        **PUSHT_INFO,
+    },
     "unitreeh1_fold_clothes": {"single_task": "Fold the sweatshirt.", **UNITREEH_INFO},
-    "unitreeh1_rearrange_objects": {"single_task": "Put the object into the bin.", **UNITREEH_INFO},
+    "unitreeh1_rearrange_objects": {
+        "single_task": "Put the object into the bin.",
+        **UNITREEH_INFO,
+    },
     "unitreeh1_two_robot_greeting": {
         "single_task": "Greet the other robot with a high five.",
         **UNITREEH_INFO,
@@ -207,13 +228,31 @@ DATASETS = {
         **UNITREEH_INFO,
     },
     "xarm_lift_medium": {"single_task": "Pick up the cube and lift it.", **XARM_INFO},
-    "xarm_lift_medium_image": {"single_task": "Pick up the cube and lift it.", **XARM_INFO},
-    "xarm_lift_medium_replay": {"single_task": "Pick up the cube and lift it.", **XARM_INFO},
-    "xarm_lift_medium_replay_image": {"single_task": "Pick up the cube and lift it.", **XARM_INFO},
+    "xarm_lift_medium_image": {
+        "single_task": "Pick up the cube and lift it.",
+        **XARM_INFO,
+    },
+    "xarm_lift_medium_replay": {
+        "single_task": "Pick up the cube and lift it.",
+        **XARM_INFO,
+    },
+    "xarm_lift_medium_replay_image": {
+        "single_task": "Pick up the cube and lift it.",
+        **XARM_INFO,
+    },
     "xarm_push_medium": {"single_task": "Push the cube onto the target.", **XARM_INFO},
-    "xarm_push_medium_image": {"single_task": "Push the cube onto the target.", **XARM_INFO},
-    "xarm_push_medium_replay": {"single_task": "Push the cube onto the target.", **XARM_INFO},
-    "xarm_push_medium_replay_image": {"single_task": "Push the cube onto the target.", **XARM_INFO},
+    "xarm_push_medium_image": {
+        "single_task": "Push the cube onto the target.",
+        **XARM_INFO,
+    },
+    "xarm_push_medium_replay": {
+        "single_task": "Push the cube onto the target.",
+        **XARM_INFO,
+    },
+    "xarm_push_medium_replay_image": {
+        "single_task": "Push the cube onto the target.",
+        **XARM_INFO,
+    },
     "umi_cup_in_the_wild": {
         "single_task": "Put the cup on the plate.",
         "license": "apache-2.0",

lerobot/common/datasets/v2/convert_dataset_v1_to_v2.py

@@ -379,7 +379,12 @@ def fix_lfs_video_files_tracking(work_dir: Path, lfs_untracked_videos: list[str]
     for i in range(0, len(lfs_untracked_videos), 100):
         files = lfs_untracked_videos[i : i + 100]
         try:
-            subprocess.run(["git", "rm", "--cached", *files], cwd=work_dir, capture_output=True, check=True)
+            subprocess.run(
+                ["git", "rm", "--cached", *files],
+                cwd=work_dir,
+                capture_output=True,
+                check=True,
+            )
         except subprocess.CalledProcessError as e:
             print("git rm --cached ERROR:")
             print(e.stderr)
@@ -402,7 +407,17 @@ def _lfs_clone(repo_id: str, work_dir: Path, branch: str) -> None:
     repo_url = f"https://huggingface.co/datasets/{repo_id}"
     env = {"GIT_LFS_SKIP_SMUDGE": "1"}  # Prevent downloading LFS files
     subprocess.run(
-        ["git", "clone", "--branch", branch, "--single-branch", "--depth", "1", repo_url, str(work_dir)],
+        [
+            "git",
+            "clone",
+            "--branch",
+            branch,
+            "--single-branch",
+            "--depth",
+            "1",
+            repo_url,
+            str(work_dir),
+        ],
         check=True,
         env=env,
     )
@@ -410,7 +425,11 @@ def _lfs_clone(repo_id: str, work_dir: Path, branch: str) -> None:
 
 def _get_lfs_untracked_videos(work_dir: Path, video_files: list[str]) -> list[str]:
     lfs_tracked_files = subprocess.run(
-        ["git", "lfs", "ls-files", "-n"], cwd=work_dir, capture_output=True, text=True, check=True
+        ["git", "lfs", "ls-files", "-n"],
+        cwd=work_dir,
+        capture_output=True,
+        text=True,
+        check=True,
     )
     lfs_tracked_files = set(lfs_tracked_files.stdout.splitlines())
     return [f for f in video_files if f not in lfs_tracked_files]
@@ -424,7 +443,11 @@ def get_videos_info(repo_id: str, local_dir: Path, video_keys: list[str], branch
     ]
     hub_api = HfApi()
     hub_api.snapshot_download(
-        repo_id=repo_id, repo_type="dataset", local_dir=local_dir, revision=branch, allow_patterns=video_files
+        repo_id=repo_id,
+        repo_type="dataset",
+        local_dir=local_dir,
+        revision=branch,
+        allow_patterns=video_files,
     )
     videos_info_dict = {}
     for vid_key, vid_path in zip(video_keys, video_files, strict=True):
@@ -451,7 +474,11 @@ def convert_dataset(
 
     hub_api = HfApi()
     hub_api.snapshot_download(
-        repo_id=repo_id, repo_type="dataset", revision=v1, local_dir=v1x_dir, ignore_patterns="videos*/"
+        repo_id=repo_id,
+        repo_type="dataset",
+        revision=v1,
+        local_dir=v1x_dir,
+        ignore_patterns="videos*/",
     )
     branch = "main"
     if test_branch:
@@ -481,7 +508,7 @@ def convert_dataset(
 
     # Tasks
     if single_task:
-        tasks_by_episodes = dict.fromkeys(episode_indices, single_task)
+        tasks_by_episodes = {ep_idx: single_task for ep_idx in episode_indices}
         dataset, tasks = add_task_index_by_episodes(dataset, tasks_by_episodes)
         tasks_by_episodes = {ep_idx: [task] for ep_idx, task in tasks_by_episodes.items()}
     elif tasks_path:
@@ -509,12 +536,21 @@ def convert_dataset(
         dataset = dataset.remove_columns(video_keys)
         clean_gitattr = Path(
             hub_api.hf_hub_download(
-                repo_id=GITATTRIBUTES_REF, repo_type="dataset", local_dir=local_dir, filename=".gitattributes"
+                repo_id=GITATTRIBUTES_REF,
+                repo_type="dataset",
+                local_dir=local_dir,
+                filename=".gitattributes",
             )
         ).absolute()
         with tempfile.TemporaryDirectory() as tmp_video_dir:
             move_videos(
-                repo_id, video_keys, total_episodes, total_chunks, Path(tmp_video_dir), clean_gitattr, branch
+                repo_id,
+                video_keys,
+                total_episodes,
+                total_chunks,
+                Path(tmp_video_dir),
+                clean_gitattr,
+                branch,
             )
         videos_info = get_videos_info(repo_id, v1x_dir, video_keys=video_keys, branch=branch)
         for key in video_keys:
@@ -543,7 +579,11 @@ def convert_dataset(
 
     # Episodes
     episodes = [
-        {"episode_index": ep_idx, "tasks": tasks_by_episodes[ep_idx], "length": episode_lengths[ep_idx]}
+        {
+            "episode_index": ep_idx,
+            "tasks": tasks_by_episodes[ep_idx],
+            "length": episode_lengths[ep_idx],
+        }
         for ep_idx in episode_indices
     ]
     write_jsonlines(episodes, v20_dir / EPISODES_PATH)
@@ -572,7 +612,12 @@ def convert_dataset(
         hub_api.delete_folder(repo_id=repo_id, path_in_repo="data", repo_type="dataset", revision=branch)
 
     with contextlib.suppress(EntryNotFoundError, HfHubHTTPError):
-        hub_api.delete_folder(repo_id=repo_id, path_in_repo="meta_data", repo_type="dataset", revision=branch)
+        hub_api.delete_folder(
+            repo_id=repo_id,
+            path_in_repo="meta_data",
+            repo_type="dataset",
+            revision=branch,
+        )
 
     with contextlib.suppress(EntryNotFoundError, HfHubHTTPError):
         hub_api.delete_folder(repo_id=repo_id, path_in_repo="meta", repo_type="dataset", revision=branch)

lerobot/common/datasets/v21/convert_dataset_v20_to_v21.py

@@ -37,8 +37,16 @@ import logging
 from huggingface_hub import HfApi
 
 from lerobot.common.datasets.lerobot_dataset import CODEBASE_VERSION, LeRobotDataset
-from lerobot.common.datasets.utils import EPISODES_STATS_PATH, STATS_PATH, load_stats, write_info
-from lerobot.common.datasets.v21.convert_stats import check_aggregate_stats, convert_stats
+from lerobot.common.datasets.utils import (
+    EPISODES_STATS_PATH,
+    STATS_PATH,
+    load_stats,
+    write_info,
+)
+from lerobot.common.datasets.v21.convert_stats import (
+    check_aggregate_stats,
+    convert_stats,
+)
 
 V20 = "v2.0"
 V21 = "v2.1"
@@ -79,10 +87,16 @@ def convert_dataset(
 
     hub_api = HfApi()
     if hub_api.file_exists(
-        repo_id=dataset.repo_id, filename=STATS_PATH, revision=branch, repo_type="dataset"
+        repo_id=dataset.repo_id,
+        filename=STATS_PATH,
+        revision=branch,
+        repo_type="dataset",
     ):
         hub_api.delete_file(
-            path_in_repo=STATS_PATH, repo_id=dataset.repo_id, revision=branch, repo_type="dataset"
+            path_in_repo=STATS_PATH,
+            repo_id=dataset.repo_id,
+            revision=branch,
+            repo_type="dataset",
         )
 
     hub_api.create_tag(repo_id, tag=CODEBASE_VERSION, revision=branch, repo_type="dataset")

lerobot/common/datasets/v21/convert_stats.py

@@ -17,7 +17,11 @@ from concurrent.futures import ThreadPoolExecutor, as_completed
 import numpy as np
 from tqdm import tqdm
 
-from lerobot.common.datasets.compute_stats import aggregate_stats, get_feature_stats, sample_indices
+from lerobot.common.datasets.compute_stats import (
+    aggregate_stats,
+    get_feature_stats,
+    sample_indices,
+)
 from lerobot.common.datasets.lerobot_dataset import LeRobotDataset
 from lerobot.common.datasets.utils import write_episode_stats
 
@@ -95,5 +99,9 @@ def check_aggregate_stats(
             if key in reference_stats and stat in reference_stats[key]:
                 err_msg = f"feature='{key}' stats='{stat}'"
                 np.testing.assert_allclose(
-                    val, reference_stats[key][stat], rtol=rtol, atol=atol, err_msg=err_msg
+                    val,
+                    reference_stats[key][stat],
+                    rtol=rtol,
+                    atol=atol,
+                    err_msg=err_msg,
                 )

lerobot/common/datasets/video_utils.py

@@ -13,15 +13,16 @@
 # 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 glob
 import importlib
+import json
 import logging
+import subprocess
 import warnings
+from collections import OrderedDict
 from dataclasses import dataclass, field
 from pathlib import Path
 from typing import Any, ClassVar
 
-import av
 import pyarrow as pa
 import torch
 import torchvision
@@ -101,7 +102,7 @@ def decode_video_frames_torchvision(
     keyframes_only = False
     torchvision.set_video_backend(backend)
     if backend == "pyav":
-        keyframes_only = True  # pyav doesn't support accurate seek
+        keyframes_only = True  # pyav doesnt support accuracte seek
 
     # set a video stream reader
     # TODO(rcadene): also load audio stream at the same time
@@ -251,83 +252,51 @@ def encode_video_frames(
     g: int | None = 2,
     crf: int | None = 30,
     fast_decode: int = 0,
-    log_level: int | None = av.logging.ERROR,
+    log_level: str | None = "error",
     overwrite: bool = False,
 ) -> None:
     """More info on ffmpeg arguments tuning on `benchmark/video/README.md`"""
-    # 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)
+    video_path.parent.mkdir(parents=True, exist_ok=True)
 
-    video_path.parent.mkdir(parents=True, exist_ok=overwrite)
-
-    # Encoders/pixel formats incompatibility check
-    if (vcodec == "libsvtav1" or vcodec == "hevc") and pix_fmt == "yuv444p":
-        logging.warning(
-            f"Incompatible pixel format 'yuv444p' for codec {vcodec}, auto-selecting format 'yuv420p'"
-        )
-        pix_fmt = "yuv420p"
-
-    # Get input frames
-    template = "frame_" + ("[0-9]" * 6) + ".png"
-    input_list = sorted(
-        glob.glob(str(imgs_dir / template)), key=lambda x: int(x.split("_")[-1].split(".")[0])
+    ffmpeg_args = OrderedDict(
+        [
+            ("-f", "image2"),
+            ("-r", str(fps)),
+            ("-i", str(imgs_dir / "frame_%06d.png")),
+            ("-vcodec", vcodec),
+            ("-pix_fmt", pix_fmt),
+        ]
     )
 
-    # Define video output frame size (assuming all input frames are the same size)
-    if len(input_list) == 0:
-        raise FileNotFoundError(f"No images found in {imgs_dir}.")
-    dummy_image = Image.open(input_list[0])
-    width, height = dummy_image.size
-
-    # Define video codec options
-    video_options = {}
-
     if g is not None:
-        video_options["g"] = str(g)
+        ffmpeg_args["-g"] = str(g)
 
     if crf is not None:
-        video_options["crf"] = str(crf)
+        ffmpeg_args["-crf"] = str(crf)
 
     if fast_decode:
-        key = "svtav1-params" if vcodec == "libsvtav1" else "tune"
+        key = "-svtav1-params" if vcodec == "libsvtav1" else "-tune"
         value = f"fast-decode={fast_decode}" if vcodec == "libsvtav1" else "fastdecode"
-        video_options[key] = value
+        ffmpeg_args[key] = value
 
-    # Set logging level
     if log_level is not None:
-        # "While less efficient, it is generally preferable to modify logging with Python’s logging"
-        logging.getLogger("libav").setLevel(log_level)
+        ffmpeg_args["-loglevel"] = str(log_level)
 
-    # Create and open output file (overwrite by default)
-    with av.open(str(video_path), "w") as output:
-        output_stream = output.add_stream(vcodec, fps, options=video_options)
-        output_stream.pix_fmt = pix_fmt
-        output_stream.width = width
-        output_stream.height = height
+    ffmpeg_args = [item for pair in ffmpeg_args.items() for item in pair]
+    if overwrite:
+        ffmpeg_args.append("-y")
 
-        # Loop through input frames and encode them
-        for input_data in input_list:
-            input_image = Image.open(input_data).convert("RGB")
-            input_frame = av.VideoFrame.from_image(input_image)
-            packet = output_stream.encode(input_frame)
-            if packet:
-                output.mux(packet)
-
-        # Flush the encoder
-        packet = output_stream.encode()
-        if packet:
-            output.mux(packet)
-
-    # Reset logging level
-    if log_level is not None:
-        av.logging.restore_default_callback()
+    ffmpeg_cmd = ["ffmpeg"] + ffmpeg_args + [str(video_path)]
+    # redirect stdin to subprocess.DEVNULL to prevent reading random keyboard inputs from terminal
+    subprocess.run(ffmpeg_cmd, check=True, stdin=subprocess.DEVNULL)
 
     if not video_path.exists():
-        raise OSError(f"Video encoding did not work. File not found: {video_path}.")
+        raise OSError(
+            f"Video encoding did not work. File not found: {video_path}. "
+            f"Try running the command manually to debug: `{''.join(ffmpeg_cmd)}`"
+        )
 
 
 @dataclass
@@ -363,68 +332,78 @@ with warnings.catch_warnings():
 
 
 def get_audio_info(video_path: Path | str) -> dict:
-    # Set logging level
-    logging.getLogger("libav").setLevel(av.logging.ERROR)
+    ffprobe_audio_cmd = [
+        "ffprobe",
+        "-v",
+        "error",
+        "-select_streams",
+        "a:0",
+        "-show_entries",
+        "stream=channels,codec_name,bit_rate,sample_rate,bit_depth,channel_layout,duration",
+        "-of",
+        "json",
+        str(video_path),
+    ]
+    result = subprocess.run(ffprobe_audio_cmd, stdout=subprocess.PIPE, stderr=subprocess.PIPE, text=True)
+    if result.returncode != 0:
+        raise RuntimeError(f"Error running ffprobe: {result.stderr}")
 
-    # Getting audio stream information
-    audio_info = {}
-    with av.open(str(video_path), "r") as audio_file:
-        try:
-            audio_stream = audio_file.streams.audio[0]
-        except IndexError:
-            # Reset logging level
-            av.logging.restore_default_callback()
-            return {"has_audio": False}
+    info = json.loads(result.stdout)
+    audio_stream_info = info["streams"][0] if info.get("streams") else None
+    if audio_stream_info is None:
+        return {"has_audio": False}
 
-        audio_info["audio.channels"] = audio_stream.channels
-        audio_info["audio.codec"] = audio_stream.codec.canonical_name
-        # In an ideal loseless case : bit depth x sample rate x channels = bit rate.
-        # In an actual compressed case, the bit rate is set according to the compression level : the lower the bit rate, the more compression is applied.
-        audio_info["audio.bit_rate"] = audio_stream.bit_rate
-        audio_info["audio.sample_rate"] = audio_stream.sample_rate  # Number of samples per second
-        # In an ideal loseless case : fixed number of bits per sample.
-        # In an actual compressed case : variable number of bits per sample (often reduced to match a given depth rate).
-        audio_info["audio.bit_depth"] = audio_stream.format.bits
-        audio_info["audio.channel_layout"] = audio_stream.layout.name
-        audio_info["has_audio"] = True
-
-    # Reset logging level
-    av.logging.restore_default_callback()
-
-    return audio_info
+    # Return the information, defaulting to None if no audio stream is present
+    return {
+        "has_audio": True,
+        "audio.channels": audio_stream_info.get("channels", None),
+        "audio.codec": audio_stream_info.get("codec_name", None),
+        "audio.bit_rate": int(audio_stream_info["bit_rate"]) if audio_stream_info.get("bit_rate") else None,
+        "audio.sample_rate": int(audio_stream_info["sample_rate"])
+        if audio_stream_info.get("sample_rate")
+        else None,
+        "audio.bit_depth": audio_stream_info.get("bit_depth", None),
+        "audio.channel_layout": audio_stream_info.get("channel_layout", None),
+    }
 
 
 def get_video_info(video_path: Path | str) -> dict:
-    # Set logging level
-    logging.getLogger("libav").setLevel(av.logging.ERROR)
+    ffprobe_video_cmd = [
+        "ffprobe",
+        "-v",
+        "error",
+        "-select_streams",
+        "v:0",
+        "-show_entries",
+        "stream=r_frame_rate,width,height,codec_name,nb_frames,duration,pix_fmt",
+        "-of",
+        "json",
+        str(video_path),
+    ]
+    result = subprocess.run(ffprobe_video_cmd, stdout=subprocess.PIPE, stderr=subprocess.PIPE, text=True)
+    if result.returncode != 0:
+        raise RuntimeError(f"Error running ffprobe: {result.stderr}")
 
-    # Getting video stream information
-    video_info = {}
-    with av.open(str(video_path), "r") as video_file:
-        try:
-            video_stream = video_file.streams.video[0]
-        except IndexError:
-            # Reset logging level
-            av.logging.restore_default_callback()
-            return {}
+    info = json.loads(result.stdout)
+    video_stream_info = info["streams"][0]
 
-        video_info["video.height"] = video_stream.height
-        video_info["video.width"] = video_stream.width
-        video_info["video.codec"] = video_stream.codec.canonical_name
-        video_info["video.pix_fmt"] = video_stream.pix_fmt
-        video_info["video.is_depth_map"] = False
+    # Calculate fps from r_frame_rate
+    r_frame_rate = video_stream_info["r_frame_rate"]
+    num, denom = map(int, r_frame_rate.split("/"))
+    fps = num / denom
 
-        # Calculate fps from r_frame_rate
-        video_info["video.fps"] = int(video_stream.base_rate)
+    pixel_channels = get_video_pixel_channels(video_stream_info["pix_fmt"])
 
-        pixel_channels = get_video_pixel_channels(video_stream.pix_fmt)
-        video_info["video.channels"] = pixel_channels
-
-    # Reset logging level
-    av.logging.restore_default_callback()
-
-    # Adding audio stream information
-    video_info.update(**get_audio_info(video_path))
+    video_info = {
+        "video.fps": fps,
+        "video.height": video_stream_info["height"],
+        "video.width": video_stream_info["width"],
+        "video.channels": pixel_channels,
+        "video.codec": video_stream_info["codec_name"],
+        "video.pix_fmt": video_stream_info["pix_fmt"],
+        "video.is_depth_map": False,
+        **get_audio_info(video_path),
+    }
 
     return video_info
 

lerobot/common/envs/configs.py

@@ -14,10 +14,12 @@
 
 import abc
 from dataclasses import dataclass, field
+from typing import Any, Dict, Optional, Tuple
 
 import draccus
 
 from lerobot.common.constants import ACTION, OBS_ENV, OBS_IMAGE, OBS_IMAGES, OBS_ROBOT
+from lerobot.common.robot_devices.robots.configs import RobotConfig
 from lerobot.configs.types import FeatureType, PolicyFeature
 
 
@@ -154,3 +156,135 @@ class XarmEnv(EnvConfig):
             "visualization_height": self.visualization_height,
             "max_episode_steps": self.episode_length,
         }
+
+
+@dataclass
+class VideoRecordConfig:
+    """Configuration for video recording in ManiSkill environments."""
+
+    enabled: bool = False
+    record_dir: str = "videos"
+    trajectory_name: str = "trajectory"
+
+
+@dataclass
+class WrapperConfig:
+    """Configuration for environment wrappers."""
+
+    joint_masking_action_space: list[bool] | None = None
+
+
+@dataclass
+class EEActionSpaceConfig:
+    """Configuration parameters for end-effector action space."""
+
+    x_step_size: float
+    y_step_size: float
+    z_step_size: float
+    bounds: Dict[str, Any]  # Contains 'min' and 'max' keys with position bounds
+    use_gamepad: bool = False
+
+
+@dataclass
+class EnvWrapperConfig:
+    """Configuration for environment wrappers."""
+
+    display_cameras: bool = False
+    use_relative_joint_positions: bool = True
+    add_joint_velocity_to_observation: bool = False
+    add_ee_pose_to_observation: bool = False
+    crop_params_dict: Optional[Dict[str, Tuple[int, int, int, int]]] = None
+    resize_size: Optional[Tuple[int, int]] = None
+    control_time_s: float = 20.0
+    fixed_reset_joint_positions: Optional[Any] = None
+    reset_time_s: float = 5.0
+    joint_masking_action_space: Optional[Any] = None
+    ee_action_space_params: Optional[EEActionSpaceConfig] = None
+    use_gripper: bool = False
+    gripper_quantization_threshold: float | None = 0.8
+    gripper_penalty: float = 0.0
+    gripper_penalty_in_reward: bool = False
+    open_gripper_on_reset: bool = False
+
+
+@EnvConfig.register_subclass(name="gym_manipulator")
+@dataclass
+class HILSerlRobotEnvConfig(EnvConfig):
+    """Configuration for the HILSerlRobotEnv environment."""
+
+    robot: Optional[RobotConfig] = None
+    wrapper: Optional[EnvWrapperConfig] = None
+    fps: int = 10
+    name: str = "real_robot"
+    mode: str = None  # Either "record", "replay", None
+    repo_id: Optional[str] = None
+    dataset_root: Optional[str] = None
+    task: str = ""
+    num_episodes: int = 10  # only for record mode
+    episode: int = 0
+    device: str = "cuda"
+    push_to_hub: bool = True
+    pretrained_policy_name_or_path: Optional[str] = None
+    reward_classifier: dict[str, str | None] = field(
+        default_factory=lambda: {
+            "pretrained_path": None,
+            "config_path": None,
+        }
+    )
+
+    def gym_kwargs(self) -> dict:
+        return {}
+
+
+@EnvConfig.register_subclass("maniskill_push")
+@dataclass
+class ManiskillEnvConfig(EnvConfig):
+    """Configuration for the ManiSkill environment."""
+
+    name: str = "maniskill/pushcube"
+    task: str = "PushCube-v1"
+    image_size: int = 64
+    control_mode: str = "pd_ee_delta_pose"
+    state_dim: int = 25
+    action_dim: int = 7
+    fps: int = 200
+    episode_length: int = 50
+    obs_type: str = "rgb"
+    render_mode: str = "rgb_array"
+    render_size: int = 64
+    device: str = "cuda"
+    robot: str = "so100"  # This is a hack to make the robot config work
+    video_record: VideoRecordConfig = field(default_factory=VideoRecordConfig)
+    wrapper: WrapperConfig = field(default_factory=WrapperConfig)
+    mock_gripper: bool = False
+    features: dict[str, PolicyFeature] = field(
+        default_factory=lambda: {
+            "action": PolicyFeature(type=FeatureType.ACTION, shape=(7,)),
+            "observation.image": PolicyFeature(type=FeatureType.VISUAL, shape=(3, 64, 64)),
+            "observation.state": PolicyFeature(type=FeatureType.STATE, shape=(25,)),
+        }
+    )
+    features_map: dict[str, str] = field(
+        default_factory=lambda: {
+            "action": ACTION,
+            "observation.image": OBS_IMAGE,
+            "observation.state": OBS_ROBOT,
+        }
+    )
+    reward_classifier: dict[str, str | None] = field(
+        default_factory=lambda: {
+            "pretrained_path": None,
+            "config_path": None,
+        }
+    )
+
+    @property
+    def gym_kwargs(self) -> dict:
+        return {
+            "obs_type": self.obs_type,
+            "render_mode": self.render_mode,
+            "max_episode_steps": self.episode_length,
+            "control_mode": self.control_mode,
+            "sensor_configs": {"width": self.image_size, "height": self.image_size},
+            "num_envs": 1,
+        }

lerobot/common/envs/utils.py

@@ -13,11 +13,7 @@
 # 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 warnings
-from typing import Any
-
 import einops
-import gymnasium as gym
 import numpy as np
 import torch
 from torch import Tensor
@@ -37,29 +33,35 @@ def preprocess_observation(observations: dict[str, np.ndarray]) -> dict[str, Ten
     """
     # map to expected inputs for the policy
     return_observations = {}
-    if "pixels" in observations:
-        if isinstance(observations["pixels"], dict):
-            imgs = {f"observation.images.{key}": img for key, img in observations["pixels"].items()}
-        else:
-            imgs = {"observation.image": observations["pixels"]}
+    # TODO: You have to merge all tensors from agent key and extra key
+    # You don't keep sensor param key in the observation
+    # And you keep sensor data rgb
+    for key, img in observations.items():
+        if "images" not in key:
+            continue
 
-        for imgkey, img in imgs.items():
-            # TODO(aliberts, rcadene): use transforms.ToTensor()?
+        # TODO(aliberts, rcadene): use transforms.ToTensor()?
+        if not torch.is_tensor(img):
             img = torch.from_numpy(img)
 
-            # sanity check that images are channel last
-            _, h, w, c = img.shape
-            assert c < h and c < w, f"expect channel last images, but instead got {img.shape=}"
+        if img.ndim == 3:
+            img = img.unsqueeze(0)
 
-            # sanity check that images are uint8
-            assert img.dtype == torch.uint8, f"expect torch.uint8, but instead {img.dtype=}"
+        # sanity check that images are channel last
+        _, h, w, c = img.shape
+        assert c < h and c < w, f"expect channel last images, but instead got {img.shape=}"
 
-            # convert to channel first of type float32 in range [0,1]
-            img = einops.rearrange(img, "b h w c -> b c h w").contiguous()
-            img = img.type(torch.float32)
-            img /= 255
+        # sanity check that images are uint8
+        assert img.dtype == torch.uint8, f"expect torch.uint8, but instead {img.dtype=}"
 
-            return_observations[imgkey] = img
+        # convert to channel first of type float32 in range [0,1]
+        img = einops.rearrange(img, "b h w c -> b c h w").contiguous()
+        img = img.type(torch.float32)
+        img /= 255
+
+        return_observations[key] = img
+        # obs state agent qpos and qvel
+        # image
 
     if "environment_state" in observations:
         return_observations["observation.environment_state"] = torch.from_numpy(
@@ -68,7 +70,8 @@ def preprocess_observation(observations: dict[str, np.ndarray]) -> dict[str, Ten
 
     # TODO(rcadene): enable pixels only baseline with `obs_type="pixels"` in environment by removing
     # requirement for "agent_pos"
-    return_observations["observation.state"] = torch.from_numpy(observations["agent_pos"]).float()
+    # return_observations["observation.state"] = torch.from_numpy(observations["agent_pos"]).float()
+    return_observations["observation.state"] = observations["observation.state"].float()
     return return_observations
 
 
@@ -86,42 +89,44 @@ def env_to_policy_features(env_cfg: EnvConfig) -> dict[str, PolicyFeature]:
         else:
             feature = ft
 
-        policy_key = env_cfg.features_map[key]
+        policy_key = env_cfg.features_map.get(key, key)
         policy_features[policy_key] = feature
 
     return policy_features
 
 
-def are_all_envs_same_type(env: gym.vector.VectorEnv) -> bool:
-    first_type = type(env.envs[0])  # Get type of first env
-    return all(type(e) is first_type for e in env.envs)  # Fast type check
+def preprocess_maniskill_observation(
+    observations: dict[str, np.ndarray],
+) -> dict[str, Tensor]:
+    """Convert environment observation to LeRobot format observation.
+    Args:
+        observation: Dictionary of observation batches from a Gym vector environment.
+    Returns:
+        Dictionary of observation batches with keys renamed to LeRobot format and values as tensors.
+    """
+    # map to expected inputs for the policy
+    return_observations = {}
+    # TODO: You have to merge all tensors from agent key and extra key
+    # You don't keep sensor param key in the observation
+    # And you keep sensor data rgb
+    q_pos = observations["agent"]["qpos"]
+    q_vel = observations["agent"]["qvel"]
+    tcp_pos = observations["extra"]["tcp_pose"]
+    img = observations["sensor_data"]["base_camera"]["rgb"]
 
+    _, h, w, c = img.shape
+    assert c < h and c < w, f"expect channel last images, but instead got {img.shape=}"
 
-def check_env_attributes_and_types(env: gym.vector.VectorEnv) -> None:
-    with warnings.catch_warnings():
-        warnings.simplefilter("once", UserWarning)  # Apply filter only in this function
+    # sanity check that images are uint8
+    assert img.dtype == torch.uint8, f"expect torch.uint8, but instead {img.dtype=}"
 
-        if not (hasattr(env.envs[0], "task_description") and hasattr(env.envs[0], "task")):
-            warnings.warn(
-                "The environment does not have 'task_description' and 'task'. Some policies require these features.",
-                UserWarning,
-                stacklevel=2,
-            )
-        if not are_all_envs_same_type(env):
-            warnings.warn(
-                "The environments have different types. Make sure you infer the right task from each environment. Empty task will be passed instead.",
-                UserWarning,
-                stacklevel=2,
-            )
+    # convert to channel first of type float32 in range [0,1]
+    img = einops.rearrange(img, "b h w c -> b c h w").contiguous()
+    img = img.type(torch.float32)
+    img /= 255
 
+    state = torch.cat([q_pos, q_vel, tcp_pos], dim=-1)
 
-def add_envs_task(env: gym.vector.VectorEnv, observation: dict[str, Any]) -> dict[str, Any]:
-    """Adds task feature to the observation dict with respect to the first environment attribute."""
-    if hasattr(env.envs[0], "task_description"):
-        observation["task"] = env.call("task_description")
-    elif hasattr(env.envs[0], "task"):
-        observation["task"] = env.call("task")
-    else:  #  For envs without language instructions, e.g. aloha transfer cube and etc.
-        num_envs = observation[list(observation.keys())[0]].shape[0]
-        observation["task"] = ["" for _ in range(num_envs)]
-    return observation
+    return_observations["observation.image"] = img
+    return_observations["observation.state"] = state
+    return return_observations

lerobot/common/optim/optimizers.py

@@ -14,8 +14,9 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 import abc
-from dataclasses import asdict, dataclass
+from dataclasses import asdict, dataclass, field
 from pathlib import Path
+from typing import Any
 
 import draccus
 import torch
@@ -44,7 +45,7 @@ class OptimizerConfig(draccus.ChoiceRegistry, abc.ABC):
         return "adam"
 
     @abc.abstractmethod
-    def build(self) -> torch.optim.Optimizer:
+    def build(self) -> torch.optim.Optimizer | dict[str, torch.optim.Optimizer]:
         raise NotImplementedError
 
 
@@ -94,7 +95,76 @@ class SGDConfig(OptimizerConfig):
         return torch.optim.SGD(params, **kwargs)
 
 
-def save_optimizer_state(optimizer: torch.optim.Optimizer, save_dir: Path) -> None:
+@OptimizerConfig.register_subclass("multi_adam")
+@dataclass
+class MultiAdamConfig(OptimizerConfig):
+    """Configuration for multiple Adam optimizers with different parameter groups.
+
+    This creates a dictionary of Adam optimizers, each with its own hyperparameters.
+
+    Args:
+        lr: Default learning rate (used if not specified for a group)
+        weight_decay: Default weight decay (used if not specified for a group)
+        optimizer_groups: Dictionary mapping parameter group names to their hyperparameters
+        grad_clip_norm: Gradient clipping norm
+    """
+
+    lr: float = 1e-3
+    weight_decay: float = 0.0
+    grad_clip_norm: float = 10.0
+    optimizer_groups: dict[str, dict[str, Any]] = field(default_factory=dict)
+
+    def build(self, params_dict: dict[str, list]) -> dict[str, torch.optim.Optimizer]:
+        """Build multiple Adam optimizers.
+
+        Args:
+            params_dict: Dictionary mapping parameter group names to lists of parameters
+                         The keys should match the keys in optimizer_groups
+
+        Returns:
+            Dictionary mapping parameter group names to their optimizers
+        """
+        optimizers = {}
+
+        for name, params in params_dict.items():
+            # Get group-specific hyperparameters or use defaults
+            group_config = self.optimizer_groups.get(name, {})
+
+            # Create optimizer with merged parameters (defaults + group-specific)
+            optimizer_kwargs = {
+                "lr": group_config.get("lr", self.lr),
+                "betas": group_config.get("betas", (0.9, 0.999)),
+                "eps": group_config.get("eps", 1e-5),
+                "weight_decay": group_config.get("weight_decay", self.weight_decay),
+            }
+
+            optimizers[name] = torch.optim.Adam(params, **optimizer_kwargs)
+
+        return optimizers
+
+
+def save_optimizer_state(
+    optimizer: torch.optim.Optimizer | dict[str, torch.optim.Optimizer], save_dir: Path
+) -> None:
+    """Save optimizer state to disk.
+
+    Args:
+        optimizer: Either a single optimizer or a dictionary of optimizers.
+        save_dir: Directory to save the optimizer state.
+    """
+    if isinstance(optimizer, dict):
+        # Handle dictionary of optimizers
+        for name, opt in optimizer.items():
+            optimizer_dir = save_dir / name
+            optimizer_dir.mkdir(exist_ok=True, parents=True)
+            _save_single_optimizer_state(opt, optimizer_dir)
+    else:
+        # Handle single optimizer
+        _save_single_optimizer_state(optimizer, save_dir)
+
+
+def _save_single_optimizer_state(optimizer: torch.optim.Optimizer, save_dir: Path) -> None:
+    """Save a single optimizer's state to disk."""
     state = optimizer.state_dict()
     param_groups = state.pop("param_groups")
     flat_state = flatten_dict(state)
@@ -102,11 +172,44 @@ def save_optimizer_state(optimizer: torch.optim.Optimizer, save_dir: Path) -> No
     write_json(param_groups, save_dir / OPTIMIZER_PARAM_GROUPS)
 
 
-def load_optimizer_state(optimizer: torch.optim.Optimizer, save_dir: Path) -> torch.optim.Optimizer:
+def load_optimizer_state(
+    optimizer: torch.optim.Optimizer | dict[str, torch.optim.Optimizer], save_dir: Path
+) -> torch.optim.Optimizer | dict[str, torch.optim.Optimizer]:
+    """Load optimizer state from disk.
+
+    Args:
+        optimizer: Either a single optimizer or a dictionary of optimizers.
+        save_dir: Directory to load the optimizer state from.
+
+    Returns:
+        The updated optimizer(s) with loaded state.
+    """
+    if isinstance(optimizer, dict):
+        # Handle dictionary of optimizers
+        loaded_optimizers = {}
+        for name, opt in optimizer.items():
+            optimizer_dir = save_dir / name
+            if optimizer_dir.exists():
+                loaded_optimizers[name] = _load_single_optimizer_state(opt, optimizer_dir)
+            else:
+                loaded_optimizers[name] = opt
+        return loaded_optimizers
+    else:
+        # Handle single optimizer
+        return _load_single_optimizer_state(optimizer, save_dir)
+
+
+def _load_single_optimizer_state(optimizer: torch.optim.Optimizer, save_dir: Path) -> torch.optim.Optimizer:
+    """Load a single optimizer's state from disk."""
     current_state_dict = optimizer.state_dict()
     flat_state = load_file(save_dir / OPTIMIZER_STATE)
     state = unflatten_dict(flat_state)
-    loaded_state_dict = {"state": {int(k): v for k, v in state["state"].items()}}
+
+    # Handle case where 'state' key might not exist (for newly created optimizers)
+    if "state" in state:
+        loaded_state_dict = {"state": {int(k): v for k, v in state["state"].items()}}
+    else:
+        loaded_state_dict = {"state": {}}
 
     if "param_groups" in current_state_dict:
         param_groups = deserialize_json_into_object(

lerobot/common/optim/schedulers.py

@@ -49,7 +49,11 @@ class DiffuserSchedulerConfig(LRSchedulerConfig):
     def build(self, optimizer: Optimizer, num_training_steps: int) -> LambdaLR:
         from diffusers.optimization import get_scheduler
 
-        kwargs = {**asdict(self), "num_training_steps": num_training_steps, "optimizer": optimizer}
+        kwargs = {
+            **asdict(self),
+            "num_training_steps": num_training_steps,
+            "optimizer": optimizer,
+        }
         return get_scheduler(**kwargs)
 
 
@@ -71,7 +75,10 @@ class VQBeTSchedulerConfig(LRSchedulerConfig):
                 progress = float(adjusted_step - self.num_warmup_steps) / float(
                     max(1, num_training_steps - self.num_warmup_steps)
                 )
-                return max(0.0, 0.5 * (1.0 + math.cos(math.pi * float(self.num_cycles) * 2.0 * progress)))
+                return max(
+                    0.0,
+                    0.5 * (1.0 + math.cos(math.pi * float(self.num_cycles) * 2.0 * progress)),
+                )
 
         return LambdaLR(optimizer, lr_lambda, -1)
 

lerobot/common/policies/__init__.py

@@ -15,6 +15,5 @@
 from .act.configuration_act import ACTConfig as ACTConfig
 from .diffusion.configuration_diffusion import DiffusionConfig as DiffusionConfig
 from .pi0.configuration_pi0 import PI0Config as PI0Config
-from .smolvla.configuration_smolvla import SmolVLAConfig as SmolVLAConfig
 from .tdmpc.configuration_tdmpc import TDMPCConfig as TDMPCConfig
 from .vqbet.configuration_vqbet import VQBeTConfig as VQBeTConfig

lerobot/common/policies/act/modeling_act.py

@@ -241,7 +241,9 @@ class ACTTemporalEnsembler:
             # Note: The last dimension is unsqueeze to make sure we can broadcast properly for tensor
             # operations later.
             self.ensembled_actions_count = torch.ones(
-                (self.chunk_size, 1), dtype=torch.long, device=self.ensembled_actions.device
+                (self.chunk_size, 1),
+                dtype=torch.long,
+                device=self.ensembled_actions.device,
             )
         else:
             # self.ensembled_actions will have shape (batch_size, chunk_size - 1, action_dim). Compute
@@ -253,7 +255,10 @@ class ACTTemporalEnsembler:
             # The last action, which has no prior online average, needs to get concatenated onto the end.
             self.ensembled_actions = torch.cat([self.ensembled_actions, actions[:, -1:]], dim=1)
             self.ensembled_actions_count = torch.cat(
-                [self.ensembled_actions_count, torch.ones_like(self.ensembled_actions_count[-1:])]
+                [
+                    self.ensembled_actions_count,
+                    torch.ones_like(self.ensembled_actions_count[-1:]),
+                ]
             )
         # "Consume" the first action.
         action, self.ensembled_actions, self.ensembled_actions_count = (
@@ -333,7 +338,11 @@ class ACT(nn.Module):
         # Backbone for image feature extraction.
         if self.config.image_features:
             backbone_model = getattr(torchvision.models, config.vision_backbone)(
-                replace_stride_with_dilation=[False, False, config.replace_final_stride_with_dilation],
+                replace_stride_with_dilation=[
+                    False,
+                    False,
+                    config.replace_final_stride_with_dilation,
+                ],
                 weights=config.pretrained_backbone_weights,
                 norm_layer=FrozenBatchNorm2d,
             )
@@ -427,7 +436,11 @@ class ACT(nn.Module):
             action_embed = self.vae_encoder_action_input_proj(batch["action"])  # (B, S, D)
 
             if self.config.robot_state_feature:
-                vae_encoder_input = [cls_embed, robot_state_embed, action_embed]  # (B, S+2, D)
+                vae_encoder_input = [
+                    cls_embed,
+                    robot_state_embed,
+                    action_embed,
+                ]  # (B, S+2, D)
             else:
                 vae_encoder_input = [cls_embed, action_embed]
             vae_encoder_input = torch.cat(vae_encoder_input, axis=1)
@@ -540,7 +553,10 @@ class ACTEncoder(nn.Module):
         self.norm = nn.LayerNorm(config.dim_model) if config.pre_norm else nn.Identity()
 
     def forward(
-        self, x: Tensor, pos_embed: Tensor | None = None, key_padding_mask: Tensor | None = None
+        self,
+        x: Tensor,
+        pos_embed: Tensor | None = None,
+        key_padding_mask: Tensor | None = None,
     ) -> Tensor:
         for layer in self.layers:
             x = layer(x, pos_embed=pos_embed, key_padding_mask=key_padding_mask)
@@ -603,7 +619,10 @@ class ACTDecoder(nn.Module):
     ) -> Tensor:
         for layer in self.layers:
             x = layer(
-                x, encoder_out, decoder_pos_embed=decoder_pos_embed, encoder_pos_embed=encoder_pos_embed
+                x,
+                encoder_out,
+                decoder_pos_embed=decoder_pos_embed,
+                encoder_pos_embed=encoder_pos_embed,
             )
         if self.norm is not None:
             x = self.norm(x)

lerobot/common/policies/diffusion/modeling_diffusion.py

@@ -209,7 +209,10 @@ class DiffusionModel(nn.Module):
 
     # ========= inference  ============
     def conditional_sample(
-        self, batch_size: int, global_cond: Tensor | None = None, generator: torch.Generator | None = None
+        self,
+        batch_size: int,
+        global_cond: Tensor | None = None,
+        generator: torch.Generator | None = None,
     ) -> Tensor:
         device = get_device_from_parameters(self)
         dtype = get_dtype_from_parameters(self)
@@ -254,7 +257,10 @@ class DiffusionModel(nn.Module):
                 # Separate batch and sequence dims back out. The camera index dim gets absorbed into the
                 # feature dim (effectively concatenating the camera features).
                 img_features = einops.rearrange(
-                    img_features_list, "(n b s) ... -> b s (n ...)", b=batch_size, s=n_obs_steps
+                    img_features_list,
+                    "(n b s) ... -> b s (n ...)",
+                    b=batch_size,
+                    s=n_obs_steps,
                 )
             else:
                 # Combine batch, sequence, and "which camera" dims before passing to shared encoder.
@@ -264,7 +270,10 @@ class DiffusionModel(nn.Module):
                 # Separate batch dim and sequence dim back out. The camera index dim gets absorbed into the
                 # feature dim (effectively concatenating the camera features).
                 img_features = einops.rearrange(
-                    img_features, "(b s n) ... -> b s (n ...)", b=batch_size, s=n_obs_steps
+                    img_features,
+                    "(b s n) ... -> b s (n ...)",
+                    b=batch_size,
+                    s=n_obs_steps,
                 )
             global_cond_feats.append(img_features)
 
@@ -515,7 +524,9 @@ class DiffusionRgbEncoder(nn.Module):
 
 
 def _replace_submodules(
-    root_module: nn.Module, predicate: Callable[[nn.Module], bool], func: Callable[[nn.Module], nn.Module]
+    root_module: nn.Module,
+    predicate: Callable[[nn.Module], bool],
+    func: Callable[[nn.Module], nn.Module],
 ) -> nn.Module:
     """
     Args:
@@ -633,10 +644,14 @@ class DiffusionConditionalUnet1d(nn.Module):
         self.mid_modules = nn.ModuleList(
             [
                 DiffusionConditionalResidualBlock1d(
-                    config.down_dims[-1], config.down_dims[-1], **common_res_block_kwargs
+                    config.down_dims[-1],
+                    config.down_dims[-1],
+                    **common_res_block_kwargs,
                 ),
                 DiffusionConditionalResidualBlock1d(
-                    config.down_dims[-1], config.down_dims[-1], **common_res_block_kwargs
+                    config.down_dims[-1],
+                    config.down_dims[-1],
+                    **common_res_block_kwargs,
                 ),
             ]
         )

lerobot/common/policies/factory.py

@@ -24,10 +24,9 @@ from lerobot.common.envs.configs import EnvConfig
 from lerobot.common.envs.utils import env_to_policy_features
 from lerobot.common.policies.act.configuration_act import ACTConfig
 from lerobot.common.policies.diffusion.configuration_diffusion import DiffusionConfig
+from lerobot.common.policies.hilserl.classifier.configuration_classifier import ClassifierConfig
 from lerobot.common.policies.pi0.configuration_pi0 import PI0Config
-from lerobot.common.policies.pi0fast.configuration_pi0fast import PI0FASTConfig
 from lerobot.common.policies.pretrained import PreTrainedPolicy
-from lerobot.common.policies.smolvla.configuration_smolvla import SmolVLAConfig
 from lerobot.common.policies.tdmpc.configuration_tdmpc import TDMPCConfig
 from lerobot.common.policies.vqbet.configuration_vqbet import VQBeTConfig
 from lerobot.configs.policies import PreTrainedConfig
@@ -56,14 +55,14 @@ def get_policy_class(name: str) -> PreTrainedPolicy:
         from lerobot.common.policies.pi0.modeling_pi0 import PI0Policy
 
         return PI0Policy
-    elif name == "pi0fast":
-        from lerobot.common.policies.pi0fast.modeling_pi0fast import PI0FASTPolicy
+    elif name == "sac":
+        from lerobot.common.policies.sac.modeling_sac import SACPolicy
 
-        return PI0FASTPolicy
-    elif name == "smolvla":
-        from lerobot.common.policies.smolvla.modeling_smolvla import SmolVLAPolicy
+        return SACPolicy
+    elif name == "hilserl_classifier":
+        from lerobot.common.policies.hilserl.classifier.modeling_classifier import Classifier
 
-        return SmolVLAPolicy
+        return Classifier
     else:
         raise NotImplementedError(f"Policy with name {name} is not implemented.")
 
@@ -79,10 +78,8 @@ def make_policy_config(policy_type: str, **kwargs) -> PreTrainedConfig:
         return VQBeTConfig(**kwargs)
     elif policy_type == "pi0":
         return PI0Config(**kwargs)
-    elif policy_type == "pi0fast":
-        return PI0FASTConfig(**kwargs)
-    elif policy_type == "smolvla":
-        return SmolVLAConfig(**kwargs)
+    elif policy_type == "hilserl_classifier":
+        return ClassifierConfig(**kwargs)
     else:
         raise ValueError(f"Policy type '{policy_type}' is not available.")
 

lerobot/common/policies/hilserl/classifier/configuration_classifier.py

@@ -0,0 +1,53 @@
+from dataclasses import dataclass
+from typing import List
+
+from lerobot.common.optim.optimizers import AdamWConfig, OptimizerConfig
+from lerobot.common.optim.schedulers import LRSchedulerConfig
+from lerobot.configs.policies import PreTrainedConfig
+
+
+@PreTrainedConfig.register_subclass(name="hilserl_classifier")
+@dataclass
+class ClassifierConfig(PreTrainedConfig):
+    """Configuration for the Classifier model."""
+
+    name: str = "hilserl_classifier"
+    num_classes: int = 2
+    hidden_dim: int = 256
+    dropout_rate: float = 0.1
+    model_name: str = "helper2424/resnet10"
+    device: str = "cpu"
+    model_type: str = "cnn"  # "transformer" or "cnn"
+    num_cameras: int = 2
+    learning_rate: float = 1e-4
+    normalization_mode = None
+    # output_features: Dict[str, PolicyFeature] = field(
+    #     default_factory=lambda: {"next.reward": PolicyFeature(type=FeatureType.REWARD, shape=(1,))}
+    # )
+
+    @property
+    def observation_delta_indices(self) -> List | None:
+        return None
+
+    @property
+    def action_delta_indices(self) -> List | None:
+        return None
+
+    @property
+    def reward_delta_indices(self) -> List | None:
+        return None
+
+    def get_optimizer_preset(self) -> OptimizerConfig:
+        return AdamWConfig(
+            lr=self.learning_rate,
+            weight_decay=0.01,
+            grad_clip_norm=1.0,
+        )
+
+    def get_scheduler_preset(self) -> LRSchedulerConfig | None:
+        return None
+
+    def validate_features(self) -> None:
+        """Validate feature configurations."""
+        # Classifier doesn't need specific feature validation
+        pass

lerobot/common/policies/hilserl/classifier/modeling_classifier.py

@@ -0,0 +1,237 @@
+import logging
+from typing import Dict, Optional, Tuple
+
+import torch
+from torch import Tensor, nn
+
+from lerobot.common.constants import OBS_IMAGE
+from lerobot.common.policies.hilserl.classifier.configuration_classifier import (
+    ClassifierConfig,
+)
+from lerobot.common.policies.normalize import Normalize, Unnormalize
+from lerobot.common.policies.pretrained import PreTrainedPolicy
+
+logging.basicConfig(level=logging.INFO, format="%(asctime)s - %(name)s - %(levelname)s - %(message)s")
+logger = logging.getLogger(__name__)
+
+
+class ClassifierOutput:
+    """Wrapper for classifier outputs with additional metadata."""
+
+    def __init__(
+        self,
+        logits: Tensor,
+        probabilities: Optional[Tensor] = None,
+        hidden_states: Optional[Tensor] = None,
+    ):
+        self.logits = logits
+        self.probabilities = probabilities
+        self.hidden_states = hidden_states
+
+    def __repr__(self):
+        return (
+            f"ClassifierOutput(logits={self.logits}, "
+            f"probabilities={self.probabilities}, "
+            f"hidden_states={self.hidden_states})"
+        )
+
+
+class Classifier(PreTrainedPolicy):
+    """Image classifier built on top of a pre-trained encoder."""
+
+    name = "hilserl_classifier"
+    config_class = ClassifierConfig
+
+    def __init__(
+        self,
+        config: ClassifierConfig,
+        dataset_stats: Dict[str, Dict[str, Tensor]] | None = None,
+    ):
+        from transformers import AutoModel
+
+        super().__init__(config)
+        self.config = config
+
+        # Initialize normalization (standardized with the policy framework)
+        self.normalize_inputs = Normalize(config.input_features, config.normalization_mapping, dataset_stats)
+        self.normalize_targets = Normalize(
+            config.output_features, config.normalization_mapping, dataset_stats
+        )
+        self.unnormalize_outputs = Unnormalize(
+            config.output_features, config.normalization_mapping, dataset_stats
+        )
+
+        # Set up encoder
+        encoder = AutoModel.from_pretrained(self.config.model_name, trust_remote_code=True)
+        # Extract vision model if we're given a multimodal model
+        if hasattr(encoder, "vision_model"):
+            logging.info("Multimodal model detected - using vision encoder only")
+            self.encoder = encoder.vision_model
+            self.vision_config = encoder.config.vision_config
+        else:
+            self.encoder = encoder
+            self.vision_config = getattr(encoder, "config", None)
+
+        # Model type from config
+        self.is_cnn = self.config.model_type == "cnn"
+
+        # For CNNs, initialize backbone
+        if self.is_cnn:
+            self._setup_cnn_backbone()
+
+        self._freeze_encoder()
+        self._build_classifier_head()
+
+    def _setup_cnn_backbone(self):
+        """Set up CNN encoder"""
+        if hasattr(self.encoder, "fc"):
+            self.feature_dim = self.encoder.fc.in_features
+            self.encoder = nn.Sequential(*list(self.encoder.children())[:-1])
+        elif hasattr(self.encoder.config, "hidden_sizes"):
+            self.feature_dim = self.encoder.config.hidden_sizes[-1]  # Last channel dimension
+        else:
+            raise ValueError("Unsupported CNN architecture")
+
+    def _freeze_encoder(self) -> None:
+        """Freeze the encoder parameters."""
+        for param in self.encoder.parameters():
+            param.requires_grad = False
+
+    def _build_classifier_head(self) -> None:
+        """Initialize the classifier head architecture."""
+        # Get input dimension based on model type
+        if self.is_cnn:
+            input_dim = self.feature_dim
+        else:  # Transformer models
+            if hasattr(self.encoder.config, "hidden_size"):
+                input_dim = self.encoder.config.hidden_size
+            else:
+                raise ValueError("Unsupported transformer architecture since hidden_size is not found")
+
+        self.classifier_head = nn.Sequential(
+            nn.Linear(input_dim * self.config.num_cameras, self.config.hidden_dim),
+            nn.Dropout(self.config.dropout_rate),
+            nn.LayerNorm(self.config.hidden_dim),
+            nn.ReLU(),
+            nn.Linear(
+                self.config.hidden_dim,
+                1 if self.config.num_classes == 2 else self.config.num_classes,
+            ),
+        )
+
+    def _get_encoder_output(self, x: torch.Tensor) -> torch.Tensor:
+        """Extract the appropriate output from the encoder."""
+        with torch.no_grad():
+            if self.is_cnn:
+                # The HF ResNet applies pooling internally
+                outputs = self.encoder(x)
+                # Get pooled output directly
+                features = outputs.pooler_output
+
+                if features.dim() > 2:
+                    features = features.squeeze(-1).squeeze(-1)
+                return features
+            else:  # Transformer models
+                outputs = self.encoder(x)
+                if hasattr(outputs, "pooler_output") and outputs.pooler_output is not None:
+                    return outputs.pooler_output
+                return outputs.last_hidden_state[:, 0, :]
+
+    def extract_images_and_labels(self, batch: Dict[str, Tensor]) -> Tuple[list, Tensor]:
+        """Extract image tensors and label tensors from batch."""
+        # Find image keys in input features
+        image_keys = [key for key in self.config.input_features if key.startswith(OBS_IMAGE)]
+
+        # Extract the images and labels
+        images = [batch[key] for key in image_keys]
+        labels = batch["next.reward"]
+
+        return images, labels
+
+    def predict(self, xs: list) -> ClassifierOutput:
+        """Forward pass of the classifier for inference."""
+        encoder_outputs = torch.hstack([self._get_encoder_output(x) for x in xs])
+        logits = self.classifier_head(encoder_outputs)
+
+        if self.config.num_classes == 2:
+            logits = logits.squeeze(-1)
+            probabilities = torch.sigmoid(logits)
+        else:
+            probabilities = torch.softmax(logits, dim=-1)
+
+        return ClassifierOutput(logits=logits, probabilities=probabilities, hidden_states=encoder_outputs)
+
+    def forward(self, batch: Dict[str, Tensor]) -> Tuple[Tensor, Dict[str, Tensor]]:
+        """Standard forward pass for training compatible with train.py."""
+        # Normalize inputs if needed
+        batch = self.normalize_inputs(batch)
+        batch = self.normalize_targets(batch)
+
+        # Extract images and labels
+        images, labels = self.extract_images_and_labels(batch)
+
+        # Get predictions
+        outputs = self.predict(images)
+
+        # Calculate loss
+        if self.config.num_classes == 2:
+            # Binary classification
+            loss = nn.functional.binary_cross_entropy_with_logits(outputs.logits, labels)
+            predictions = (torch.sigmoid(outputs.logits) > 0.5).float()
+        else:
+            # Multi-class classification
+            loss = nn.functional.cross_entropy(outputs.logits, labels.long())
+            predictions = torch.argmax(outputs.logits, dim=1)
+
+        # Calculate accuracy for logging
+        correct = (predictions == labels).sum().item()
+        total = labels.size(0)
+        accuracy = 100 * correct / total
+
+        # Return loss and metrics for logging
+        output_dict = {
+            "accuracy": accuracy,
+            "correct": correct,
+            "total": total,
+        }
+
+        return loss, output_dict
+
+    def predict_reward(self, batch, threshold=0.6):
+        """Legacy method for compatibility."""
+        images, _ = self.extract_images_and_labels(batch)
+        if self.config.num_classes == 2:
+            probs = self.predict(images).probabilities
+            logging.debug(f"Predicted reward images: {probs}")
+            return (probs > threshold).float()
+        else:
+            return torch.argmax(self.predict(images).probabilities, dim=1)
+
+    # Methods required by PreTrainedPolicy abstract class
+
+    def get_optim_params(self) -> dict:
+        """Return optimizer parameters for the policy."""
+        return {
+            "params": self.parameters(),
+            "lr": getattr(self.config, "learning_rate", 1e-4),
+            "weight_decay": getattr(self.config, "weight_decay", 0.01),
+        }
+
+    def reset(self):
+        """Reset any stateful components (required by PreTrainedPolicy)."""
+        # Classifier doesn't have stateful components that need resetting
+        pass
+
+    def select_action(self, batch: Dict[str, Tensor]) -> Tensor:
+        """Return action (class prediction) based on input observation."""
+        images, _ = self.extract_images_and_labels(batch)
+
+        with torch.no_grad():
+            outputs = self.predict(images)
+
+            if self.config.num_classes == 2:
+                # For binary classification return 0 or 1
+                return (outputs.probabilities > 0.5).float()
+            else:
+                # For multi-class return the predicted class
+                return torch.argmax(outputs.probabilities, dim=1)

lerobot/common/policies/hilserl/configuration_hilserl.py

@@ -0,0 +1,23 @@
+#!/usr/bin/env python
+
+# Copyright 2024 The HuggingFace Inc. team.
+# All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from dataclasses import dataclass
+
+
+@dataclass
+class HILSerlConfig:
+    pass

lerobot/common/policies/hilserl/modeling_hilserl.py

@@ -0,0 +1,29 @@
+#!/usr/bin/env python
+
+# Copyright 2024 The HuggingFace Inc. team.
+# All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import torch.nn as nn
+from huggingface_hub import PyTorchModelHubMixin
+
+
+class HILSerlPolicy(
+    nn.Module,
+    PyTorchModelHubMixin,
+    library_name="lerobot",
+    repo_url="https://github.com/huggingface/lerobot",
+    tags=["robotics", "hilserl"],
+):
+    pass

lerobot/common/policies/normalize.py

@@ -79,28 +79,46 @@ def create_stats_buffers(
             )
 
         # TODO(aliberts, rcadene): harmonize this to only use one framework (np or torch)
-        if stats:
-            if isinstance(stats[key]["mean"], np.ndarray):
-                if norm_mode is NormalizationMode.MEAN_STD:
+        if stats and key in stats:
+            if norm_mode is NormalizationMode.MEAN_STD:
+                if "mean" not in stats[key] or "std" not in stats[key]:
+                    raise ValueError(
+                        f"Missing 'mean' or 'std' in stats for key {key} with MEAN_STD normalization"
+                    )
+
+                if isinstance(stats[key]["mean"], np.ndarray):
                     buffer["mean"].data = torch.from_numpy(stats[key]["mean"]).to(dtype=torch.float32)
                     buffer["std"].data = torch.from_numpy(stats[key]["std"]).to(dtype=torch.float32)
-                elif norm_mode is NormalizationMode.MIN_MAX:
-                    buffer["min"].data = torch.from_numpy(stats[key]["min"]).to(dtype=torch.float32)
-                    buffer["max"].data = torch.from_numpy(stats[key]["max"]).to(dtype=torch.float32)
-            elif isinstance(stats[key]["mean"], torch.Tensor):
-                # Note: The clone is needed to make sure that the logic in save_pretrained doesn't see duplicated
-                # tensors anywhere (for example, when we use the same stats for normalization and
-                # unnormalization). See the logic here
-                # https://github.com/huggingface/safetensors/blob/079781fd0dc455ba0fe851e2b4507c33d0c0d407/bindings/python/py_src/safetensors/torch.py#L97.
-                if norm_mode is NormalizationMode.MEAN_STD:
+                elif isinstance(stats[key]["mean"], torch.Tensor):
+                    # Note: The clone is needed to make sure that the logic in save_pretrained doesn't see duplicated
+                    # tensors anywhere (for example, when we use the same stats for normalization and
+                    # unnormalization). See the logic here
+                    # https://github.com/huggingface/safetensors/blob/079781fd0dc455ba0fe851e2b4507c33d0c0d407/bindings/python/py_src/safetensors/torch.py#L97.
                     buffer["mean"].data = stats[key]["mean"].clone().to(dtype=torch.float32)
                     buffer["std"].data = stats[key]["std"].clone().to(dtype=torch.float32)
-                elif norm_mode is NormalizationMode.MIN_MAX:
+                else:
+                    type_ = type(stats[key]["mean"])
+                    raise ValueError(
+                        f"np.ndarray or torch.Tensor expected for 'mean', but type is '{type_}' instead."
+                    )
+
+            elif norm_mode is NormalizationMode.MIN_MAX:
+                if "min" not in stats[key] or "max" not in stats[key]:
+                    raise ValueError(
+                        f"Missing 'min' or 'max' in stats for key {key} with MIN_MAX normalization"
+                    )
+
+                if isinstance(stats[key]["min"], np.ndarray):
+                    buffer["min"].data = torch.from_numpy(stats[key]["min"]).to(dtype=torch.float32)
+                    buffer["max"].data = torch.from_numpy(stats[key]["max"]).to(dtype=torch.float32)
+                elif isinstance(stats[key]["min"], torch.Tensor):
                     buffer["min"].data = stats[key]["min"].clone().to(dtype=torch.float32)
                     buffer["max"].data = stats[key]["max"].clone().to(dtype=torch.float32)
-            else:
-                type_ = type(stats[key]["mean"])
-                raise ValueError(f"np.ndarray or torch.Tensor expected, but type is '{type_}' instead.")
+                else:
+                    type_ = type(stats[key]["min"])
+                    raise ValueError(
+                        f"np.ndarray or torch.Tensor expected for 'min', but type is '{type_}' instead."
+                    )
 
         stats_buffers[key] = buffer
     return stats_buffers
@@ -149,12 +167,13 @@ class Normalize(nn.Module):
             setattr(self, "buffer_" + key.replace(".", "_"), buffer)
 
     # TODO(rcadene): should we remove torch.no_grad?
-    @torch.no_grad
+    # @torch.no_grad
     def forward(self, batch: dict[str, Tensor]) -> dict[str, Tensor]:
         batch = dict(batch)  # shallow copy avoids mutating the input batch
         for key, ft in self.features.items():
             if key not in batch:
                 # FIXME(aliberts, rcadene): This might lead to silent fail!
+                # NOTE: (azouitine) This continues help us for instantiation SACPolicy
                 continue
 
             norm_mode = self.norm_map.get(ft.type, NormalizationMode.IDENTITY)
@@ -223,7 +242,7 @@ class Unnormalize(nn.Module):
             setattr(self, "buffer_" + key.replace(".", "_"), buffer)
 
     # TODO(rcadene): should we remove torch.no_grad?
-    @torch.no_grad
+    # @torch.no_grad
     def forward(self, batch: dict[str, Tensor]) -> dict[str, Tensor]:
         batch = dict(batch)  # shallow copy avoids mutating the input batch
         for key, ft in self.features.items():

lerobot/common/policies/pi0/conversion_scripts/convert_pi0_to_hf_lerobot.py

@@ -61,7 +61,11 @@ from lerobot.common.policies.pi0.conversion_scripts.conversion_utils import (
 )
 from lerobot.common.policies.pi0.modeling_pi0 import PI0Policy
 
-PRECISIONS = {"bfloat16": torch.bfloat16, "float32": torch.float32, "float16": torch.float16}
+PRECISIONS = {
+    "bfloat16": torch.bfloat16,
+    "float32": torch.float32,
+    "float16": torch.float16,
+}
 
 
 def slice_paligemma_state_dict(state_dict, config):

lerobot/common/policies/pi0/flex_attention.py

@@ -48,18 +48,32 @@ def flex_attention_forward(
 
     key_states = key_states[:, :, :, None, :]
     key_states = key_states.expand(
-        batch_size, key_states.shape[1], num_key_value_heads, num_key_value_groups, head_dim
+        batch_size,
+        key_states.shape[1],
+        num_key_value_heads,
+        num_key_value_groups,
+        head_dim,
     )
     key_states = key_states.reshape(
-        batch_size, key_states.shape[1], num_key_value_heads * num_key_value_groups, head_dim
+        batch_size,
+        key_states.shape[1],
+        num_key_value_heads * num_key_value_groups,
+        head_dim,
     )
 
     value_states = value_states[:, :, :, None, :]
     value_states = value_states.expand(
-        batch_size, value_states.shape[1], num_key_value_heads, num_key_value_groups, head_dim
+        batch_size,
+        value_states.shape[1],
+        num_key_value_heads,
+        num_key_value_groups,
+        head_dim,
     )
     value_states = value_states.reshape(
-        batch_size, value_states.shape[1], num_key_value_heads * num_key_value_groups, head_dim
+        batch_size,
+        value_states.shape[1],
+        num_key_value_heads * num_key_value_groups,
+        head_dim,
     )
 
     query_states = query_states.transpose(1, 2)

lerobot/common/policies/pi0/modeling_pi0.py

@@ -24,7 +24,7 @@ Designed by Physical Intelligence. Ported from Jax by Hugging Face.
 
 Install pi0 extra dependencies:
 ```bash
-pip install -e ".[pi0]"
+pip install --no-binary=av -e ".[pi0]"
 ```
 
 Example of finetuning the pi0 pretrained model (`pi0_base` in `openpi`):
@@ -69,7 +69,11 @@ from lerobot.common.utils.utils import get_safe_dtype
 
 
 def create_sinusoidal_pos_embedding(
-    time: torch.tensor, dimension: int, min_period: float, max_period: float, device="cpu"
+    time: torch.tensor,
+    dimension: int,
+    min_period: float,
+    max_period: float,
+    device="cpu",
 ) -> Tensor:
     """Computes sine-cosine positional embedding vectors for scalar positions."""
     if dimension % 2 != 0:
@@ -357,7 +361,7 @@ class PI0Policy(PreTrainedPolicy):
             if self.config.resize_imgs_with_padding is not None:
                 img = resize_with_pad(img, *self.config.resize_imgs_with_padding, pad_value=0)
 
-            # Normalize from range [0,1] to [-1,1] as expected by siglip
+            # Normalize from range [0,1] to [-1,1] as expacted by siglip
             img = img * 2.0 - 1.0
 
             bsize = img.shape[0]
@@ -577,7 +581,11 @@ class PI0FlowMatching(nn.Module):
 
         # Embed timestep using sine-cosine positional encoding with sensitivity in the range [0, 1]
         time_emb = create_sinusoidal_pos_embedding(
-            timestep, self.config.proj_width, min_period=4e-3, max_period=4.0, device=device
+            timestep,
+            self.config.proj_width,
+            min_period=4e-3,
+            max_period=4.0,
+            device=device,
         )
         time_emb = time_emb.type(dtype=dtype)
 
@@ -609,7 +617,15 @@ class PI0FlowMatching(nn.Module):
         return embs, pad_masks, att_masks
 
     def forward(
-        self, images, img_masks, lang_tokens, lang_masks, state, actions, noise=None, time=None
+        self,
+        images,
+        img_masks,
+        lang_tokens,
+        lang_masks,
+        state,
+        actions,
+        noise=None,
+        time=None,
     ) -> Tensor:
         """Do a full training forward pass and compute the loss (batch_size x num_steps x num_motors)"""
         if noise is None:
@@ -655,7 +671,11 @@ class PI0FlowMatching(nn.Module):
         device = state.device
 
         if noise is None:
-            actions_shape = (bsize, self.config.n_action_steps, self.config.max_action_dim)
+            actions_shape = (
+                bsize,
+                self.config.n_action_steps,
+                self.config.max_action_dim,
+            )
             noise = self.sample_noise(actions_shape, device)
 
         prefix_embs, prefix_pad_masks, prefix_att_masks = self.embed_prefix(

lerobot/common/policies/pi0/paligemma_with_expert.py

@@ -293,12 +293,18 @@ class PaliGemmaWithExpertModel(PreTrainedModel):
                     # in `transformers`. (molbap)
                     key_states = torch.cat([past_key_values[layer_idx]["key_states"], key_states], dim=1)
                     value_states = torch.cat(
-                        [past_key_values[layer_idx]["value_states"], value_states], dim=1
+                        [past_key_values[layer_idx]["value_states"], value_states],
+                        dim=1,
                     )
 
             attention_interface = self.get_attention_interface()
             att_output = attention_interface(
-                attention_mask, batch_size, head_dim, query_states, key_states, value_states
+                attention_mask,
+                batch_size,
+                head_dim,
+                query_states,
+                key_states,
+                value_states,
             )
             att_output = att_output.to(dtype=torch.bfloat16)
 
@@ -358,12 +364,24 @@ class PaliGemmaWithExpertModel(PreTrainedModel):
         return attention_interface
 
     def flash_attention_forward(
-        self, attention_mask, batch_size, head_dim, query_states, key_states, value_states
+        self,
+        attention_mask,
+        batch_size,
+        head_dim,
+        query_states,
+        key_states,
+        value_states,
     ):
         raise NotImplementedError("FA2 is not implemented (yet)")
 
     def eager_attention_forward(
-        self, attention_mask, batch_size, head_dim, query_states, key_states, value_states
+        self,
+        attention_mask,
+        batch_size,
+        head_dim,
+        query_states,
+        key_states,
+        value_states,
     ):
         num_att_heads = self.config.paligemma_config.text_config.num_attention_heads
         num_key_value_heads = self.config.paligemma_config.text_config.num_key_value_heads
@@ -375,17 +393,31 @@ class PaliGemmaWithExpertModel(PreTrainedModel):
         sequence_length = key_states.shape[1]
 
         key_states = key_states[:, :, :, None, :].expand(
-            batch_size, sequence_length, num_key_value_heads, num_key_value_groups, head_dim
+            batch_size,
+            sequence_length,
+            num_key_value_heads,
+            num_key_value_groups,
+            head_dim,
         )
         key_states = key_states.reshape(
-            batch_size, sequence_length, num_key_value_heads * num_key_value_groups, head_dim
+            batch_size,
+            sequence_length,
+            num_key_value_heads * num_key_value_groups,
+            head_dim,
         )
 
         value_states = value_states[:, :, :, None, :].expand(
-            batch_size, sequence_length, num_key_value_heads, num_key_value_groups, head_dim
+            batch_size,
+            sequence_length,
+            num_key_value_heads,
+            num_key_value_groups,
+            head_dim,
         )
         value_states = value_states.reshape(
-            batch_size, sequence_length, num_key_value_heads * num_key_value_groups, head_dim
+            batch_size,
+            sequence_length,
+            num_key_value_heads * num_key_value_groups,
+            head_dim,
         )
 
         # Attention here is upcasted to float32 to match the original eager implementation.

lerobot/common/policies/pi0fast/configuration_pi0fast.py

@@ -1,136 +0,0 @@
-from dataclasses import dataclass, field
-
-from lerobot.common.optim.optimizers import AdamWConfig
-from lerobot.common.optim.schedulers import (
-    CosineDecayWithWarmupSchedulerConfig,
-)
-from lerobot.configs.policies import PreTrainedConfig
-from lerobot.configs.types import FeatureType, NormalizationMode, PolicyFeature
-
-
-@PreTrainedConfig.register_subclass("pi0fast")
-@dataclass
-class PI0FASTConfig(PreTrainedConfig):
-    # Input / output structure.
-    n_obs_steps: int = 1
-    chunk_size: int = 10
-    n_action_steps: int = 5
-
-    normalization_mapping: dict[str, NormalizationMode] = field(
-        default_factory=lambda: {
-            "VISUAL": NormalizationMode.IDENTITY,
-            "STATE": NormalizationMode.MEAN_STD,
-            "ACTION": NormalizationMode.MEAN_STD,
-        }
-    )
-
-    # Shorter state and action vectors will be padded
-    max_state_dim: int = 32  # 32
-    max_action_dim: int = 32  # 32
-
-    # Image preprocessing
-    resize_imgs_with_padding: tuple[int, int] = (224, 224)
-    interpolate_like_pi: bool = False
-
-    # Add empty images. Used by pi0_aloha_sim which adds the empty
-    # left and right wrist cameras in addition to the top camera.
-    empty_cameras: int = 0
-
-    # Converts the joint and gripper values from the standard Aloha space to
-    # the space used by the pi internal runtime which was used to train the base model.
-    adapt_to_pi_aloha: bool = False
-
-    # Converts joint dimensions to deltas with respect to the current state before passing to the model.
-    # Gripper dimensions will remain in absolute values.
-    use_delta_joint_actions_aloha: bool = False
-
-    # Tokenizer
-    tokenizer_max_length: int = 48
-
-    # Projector
-    proj_width: int = 1024
-
-    # Decoding
-    max_decoding_steps: int = 256
-    fast_skip_tokens: int = 128  # Skip last 128 tokens in PaliGemma vocab since they are special tokens
-    max_input_seq_len: int = 256  # 512
-
-    # Utils
-    use_cache: bool = True
-
-    # Frozen parameters
-    freeze_vision_encoder: bool = True
-    freeze_lm_head: bool = True
-
-    # Training presets
-    optimizer_lr: float = 1e-4
-    optimizer_betas: tuple[float, float] = (0.9, 0.95)
-    optimizer_eps: float = 1e-8
-    optimizer_weight_decay: float = 1e-5
-
-    scheduler_warmup_steps: int = 1_000
-    scheduler_decay_steps: int = 30_000
-    scheduler_decay_lr: float = 2.5e-6
-
-    checkpoint_path: str = None
-
-    padding_side: str = "right"
-
-    precision: str = "bfloat16"
-    grad_clip_norm: float = 1
-
-    # Allows padding/truncation of generated action tokens during detokenization to ensure decoding.
-    # In the original version, tensors of 0s were generated if shapes didn't match for stable decoding.
-    relaxed_action_decoding: bool = True
-
-    def __post_init__(self):
-        super().__post_init__()
-
-        """Input validation (not exhaustive)."""
-        if self.n_action_steps > self.chunk_size:
-            raise ValueError(
-                f"The chunk size is the upper bound for the number of action steps per model invocation. Got "
-                f"{self.n_action_steps} for `n_action_steps` and {self.chunk_size} for `chunk_size`."
-            )
-        if self.n_obs_steps != 1:
-            raise ValueError(
-                f"Multiple observation steps not handled yet. Got `nobs_steps={self.n_obs_steps}`"
-            )
-
-    def validate_features(self) -> None:
-        for i in range(self.empty_cameras):
-            key = f"observation.images.empty_camera_{i}"
-            empty_camera = PolicyFeature(
-                type=FeatureType.VISUAL,
-                shape=(3, 480, 640),
-            )
-            self.input_features[key] = empty_camera
-
-    def get_optimizer_preset(self) -> AdamWConfig:
-        return AdamWConfig(
-            lr=self.optimizer_lr,
-            betas=self.optimizer_betas,
-            eps=self.optimizer_eps,
-            weight_decay=self.optimizer_weight_decay,
-            grad_clip_norm=self.grad_clip_norm,
-        )
-
-    def get_scheduler_preset(self):
-        return CosineDecayWithWarmupSchedulerConfig(
-            peak_lr=self.optimizer_lr,
-            decay_lr=self.scheduler_decay_lr,
-            num_warmup_steps=self.scheduler_warmup_steps,
-            num_decay_steps=self.scheduler_decay_steps,
-        )
-
-    @property
-    def observation_delta_indices(self) -> None:
-        return None
-
-    @property
-    def action_delta_indices(self) -> list:
-        return list(range(self.chunk_size))
-
-    @property
-    def reward_delta_indices(self) -> None:
-        return None

lerobot/common/policies/pi0fast/modeling_pi0fast.py

@@ -1,973 +0,0 @@
-#!/usr/bin/env python
-
-# Copyright 2025 Physical Intelligence and The HuggingFace Inc. team. All rights reserved.
-#
-# Licensed under the Apache License, Version 2.0 (the "License");
-# you may not use this file except in compliance with the License.
-# You may obtain a copy of the License at
-#
-#     http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing, software
-# distributed under the License is distributed on an "AS IS" BASIS,
-# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-# See the License for the specific language governing permissions and
-# limitations under the License.
-
-"""
-π0+FAST: Efficient Action Tokenization for Vision-Language-Action Models
-
-[Paper](https://arxiv.org/abs/2501.09747)
-[Jax code](https://github.com/Physical-Intelligence/openpi)
-
-Designed by Physical Intelligence. Ported from Jax by Hugging Face.
-
-Example of finetuning the pi0+FAST pretrained model (`pi0_fast_base` in `openpi`):
-```bash
-python lerobot/scripts/train.py \
---policy.path=lerobot/pi0fast_base \
---dataset.repo_id=danaaubakirova/koch_test
-```
-
-Example of training the pi0+FAST neural network with from scratch:
-```bash
-python lerobot/scripts/train.py \
---policy.type=pi0fast \
---dataset.repo_id=danaaubakirova/koch_test
-```
-
-Example of using the pi0 pretrained model outside LeRobot training framework:
-```python
-policy = PI0FASTPolicy.from_pretrained("lerobot/pi0fast_base")
-```
-
-"""
-
-from collections import deque
-from functools import partial
-
-import numpy as np
-import torch
-import torch.nn.functional as F  # noqa: N812
-from PIL import Image
-from scipy.fft import idct
-from torch import Tensor, nn
-from transformers import AutoProcessor, AutoTokenizer, PaliGemmaForConditionalGeneration
-from transformers.cache_utils import HybridCache, StaticCache
-from transformers.models.auto import CONFIG_MAPPING
-
-from lerobot.common.constants import ACTION, OBS_ROBOT
-from lerobot.common.policies.normalize import Normalize, Unnormalize
-from lerobot.common.policies.pi0fast.configuration_pi0fast import PI0FASTConfig
-from lerobot.common.policies.pretrained import PreTrainedPolicy
-
-PRECISION = {
-    "float16": torch.float16,
-    "float32": torch.float32,
-    "bfloat16": torch.bfloat16,
-}
-
-
-def normalize(x, min_val, max_val):
-    return (x - min_val) / (max_val - min_val)
-
-
-def unnormalize(x, min_val, max_val):
-    return x * (max_val - min_val) + min_val
-
-
-def safe_arcsin(value):
-    # This ensures that the input stays within
-    # [−1,1] to avoid invalid values for arcsin
-    return torch.arcsin(torch.clamp(value, -1.0, 1.0))
-
-
-def aloha_gripper_to_angular(value):
-    # Aloha transforms the gripper positions into a linear space. The following code
-    # reverses this transformation to be consistent with pi0 which is pretrained in
-    # angular space.
-    #
-    # These values are coming from the Aloha code:
-    # PUPPET_GRIPPER_POSITION_OPEN, PUPPET_GRIPPER_POSITION_CLOSED
-    value = unnormalize(value, min_val=0.01844, max_val=0.05800)
-
-    # This is the inverse of the angular to linear transformation inside the Interbotix code.
-    def linear_to_radian(linear_position, arm_length, horn_radius):
-        value = (horn_radius**2 + linear_position**2 - arm_length**2) / (2 * horn_radius * linear_position)
-        return safe_arcsin(value)
-
-    # The constants are taken from the Interbotix code.
-    value = linear_to_radian(value, arm_length=0.036, horn_radius=0.022)
-
-    # Normalize to [0, 1].
-    # The values 0.4 and 1.5 were measured on an actual Trossen robot.
-    return normalize(value, min_val=0.4, max_val=1.5)
-
-
-def aloha_gripper_from_angular(value):
-    # Convert from the gripper position used by pi0 to the gripper position that is used by Aloha.
-    # Note that the units are still angular but the range is different.
-
-    # The values 0.4 and 1.5 were measured on an actual Trossen robot.
-    value = unnormalize(value, min_val=0.4, max_val=1.5)
-
-    # These values are coming from the Aloha code:
-    # PUPPET_GRIPPER_JOINT_OPEN, PUPPET_GRIPPER_JOINT_CLOSE
-    return normalize(value, min_val=-0.6213, max_val=1.4910)
-
-
-def aloha_gripper_from_angular_inv(value):
-    # Directly inverts the gripper_from_angular function.
-    value = unnormalize(value, min_val=-0.6213, max_val=1.4910)
-    return normalize(value, min_val=0.4, max_val=1.5)
-
-
-class PI0FASTPolicy(PreTrainedPolicy):
-    """Wrapper class around PI0FAST tokenizer and model to train and run inference within LeRobot."""
-
-    config_class = PI0FASTConfig
-    name = "pi0fast"
-
-    def __init__(
-        self,
-        config: PI0FASTConfig,
-        dataset_stats: dict[str, dict[str, Tensor]] | None = None,
-    ):
-        """
-        Args:
-            config: Policy configuration class instance or None, in which case the default instantiation of
-                    the configuration class is used.
-            dataset_stats: Dataset statistics to be used for normalization. If not passed here, it is expected
-                that they will be passed with a call to `load_state_dict` before the policy is used.
-        """
-
-        super().__init__(config)
-        config.validate_features()
-        self.config = config
-
-        self.normalize_inputs = Normalize(config.input_features, config.normalization_mapping, dataset_stats)
-        self.normalize_targets = Normalize(
-            config.output_features, config.normalization_mapping, dataset_stats
-        )
-        self.unnormalize_outputs = Unnormalize(
-            config.output_features, config.normalization_mapping, dataset_stats
-        )
-
-        self.language_tokenizer = AutoProcessor.from_pretrained("google/paligemma-3b-pt-224")
-        self.model = PI0FAST(config)
-
-        self.reset()
-
-    def reset(self):
-        """This should be called whenever the environment is reset."""
-        self._action_queue = deque([], maxlen=self.config.n_action_steps)
-
-    def get_optim_params(self) -> dict:
-        return self.parameters()
-
-    def _pi_aloha_decode_state(self, state):
-        # Flip the joints.
-        for motor_idx in [1, 2, 8, 9]:
-            state[:, motor_idx] *= -1
-        # Reverse the gripper transformation that is being applied by the Aloha runtime.
-        for motor_idx in [6, 13]:
-            state[:, motor_idx] = aloha_gripper_to_angular(state[:, motor_idx])
-        return state
-
-    def _pi_aloha_encode_actions(self, actions):
-        # Flip the joints.
-        for motor_idx in [1, 2, 8, 9]:
-            actions[:, :, motor_idx] *= -1
-        # Reverse the gripper transformation that is being applied by the Aloha runtime.
-        for motor_idx in [6, 13]:
-            actions[:, :, motor_idx] = aloha_gripper_from_angular(actions[:, :, motor_idx])
-        return actions
-
-    def _pi_aloha_encode_actions_inv(self, actions):
-        # Flip the joints again.
-        for motor_idx in [1, 2, 8, 9]:
-            actions[:, :, motor_idx] *= -1
-        # Reverse the gripper transformation that is being applied by the Aloha runtime.
-        for motor_idx in [6, 13]:
-            actions[:, :, motor_idx] = aloha_gripper_from_angular_inv(actions[:, :, motor_idx])
-        return actions
-
-    @torch.no_grad
-    def select_action(self, batch: dict[str, Tensor]) -> Tensor:
-        """Select a single action given environment observations.
-
-        This method wraps `select_actions` in order to return one action at a time for execution in the
-        environment. It works by managing the actions in a queue and only calling `select_actions` when the
-        queue is empty.
-        """
-        self.eval()
-
-        if self.config.adapt_to_pi_aloha:
-            batch[OBS_ROBOT] = self._pi_aloha_decode_state(batch[OBS_ROBOT])
-
-        batch = self.normalize_inputs(batch)
-
-        # Action queue logic for n_action_steps > 1. When the action_queue is depleted, populate it by
-        # querying the policy.
-        if len(self._action_queue) == 0:
-            actions = self.model.generate_actions(batch)
-
-            actions = actions[:, : self.config.n_action_steps]
-
-            original_action_dim = self.config.action_feature.shape[
-                0
-            ]  # self.config.max_action_dim  # self.config.action_feature.shape[0]
-            actions = actions[:, :, :original_action_dim]
-
-            actions = self.unnormalize_outputs({"action": actions})["action"]
-
-            if self.config.adapt_to_pi_aloha:
-                actions = self._pi_aloha_encode_actions(actions)
-
-            # `self.model.forward` returns a (batch_size, n_action_steps, action_dim) tensor, but the queue
-            # effectively has shape (n_action_steps, batch_size, *), hence the transpose.
-            self._action_queue.extend(actions.transpose(0, 1))
-        return self._action_queue.popleft()
-
-    def forward(self, batch: dict[str, Tensor]) -> dict[str, Tensor]:
-        if self.config.adapt_to_pi_aloha:
-            batch[OBS_ROBOT] = self._pi_aloha_decode_state(batch[OBS_ROBOT])
-            batch[ACTION] = self._pi_aloha_encode_actions_inv(batch[ACTION])
-        batch = self.normalize_inputs(batch)
-        batch = self.normalize_targets(batch)
-        loss_dict = self.model.forward(batch)
-        return loss_dict["loss"], loss_dict
-
-
-def block_causal_update_causal_mask(
-    attention_mask,
-    token_type_ids=None,
-    past_key_values=None,
-    cache_position=None,
-    input_tensor=None,
-    attn_implementation: str = "eager",
-    dtype: torch.dtype = "float32",
-):
-    """
-    Update the causal mask during training and generation. It can be customized to different attention masks.
-    """
-    if attn_implementation == "flash_attention_2":
-        if attention_mask is not None and 0.0 in attention_mask:
-            return attention_mask
-        return None
-    using_static_cache = isinstance(past_key_values, StaticCache)
-    min_dtype = torch.finfo(dtype).min
-
-    if input_tensor is None:
-        input_tensor = attention_mask
-
-    inputs_lead_dim, sequence_length = input_tensor.shape[:2]
-
-    if using_static_cache or isinstance(past_key_values, HybridCache):
-        target_length = past_key_values.get_max_cache_shape()
-    else:
-        target_length = (
-            attention_mask.shape[-1]
-            if isinstance(attention_mask, torch.Tensor)
-            else cache_position[0] + sequence_length + 1
-        )
-
-    # Handle precomputed attention masks
-    if attention_mask is not None and attention_mask.dim() == 4:
-        return attention_mask
-
-    # Causal mask initialization
-    causal_mask = torch.full(
-        (sequence_length, target_length), fill_value=min_dtype, dtype=dtype, device=cache_position.device
-    )
-
-    # Standard causal masking (triu ensures tokens can only attend to past)
-    if sequence_length != 1:
-        causal_mask = torch.triu(causal_mask, diagonal=1)
-
-        # Apply block causal mask
-        if token_type_ids is not None:
-            token_type_ids = token_type_ids.to(causal_mask.device).bool()
-            cumsum = torch.cumsum(token_type_ids, dim=1)
-            block_causal_mask = cumsum[:, None, :] <= cumsum[:, :, None]
-
-            # Combine causal_mask with block-wise attention mask
-            causal_mask = torch.where(block_causal_mask, 0.0, causal_mask)
-            causal_mask = causal_mask[:, None, :, :]
-        else:
-            # Apply past cache position constraint
-            causal_mask *= torch.arange(target_length, device=cache_position.device) > cache_position.reshape(
-                -1, 1
-            )
-            causal_mask = causal_mask[None, None, :, :].expand(inputs_lead_dim, 1, -1, -1)
-    else:
-        # Apply past cache position constraint
-        causal_mask *= torch.arange(target_length, device=cache_position.device) > cache_position.reshape(
-            -1, 1
-        )
-        causal_mask = causal_mask[None, None, :, :].expand(inputs_lead_dim, 1, -1, -1)
-
-    if attention_mask is not None:
-        causal_mask = causal_mask.clone()  # Copy to contiguous memory for in-place edits
-        mask_length = attention_mask.shape[-1]
-
-        # Apply padding mask
-        padding_mask = causal_mask[:, :, :, :mask_length] + attention_mask[:, None, None, :].to(
-            causal_mask.device
-        )
-        padding_mask = padding_mask == 0
-        causal_mask[:, :, :, :mask_length] = causal_mask[:, :, :, :mask_length].masked_fill(
-            padding_mask, min_dtype
-        )
-
-    return causal_mask
-
-
-def prepare_inputs_for_generation(
-    # self,
-    input_ids,
-    past_key_values=None,
-    inputs_embeds=None,
-    cache_position=None,
-    position_ids=None,
-    pixel_values=None,
-    attention_mask=None,
-    token_type_ids=None,
-    use_cache=True,
-    num_logits_to_keep=None,
-    labels=None,
-    self=None,
-    **kwargs,
-):
-    # create block causal attention
-    if cache_position[0] > 0 and input_ids.shape[1] > 0:
-        input_tensor = input_ids[:, -1:]
-        new_positions = (
-            torch.ones(
-                (position_ids.shape[0], input_ids.shape[1]),
-                dtype=position_ids.dtype,
-                device=position_ids.device,
-            ).cumsum(-1)
-            + position_ids[:, -1:]
-        )
-        position_ids = torch.cat([position_ids, new_positions], dim=-1)
-    else:
-        input_tensor = inputs_embeds
-    attention_mask = block_causal_update_causal_mask(
-        attention_mask=attention_mask,
-        past_key_values=past_key_values,
-        cache_position=cache_position,
-        input_tensor=input_tensor,
-        token_type_ids=token_type_ids,
-        dtype=self.dtype,
-        attn_implementation=self.config.text_config._attn_implementation,
-    )
-    # Overwritten -- custom `position_ids` and `pixel_values` handling
-    model_inputs = self.language_model.prepare_inputs_for_generation(
-        input_ids,
-        past_key_values=past_key_values,
-        inputs_embeds=inputs_embeds,
-        attention_mask=attention_mask,
-        position_ids=position_ids,
-        cache_position=cache_position,
-        use_cache=use_cache,
-        num_logits_to_keep=num_logits_to_keep,
-        token_type_ids=token_type_ids,
-        **kwargs,
-    )
-
-    # Position_ids in Paligemma are 1-indexed
-    if model_inputs.get("position_ids") is not None:
-        model_inputs["position_ids"] += 1
-    # If we're in cached decoding stage, pixel values should be None because input ids do not contain special image token anymore
-    # Otherwise we need pixel values to be passed to model. NOTE: use_cache=False needs pixel_values always
-    if cache_position[0] == 0:
-        model_inputs["pixel_values"] = pixel_values
-    is_training = token_type_ids is not None and labels is not None
-    if cache_position[0] == 0 and isinstance(past_key_values, HybridCache):
-        input_tensor = inputs_embeds if inputs_embeds is not None else input_ids
-        causal_mask = self._update_causal_mask(
-            attention_mask, token_type_ids, past_key_values, cache_position, input_tensor, is_training
-        )
-        model_inputs["attention_mask"] = causal_mask
-
-    return model_inputs
-
-
-class PI0FAST(nn.Module):
-    def __init__(self, config: PI0FASTConfig):
-        super().__init__()
-        self.config = config
-
-        # TODO: move tokenizers in Policy
-        fast_tokenizer_path = "physical-intelligence/fast"
-        pi0_paligemma_path = "google/paligemma-3b-pt-224"
-        self.paligemma_tokenizer = AutoTokenizer.from_pretrained(pi0_paligemma_path)
-        self.processor = AutoProcessor.from_pretrained(pi0_paligemma_path)
-        self.fast_tokenizer = AutoProcessor.from_pretrained(fast_tokenizer_path, trust_remote_code=True)
-        self.fast_skip_tokens = self.config.fast_skip_tokens
-        self.max_input_seq_len = self.config.max_input_seq_len
-        self.action_horizon = self.config.chunk_size
-        self.action_dim = self.config.action_feature.shape[
-            0
-        ]  # self.config.max_action_dim  # self.config.action_feature.shape[0]
-        precision = config.precision
-        torch_precision = PRECISION.get(precision, torch.float32)
-        self.pad_token_id = (
-            self.paligemma_tokenizer.pad_token_id
-            if hasattr(self.paligemma_tokenizer, "pad_token_id")
-            else self.paligemma_tokenizer.eos_token_id
-        )
-
-        paligemma_config = CONFIG_MAPPING["paligemma"](
-            transformers_version="4.48.1",
-            _vocab_size=257152,
-            bos_token_id=2,
-            eos_token_id=1,
-            hidden_size=2048,
-            image_token_index=257152,
-            model_type="paligemma",
-            pad_token_id=0,
-            projection_dim=2048,
-            text_config={
-                "hidden_activation": "gelu_pytorch_tanh",
-                "hidden_size": 2048,
-                "intermediate_size": 16384,
-                "model_type": "gemma",
-                "num_attention_heads": 8,
-                "num_hidden_layers": 18,
-                "num_image_tokens": 256,
-                "num_key_value_heads": 1,
-                "torch_dtype": precision,
-                "vocab_size": 257152,
-                "_attn_implementation": "eager",
-            },
-            vision_config={
-                "hidden_size": 1152,
-                "intermediate_size": 4304,
-                "model_type": "siglip_vision_model",
-                "num_attention_heads": 16,
-                "num_hidden_layers": 27,
-                "num_image_tokens": 256,
-                "patch_size": 14,
-                "projection_dim": 2048,
-                "projector_hidden_act": "gelu_pytorch_tanh",
-                "torch_dtype": precision,
-                "vision_use_head": False,
-            },
-        )
-        self.pi0_paligemma = PaliGemmaForConditionalGeneration(config=paligemma_config)
-
-        self.pi0_paligemma.prepare_inputs_for_generation = partial(
-            prepare_inputs_for_generation, self=self.pi0_paligemma
-        )
-        # change important stuff in bf16
-        params_to_change_dtype = [
-            "language_model",
-            "vision_tower",
-            "multi_modal",
-        ]
-        for name, param in self.pi0_paligemma.named_parameters():
-            if any(selector in name for selector in params_to_change_dtype):
-                param.data = param.data.to(dtype=torch_precision)
-        self.set_requires_grad()
-        self.image_keys = self.config.image_features.keys()
-        self.ignore_index = self.pi0_paligemma.config.ignore_index
-        self.padding_side = self.config.padding_side
-
-    def set_requires_grad(self):
-        if self.config.freeze_vision_encoder:
-            self.pi0_paligemma.vision_tower.eval()
-            for params in self.pi0_paligemma.vision_tower.parameters():
-                params.requires_grad = False
-        # To avoid unused params issue with distributed training
-        if self.config.freeze_lm_head:
-            for name, params in self.pi0_paligemma.named_parameters():
-                if "embed_tokens" in name:  # lm heads and embedding layer are tied
-                    params.requires_grad = False
-
-    def embed_tokens(self, tokens: torch.Tensor):
-        return self.pi0_paligemma.language_model.model.embed_tokens(tokens)
-
-    def prepare_inputs_for_generation(self, *args, **kwargs):
-        return self.pi0_paligemma.prepare_inputs_for_generation(*args, **kwargs)
-
-    def prepare_images(self, batch):
-        """Preprocess LeRobot batch into Pi0 inputs"""
-        images = []
-        img_masks = []
-        present_img_keys = [key for key in self.image_keys if key in batch]
-        if len(present_img_keys) == 0:
-            raise ValueError(
-                f"All image features are missing from the batch. At least one expected. (batch: {batch.keys()}) (image_features:{self.config.image_features})"
-            )
-
-        # Preprocess image features present in the batch
-        num_empty_cameras = 0
-        for key in self.image_keys:
-            if key in present_img_keys:
-                img = batch[key]
-
-                if self.config.resize_imgs_with_padding is not None:
-                    img = resize_with_pad(
-                        img,
-                        *self.config.resize_imgs_with_padding,
-                        pad_value=0,
-                        interpolate_like_pi=self.config.interpolate_like_pi,
-                    )
-
-                # Normalize from range [0,1] to [-1,1] as expected by siglip
-                img = img * 2.0 - 1.0
-
-                bsize = img.shape[0]
-                device = img.device
-                mask = torch.ones(bsize, dtype=torch.bool, device=device)
-            else:
-                if num_empty_cameras >= self.config.empty_cameras:
-                    continue
-                img = torch.ones_like(img) * -1
-                bsize = img.shape[0]
-                device = img.device
-                mask = torch.ones(bsize, dtype=torch.bool, device=device)
-                num_empty_cameras += 1
-
-            images.append(img)
-            img_masks.append(mask)
-        return images, img_masks
-
-    def normalize_actions(self, actions: torch.Tensor) -> torch.Tensor:
-        mins = actions.amin(dim=(1, 2), keepdim=True)  # [0]
-        maxs = actions.amax(dim=(1, 2), keepdim=True)  # [0]
-        return 2 * (actions - mins) / (maxs - mins + 1e-8) - 1
-
-    def _act_tokens_to_paligemma_tokens(self, tokens: torch.Tensor) -> torch.Tensor:
-        out = self.paligemma_tokenizer.vocab_size - 1 - self.fast_skip_tokens - tokens
-        return out
-
-    def fast_tokenizer_wrapper(self, actions_norm):
-        """
-        A wrapper for self.fast_tokenizer that ensures batch processing,
-        conversion to PyTorch tensors, and returns a dictionary without padding.
-        """
-        batch_tokens = self.fast_tokenizer(actions_norm)
-        fast_out = self.processor.tokenizer.pad({"input_ids": batch_tokens}, return_tensors="pt")
-
-        return fast_out
-
-    def create_token_type_ids(self, padded_mask: torch.Tensor, prefix_len: int) -> torch.Tensor:
-        token_type_ids = torch.zeros_like(padded_mask, dtype=torch.bool)
-        # Compute cumulative sum mask
-        cumsum_mask = (padded_mask != 0).cumsum(dim=1)
-        # Suffix block (everything after prefix_len)
-        suffix_mask = cumsum_mask > prefix_len
-        token_type_ids = suffix_mask
-        return token_type_ids
-
-    def create_input_tokens(self, state, lang_text, actions=None):
-        bsize = state.shape[0]
-        device = state.device
-        bins = torch.linspace(-1, 1, 256 + 1, device=device)[:-1]
-        discretized = torch.bucketize(state, bins) - 1
-        discretized = discretized[:, :32]
-
-        prefix_texts = []
-        state_text = []
-        for txt, disc in zip(lang_text, discretized, strict=False):
-            cleaned = txt.lower().strip().replace("_", " ")
-            state_str = " ".join(str(val.item()) for val in disc)
-            prefix_texts.append(f"Task: {cleaned}, State: {state_str};\n")
-            state_text.append(f"State: {state_str};\n")
-
-        prefix_out = self.paligemma_tokenizer(
-            prefix_texts, add_special_tokens=True, return_tensors="pt", padding="longest", truncation=False
-        )
-        prefix_ids = prefix_out["input_ids"].to(device)
-        prefix_mask = prefix_out["attention_mask"].to(device)
-        prefix_lens = prefix_mask.sum(dim=1)[:, None].cpu()
-
-        if actions is not None:
-            actions_norm = self.normalize_actions(actions)
-            actions_pad = F.pad(
-                actions_norm, (0, max(0, self.config.max_action_dim - actions_norm.shape[2])), value=0
-            )[:, :, : self.config.max_action_dim]
-            fast_out = self.fast_tokenizer_wrapper(
-                actions_pad.cpu(),
-            )
-            act_ids = fast_out["input_ids"]
-            act_mask = fast_out["attention_mask"].to(device)
-
-            act_ids = self._act_tokens_to_paligemma_tokens(act_ids).to(device)
-            # Replace action with 0 to pad tokens
-            act_ids = torch.where(
-                act_ids == self.paligemma_tokenizer.vocab_size - 1 - self.fast_skip_tokens,
-                self.pad_token_id,
-                act_ids,
-            )
-
-            eos_token = torch.tensor(
-                [self.paligemma_tokenizer.eos_token_id], dtype=torch.long, device=device
-            ).expand(bsize, -1)
-            eos_mask = torch.tensor([1], dtype=torch.long, device=device).expand(bsize, -1)
-            bos = self.paligemma_tokenizer("Action: ", add_special_tokens=False, return_tensors="pt")
-            bos_token = bos["input_ids"].expand(act_ids.shape[0], -1).to(device)
-            bos_mask = bos["attention_mask"].expand(act_ids.shape[0], -1).to(device)
-            act_ids = torch.cat([bos_token, act_ids, eos_token], dim=1)
-            act_mask = torch.cat([bos_mask, act_mask, eos_mask], dim=1)
-            act_mask = act_mask.to(device)
-        else:
-            act_ids = torch.empty(bsize, self.pad_token_id, dtype=torch.long, device=device)
-            act_mask = torch.empty(bsize, 0, dtype=torch.long, device=device)
-        final_ids = torch.cat([prefix_ids, act_ids], dim=1)
-
-        final_mask = torch.cat([prefix_mask, act_mask], dim=1)
-        batch_inputs = {"input_ids": final_ids.tolist(), "attention_mask": final_mask.tolist()}
-
-        # Use tokenizer pad function
-        padded_output = self.paligemma_tokenizer.pad(
-            batch_inputs, padding="longest", max_length=180, return_tensors="pt"
-        )
-        padded_mask = padded_output["attention_mask"]
-
-        # define tensor of padding lengths
-        att_mask = (padded_mask != 0).cumsum(dim=1) > prefix_lens
-
-        token_type_ids = self.create_token_type_ids(padded_mask=padded_mask, prefix_len=prefix_lens)
-
-        padded_output["padded_mask"] = padded_output.pop("attention_mask")
-        padded_output["attention_mask"] = att_mask
-        # loss is computed not on prefix, and not on padding
-        padded_output["loss_mask"] = att_mask & padded_output["padded_mask"]
-        padded_output["token_type_ids"] = token_type_ids
-        return padded_output
-
-    def shift_padding_side(
-        self,
-        tokens: torch.Tensor,
-        ar_mask: torch.Tensor,
-        padding_mask: torch.Tensor,
-        loss_mask: torch.Tensor,
-        targets: torch.Tensor,
-        token_type_ids: torch.Tensor,
-        padding_side: str = "right",
-    ) -> tuple[torch.Tensor]:
-        if padding_side not in ["right", "left"]:
-            return tokens, ar_mask, padding_mask, loss_mask, targets, token_type_ids
-
-        new_tokens = torch.empty_like(tokens)
-        new_ar_masks = torch.empty_like(ar_mask)
-        new_padding_mask = torch.empty_like(padding_mask)
-        new_loss_mask = torch.empty_like(loss_mask)
-        new_targets = torch.empty_like(targets)
-        new_token_type_ids = torch.empty_like(token_type_ids)
-        batch_size = tokens.shape[0]
-        for i in range(batch_size):
-            padding_indices = torch.where(padding_mask[i] == 0)[0]
-            non_padding_indices = torch.where(padding_mask[i] == 1)[0]
-            if padding_side == "left":
-                new_indices = torch.cat((padding_indices, non_padding_indices), dim=0)
-            else:
-                new_indices = torch.cat((non_padding_indices, padding_indices), dim=0)
-            new_tokens[i] = tokens[i].index_select(0, new_indices)
-            new_ar_masks[i] = ar_mask[i].index_select(0, new_indices)
-            new_padding_mask[i] = padding_mask[i].index_select(0, new_indices)
-            new_loss_mask[i] = loss_mask[i].index_select(0, new_indices)
-            new_targets[i] = targets[i].index_select(0, new_indices)
-            new_token_type_ids[i] = token_type_ids[i].index_select(0, new_indices)
-
-        return new_tokens, new_ar_masks, new_padding_mask, new_loss_mask, new_targets, new_token_type_ids
-
-    def forward(self, batch: dict[str, Tensor]):
-        device = batch[OBS_ROBOT].device
-        # TODO: keep like this or move to the policy .forward
-        images, img_masks = self.prepare_images(batch)
-
-        padded_outs = self.create_input_tokens(
-            state=batch[OBS_ROBOT],
-            lang_text=batch["task"],
-            actions=batch[ACTION],
-        )
-
-        embs, pad_masks, _, targets, loss_mask, token_type_ids = self.embed_inputs(
-            images,
-            img_masks,
-            padded_outs["input_ids"],
-            padded_outs["padded_mask"],
-            padded_outs["attention_mask"],
-            padded_outs["loss_mask"],
-            padded_outs["token_type_ids"],
-            padding_side=self.padding_side,
-        )
-        position_ids = torch.cumsum(pad_masks, dim=1) - 1
-        token_type_ids = token_type_ids.to(dtype=torch.int64)
-        past_seen_tokens = 0
-        cache_position = torch.arange(past_seen_tokens, past_seen_tokens + embs.shape[1], device=embs.device)
-        pad_masks = block_causal_update_causal_mask(
-            attention_mask=pad_masks,
-            past_key_values=None,
-            cache_position=cache_position,
-            input_tensor=embs,
-            token_type_ids=token_type_ids,
-            dtype=self.pi0_paligemma.dtype,
-            attn_implementation=self.pi0_paligemma.config.text_config._attn_implementation,
-        )
-        outputs = self.pi0_paligemma.forward(
-            input_ids=None,
-            token_type_ids=None,
-            attention_mask=pad_masks,
-            position_ids=position_ids,
-            past_key_values=None,
-            inputs_embeds=embs,
-            use_cache=False,
-            labels=None,
-        )
-
-        logits = outputs.logits
-
-        loss_fct = nn.CrossEntropyLoss(reduction="none")
-
-        # Shift left for next-step prediction
-        logits = logits[:, :-1, :]
-        targets = targets[:, 1:].to(device)  # Shift targets
-        loss_mask = loss_mask[:, 1:].to(device)  # Ensure correct shape
-
-        # Compute per-token loss
-        token_loss = loss_fct(logits.reshape(-1, logits.shape[-1]), targets.reshape(-1))
-
-        # Apply loss mask
-        token_loss = token_loss * loss_mask.reshape(-1)
-
-        # Compute final loss
-        loss = token_loss.sum() / torch.clamp(loss_mask.sum(), min=1)
-
-        # Return loss dictionary
-        loss_dict = {"ce_loss": loss.item(), "loss": loss}
-        return loss_dict
-
-    def decode_actions_with_fast(
-        self,
-        tokens: list[list[int]],
-        *,
-        time_horizon: int | None = None,
-        action_dim: int | None = None,
-        relaxed_decoding: bool = True,
-    ) -> np.array:
-        """
-        Adapt original decoding in FAST to always return actions instead of zeros.
-        """
-        self.time_horizon = (
-            time_horizon or self.fast_tokenizer.time_horizon or self.fast_tokenizer.called_time_horizon
-        )
-        self.action_dim = (
-            action_dim or self.fast_tokenizer.action_dim or self.fast_tokenizer.called_action_dim
-        )
-
-        # Cache the time horizon and action dimension for the next call
-        self.called_time_horizon = self.time_horizon
-        self.called_action_dim = self.action_dim
-
-        assert self.time_horizon is not None and self.action_dim is not None, (
-            "Tokenizer not initialized, call encode() once or pass in time_horizon and action_dim."
-        )
-
-        decoded_actions = []
-        for token in tokens:
-            try:
-                decoded_tokens = self.fast_tokenizer.bpe_tokenizer.decode(token)
-                decoded_dct_coeff = np.array(list(map(ord, decoded_tokens))) + self.fast_tokenizer.min_token
-                if relaxed_decoding:
-                    # Expected sequence length
-                    expected_seq_len = self.time_horizon * self.action_dim
-                    diff = expected_seq_len - decoded_dct_coeff.shape[0]
-                    # Apply truncation if too long
-                    if diff < 0:
-                        decoded_dct_coeff = decoded_dct_coeff[:expected_seq_len]  # Truncate on the right
-                    # Apply padding if too short
-                    elif diff > 0:
-                        decoded_dct_coeff = np.pad(
-                            decoded_dct_coeff, (0, diff), mode="constant", constant_values=0
-                        )
-
-                decoded_dct_coeff = decoded_dct_coeff.reshape(-1, self.action_dim)
-                assert decoded_dct_coeff.shape == (
-                    self.time_horizon,
-                    self.action_dim,
-                ), (
-                    f"Decoded DCT coefficients have shape {decoded_dct_coeff.shape}, expected ({self.time_horizon}, {self.action_dim})"
-                )
-            except Exception as e:
-                print(f"Error decoding tokens: {e}")
-                print(f"Tokens: {token}")
-                decoded_dct_coeff = np.zeros((self.time_horizon, self.action_dim))
-            decoded_actions.append(idct(decoded_dct_coeff / self.fast_tokenizer.scale, axis=0, norm="ortho"))
-        return np.stack(decoded_actions)
-
-    def extract_actions(self, tokens: torch.Tensor, action_horizon: int, action_dim: int) -> torch.Tensor:
-        """
-        Extracts actions from predicted output tokens using the FAST model.
-
-        Args:
-            tokens (torch.Tensor): The input tensor of tokenized outputs.
-            action_horizon (int): The number of timesteps for actions.
-            action_dim (int): The dimensionality of each action.
-
-        Returns:
-            torch.Tensor: The extracted actions as a tensor of shape (action_horizon, action_dim).
-        """
-        # Decode predicted output tokens
-        decoded_tokens = self.paligemma_tokenizer.batch_decode(tokens, skip_special_tokens=True)
-        cleaned_tokens = [
-            tokens_sequence.replace("Action:", "").replace(":", "").strip().split("|")[0].strip()
-            for tokens_sequence in decoded_tokens
-        ]
-        raw_action_tokens = [
-            self.processor.tokenizer.encode(sample_tokens, return_tensors="pt", padding=False)
-            for sample_tokens in cleaned_tokens
-        ]  # something like this should be robust #looks good
-        action_tokens = [
-            self._act_tokens_to_paligemma_tokens(raw_action_token) for raw_action_token in raw_action_tokens
-        ]
-        # returns the tensor of decoded actions per sample in a list
-        decoded_actions = [
-            torch.tensor(
-                self.decode_actions_with_fast(
-                    tok.tolist(),
-                    time_horizon=action_horizon,
-                    action_dim=action_dim,
-                    relaxed_decoding=self.config.relaxed_action_decoding,
-                ),
-                device=tokens.device,
-            ).squeeze(0)
-            for tok in action_tokens
-        ]
-
-        return torch.stack(
-            decoded_actions,
-            dim=0,
-        )
-
-    def generate_actions(self, batch: dict[str, Tensor]):
-        # TODO: keep like this or move to the policy .forward
-        images, img_masks = self.prepare_images(batch)
-
-        padded_outs = self.create_input_tokens(state=batch[OBS_ROBOT], lang_text=batch["task"], actions=None)
-        embs, pad_masks, att_masks2, targets, loss_mask, token_type_ids = self.embed_inputs(
-            images,
-            img_masks,
-            padded_outs["input_ids"],
-            padded_outs["padded_mask"],
-            padded_outs["attention_mask"],
-            padded_outs["loss_mask"],
-            padded_outs["token_type_ids"],
-            padding_side="left",
-        )
-        token_type_ids = token_type_ids.to(dtype=torch.int64)
-        prefix_position_ids = torch.cumsum(pad_masks, dim=1) - 1
-        output_tokens = self.pi0_paligemma.generate(
-            input_ids=None,
-            attention_mask=pad_masks,
-            position_ids=prefix_position_ids,
-            past_key_values=None,
-            inputs_embeds=embs,
-            use_cache=self.config.use_cache,
-            max_new_tokens=self.config.max_decoding_steps,
-            do_sample=False,
-            num_beams=1,
-            token_type_ids=token_type_ids,
-        )
-        actions = self.extract_actions(output_tokens, self.action_horizon, self.action_dim)
-        return actions
-
-    def embed_image(self, image: torch.Tensor):
-        return self.pi0_paligemma.get_image_features(image)
-
-    def embed_inputs(
-        self,
-        images,
-        img_masks,
-        tokens,
-        pad_mask,
-        ar_mask,
-        loss_mask,
-        token_type_ids,
-        padding_side: str = "right",
-    ):
-        # TODO: avoid list in python and torch.cat ; prefer pre-allocation with torch.empty
-        # images are a list of same size
-        # vectorizing everything!
-        device = images[0].device
-        image_embedding_dim = images[0].shape[-1]  # TODO should be from self.config
-        all_images = torch.stack(images, dim=1).to(device)
-        b, n, c, h, w = all_images.shape
-        all_images = all_images.view(b * n, c, h, w)
-        embedded = self.embed_image(all_images).to(device)
-        b_n, p, image_embedding_dim = embedded.shape  # Extract current dimensions
-        m = b_n // b  # Compute the number of images per sample dynamically
-
-        # Reshape dynamically
-        embedded = embedded.view(b, m, p, image_embedding_dim)
-        tokens_embs = self.embed_tokens(tokens.to(device))
-
-        img_masks = torch.stack(img_masks, dim=1).unsqueeze(-1).to(device)
-        num_img_emb = embedded.shape[2]
-        img_pad_masks = img_masks.repeat(1, 1, num_img_emb).view(b, -1)
-        img_att_masks = torch.zeros((b, n, num_img_emb), dtype=torch.long, device=device).reshape(b, -1)
-
-        image_target_tokens = (
-            torch.ones((b, n, num_img_emb), dtype=torch.long, device=device) * self.pad_token_id
-        ).reshape(b, -1)
-        image_loss_mask = torch.zeros((b, n, num_img_emb), dtype=torch.long, device=device).reshape(b, -1)
-
-        embedded = embedded.reshape(b, n * num_img_emb, image_embedding_dim)  # Shape: (B, N*P, D)
-
-        embs = torch.cat([embedded, tokens_embs], dim=1).to(device)
-        pad_masks = torch.cat([img_pad_masks, pad_mask.to(device)], dim=1)
-        att_masks = torch.cat([img_att_masks, ar_mask.to(device)], dim=1)
-        loss_masks = torch.cat([image_loss_mask, loss_mask.to(device)], dim=1)
-        targets = torch.cat([image_target_tokens, tokens.to(device)], dim=1)
-        token_type_ids = torch.cat([img_att_masks, token_type_ids.to(device)], dim=1)
-
-        # Shift pad tokens to the left (.generate()) or right (.train())
-        embs, att_masks, pad_masks, loss_masks, targets, token_type_ids = self.shift_padding_side(
-            embs, att_masks, pad_masks, loss_masks, targets, token_type_ids, padding_side=padding_side
-        )
-
-        targets = torch.where(targets == self.pad_token_id, self.ignore_index, targets)
-        return embs, pad_masks, att_masks, targets, loss_masks, token_type_ids
-
-
-def resize_with_pad(img, width, height, pad_value=0, interpolate_like_pi=True):
-    # assume no-op when width height fits already
-    if img.ndim != 4:
-        raise ValueError(f"(b,c,h,w) expected, but {img.shape}")
-
-    cur_height, cur_width = img.shape[2:]
-
-    ratio = max(cur_width / width, cur_height / height)
-    resized_height = int(cur_height / ratio)
-    resized_width = int(cur_width / ratio)
-
-    if interpolate_like_pi:
-        img = (img * 255.0).to(dtype=torch.uint8)
-        img = img.permute(0, 2, 3, 1)
-        original_device = img.device
-        img = img.to(device="cpu").numpy()
-        imgs = []
-        for sub_img in img:
-            sub_img = Image.fromarray(sub_img)
-            resized_img = sub_img.resize((resized_width, resized_height), resample=2)
-            resized_img = torch.from_numpy(np.array(resized_img))
-            imgs.append(resized_img)
-        img = torch.stack(imgs, dim=0)
-        img = img.permute(0, 3, 1, 2)
-        resized_img = img.to(device=original_device, dtype=torch.float32) / 255.0
-    else:
-        resized_img = F.interpolate(
-            img, size=(resized_height, resized_width), mode="bilinear", align_corners=False
-        )
-
-    pad_height = max(0, int(height - resized_height))
-    pad_width = max(0, int(width - resized_width))
-
-    # pad on left and top of image
-    padded_img = F.pad(resized_img, (pad_width, 0, pad_height, 0), value=pad_value)
-    return padded_img

lerobot/common/policies/sac/configuration_sac.py

@@ -0,0 +1,229 @@
+#!/usr/bin/env python
+
+# Copyright 2024 The HuggingFace Inc. team.
+# All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from dataclasses import dataclass, field
+
+from lerobot.common.optim.optimizers import MultiAdamConfig
+from lerobot.configs.policies import PreTrainedConfig
+from lerobot.configs.types import NormalizationMode
+
+
+@dataclass
+class ConcurrencyConfig:
+    actor: str = "threads"
+    learner: str = "threads"
+
+
+@dataclass
+class ActorLearnerConfig:
+    learner_host: str = "127.0.0.1"
+    learner_port: int = 50051
+    policy_parameters_push_frequency: int = 4
+
+
+@dataclass
+class CriticNetworkConfig:
+    hidden_dims: list[int] = field(default_factory=lambda: [256, 256])
+    activate_final: bool = True
+    final_activation: str | None = None
+
+
+@dataclass
+class ActorNetworkConfig:
+    hidden_dims: list[int] = field(default_factory=lambda: [256, 256])
+    activate_final: bool = True
+
+
+@dataclass
+class PolicyConfig:
+    use_tanh_squash: bool = True
+    log_std_min: float = 1e-5
+    log_std_max: float = 10.0
+    init_final: float = 0.05
+
+
+@PreTrainedConfig.register_subclass("sac")
+@dataclass
+class SACConfig(PreTrainedConfig):
+    """Soft Actor-Critic (SAC) configuration.
+
+    SAC is an off-policy actor-critic deep RL algorithm based on the maximum entropy
+    reinforcement learning framework. It learns a policy and a Q-function simultaneously
+    using experience collected from the environment.
+
+    This configuration class contains all the parameters needed to define a SAC agent,
+    including network architectures, optimization settings, and algorithm-specific
+    hyperparameters.
+
+    Args:
+        actor_network: Configuration for the actor network architecture.
+        critic_network: Configuration for the critic network architecture.
+        policy: Configuration for the policy parameters.
+        n_obs_steps: Number of observation steps to consider.
+        normalization_mapping: Mapping of feature types to normalization modes.
+        dataset_stats: Statistics for normalizing different types of inputs.
+        input_features: Dictionary of input features with their types and shapes.
+        output_features: Dictionary of output features with their types and shapes.
+        camera_number: Number of cameras used for visual observations.
+        device: Device to run the model on (e.g., "cuda", "cpu").
+        storage_device: Device to store the model on.
+        vision_encoder_name: Name of the vision encoder model.
+        freeze_vision_encoder: Whether to freeze the vision encoder during training.
+        image_encoder_hidden_dim: Hidden dimension size for the image encoder.
+        shared_encoder: Whether to use a shared encoder for actor and critic.
+        num_discrete_actions: Number of discrete actions, eg for gripper actions.
+        image_embedding_pooling_dim: Dimension of the image embedding pooling.
+        concurrency: Configuration for concurrency settings.
+        actor_learner: Configuration for actor-learner architecture.
+        online_steps: Number of steps for online training.
+        online_env_seed: Seed for the online environment.
+        online_buffer_capacity: Capacity of the online replay buffer.
+        offline_buffer_capacity: Capacity of the offline replay buffer.
+        async_prefetch: Whether to use asynchronous prefetching for the buffers.
+        online_step_before_learning: Number of steps before learning starts.
+        policy_update_freq: Frequency of policy updates.
+        discount: Discount factor for the SAC algorithm.
+        temperature_init: Initial temperature value.
+        num_critics: Number of critics in the ensemble.
+        num_subsample_critics: Number of subsampled critics for training.
+        critic_lr: Learning rate for the critic network.
+        actor_lr: Learning rate for the actor network.
+        temperature_lr: Learning rate for the temperature parameter.
+        critic_target_update_weight: Weight for the critic target update.
+        utd_ratio: Update-to-data ratio for the UTD algorithm.
+        state_encoder_hidden_dim: Hidden dimension size for the state encoder.
+        latent_dim: Dimension of the latent space.
+        target_entropy: Target entropy for the SAC algorithm.
+        use_backup_entropy: Whether to use backup entropy for the SAC algorithm.
+        grad_clip_norm: Gradient clipping norm for the SAC algorithm.
+    """
+
+    normalization_mapping: dict[str, NormalizationMode] = field(
+        default_factory=lambda: {
+            "VISUAL": NormalizationMode.MEAN_STD,
+            "STATE": NormalizationMode.MIN_MAX,
+            "ENV": NormalizationMode.MIN_MAX,
+            "ACTION": NormalizationMode.MIN_MAX,
+        }
+    )
+
+    dataset_stats: dict[str, dict[str, list[float]]] | None = field(
+        default_factory=lambda: {
+            "observation.image": {
+                "mean": [0.485, 0.456, 0.406],
+                "std": [0.229, 0.224, 0.225],
+            },
+            "observation.state": {
+                "min": [0.0, 0.0],
+                "max": [1.0, 1.0],
+            },
+            "action": {
+                "min": [0.0, 0.0, 0.0],
+                "max": [1.0, 1.0, 1.0],
+            },
+        }
+    )
+
+    # Architecture specifics
+    camera_number: int = 1
+    device: str = "cuda"
+    storage_device: str = "cpu"
+    # Set to "helper2424/resnet10" for hil serl
+    vision_encoder_name: str | None = None
+    freeze_vision_encoder: bool = True
+    image_encoder_hidden_dim: int = 32
+    shared_encoder: bool = True
+    num_discrete_actions: int | None = None
+    image_embedding_pooling_dim: int = 8
+
+    # Training parameter
+    online_steps: int = 1000000
+    online_env_seed: int = 10000
+    online_buffer_capacity: int = 100000
+    offline_buffer_capacity: int = 100000
+    async_prefetch: bool = False
+    online_step_before_learning: int = 100
+    policy_update_freq: int = 1
+
+    # SAC algorithm parameters
+    discount: float = 0.99
+    temperature_init: float = 1.0
+    num_critics: int = 2
+    num_subsample_critics: int | None = None
+    critic_lr: float = 3e-4
+    actor_lr: float = 3e-4
+    temperature_lr: float = 3e-4
+    critic_target_update_weight: float = 0.005
+    utd_ratio: int = 1  # If you want enable utd_ratio, you need to set it to >1
+    state_encoder_hidden_dim: int = 256
+    latent_dim: int = 256
+    target_entropy: float | None = None
+    use_backup_entropy: bool = True
+    grad_clip_norm: float = 40.0
+
+    # Network configuration
+    critic_network_kwargs: CriticNetworkConfig = field(default_factory=CriticNetworkConfig)
+    actor_network_kwargs: ActorNetworkConfig = field(default_factory=ActorNetworkConfig)
+    policy_kwargs: PolicyConfig = field(default_factory=PolicyConfig)
+    grasp_critic_network_kwargs: CriticNetworkConfig = field(default_factory=CriticNetworkConfig)
+    actor_learner_config: ActorLearnerConfig = field(default_factory=ActorLearnerConfig)
+    concurrency: ConcurrencyConfig = field(default_factory=ConcurrencyConfig)
+
+    def __post_init__(self):
+        super().__post_init__()
+        # Any validation specific to SAC configuration
+
+    def get_optimizer_preset(self) -> MultiAdamConfig:
+        return MultiAdamConfig(
+            weight_decay=0.0,
+            optimizer_groups={
+                "actor": {"lr": self.actor_lr},
+                "critic": {"lr": self.critic_lr},
+                "temperature": {"lr": self.temperature_lr},
+            },
+        )
+
+    def get_scheduler_preset(self) -> None:
+        return None
+
+    def validate_features(self) -> None:
+        has_image = any(key.startswith("observation.image") for key in self.input_features)
+        has_state = "observation.state" in self.input_features
+
+        if not (has_state or has_image):
+            raise ValueError(
+                "You must provide either 'observation.state' or an image observation (key starting with 'observation.image') in the input features"
+            )
+
+        if "action" not in self.output_features:
+            raise ValueError("You must provide 'action' in the output features")
+
+    @property
+    def image_features(self) -> list[str]:
+        return [key for key in self.input_features if "image" in key]
+
+    @property
+    def observation_delta_indices(self) -> list:
+        return None
+
+    @property
+    def action_delta_indices(self) -> list:
+        return None  # SAC typically predicts one action at a time
+
+    @property
+    def reward_delta_indices(self) -> None:
+        return None

lerobot/common/policies/smolvla/configuration_smolvla.py

@@ -1,154 +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.
-
-from dataclasses import dataclass, field
-
-from lerobot.common.optim.optimizers import AdamWConfig
-from lerobot.common.optim.schedulers import (
-    CosineDecayWithWarmupSchedulerConfig,
-)
-from lerobot.configs.policies import PreTrainedConfig
-from lerobot.configs.types import FeatureType, NormalizationMode, PolicyFeature
-
-
-@PreTrainedConfig.register_subclass("smolvla")
-@dataclass
-class SmolVLAConfig(PreTrainedConfig):
-    # Input / output structure.
-    n_obs_steps: int = 1
-    chunk_size: int = 50
-    n_action_steps: int = 50
-
-    normalization_mapping: dict[str, NormalizationMode] = field(
-        default_factory=lambda: {
-            "VISUAL": NormalizationMode.IDENTITY,
-            "STATE": NormalizationMode.MEAN_STD,
-            "ACTION": NormalizationMode.MEAN_STD,
-        }
-    )
-
-    # Shorter state and action vectors will be padded
-    max_state_dim: int = 32
-    max_action_dim: int = 32
-
-    # Image preprocessing
-    resize_imgs_with_padding: tuple[int, int] = (512, 512)
-
-    # Add empty images. Used by smolvla_aloha_sim which adds the empty
-    # left and right wrist cameras in addition to the top camera.
-    empty_cameras: int = 0
-
-    # Converts the joint and gripper values from the standard Aloha space to
-    # the space used by the pi internal runtime which was used to train the base model.
-    adapt_to_pi_aloha: bool = False
-
-    # Converts joint dimensions to deltas with respect to the current state before passing to the model.
-    # Gripper dimensions will remain in absolute values.
-    use_delta_joint_actions_aloha: bool = False
-
-    # Tokenizer
-    tokenizer_max_length: int = 48
-
-    # Decoding
-    num_steps: int = 10
-
-    # Attention utils
-    use_cache: bool = True
-
-    # Finetuning settings
-    freeze_vision_encoder: bool = True
-    train_expert_only: bool = True
-    train_state_proj: bool = True
-
-    # Training presets
-    optimizer_lr: float = 1e-4
-    optimizer_betas: tuple[float, float] = (0.9, 0.95)
-    optimizer_eps: float = 1e-8
-    optimizer_weight_decay: float = 1e-10
-    optimizer_grad_clip_norm: float = 10
-
-    scheduler_warmup_steps: int = 1_000
-    scheduler_decay_steps: int = 30_000
-    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 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.
-
-    attention_mode: str = "cross_attn"
-
-    prefix_length: int = -1
-
-    pad_language_to: str = "longest"  # "max_length"
-
-    num_expert_layers: int = -1  # Less or equal to 0 is the default where the action expert has the same number of layers of VLM. Otherwise the expert have less layers.
-    num_vlm_layers: int = 16  # Number of layers used in the VLM (first num_vlm_layers layers)
-    self_attn_every_n_layers: int = 2  # Interleave SA layers each self_attn_every_n_layers
-    expert_width_multiplier: float = 0.75  # The action expert hidden size (wrt to the VLM)
-
-    min_period: float = 4e-3  # sensitivity range for the timestep used in sine-cosine positional encoding
-    max_period: float = 4.0
-
-    def __post_init__(self):
-        super().__post_init__()
-
-        """Input validation (not exhaustive)."""
-        if self.n_action_steps > self.chunk_size:
-            raise ValueError(
-                f"The chunk size is the upper bound for the number of action steps per model invocation. Got "
-                f"{self.n_action_steps} for `n_action_steps` and {self.chunk_size} for `chunk_size`."
-            )
-        if self.use_delta_joint_actions_aloha:
-            raise NotImplementedError(
-                "`use_delta_joint_actions_aloha` is used by smolvla for aloha real models. It is not ported yet in LeRobot."
-            )
-
-    def validate_features(self) -> None:
-        for i in range(self.empty_cameras):
-            key = f"observation.images.empty_camera_{i}"
-            empty_camera = PolicyFeature(
-                type=FeatureType.VISUAL,
-                shape=(3, 480, 640),
-            )
-            self.input_features[key] = empty_camera
-
-    def get_optimizer_preset(self) -> AdamWConfig:
-        return AdamWConfig(
-            lr=self.optimizer_lr,
-            betas=self.optimizer_betas,
-            eps=self.optimizer_eps,
-            weight_decay=self.optimizer_weight_decay,
-            grad_clip_norm=self.optimizer_grad_clip_norm,
-        )
-
-    def get_scheduler_preset(self):
-        return CosineDecayWithWarmupSchedulerConfig(
-            peak_lr=self.optimizer_lr,
-            decay_lr=self.scheduler_decay_lr,
-            num_warmup_steps=self.scheduler_warmup_steps,
-            num_decay_steps=self.scheduler_decay_steps,
-        )
-
-    @property
-    def observation_delta_indices(self) -> list:
-        return [0]
-
-    @property
-    def action_delta_indices(self) -> list:
-        return list(range(self.chunk_size))
-
-    @property
-    def reward_delta_indices(self) -> None:
-        return None

lerobot/common/policies/smolvla/modeling_smolvla.py

@@ -1,801 +0,0 @@
-#!/usr/bin/env python
-
-# Copyright 2025 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.
-
-"""
-SmolVLA:
-
-[Paper](https://huggingface.co/papers/2506.01844)
-
-Designed by Hugging Face.
-
-Install smolvla extra dependencies:
-```bash
-pip install -e ".[smolvla]"
-```
-
-Example of finetuning the smolvla pretrained model (`smolvla_base`):
-```bash
-python lerobot/scripts/train.py \
---policy.path=lerobot/smolvla_base \
---dataset.repo_id=danaaubakirova/svla_so100_task1_v3 \
---batch_size=64 \
---steps=200000
-```
-
-Example of finetuning a smolVLA. SmolVLA is composed of a pretrained VLM,
-and an action expert.
-```bash
-python lerobot/scripts/train.py \
---policy.type=smolvla \
---dataset.repo_id=danaaubakirova/svla_so100_task1_v3 \
---batch_size=64 \
---steps=200000
-```
-
-Example of using the smolvla pretrained model outside LeRobot training framework:
-```python
-policy = SmolVLAPolicy.from_pretrained("lerobot/smolvla_base")
-```
-
-"""
-
-import math
-from collections import deque
-
-import torch
-import torch.nn.functional as F  # noqa: N812
-from torch import Tensor, nn
-from transformers import AutoProcessor
-
-from lerobot.common.constants import ACTION, OBS_ROBOT
-from lerobot.common.policies.normalize import (
-    Normalize,
-    Unnormalize,
-)
-from lerobot.common.policies.pretrained import PreTrainedPolicy
-from lerobot.common.policies.smolvla.configuration_smolvla import SmolVLAConfig
-from lerobot.common.policies.smolvla.smolvlm_with_expert import SmolVLMWithExpertModel
-from lerobot.common.policies.utils import (
-    populate_queues,
-)
-from lerobot.common.utils.utils import get_safe_dtype
-
-
-def create_sinusoidal_pos_embedding(
-    time: torch.tensor, dimension: int, min_period: float, max_period: float, device="cpu"
-) -> Tensor:
-    """Computes sine-cosine positional embedding vectors for scalar positions."""
-    if dimension % 2 != 0:
-        raise ValueError(f"dimension ({dimension}) must be divisible by 2")
-
-    if time.ndim != 1:
-        raise ValueError("The time tensor is expected to be of shape `(batch_size, )`.")
-
-    dtype = get_safe_dtype(torch.float64, device.type)
-    fraction = torch.linspace(0.0, 1.0, dimension // 2, dtype=dtype, device=device)
-    period = min_period * (max_period / min_period) ** fraction
-
-    # Compute the outer product
-    scaling_factor = 1.0 / period * 2 * math.pi
-    sin_input = scaling_factor[None, :] * time[:, None]
-    pos_emb = torch.cat([torch.sin(sin_input), torch.cos(sin_input)], dim=1)
-    return pos_emb
-
-
-def sample_beta(alpha, beta, bsize, device):
-    gamma1 = torch.empty((bsize,), device=device).uniform_(0, 1).pow(1 / alpha)
-    gamma2 = torch.empty((bsize,), device=device).uniform_(0, 1).pow(1 / beta)
-    return gamma1 / (gamma1 + gamma2)
-
-
-def make_att_2d_masks(pad_masks, att_masks):
-    """Copied from big_vision.
-
-    Tokens can attend to valid inputs tokens which have a cumulative mask_ar
-    smaller or equal to theirs. This way `mask_ar` int[B, N] can be used to
-    setup several types of attention, for example:
-
-      [[1 1 1 1 1 1]]: pure causal attention.
-
-      [[0 0 0 1 1 1]]: prefix-lm attention. The first 3 tokens can attend between
-          themselves and the last 3 tokens have a causal attention. The first
-          entry could also be a 1 without changing behaviour.
-
-      [[1 0 1 0 1 0 0 1 0 0]]: causal attention between 4 blocks. Tokens of a
-          block can attend all previous blocks and all tokens on the same block.
-
-    Args:
-      input_mask: bool[B, N] true if its part of the input, false if padding.
-      mask_ar: int32[B, N] mask that's 1 where previous tokens cannot depend on
-        it and 0 where it shares the same attention mask as the previous token.
-    """
-    if att_masks.ndim != 2:
-        raise ValueError(att_masks.ndim)
-    if pad_masks.ndim != 2:
-        raise ValueError(pad_masks.ndim)
-
-    cumsum = torch.cumsum(att_masks, dim=1)
-    att_2d_masks = cumsum[:, None, :] <= cumsum[:, :, None]
-    pad_2d_masks = pad_masks[:, None, :] * pad_masks[:, :, None]
-    att_2d_masks = att_2d_masks & pad_2d_masks
-    return att_2d_masks
-
-
-def resize_with_pad(img, width, height, pad_value=-1):
-    # assume no-op when width height fits already
-    if img.ndim != 4:
-        raise ValueError(f"(b,c,h,w) expected, but {img.shape}")
-
-    cur_height, cur_width = img.shape[2:]
-
-    ratio = max(cur_width / width, cur_height / height)
-    resized_height = int(cur_height / ratio)
-    resized_width = int(cur_width / ratio)
-    resized_img = F.interpolate(
-        img, size=(resized_height, resized_width), mode="bilinear", align_corners=False
-    )
-
-    pad_height = max(0, int(height - resized_height))
-    pad_width = max(0, int(width - resized_width))
-
-    # pad on left and top of image
-    padded_img = F.pad(resized_img, (pad_width, 0, pad_height, 0), value=pad_value)
-    return padded_img
-
-
-def pad_vector(vector, new_dim):
-    """Can be (batch_size x sequence_length x features_dimension)
-    or (batch_size x features_dimension)
-    """
-    if vector.shape[-1] == new_dim:
-        return vector
-    shape = list(vector.shape)
-    current_dim = shape[-1]
-    shape[-1] = new_dim
-    new_vector = torch.zeros(*shape, dtype=vector.dtype, device=vector.device)
-    new_vector[..., :current_dim] = vector
-    return new_vector
-
-
-def normalize(x, min_val, max_val):
-    return (x - min_val) / (max_val - min_val)
-
-
-def unnormalize(x, min_val, max_val):
-    return x * (max_val - min_val) + min_val
-
-
-def safe_arcsin(value):
-    # This ensures that the input stays within
-    # [−1,1] to avoid invalid values for arcsin
-    return torch.arcsin(torch.clamp(value, -1.0, 1.0))
-
-
-def aloha_gripper_to_angular(value):
-    # Aloha transforms the gripper positions into a linear space. The following code
-    # reverses this transformation to be consistent with smolvla which is pretrained in
-    # angular space.
-    #
-    # These values are coming from the Aloha code:
-    # PUPPET_GRIPPER_POSITION_OPEN, PUPPET_GRIPPER_POSITION_CLOSED
-    value = unnormalize(value, min_val=0.01844, max_val=0.05800)
-
-    # This is the inverse of the angular to linear transformation inside the Interbotix code.
-    def linear_to_radian(linear_position, arm_length, horn_radius):
-        value = (horn_radius**2 + linear_position**2 - arm_length**2) / (2 * horn_radius * linear_position)
-        return safe_arcsin(value)
-
-    # The constants are taken from the Interbotix code.
-    value = linear_to_radian(value, arm_length=0.036, horn_radius=0.022)
-
-    # Normalize to [0, 1].
-    # The values 0.4 and 1.5 were measured on an actual Trossen robot.
-    return normalize(value, min_val=0.4, max_val=1.5)
-
-
-def aloha_gripper_from_angular(value):
-    # Convert from the gripper position used by smolvla to the gripper position that is used by Aloha.
-    # Note that the units are still angular but the range is different.
-
-    # The values 0.4 and 1.5 were measured on an actual Trossen robot.
-    value = unnormalize(value, min_val=0.4, max_val=1.5)
-
-    # These values are coming from the Aloha code:
-    # PUPPET_GRIPPER_JOINT_OPEN, PUPPET_GRIPPER_JOINT_CLOSE
-    return normalize(value, min_val=-0.6213, max_val=1.4910)
-
-
-def aloha_gripper_from_angular_inv(value):
-    # Directly inverts the gripper_from_angular function.
-    value = unnormalize(value, min_val=-0.6213, max_val=1.4910)
-    return normalize(value, min_val=0.4, max_val=1.5)
-
-
-class SmolVLAPolicy(PreTrainedPolicy):
-    """Wrapper class around VLAFlowMatching model to train and run inference within LeRobot."""
-
-    config_class = SmolVLAConfig
-    name = "smolvla"
-
-    def __init__(
-        self,
-        config: SmolVLAConfig,
-        dataset_stats: dict[str, dict[str, Tensor]] | None = None,
-    ):
-        """
-        Args:
-            config: Policy configuration class instance or None, in which case the default instantiation of
-                    the configuration class is used.
-            dataset_stats: Dataset statistics to be used for normalization. If not passed here, it is expected
-                that they will be passed with a call to `load_state_dict` before the policy is used.
-        """
-
-        super().__init__(config)
-        config.validate_features()
-        self.config = config
-        self.normalize_inputs = Normalize(config.input_features, config.normalization_mapping, dataset_stats)
-        self.normalize_targets = Normalize(
-            config.output_features, config.normalization_mapping, dataset_stats
-        )
-        self.unnormalize_outputs = Unnormalize(
-            config.output_features, config.normalization_mapping, dataset_stats
-        )
-
-        self.language_tokenizer = AutoProcessor.from_pretrained(self.config.vlm_model_name).tokenizer
-        self.model = VLAFlowMatching(config)
-        self.reset()
-
-    def reset(self):
-        """This should be called whenever the environment is reset."""
-        self._queues = {
-            ACTION: deque(maxlen=self.config.n_action_steps),
-        }
-
-    def get_optim_params(self) -> dict:
-        return self.parameters()
-
-    @torch.no_grad
-    def select_action(self, batch: dict[str, Tensor], noise: Tensor | None = None) -> Tensor:
-        """Select a single action given environment observations.
-
-        This method wraps `select_actions` in order to return one action at a time for execution in the
-        environment. It works by managing the actions in a queue and only calling `select_actions` when the
-        queue is empty.
-        """
-        self.eval()
-
-        if self.config.adapt_to_pi_aloha:
-            batch[OBS_ROBOT] = self._pi_aloha_decode_state(batch[OBS_ROBOT])
-
-        batch = self.normalize_inputs(batch)
-
-        self._queues = populate_queues(self._queues, batch, exclude_keys=[ACTION])
-        # Action queue logic for n_action_steps > 1. When the action_queue is depleted, populate it by
-        # querying the policy.
-        if len(self._queues[ACTION]) == 0:
-            for k in batch:
-                if k in self._queues:
-                    batch[k] = torch.stack(list(self._queues[k]), dim=1)
-            images, img_masks = self.prepare_images(batch)
-            state = self.prepare_state(batch)
-            lang_tokens, lang_masks = self.prepare_language(batch)
-
-            actions = self.model.sample_actions(
-                images, img_masks, lang_tokens, lang_masks, state, noise=noise
-            )
-            # Unpad actions
-            original_action_dim = self.config.action_feature.shape[0]
-            actions = actions[:, :, :original_action_dim]
-
-            actions = self.unnormalize_outputs({"action": actions})["action"]
-
-            if self.config.adapt_to_pi_aloha:
-                actions = self._pi_aloha_encode_actions(actions)
-
-            # `self.model.forward` returns a (batch_size, n_action_steps, action_dim) tensor, but the queue
-            # effectively has shape (n_action_steps, batch_size, *), hence the transpose.
-            self._queues[ACTION].extend(actions.transpose(0, 1)[: self.config.n_action_steps])
-        return self._queues[ACTION].popleft()
-
-    def forward(self, batch: dict[str, Tensor], noise=None, time=None) -> dict[str, Tensor]:
-        """Do a full training forward pass to compute the loss"""
-        if self.config.adapt_to_pi_aloha:
-            batch[OBS_ROBOT] = self._pi_aloha_decode_state(batch[OBS_ROBOT])
-            batch[ACTION] = self._pi_aloha_encode_actions_inv(batch[ACTION])
-        batch = self.normalize_inputs(batch)
-        batch = self.normalize_targets(batch)
-        images, img_masks = self.prepare_images(batch)
-        state = self.prepare_state(batch)
-        lang_tokens, lang_masks = self.prepare_language(batch)
-        actions = self.prepare_action(batch)
-        actions_is_pad = batch.get("actions_id_pad")
-        loss_dict = {}
-        losses = self.model.forward(images, img_masks, lang_tokens, lang_masks, state, actions, noise, time)
-        loss_dict["losses_after_forward"] = losses.clone()
-
-        if actions_is_pad is not None:
-            in_episode_bound = ~actions_is_pad
-            losses = losses * in_episode_bound.unsqueeze(-1)
-            loss_dict["losses_after_in_ep_bound"] = losses.clone()
-
-        # Remove padding
-        losses = losses[:, :, : self.config.max_action_dim]
-        loss_dict["losses_after_rm_padding"] = losses.clone()
-
-        # For backward pass
-        loss = losses.mean()
-        # For backward pass
-        loss_dict["loss"] = loss
-        return loss, loss_dict
-
-    def prepare_images(self, batch):
-        """Apply SmolVLA preprocessing to the images, like resizing to 224x224 and padding to keep aspect ratio, and
-        convert pixel range from [0.0, 1.0] to [-1.0, 1.0] as requested by SigLIP.
-        """
-        images = []
-        img_masks = []
-        present_img_keys = [key for key in self.config.image_features if key in batch]
-        missing_img_keys = [key for key in self.config.image_features if key not in batch]
-
-        if len(present_img_keys) == 0:
-            raise ValueError(
-                f"All image features are missing from the batch. At least one expected. (batch: {batch.keys()}) (image_features:{self.config.image_features})"
-            )
-        # Preprocess image features present in the batch
-        for key in present_img_keys:
-            img = batch[key][:, -1, :, :, :] if batch[key].ndim == 5 else batch[key]
-            if self.config.resize_imgs_with_padding is not None:
-                img = resize_with_pad(img, *self.config.resize_imgs_with_padding, pad_value=0)
-
-            # Normalize from range [0,1] to [-1,1] as expacted by siglip
-            img = img * 2.0 - 1.0
-
-            bsize = img.shape[0]
-            device = img.device
-            if f"{key}_padding_mask" in batch:
-                mask = batch[f"{key}_padding_mask"].bool()
-            else:
-                mask = torch.ones(bsize, dtype=torch.bool, device=device)
-            images.append(img)
-            img_masks.append(mask)
-
-        # Create image features not present in the batch
-        # as fully 0 padded images.
-        for num_empty_cameras in range(len(missing_img_keys)):
-            if num_empty_cameras >= self.config.empty_cameras:
-                break
-            img = torch.ones_like(img) * -1
-            mask = torch.zeros_like(mask)
-            images.append(img)
-            img_masks.append(mask)
-        return images, img_masks
-
-    def prepare_language(self, batch) -> tuple[Tensor, Tensor]:
-        """Tokenize the text input"""
-        device = batch[OBS_ROBOT].device
-        tasks = batch["task"]
-        if len(tasks) == 1:
-            tasks = [tasks[0] for _ in range(batch[OBS_ROBOT].shape[0])]
-
-        tasks = [task if task.endswith("\n") else f"{task}\n" for task in tasks]
-        tokenized_prompt = self.language_tokenizer.__call__(
-            tasks,
-            padding=self.config.pad_language_to,
-            padding_side="right",
-            max_length=self.config.tokenizer_max_length,
-            return_tensors="pt",
-        )
-        lang_tokens = tokenized_prompt["input_ids"].to(device=device)
-        lang_masks = tokenized_prompt["attention_mask"].to(device=device, dtype=torch.bool)
-
-        return lang_tokens, lang_masks
-
-    def _pi_aloha_decode_state(self, state):
-        # Flip the joints.
-        for motor_idx in [1, 2, 8, 9]:
-            state[:, motor_idx] *= -1
-        # Reverse the gripper transformation that is being applied by the Aloha runtime.
-        for motor_idx in [6, 13]:
-            state[:, motor_idx] = aloha_gripper_to_angular(state[:, motor_idx])
-        return state
-
-    def _pi_aloha_encode_actions(self, actions):
-        # Flip the joints.
-        for motor_idx in [1, 2, 8, 9]:
-            actions[:, :, motor_idx] *= -1
-        # Reverse the gripper transformation that is being applied by the Aloha runtime.
-        for motor_idx in [6, 13]:
-            actions[:, :, motor_idx] = aloha_gripper_from_angular(actions[:, :, motor_idx])
-        return actions
-
-    def _pi_aloha_encode_actions_inv(self, actions):
-        # Flip the joints again.
-        for motor_idx in [1, 2, 8, 9]:
-            actions[:, :, motor_idx] *= -1
-        # Reverse the gripper transformation that is being applied by the Aloha runtime.
-        for motor_idx in [6, 13]:
-            actions[:, :, motor_idx] = aloha_gripper_from_angular_inv(actions[:, :, motor_idx])
-        return actions
-
-    def prepare_state(self, batch):
-        """Pad state"""
-        state = batch[OBS_ROBOT][:, -1, :] if batch[OBS_ROBOT].ndim > 2 else batch[OBS_ROBOT]
-        state = pad_vector(state, self.config.max_state_dim)
-        return state
-
-    def prepare_action(self, batch):
-        """Pad action"""
-        actions = pad_vector(batch[ACTION], self.config.max_action_dim)
-        return actions
-
-
-def pad_tensor(tensor, max_len, pad_value=0):
-    """
-    Efficiently pads a tensor along sequence dimension to match max_len.
-
-    Args:
-        tensor (torch.Tensor): Shape (B, L, ...) or (B, L).
-        max_len (int): Fixed sequence length.
-        pad_value (int/float): Value for padding.
-
-    Returns:
-        torch.Tensor: Shape (B, max_len, ...) or (B, max_len).
-    """
-    b, d = tensor.shape[:2]
-
-    # Create a padded tensor of max_len and copy the existing values
-    padded_tensor = torch.full(
-        (b, max_len, *tensor.shape[2:]), pad_value, dtype=tensor.dtype, device=tensor.device
-    )
-    padded_tensor[:, :d] = tensor  # Efficient in-place copy
-
-    return padded_tensor
-
-
-class VLAFlowMatching(nn.Module):
-    """
-    SmolVLA
-
-    [Paper]()
-
-    Designed by Hugging Face.
-    ┌──────────────────────────────┐
-    │                 actions      │
-    │                    ▲         │
-    │ ┌─────────┐      ┌─|────┐    │
-    │ |         │────► │      │    │
-    │ |         │ kv   │      │    │
-    │ |         │────► │Action│    │
-    │ |   VLM   │cache │Expert│    |
-    │ │         │────► |      │    │
-    │ │         │      │      │    │
-    │ └▲──▲───▲─┘      └───▲──┘    |
-    │  │  |   |            │       |
-    │  |  |   |          noise     │
-    │  │  │ state                  │
-    │  │ language tokens           │
-    │  image(s)                    │
-    └──────────────────────────────┘
-    """
-
-    def __init__(self, config):
-        super().__init__()
-        self.config = config
-
-        self.vlm_with_expert = SmolVLMWithExpertModel(
-            model_id=self.config.vlm_model_name,
-            freeze_vision_encoder=self.config.freeze_vision_encoder,
-            train_expert_only=self.config.train_expert_only,
-            load_vlm_weights=self.config.load_vlm_weights,
-            attention_mode=self.config.attention_mode,
-            num_expert_layers=self.config.num_expert_layers,
-            num_vlm_layers=self.config.num_vlm_layers,
-            self_attn_every_n_layers=self.config.self_attn_every_n_layers,
-            expert_width_multiplier=self.config.expert_width_multiplier,
-        )
-        self.state_proj = nn.Linear(
-            self.config.max_state_dim, self.vlm_with_expert.config.text_config.hidden_size
-        )
-        self.action_in_proj = nn.Linear(self.config.max_action_dim, self.vlm_with_expert.expert_hidden_size)
-        self.action_out_proj = nn.Linear(self.vlm_with_expert.expert_hidden_size, self.config.max_action_dim)
-
-        self.action_time_mlp_in = nn.Linear(
-            self.vlm_with_expert.expert_hidden_size * 2, self.vlm_with_expert.expert_hidden_size
-        )
-        self.action_time_mlp_out = nn.Linear(
-            self.vlm_with_expert.expert_hidden_size, self.vlm_with_expert.expert_hidden_size
-        )
-
-        self.set_requires_grad()
-        self.fake_image_token = self.vlm_with_expert.processor.tokenizer.fake_image_token_id
-        self.global_image_token = self.vlm_with_expert.processor.tokenizer.global_image_token_id
-        self.global_image_start_token = torch.tensor(
-            [self.fake_image_token, self.global_image_token], dtype=torch.long
-        )
-
-        self.add_image_special_tokens = self.config.add_image_special_tokens
-        self.image_end_token = torch.tensor([self.fake_image_token], dtype=torch.long)
-        self.prefix_length = self.config.prefix_length
-
-    def set_requires_grad(self):
-        for params in self.state_proj.parameters():
-            params.requires_grad = self.config.train_state_proj
-
-    def sample_noise(self, shape, device):
-        noise = torch.normal(
-            mean=0.0,
-            std=1.0,
-            size=shape,
-            dtype=torch.float32,
-            device=device,
-        )
-        return noise
-
-    def sample_time(self, bsize, device):
-        time_beta = sample_beta(1.5, 1.0, bsize, device)
-        time = time_beta * 0.999 + 0.001
-        return time.to(dtype=torch.float32, device=device)
-
-    def embed_prefix(
-        self, images, img_masks, lang_tokens, lang_masks, state: torch.Tensor = None
-    ) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
-        """Embed images with SigLIP and language tokens with embedding layer to prepare
-        for SmolVLM transformer processing.
-        """
-        embs = []
-        pad_masks = []
-        att_masks = []
-        for _img_idx, (
-            img,
-            img_mask,
-        ) in enumerate(zip(images, img_masks, strict=False)):
-            if self.add_image_special_tokens:
-                image_start_token = (
-                    self.vlm_with_expert.embed_language_tokens(
-                        self.global_image_start_token.to(device=self.vlm_with_expert.vlm.device)
-                    )
-                    .unsqueeze(0)
-                    .expand(img.shape[0], -1, -1)
-                )
-                image_start_mask = torch.ones_like(
-                    image_start_token[:, :, 0], dtype=torch.bool, device=image_start_token.device
-                )
-                att_masks += [0] * (image_start_mask.shape[-1])
-                embs.append(image_start_token)
-                pad_masks.append(image_start_mask)
-
-            img_emb = self.vlm_with_expert.embed_image(img)
-            img_emb = img_emb
-
-            # Normalize image embeddings
-            img_emb_dim = img_emb.shape[-1]
-            img_emb = img_emb * torch.tensor(img_emb_dim**0.5, dtype=img_emb.dtype, device=img_emb.device)
-
-            bsize, num_img_embs = img_emb.shape[:2]
-            img_mask = img_mask[:, None].expand(bsize, num_img_embs)
-
-            embs.append(img_emb)
-            pad_masks.append(img_mask)
-
-            att_masks += [0] * (num_img_embs)
-            if self.add_image_special_tokens:
-                image_end_token = (
-                    self.vlm_with_expert.embed_language_tokens(
-                        self.image_end_token.to(device=self.vlm_with_expert.vlm.device)
-                    )
-                    .unsqueeze(0)
-                    .expand(img.shape[0], -1, -1)
-                )
-                image_end_mask = torch.ones_like(
-                    image_end_token[:, :, 0], dtype=torch.bool, device=image_end_token.device
-                )
-                embs.append(image_end_token)
-                pad_masks.append(image_end_mask)
-                att_masks += [0] * (image_end_mask.shape[1])
-        lang_emb = self.vlm_with_expert.embed_language_tokens(lang_tokens)
-        # Normalize language embeddings
-        lang_emb_dim = lang_emb.shape[-1]
-        lang_emb = lang_emb * math.sqrt(lang_emb_dim)
-
-        embs.append(lang_emb)
-        pad_masks.append(lang_masks)
-
-        num_lang_embs = lang_emb.shape[1]
-        att_masks += [0] * num_lang_embs
-
-        state_emb = self.state_proj(state)
-        state_emb = state_emb[:, None, :] if state_emb.ndim == 2 else state_emb
-        embs.append(state_emb)
-        bsize = state_emb.shape[0]
-        device = state_emb.device
-
-        states_seq_len = state_emb.shape[1]
-        state_mask = torch.ones(bsize, states_seq_len, dtype=torch.bool, device=device)
-        pad_masks.append(state_mask)
-
-        # Set attention masks so that image and language inputs do not attend to state or actions
-        att_masks += [1] * (states_seq_len)
-        embs = torch.cat(embs, dim=1)
-        pad_masks = torch.cat(pad_masks, dim=1)
-        att_masks = torch.tensor(att_masks, dtype=torch.bool, device=pad_masks.device)
-        att_masks = att_masks[None, :]
-
-        seq_len = pad_masks.shape[1]
-        if seq_len < self.prefix_length:
-            embs = pad_tensor(embs, self.prefix_length, pad_value=0)
-            pad_masks = pad_tensor(pad_masks, self.prefix_length, pad_value=0)
-            att_masks = pad_tensor(att_masks, self.prefix_length, pad_value=0)
-
-        att_masks = att_masks.expand(bsize, -1)
-
-        return embs, pad_masks, att_masks
-
-    def embed_suffix(self, noisy_actions, timestep):
-        """Embed state, noisy_actions, timestep to prepare for Expert Gemma processing."""
-        embs = []
-        pad_masks = []
-        att_masks = []
-
-        # Fuse timestep + action information using an MLP
-        action_emb = self.action_in_proj(noisy_actions)
-        device = action_emb.device
-        bsize = action_emb.shape[0]
-        dtype = action_emb.dtype
-        # Embed timestep using sine-cosine positional encoding with sensitivity in the range [0, 1]
-        time_emb = create_sinusoidal_pos_embedding(
-            timestep,
-            self.vlm_with_expert.expert_hidden_size,
-            self.config.min_period,
-            self.config.max_period,
-            device=device,
-        )
-        time_emb = time_emb.type(dtype=dtype)
-
-        time_emb = time_emb[:, None, :].expand_as(action_emb)
-        action_time_emb = torch.cat([action_emb, time_emb], dim=2)
-
-        action_time_emb = self.action_time_mlp_in(action_time_emb)
-        action_time_emb = F.silu(action_time_emb)  # swish == silu
-        action_time_emb = self.action_time_mlp_out(action_time_emb)
-
-        # Add to input tokens
-        embs.append(action_time_emb)
-
-        bsize, action_time_dim = action_time_emb.shape[:2]
-        action_time_mask = torch.ones(bsize, action_time_dim, dtype=torch.bool, device=device)
-        pad_masks.append(action_time_mask)
-
-        # Set attention masks so that image, language and state inputs do not attend to action tokens
-        att_masks += [1] * self.config.chunk_size
-        embs = torch.cat(embs, dim=1)
-        pad_masks = torch.cat(pad_masks, dim=1)
-        att_masks = torch.tensor(att_masks, dtype=embs.dtype, device=embs.device)
-        att_masks = att_masks[None, :].expand(bsize, len(att_masks))
-        return embs, pad_masks, att_masks
-
-    def forward(
-        self, images, img_masks, lang_tokens, lang_masks, state, actions, noise=None, time=None
-    ) -> Tensor:
-        """Do a full training forward pass and compute the loss (batch_size x num_steps x num_motors)"""
-        if noise is None:
-            noise = self.sample_noise(actions.shape, actions.device)
-
-        if time is None:
-            time = self.sample_time(actions.shape[0], actions.device)
-
-        time_expanded = time[:, None, None]
-        x_t = time_expanded * noise + (1 - time_expanded) * actions
-        u_t = noise - actions
-        prefix_embs, prefix_pad_masks, prefix_att_masks = self.embed_prefix(
-            images, img_masks, lang_tokens, lang_masks, state=state
-        )
-        suffix_embs, suffix_pad_masks, suffix_att_masks = self.embed_suffix(x_t, time)
-
-        pad_masks = torch.cat([prefix_pad_masks, suffix_pad_masks], dim=1)
-        att_masks = torch.cat([prefix_att_masks, suffix_att_masks], dim=1)
-
-        att_2d_masks = make_att_2d_masks(pad_masks, att_masks)
-        position_ids = torch.cumsum(pad_masks, dim=1) - 1
-        (_, suffix_out), _ = self.vlm_with_expert.forward(
-            attention_mask=att_2d_masks,
-            position_ids=position_ids,
-            past_key_values=None,
-            inputs_embeds=[prefix_embs, suffix_embs],
-            use_cache=False,
-            fill_kv_cache=False,
-        )
-        suffix_out = suffix_out[:, -self.config.chunk_size :]
-        # Original openpi code, upcast attention output
-        suffix_out = suffix_out.to(dtype=torch.float32)
-        v_t = self.action_out_proj(suffix_out)
-        losses = F.mse_loss(u_t, v_t, reduction="none")
-        return losses
-
-    def sample_actions(self, images, img_masks, lang_tokens, lang_masks, state, noise=None) -> Tensor:
-        """Do a full inference forward and compute the action (batch_size x num_steps x num_motors)"""
-        bsize = state.shape[0]
-        device = state.device
-
-        if noise is None:
-            actions_shape = (bsize, self.config.chunk_size, self.config.max_action_dim)
-            noise = self.sample_noise(actions_shape, device)
-
-        prefix_embs, prefix_pad_masks, prefix_att_masks = self.embed_prefix(
-            images, img_masks, lang_tokens, lang_masks, state=state
-        )
-        prefix_att_2d_masks = make_att_2d_masks(prefix_pad_masks, prefix_att_masks)
-        prefix_position_ids = torch.cumsum(prefix_pad_masks, dim=1) - 1
-        # Compute image and language key value cache
-        _, past_key_values = self.vlm_with_expert.forward(
-            attention_mask=prefix_att_2d_masks,
-            position_ids=prefix_position_ids,
-            past_key_values=None,
-            inputs_embeds=[prefix_embs, None],
-            use_cache=self.config.use_cache,
-            fill_kv_cache=True,
-        )
-        dt = -1.0 / self.config.num_steps
-        dt = torch.tensor(dt, dtype=torch.float32, device=device)
-
-        x_t = noise
-        time = torch.tensor(1.0, dtype=torch.float32, device=device)
-        while time >= -dt / 2:
-            expanded_time = time.expand(bsize)
-            v_t = self.denoise_step(
-                prefix_pad_masks,
-                past_key_values,
-                x_t,
-                expanded_time,
-            )
-            # Euler step
-            x_t += dt * v_t
-            time += dt
-        return x_t
-
-    def denoise_step(
-        self,
-        prefix_pad_masks,
-        past_key_values,
-        x_t,
-        timestep,
-    ):
-        """Apply one denoising step of the noise `x_t` at a given timestep."""
-        suffix_embs, suffix_pad_masks, suffix_att_masks = self.embed_suffix(x_t, timestep)
-
-        suffix_len = suffix_pad_masks.shape[1]
-        batch_size = prefix_pad_masks.shape[0]
-        prefix_len = prefix_pad_masks.shape[1]
-        prefix_pad_2d_masks = prefix_pad_masks[:, None, :].expand(batch_size, suffix_len, prefix_len)
-
-        suffix_att_2d_masks = make_att_2d_masks(suffix_pad_masks, suffix_att_masks)
-
-        full_att_2d_masks = torch.cat([prefix_pad_2d_masks, suffix_att_2d_masks], dim=2)
-        prefix_offsets = torch.sum(prefix_pad_masks, dim=-1)[:, None]
-        position_ids = prefix_offsets + torch.cumsum(suffix_pad_masks, dim=1) - 1
-
-        outputs_embeds, _ = self.vlm_with_expert.forward(
-            attention_mask=full_att_2d_masks,
-            position_ids=position_ids,
-            past_key_values=past_key_values,
-            inputs_embeds=[None, suffix_embs],
-            use_cache=self.config.use_cache,
-            fill_kv_cache=False,
-        )
-        suffix_out = outputs_embeds[1]
-        suffix_out = suffix_out[:, -self.config.chunk_size :]
-        suffix_out = suffix_out.to(dtype=torch.float32)
-        v_t = self.action_out_proj(suffix_out)
-        return v_t

lerobot/common/policies/smolvla/smolvlm_with_expert.py

@@ -1,550 +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.
-
-import copy
-from typing import List, Optional
-
-import torch
-from torch import nn
-from transformers import (
-    AutoConfig,
-    AutoModel,
-    AutoModelForImageTextToText,
-    AutoProcessor,
-    SmolVLMForConditionalGeneration,
-)
-
-
-def apply_rope(x, positions, max_wavelength=10_000):
-    """
-    Applies RoPE positions [B, L] to x [B, L, H, D].
-    """
-    d_half = x.shape[-1] // 2
-    device = x.device
-    dtype = x.dtype
-    x = x.to(torch.float32)
-
-    freq_exponents = (2.0 / x.shape[-1]) * torch.arange(d_half, dtype=torch.float32, device=device)
-    timescale = max_wavelength**freq_exponents
-    radians = positions[..., None].to(torch.float32) / timescale[None, None, :].to(torch.float32)
-
-    radians = radians[..., None, :]
-
-    sin = torch.sin(radians)  # .to(dtype=dtype)
-    cos = torch.cos(radians)  # .to(dtype=dtype)
-
-    x1, x2 = x.split(d_half, dim=-1)
-    res = torch.empty_like(x)
-    res[..., :d_half] = x1 * cos - x2 * sin
-    res[..., d_half:] = x2 * cos + x1 * sin
-
-    return res.to(dtype)
-
-
-def get_intermediate_size(hidden_dim, ffn_dim_multiplier=4, multiple_of=256):
-    hidden_dim = int(2 * hidden_dim / 3)
-    hidden_dim = int(ffn_dim_multiplier * hidden_dim)
-    hidden_dim = multiple_of * ((hidden_dim + multiple_of - 1) // multiple_of)
-    return hidden_dim
-
-
-class SmolVLMWithExpertModel(nn.Module):
-    def __init__(
-        self,
-        model_id: str = "HuggingFaceTB/SmolVLM2-500M-Video-Instruct",
-        load_vlm_weights: bool = True,
-        train_expert_only: bool = True,
-        freeze_vision_encoder: bool = False,
-        attention_mode: str = "self_attn",
-        num_expert_layers: int = -1,
-        num_vlm_layers: int = -1,
-        self_attn_every_n_layers: int = -1,
-        expert_width_multiplier: float = 0.5,
-    ):
-        super().__init__()
-        if load_vlm_weights:
-            print(f"Loading  {model_id} weights ...")
-            self.vlm = AutoModelForImageTextToText.from_pretrained(
-                model_id,
-                device_map="auto",
-                torch_dtype="bfloat16",
-                low_cpu_mem_usage=True,
-            )
-            config = self.vlm.config
-        else:
-            config = AutoConfig.from_pretrained(model_id)
-            self.vlm = SmolVLMForConditionalGeneration(config=config)
-        self.processor = AutoProcessor.from_pretrained(model_id)
-        if num_vlm_layers > 0:
-            print(f"Reducing the number of VLM layers to {num_vlm_layers} ...")
-            self.get_vlm_model().text_model.layers = self.get_vlm_model().text_model.layers[:num_vlm_layers]
-        self.num_vlm_layers = len(self.get_vlm_model().text_model.layers)
-        self.config = config
-        # Smaller lm expert
-        lm_expert_config = copy.deepcopy(config.text_config)
-        hidden_size = lm_expert_config.hidden_size
-        lm_expert_config.hidden_size = int(hidden_size * expert_width_multiplier)  # hidden_size // 2
-        lm_expert_config.intermediate_size = get_intermediate_size(int(hidden_size * expert_width_multiplier))
-        lm_expert_config.num_hidden_layers = self.num_vlm_layers
-        if num_expert_layers > 0:
-            assert len(self.get_vlm_model().text_model.layers) % num_expert_layers == 0, (
-                f"Number of layers in the VLM {len(self.get_vlm_model().text_model.layers)} are not multiple of num_expert_layers {num_expert_layers}"
-            )
-            lm_expert_config.num_hidden_layers = num_expert_layers
-        self.lm_expert = AutoModel.from_config(lm_expert_config)
-
-        self.num_expert_layers = len(self.lm_expert.layers)
-        self.self_attn_every_n_layers = self_attn_every_n_layers
-        if "cross" in attention_mode:
-            # Reshape qkv projections to have the same input dimension as the vlm
-            for layer_idx in range(len(self.lm_expert.layers)):
-                if self.self_attn_every_n_layers > 0 and layer_idx % self.self_attn_every_n_layers == 0:
-                    continue
-                self.lm_expert.layers[layer_idx].self_attn.k_proj = nn.Linear(
-                    config.text_config.num_key_value_heads * config.text_config.head_dim,
-                    lm_expert_config.num_key_value_heads * lm_expert_config.head_dim,
-                    bias=lm_expert_config.attention_bias,
-                )
-                self.lm_expert.layers[layer_idx].self_attn.v_proj = nn.Linear(
-                    config.text_config.num_key_value_heads * config.text_config.head_dim,
-                    lm_expert_config.num_key_value_heads * lm_expert_config.head_dim,
-                    bias=lm_expert_config.attention_bias,
-                )
-        # Remove unused embed_tokens
-        self.lm_expert.embed_tokens = None
-
-        self.num_attention_heads = self.config.text_config.num_attention_heads
-        self.num_key_value_heads = self.config.text_config.num_key_value_heads
-
-        self.freeze_vision_encoder = freeze_vision_encoder
-        self.train_expert_only = train_expert_only
-        self.attention_mode = attention_mode
-        self.expert_hidden_size = lm_expert_config.hidden_size
-        self.set_requires_grad()
-
-    def get_vlm_model(self):
-        return self.vlm.model
-
-    def set_requires_grad(self):
-        if self.freeze_vision_encoder:
-            self.get_vlm_model().vision_model.eval()
-            for params in self.get_vlm_model().vision_model.parameters():
-                params.requires_grad = False
-        if self.train_expert_only:
-            self.vlm.eval()
-            for params in self.vlm.parameters():
-                params.requires_grad = False
-        else:
-            # To avoid unused params issue with distributed training
-            last_layers = [self.num_vlm_layers - 1]
-            if (
-                self.num_vlm_layers != self.num_expert_layers
-                and self.num_vlm_layers % self.num_expert_layers == 0
-            ):
-                last_layers.append(self.num_vlm_layers - 2)
-            frozen_layers = [
-                "lm_head",
-                "text_model.model.norm.weight",
-            ]
-            for layer in last_layers:
-                frozen_layers.append(f"text_model.model.layers.{layer}.")
-
-            for name, params in self.vlm.named_parameters():
-                if any(k in name for k in frozen_layers):
-                    params.requires_grad = False
-        # To avoid unused params issue with distributed training
-        for name, params in self.lm_expert.named_parameters():
-            if "lm_head" in name:
-                params.requires_grad = False
-
-    def train(self, mode: bool = True):
-        super().train(mode)
-
-        if self.freeze_vision_encoder:
-            self.get_vlm_model().vision_model.eval()
-
-        if self.train_expert_only:
-            self.vlm.eval()
-
-    def embed_image(self, image: torch.Tensor):
-        patch_attention_mask = None
-        # Get sequence from the vision encoder
-        image_hidden_states = (
-            self.get_vlm_model()
-            .vision_model(
-                pixel_values=image.to(dtype=self.get_vlm_model().vision_model.dtype),
-                patch_attention_mask=patch_attention_mask,
-            )
-            .last_hidden_state
-        )
-        # Modality projection & resampling
-        image_hidden_states = self.get_vlm_model().connector(image_hidden_states)
-        return image_hidden_states
-
-    def embed_language_tokens(self, tokens: torch.Tensor):
-        return self.get_vlm_model().text_model.get_input_embeddings()(tokens)
-
-    def forward_attn_layer(
-        self,
-        model_layers,
-        inputs_embeds,
-        layer_idx,
-        position_ids,
-        attention_mask,
-        batch_size,
-        head_dim,
-        use_cache: bool = True,
-        fill_kv_cache: bool = True,
-        past_key_values=None,
-    ) -> list[torch.Tensor]:
-        query_states = []
-        key_states = []
-        value_states = []
-        for i, hidden_states in enumerate(inputs_embeds):
-            layer = model_layers[i][layer_idx]
-            if hidden_states is None or layer is None:
-                continue
-            hidden_states = layer.input_layernorm(hidden_states)
-
-            input_shape = hidden_states.shape[:-1]
-            hidden_shape = (*input_shape, -1, layer.self_attn.head_dim)
-
-            hidden_states = hidden_states.to(dtype=layer.self_attn.q_proj.weight.dtype)
-            query_state = layer.self_attn.q_proj(hidden_states).view(hidden_shape)
-            key_state = layer.self_attn.k_proj(hidden_states).view(hidden_shape)
-            value_state = layer.self_attn.v_proj(hidden_states).view(hidden_shape)
-
-            query_states.append(query_state)
-            key_states.append(key_state)
-            value_states.append(value_state)
-
-        # B,L,H,D with L sequence length, H number of heads, D head dim
-        # concatenate on the number of embeddings/tokens
-        query_states = torch.cat(query_states, dim=1)
-        key_states = torch.cat(key_states, dim=1)
-        value_states = torch.cat(value_states, dim=1)
-        seq_len = query_states.shape[1]
-        if seq_len < position_ids.shape[1]:
-            _position_ids = position_ids[:, :seq_len]
-            _attention_mask = attention_mask[:, :seq_len, :seq_len]
-        else:
-            _position_ids = position_ids
-            _attention_mask = attention_mask
-
-        attention_mask_ = _attention_mask
-        position_ids_ = _position_ids
-
-        query_states = apply_rope(query_states, position_ids_)
-        key_states = apply_rope(key_states, position_ids_)
-
-        if use_cache and past_key_values is None:
-            past_key_values = {}
-
-        if use_cache:
-            if fill_kv_cache:
-                past_key_values[layer_idx] = {
-                    "key_states": key_states,
-                    "value_states": value_states,
-                }
-            else:
-                # TODO here, some optimization can be done - similar to a `StaticCache` we can declare the `max_len` before.
-                # so we create an empty cache, with just one cuda malloc, and if (in autoregressive case) we reach
-                # the max len, then we (for instance) double the cache size. This implementation already exists
-                # in `transformers`. (molbap)
-                key_states = torch.cat([past_key_values[layer_idx]["key_states"], key_states], dim=1)
-                value_states = torch.cat([past_key_values[layer_idx]["value_states"], value_states], dim=1)
-
-        attention_interface = self.get_attention_interface()
-
-        att_output = attention_interface(
-            attention_mask_, batch_size, head_dim, query_states, key_states, value_states
-        )
-        return [att_output], past_key_values
-
-    def forward_cross_attn_layer(
-        self,
-        model_layers,
-        inputs_embeds,
-        layer_idx,
-        position_ids,
-        attention_mask,
-        batch_size,
-        head_dim,
-        use_cache: bool = True,
-        fill_kv_cache: bool = True,
-        past_key_values=None,
-    ) -> list[torch.Tensor]:
-        attention_interface = self.get_attention_interface()
-
-        att_outputs = []
-        assert len(inputs_embeds) == 2 or (use_cache and past_key_values is not None and not fill_kv_cache), (
-            f"Both len(inputs_embeds) == {len(inputs_embeds)} and past_key_values is {past_key_values}"
-        )
-
-        if len(inputs_embeds) == 2 and not past_key_values:
-            # Prefix attention
-            seq_len = inputs_embeds[0].shape[1]
-            position_id, expert_position_id = position_ids[:, :seq_len], position_ids[:, seq_len:]
-            prefix_attention_mask = attention_mask[:, :seq_len, :seq_len]
-
-            layer = model_layers[0][layer_idx]
-
-            hidden_states = layer.input_layernorm(inputs_embeds[0])
-
-            input_shape = hidden_states.shape[:-1]
-            hidden_shape = (*input_shape, -1, layer.self_attn.head_dim)
-
-            hidden_states = hidden_states.to(dtype=layer.self_attn.q_proj.weight.dtype)
-            query_state = layer.self_attn.q_proj(hidden_states).view(hidden_shape)
-            key_state = layer.self_attn.k_proj(hidden_states).view(hidden_shape)
-            value_states = layer.self_attn.v_proj(hidden_states).view(hidden_shape)
-
-            # B,L,H,D with L sequence length, H number of heads, D head dim
-            query_states = apply_rope(query_state, position_id)
-            key_states = apply_rope(key_state, position_id)
-
-            att_output = attention_interface(
-                prefix_attention_mask, batch_size, head_dim, query_states, key_states, value_states
-            )
-            att_outputs.append(att_output)
-        else:
-            expert_position_id = position_ids
-
-        if use_cache and past_key_values is None:
-            past_key_values = {}
-
-        if use_cache:
-            if fill_kv_cache:
-                past_key_values[layer_idx] = {
-                    "key_states": key_states,
-                    "value_states": value_states,
-                }
-            else:
-                # TODO here, some optimization can be done - similar to a `StaticCache` we can declare the `max_len` before.
-                # so we create an empty cache, with just one cuda malloc, and if (in autoregressive case) we reach
-                # the max len, then we (for instance) double the cache size. This implementation already exists
-                # in `transformers`. (molbap)
-                key_states = past_key_values[layer_idx]["key_states"]
-                value_states = past_key_values[layer_idx]["value_states"]
-
-        # Expert
-        expert_layer = model_layers[1][layer_idx]
-        if expert_layer is not None:
-            expert_hidden_states = expert_layer.input_layernorm(inputs_embeds[1])
-
-            expert_input_shape = expert_hidden_states.shape[:-1]
-            expert_hidden_shape = (*expert_input_shape, -1, expert_layer.self_attn.head_dim)
-
-            expert_hidden_states = expert_hidden_states.to(dtype=expert_layer.self_attn.q_proj.weight.dtype)
-            expert_query_state = expert_layer.self_attn.q_proj(expert_hidden_states).view(expert_hidden_shape)
-
-            _key_states = key_states.to(dtype=expert_layer.self_attn.k_proj.weight.dtype).view(
-                *key_states.shape[:2], -1
-            )
-            expert_key_states = expert_layer.self_attn.k_proj(_key_states).view(
-                *_key_states.shape[:-1], -1, expert_layer.self_attn.head_dim
-            )  # k_proj should have same dim as kv
-
-            _value_states = value_states.to(dtype=expert_layer.self_attn.v_proj.weight.dtype).view(
-                *value_states.shape[:2], -1
-            )
-            expert_value_states = expert_layer.self_attn.v_proj(_value_states).view(
-                *_value_states.shape[:-1], -1, expert_layer.self_attn.head_dim
-            )
-
-            expert_position_id = (
-                expert_position_id - torch.min(expert_position_id, dim=1, keepdim=True).values
-            )  # start from 0
-            expert_attention_mask = attention_mask[
-                :, -inputs_embeds[1].shape[1] :, : expert_key_states.shape[1] :
-            ]  # take into account kv
-
-            expert_query_states = apply_rope(expert_query_state, expert_position_id)
-
-            att_output = attention_interface(
-                expert_attention_mask,
-                batch_size,
-                head_dim,
-                expert_query_states,
-                expert_key_states,
-                expert_value_states,
-            )
-            att_outputs.append(att_output)
-        else:
-            att_outputs.append(None)
-
-        # att_output = att_output.to(dtype=models[i].dtype)
-        return att_outputs, past_key_values
-
-    def get_model_layers(self, models: list) -> list:
-        vlm_layers = []
-        expert_layers = []
-        multiple_of = self.num_vlm_layers // self.num_expert_layers
-        for i in range(self.num_vlm_layers):
-            if multiple_of > 0 and i > 0 and i % multiple_of != 0:
-                expert_layer = None
-            else:
-                expert_layer_index = i // multiple_of if multiple_of > 0 else i
-                expert_layer = models[1].layers[expert_layer_index]
-            vlm_layers.append(models[0].layers[i])
-            expert_layers.append(expert_layer)
-        return [vlm_layers, expert_layers]
-
-    def forward(
-        self,
-        attention_mask: Optional[torch.Tensor] = None,
-        position_ids: Optional[torch.LongTensor] = None,
-        past_key_values: Optional[List[torch.FloatTensor]] = None,
-        inputs_embeds: List[torch.FloatTensor] = None,
-        use_cache: Optional[bool] = None,
-        fill_kv_cache: Optional[bool] = None,
-    ):
-        models = [self.get_vlm_model().text_model, self.lm_expert]
-        model_layers = self.get_model_layers(models)
-        for hidden_states in inputs_embeds:
-            # TODO this is very inefficient
-            # dtype is always the same, batch size too (if > 1 len)
-            # device could be trickier in multi gpu edge cases but that's it
-            if hidden_states is None:
-                continue
-            batch_size = hidden_states.shape[0]
-
-        # RMSNorm
-        num_layers = self.num_vlm_layers
-        head_dim = self.vlm.config.text_config.head_dim
-        for layer_idx in range(num_layers):
-            if (
-                fill_kv_cache
-                or "cross" not in self.attention_mode
-                or (self.self_attn_every_n_layers > 0 and layer_idx % self.self_attn_every_n_layers == 0)
-            ):
-                att_outputs, past_key_values = self.forward_attn_layer(
-                    model_layers,
-                    inputs_embeds,
-                    layer_idx,
-                    position_ids,
-                    attention_mask,
-                    batch_size,
-                    head_dim,
-                    use_cache=use_cache,
-                    fill_kv_cache=fill_kv_cache,
-                    past_key_values=past_key_values,
-                )
-            else:
-                att_outputs, past_key_values = self.forward_cross_attn_layer(
-                    model_layers,
-                    inputs_embeds,
-                    layer_idx,
-                    position_ids,
-                    attention_mask,
-                    batch_size,
-                    head_dim,
-                    use_cache=use_cache,
-                    fill_kv_cache=fill_kv_cache,
-                    past_key_values=past_key_values,
-                )
-            outputs_embeds = []
-            start = 0
-            for i, hidden_states in enumerate(inputs_embeds):
-                layer = model_layers[i][layer_idx]
-                att_output = (
-                    att_outputs[i] if i < len(att_outputs) else att_outputs[0]
-                )  # in case of self_attn
-                if hidden_states is not None:
-                    if layer is None:
-                        outputs_embeds.append(hidden_states)
-                        continue
-                    end = start + hidden_states.shape[1]
-
-                    if att_output.dtype != layer.self_attn.o_proj.weight.dtype:
-                        att_output = att_output.to(layer.self_attn.o_proj.weight.dtype)
-                    att_out = att_output[:, start:end]
-                    out_emb = layer.self_attn.o_proj(att_out)
-
-                    out_emb += hidden_states
-                    after_first_residual = out_emb.clone()
-
-                    out_emb = layer.post_attention_layernorm(out_emb)
-                    out_emb = layer.mlp(out_emb)
-
-                    out_emb += after_first_residual
-
-                    outputs_embeds.append(out_emb)
-
-                    start = end if len(att_outputs) == 1 else 0
-                else:
-                    outputs_embeds.append(None)
-
-            inputs_embeds = outputs_embeds
-
-        # final norm
-        outputs_embeds = []
-        for i, hidden_states in enumerate(inputs_embeds):
-            if hidden_states is not None:
-                out_emb = models[i].norm(hidden_states)
-                outputs_embeds.append(out_emb)
-            else:
-                outputs_embeds.append(None)
-        return outputs_embeds, past_key_values
-
-    def get_attention_interface(self):
-        attention_interface = self.eager_attention_forward
-        return attention_interface
-
-    def eager_attention_forward(
-        self, attention_mask, batch_size, head_dim, query_states, key_states, value_states
-    ):
-        num_att_heads = self.num_attention_heads
-        num_key_value_heads = self.num_key_value_heads
-        num_key_value_groups = num_att_heads // num_key_value_heads
-
-        sequence_length = key_states.shape[1]
-
-        key_states = key_states[:, :, :, None, :].expand(
-            batch_size, sequence_length, num_key_value_heads, num_key_value_groups, head_dim
-        )
-        key_states = key_states.reshape(
-            batch_size, sequence_length, num_key_value_heads * num_key_value_groups, head_dim
-        )
-
-        value_states = value_states[:, :, :, None, :].expand(
-            batch_size, sequence_length, num_key_value_heads, num_key_value_groups, head_dim
-        )
-        value_states = value_states.reshape(
-            batch_size, sequence_length, num_key_value_heads * num_key_value_groups, head_dim
-        )
-
-        # Attention here is upcasted to float32 to match the original eager implementation.
-        query_states = query_states.to(dtype=torch.float32)
-        key_states = key_states.to(dtype=torch.float32)
-
-        query_states = query_states.transpose(1, 2)
-        key_states = key_states.transpose(1, 2)
-
-        att_weights = torch.matmul(query_states, key_states.transpose(2, 3))
-        att_weights *= head_dim**-0.5
-
-        att_weights = att_weights.to(dtype=torch.float32)
-        big_neg = torch.finfo(att_weights.dtype).min  # -2.3819763e38  # See gemma/modules.py
-        masked_att_weights = torch.where(attention_mask[:, None, :, :], att_weights, big_neg)
-        probs = nn.functional.softmax(masked_att_weights, dim=-1)
-        probs = probs.to(dtype=value_states.dtype)
-
-        att_output = torch.matmul(probs, value_states.permute(0, 2, 1, 3))
-
-        att_output = att_output.permute(0, 2, 1, 3)
-        # we use -1 because sequence length can change
-        att_output = att_output.reshape(batch_size, -1, num_key_value_heads * num_key_value_groups * head_dim)
-
-        return att_output

lerobot/common/policies/tdmpc/modeling_tdmpc.py

@@ -39,7 +39,11 @@ from lerobot.common.constants import OBS_ENV, OBS_ROBOT
 from lerobot.common.policies.normalize import Normalize, Unnormalize
 from lerobot.common.policies.pretrained import PreTrainedPolicy
 from lerobot.common.policies.tdmpc.configuration_tdmpc import TDMPCConfig
-from lerobot.common.policies.utils import get_device_from_parameters, get_output_shape, populate_queues
+from lerobot.common.policies.utils import (
+    get_device_from_parameters,
+    get_output_shape,
+    populate_queues,
+)
 
 
 class TDMPCPolicy(PreTrainedPolicy):
@@ -63,7 +67,11 @@ class TDMPCPolicy(PreTrainedPolicy):
     config_class = TDMPCConfig
     name = "tdmpc"
 
-    def __init__(self, config: TDMPCConfig, dataset_stats: dict[str, dict[str, Tensor]] | None = None):
+    def __init__(
+        self,
+        config: TDMPCConfig,
+        dataset_stats: dict[str, dict[str, Tensor]] | None = None,
+    ):
         """
         Args:
             config: Policy configuration class instance or None, in which case the default instantiation of
@@ -122,7 +130,7 @@ class TDMPCPolicy(PreTrainedPolicy):
 
         # When the action queue is depleted, populate it again by querying the policy.
         if len(self._queues["action"]) == 0:
-            batch = {key: torch.stack(list(self._queues[key]), dim=1) for key in batch if key in self._queues}
+            batch = {key: torch.stack(list(self._queues[key]), dim=1) for key in batch}
 
             # Remove the time dimensions as it is not handled yet.
             for key in batch:
@@ -189,13 +197,20 @@ class TDMPCPolicy(PreTrainedPolicy):
 
         # In the CEM loop we will need this for a call to estimate_value with the gaussian sampled
         # trajectories.
-        z = einops.repeat(z, "b d -> n b d", n=self.config.n_gaussian_samples + self.config.n_pi_samples)
+        z = einops.repeat(
+            z,
+            "b d -> n b d",
+            n=self.config.n_gaussian_samples + self.config.n_pi_samples,
+        )
 
         # Model Predictive Path Integral (MPPI) with the cross-entropy method (CEM) as the optimization
         # algorithm.
         # The initial mean and standard deviation for the cross-entropy method (CEM).
         mean = torch.zeros(
-            self.config.horizon, batch_size, self.config.action_feature.shape[0], device=device
+            self.config.horizon,
+            batch_size,
+            self.config.action_feature.shape[0],
+            device=device,
         )
         # Maybe warm start CEM with the mean from the previous step.
         if self._prev_mean is not None:
@@ -291,9 +306,10 @@ class TDMPCPolicy(PreTrainedPolicy):
         if self.config.q_ensemble_size > 2:
             G += (
                 running_discount
-                * torch.min(terminal_values[torch.randint(0, self.config.q_ensemble_size, size=(2,))], dim=0)[
-                    0
-                ]
+                * torch.min(
+                    terminal_values[torch.randint(0, self.config.q_ensemble_size, size=(2,))],
+                    dim=0,
+                )[0]
             )
         else:
             G += running_discount * torch.min(terminal_values, dim=0)[0]
@@ -329,7 +345,10 @@ class TDMPCPolicy(PreTrainedPolicy):
         # Apply random image augmentations.
         if self.config.image_features and self.config.max_random_shift_ratio > 0:
             observations["observation.image"] = flatten_forward_unflatten(
-                partial(random_shifts_aug, max_random_shift_ratio=self.config.max_random_shift_ratio),
+                partial(
+                    random_shifts_aug,
+                    max_random_shift_ratio=self.config.max_random_shift_ratio,
+                ),
                 observations["observation.image"],
             )
 
@@ -553,7 +572,10 @@ class TDMPCTOLD(nn.Module):
         self._Qs = nn.ModuleList(
             [
                 nn.Sequential(
-                    nn.Linear(config.latent_dim + config.action_feature.shape[0], config.mlp_dim),
+                    nn.Linear(
+                        config.latent_dim + config.action_feature.shape[0],
+                        config.mlp_dim,
+                    ),
                     nn.LayerNorm(config.mlp_dim),
                     nn.Tanh(),
                     nn.Linear(config.mlp_dim, config.mlp_dim),
@@ -702,11 +724,26 @@ class TDMPCObservationEncoder(nn.Module):
                     stride=2,
                 ),
                 nn.ReLU(),
-                nn.Conv2d(config.image_encoder_hidden_dim, config.image_encoder_hidden_dim, 5, stride=2),
+                nn.Conv2d(
+                    config.image_encoder_hidden_dim,
+                    config.image_encoder_hidden_dim,
+                    5,
+                    stride=2,
+                ),
                 nn.ReLU(),
-                nn.Conv2d(config.image_encoder_hidden_dim, config.image_encoder_hidden_dim, 3, stride=2),
+                nn.Conv2d(
+                    config.image_encoder_hidden_dim,
+                    config.image_encoder_hidden_dim,
+                    3,
+                    stride=2,
+                ),
                 nn.ReLU(),
-                nn.Conv2d(config.image_encoder_hidden_dim, config.image_encoder_hidden_dim, 3, stride=2),
+                nn.Conv2d(
+                    config.image_encoder_hidden_dim,
+                    config.image_encoder_hidden_dim,
+                    3,
+                    stride=2,
+                ),
                 nn.ReLU(),
             )
             dummy_shape = (1, *next(iter(config.image_features.values())).shape)
@@ -749,7 +786,8 @@ class TDMPCObservationEncoder(nn.Module):
         if self.config.image_features:
             feat.append(
                 flatten_forward_unflatten(
-                    self.image_enc_layers, obs_dict[next(iter(self.config.image_features))]
+                    self.image_enc_layers,
+                    obs_dict[next(iter(self.config.image_features))],
                 )
             )
         if self.config.env_state_feature:
@@ -796,7 +834,9 @@ def update_ema_parameters(ema_net: nn.Module, net: nn.Module, alpha: float):
     """Update EMA parameters in place with ema_param <- alpha * ema_param + (1 - alpha) * param."""
     for ema_module, module in zip(ema_net.modules(), net.modules(), strict=True):
         for (n_p_ema, p_ema), (n_p, p) in zip(
-            ema_module.named_parameters(recurse=False), module.named_parameters(recurse=False), strict=True
+            ema_module.named_parameters(recurse=False),
+            module.named_parameters(recurse=False),
+            strict=True,
         ):
             assert n_p_ema == n_p, "Parameter names don't match for EMA model update"
             if isinstance(p, dict):

lerobot/common/policies/vqbet/configuration_vqbet.py

@@ -193,7 +193,12 @@ class VQBeTConfig(PreTrainedConfig):
 
     @property
     def action_delta_indices(self) -> list:
-        return list(range(1 - self.n_obs_steps, self.n_action_pred_token + self.action_chunk_size - 1))
+        return list(
+            range(
+                1 - self.n_obs_steps,
+                self.n_action_pred_token + self.action_chunk_size - 1,
+            )
+        )
 
     @property
     def reward_delta_indices(self) -> None:

lerobot/common/policies/vqbet/modeling_vqbet.py

@@ -29,7 +29,11 @@ from torch import Tensor, nn
 
 from lerobot.common.policies.normalize import Normalize, Unnormalize
 from lerobot.common.policies.pretrained import PreTrainedPolicy
-from lerobot.common.policies.utils import get_device_from_parameters, get_output_shape, populate_queues
+from lerobot.common.policies.utils import (
+    get_device_from_parameters,
+    get_output_shape,
+    populate_queues,
+)
 from lerobot.common.policies.vqbet.configuration_vqbet import VQBeTConfig
 from lerobot.common.policies.vqbet.vqbet_utils import GPT, ResidualVQ
 
@@ -324,7 +328,8 @@ class VQBeTModel(nn.Module):
 
         # To input state and observation features into GPT layers, we first project the features to fit the shape of input size of GPT.
         self.state_projector = MLP(
-            config.robot_state_feature.shape[0], hidden_channels=[self.config.gpt_input_dim]
+            config.robot_state_feature.shape[0],
+            hidden_channels=[self.config.gpt_input_dim],
         )
         self.rgb_feature_projector = MLP(
             self.rgb_encoder.feature_dim, hidden_channels=[self.config.gpt_input_dim]
@@ -354,7 +359,11 @@ class VQBeTModel(nn.Module):
         )
         # Separate batch and sequence dims.
         img_features = einops.rearrange(
-            img_features, "(b s n) ... -> b s n ...", b=batch_size, s=n_obs_steps, n=self.num_images
+            img_features,
+            "(b s n) ... -> b s n ...",
+            b=batch_size,
+            s=n_obs_steps,
+            n=self.num_images,
         )
 
         # Arrange prior and current observation step tokens as shown in the class docstring.
@@ -391,7 +400,11 @@ class VQBeTModel(nn.Module):
         # Thus, it predicts a historical action sequence, in addition to current and future actions (predicting future actions : optional).
         if len_additional_action_token > 0:
             features = torch.cat(
-                [features[:, historical_act_pred_index], features[:, -len_additional_action_token:]], dim=1
+                [
+                    features[:, historical_act_pred_index],
+                    features[:, -len_additional_action_token:],
+                ],
+                dim=1,
             )
         else:
             features = features[:, historical_act_pred_index]
@@ -514,7 +527,13 @@ class VQBeTHead(nn.Module):
 
             cbet_secondary_logits = self.map_to_cbet_preds_secondary_bin(
                 torch.cat(
-                    (x, F.one_hot(sampled_primary_centers, num_classes=self.config.vqvae_n_embed)),
+                    (
+                        x,
+                        F.one_hot(
+                            sampled_primary_centers,
+                            num_classes=self.config.vqvae_n_embed,
+                        ),
+                    ),
                     axis=1,
                 )
             )
@@ -532,7 +551,9 @@ class VQBeTHead(nn.Module):
         else:
             cbet_logits = self.map_to_cbet_preds_bin(x)
             cbet_logits = einops.rearrange(
-                cbet_logits, "(NT) (G C) -> (NT) G C", G=self.vqvae_model.vqvae_num_layers
+                cbet_logits,
+                "(NT) (G C) -> (NT) G C",
+                G=self.vqvae_model.vqvae_num_layers,
             )
             cbet_probs = torch.softmax(cbet_logits / self.config.bet_softmax_temperature, dim=-1)
             NT, G, choices = cbet_probs.shape
@@ -730,7 +751,9 @@ class VQBeTRgbEncoder(nn.Module):
 
 
 def _replace_submodules(
-    root_module: nn.Module, predicate: Callable[[nn.Module], bool], func: Callable[[nn.Module], nn.Module]
+    root_module: nn.Module,
+    predicate: Callable[[nn.Module], bool],
+    func: Callable[[nn.Module], nn.Module],
 ) -> nn.Module:
     """
     Args:

lerobot/common/policies/vqbet/vqbet_utils.py

@@ -377,7 +377,10 @@ class ResidualVQ(nn.Module):
         self.layers = nn.ModuleList(
             [
                 VectorQuantize(
-                    dim=codebook_dim, codebook_dim=codebook_dim, accept_image_fmap=accept_image_fmap, **kwargs
+                    dim=codebook_dim,
+                    codebook_dim=codebook_dim,
+                    accept_image_fmap=accept_image_fmap,
+                    **kwargs,
                 )
                 for _ in range(num_quantizers)
             ]

lerobot/common/robot_devices/cameras/intelrealsense.py

@@ -297,7 +297,11 @@ class IntelRealSenseCamera:
         if self.fps and self.capture_width and self.capture_height:
             # TODO(rcadene): can we set rgb8 directly?
             config.enable_stream(
-                rs.stream.color, self.capture_width, self.capture_height, rs.format.rgb8, self.fps
+                rs.stream.color,
+                self.capture_width,
+                self.capture_height,
+                rs.format.rgb8,
+                self.fps,
             )
         else:
             config.enable_stream(rs.stream.color)
@@ -305,7 +309,11 @@ class IntelRealSenseCamera:
         if self.use_depth:
             if self.fps and self.capture_width and self.capture_height:
                 config.enable_stream(
-                    rs.stream.depth, self.capture_width, self.capture_height, rs.format.z16, self.fps
+                    rs.stream.depth,
+                    self.capture_width,
+                    self.capture_height,
+                    rs.format.z16,
+                    self.fps,
                 )
             else:
                 config.enable_stream(rs.stream.depth)
@@ -512,13 +520,13 @@ if __name__ == "__main__":
     )
     parser.add_argument(
         "--width",
-        type=int,
+        type=str,
         default=640,
         help="Set the width for all cameras. If not provided, use the default width of each camera.",
     )
     parser.add_argument(
         "--height",
-        type=int,
+        type=str,
         default=480,
         help="Set the height for all cameras. If not provided, use the default height of each camera.",
     )

lerobot/common/robot_devices/cameras/opencv.py

@@ -492,13 +492,13 @@ if __name__ == "__main__":
     )
     parser.add_argument(
         "--width",
-        type=int,
+        type=str,
         default=None,
         help="Set the width for all cameras. If not provided, use the default width of each camera.",
     )
     parser.add_argument(
         "--height",
-        type=int,
+        type=str,
         default=None,
         help="Set the height for all cameras. If not provided, use the default height of each camera.",
     )

lerobot/common/robot_devices/cameras/utils.py

@@ -41,7 +41,9 @@ def make_cameras_from_configs(camera_configs: dict[str, CameraConfig]) -> list[C
             cameras[key] = OpenCVCamera(cfg)
 
         elif cfg.type == "intelrealsense":
-            from lerobot.common.robot_devices.cameras.intelrealsense import IntelRealSenseCamera
+            from lerobot.common.robot_devices.cameras.intelrealsense import (
+                IntelRealSenseCamera,
+            )
 
             cameras[key] = IntelRealSenseCamera(cfg)
         else:
@@ -58,7 +60,9 @@ def make_camera(camera_type, **kwargs) -> Camera:
         return OpenCVCamera(config)
 
     elif camera_type == "intelrealsense":
-        from lerobot.common.robot_devices.cameras.intelrealsense import IntelRealSenseCamera
+        from lerobot.common.robot_devices.cameras.intelrealsense import (
+            IntelRealSenseCamera,
+        )
 
         config = IntelRealSenseCameraConfig(**kwargs)
         return IntelRealSenseCamera(config)

lerobot/common/robot_devices/control_configs.py

@@ -41,7 +41,7 @@ class TeleoperateControlConfig(ControlConfig):
     fps: int | None = None
     teleop_time_s: float | None = None
     # Display all cameras on screen
-    display_data: bool = False
+    display_cameras: bool = True
 
 
 @ControlConfig.register_subclass("record")
@@ -82,11 +82,13 @@ class RecordControlConfig(ControlConfig):
     # Not enough threads might cause low camera fps.
     num_image_writer_threads_per_camera: int = 4
     # Display all cameras on screen
-    display_data: bool = False
+    display_cameras: bool = True
     # Use vocal synthesis to read events.
     play_sounds: bool = True
     # Resume recording on an existing dataset.
     resume: bool = False
+    # Reset follower arms to an initial configuration.
+    reset_follower_arms: bool = True
 
     def __post_init__(self):
         # HACK: We parse again the cli args here to get the pretrained path if there was one.
@@ -116,11 +118,6 @@ class ReplayControlConfig(ControlConfig):
 @dataclass
 class RemoteRobotConfig(ControlConfig):
     log_interval: int = 100
-    # Display all cameras on screen
-    display_data: bool = False
-    # Rerun configuration for remote robot (https://ref.rerun.io/docs/python/0.22.1/common/initialization_functions/#rerun.connect_tcp)
-    viewer_ip: str | None = None
-    viewer_port: str | None = None
 
 
 @dataclass

lerobot/common/robot_devices/control_utils.py

@@ -24,7 +24,8 @@ from contextlib import nullcontext
 from copy import copy
 from functools import cache
 
-import rerun as rr
+import cv2
+import numpy as np
 import torch
 from deepdiff import DeepDiff
 from termcolor import colored
@@ -58,7 +59,7 @@ def log_control_info(robot: Robot, dt_s, episode_index=None, frame_index=None, f
     log_dt("dt", dt_s)
 
     # TODO(aliberts): move robot-specific logs logic in robot.print_logs()
-    if not robot.robot_type.startswith(("stretch", "realman")):
+    if not robot.robot_type.startswith("stretch"):
         for name in robot.leader_arms:
             key = f"read_leader_{name}_pos_dt_s"
             if key in robot.logs:
@@ -109,10 +110,6 @@ def predict_action(observation, policy, device, use_amp):
     ):
         # Convert to pytorch format: channel first and float32 in [0,1] with batch dimension
         for name in observation:
-            # Skip all observations that are not tensors (e.g. text)
-            if not isinstance(observation[name], torch.Tensor):
-                continue
-
             if "image" in name:
                 observation[name] = observation[name].type(torch.float32) / 255
                 observation[name] = observation[name].permute(2, 0, 1).contiguous()
@@ -132,14 +129,22 @@ def predict_action(observation, policy, device, use_amp):
     return action
 
 
-def init_keyboard_listener():
-    # Allow to exit early while recording an episode or resetting the environment,
-    # by tapping the right arrow key '->'. This might require a sudo permission
-    # to allow your terminal to monitor keyboard events.
+def init_keyboard_listener(assign_rewards=False):
+    """
+    Initializes a keyboard listener to enable early termination of an episode
+    or environment reset by pressing the right arrow key ('->'). This may require
+    sudo permissions to allow the terminal to monitor keyboard events.
+
+    Args:
+        assign_rewards (bool): If True, allows annotating the collected trajectory
+        with a binary reward at the end of the episode to indicate success.
+    """
     events = {}
     events["exit_early"] = False
     events["rerecord_episode"] = False
     events["stop_recording"] = False
+    if assign_rewards:
+        events["next.reward"] = 0
 
     if is_headless():
         logging.warning(
@@ -164,6 +169,13 @@ def init_keyboard_listener():
                 print("Escape key pressed. Stopping data recording...")
                 events["stop_recording"] = True
                 events["exit_early"] = True
+            elif assign_rewards and key == keyboard.Key.space:
+                events["next.reward"] = 1 if events["next.reward"] == 0 else 0
+                print(
+                    "Space key pressed. Assigning new reward to the subsequent frames. New reward:",
+                    events["next.reward"],
+                )
+
         except Exception as e:
             print(f"Error handling key press: {e}")
 
@@ -178,13 +190,13 @@ def warmup_record(
     events,
     enable_teleoperation,
     warmup_time_s,
-    display_data,
+    display_cameras,
     fps,
 ):
     control_loop(
         robot=robot,
         control_time_s=warmup_time_s,
-        display_data=display_data,
+        display_cameras=display_cameras,
         events=events,
         fps=fps,
         teleoperate=enable_teleoperation,
@@ -196,7 +208,7 @@ def record_episode(
     dataset,
     events,
     episode_time_s,
-    display_data,
+    display_cameras,
     policy,
     fps,
     single_task,
@@ -204,7 +216,7 @@ def record_episode(
     control_loop(
         robot=robot,
         control_time_s=episode_time_s,
-        display_data=display_data,
+        display_cameras=display_cameras,
         dataset=dataset,
         events=events,
         policy=policy,
@@ -219,7 +231,7 @@ def control_loop(
     robot,
     control_time_s=None,
     teleoperate=False,
-    display_data=False,
+    display_cameras=False,
     dataset: LeRobotDataset | None = None,
     events=None,
     policy: PreTrainedPolicy = None,
@@ -247,24 +259,22 @@ def control_loop(
 
     timestamp = 0
     start_episode_t = time.perf_counter()
-
-    # Controls starts, if policy is given it needs cleaning up
-    if policy is not None:
-        policy.reset()
-
     while timestamp < control_time_s:
         start_loop_t = time.perf_counter()
 
+        current_joint_positions = robot.follower_arms["main"].read("Present_Position")
+
         if teleoperate:
             observation, action = robot.teleop_step(record_data=True)
         else:
             observation = robot.capture_observation()
-            action = None
-            observation["task"] = [single_task]
-            observation["robot_type"] = [policy.robot_type] if hasattr(policy, "robot_type") else [""]
+
             if policy is not None:
                 pred_action = predict_action(
-                    observation, policy, get_safe_torch_device(policy.config.device), policy.config.use_amp
+                    observation,
+                    policy,
+                    get_safe_torch_device(policy.config.device),
+                    policy.config.use_amp,
                 )
                 # Action can eventually be clipped using `max_relative_target`,
                 # so action actually sent is saved in the dataset.
@@ -272,20 +282,17 @@ def control_loop(
                 action = {"action": action}
 
         if dataset is not None:
-            observation = {k: v for k, v in observation.items() if k not in ["task", "robot_type"]}
             frame = {**observation, **action, "task": single_task}
             dataset.add_frame(frame)
 
-        # TODO(Steven): This should be more general (for RemoteRobot instead of checking the name, but anyways it will change soon)
-        if (display_data and not is_headless()) or (display_data and robot.robot_type.startswith("lekiwi")):
-            if action is not None:
-                for k, v in action.items():
-                    for i, vv in enumerate(v):
-                        rr.log(f"sent_{k}_{i}", rr.Scalar(vv.numpy()))
+            # if frame["next.done"]:
+            # break
 
+        if display_cameras and not is_headless():
             image_keys = [key for key in observation if "image" in key]
             for key in image_keys:
-                rr.log(key, rr.Image(observation[key].numpy()), static=True)
+                cv2.imshow(key, cv2.cvtColor(observation[key].numpy(), cv2.COLOR_RGB2BGR))
+            cv2.waitKey(1)
 
         if fps is not None:
             dt_s = time.perf_counter() - start_loop_t
@@ -314,11 +321,25 @@ def reset_environment(robot, events, reset_time_s, fps):
     )
 
 
-def stop_recording(robot, listener, display_data):
+def reset_follower_position(robot: Robot, target_position):
+    current_position = robot.follower_arms["main"].read("Present_Position")
+    trajectory = torch.from_numpy(
+        np.linspace(current_position, target_position, 50)
+    )  # NOTE: 30 is just an aribtrary number
+    for pose in trajectory:
+        robot.send_action(pose)
+        busy_wait(0.015)
+
+
+def stop_recording(robot, listener, display_cameras):
     robot.disconnect()
 
-    if not is_headless() and listener is not None:
-        listener.stop()
+    if not is_headless():
+        if listener is not None:
+            listener.stop()
+
+        if display_cameras:
+            cv2.destroyAllWindows()
 
 
 def sanity_check_dataset_name(repo_id, policy_cfg):
@@ -340,12 +361,20 @@ def sanity_check_dataset_name(repo_id, policy_cfg):
 
 
 def sanity_check_dataset_robot_compatibility(
-    dataset: LeRobotDataset, robot: Robot, fps: int, use_videos: bool
+    dataset: LeRobotDataset,
+    robot: Robot,
+    fps: int,
+    use_videos: bool,
+    extra_features: dict = None,
 ) -> None:
+    features_from_robot = get_features_from_robot(robot, use_videos)
+    if extra_features is not None:
+        features_from_robot.update(extra_features)
+
     fields = [
         ("robot_type", dataset.meta.robot_type, robot.robot_type),
         ("fps", dataset.fps, fps),
-        ("features", dataset.features, get_features_from_robot(robot, use_videos)),
+        ("features", dataset.features, features_from_robot),
     ]
 
     mismatches = []

lerobot/common/robot_devices/motors/configs.py

@@ -39,12 +39,3 @@ class FeetechMotorsBusConfig(MotorsBusConfig):
     port: str
     motors: dict[str, tuple[int, str]]
     mock: bool = False
-
-
-@MotorsBusConfig.register_subclass("realman")
-@dataclass
-class RealmanMotorsBusConfig(MotorsBusConfig):
-    ip: str
-    port: int
-    motors: dict[str, tuple[int, str]]
-    init_joint: dict[str, list]

lerobot/common/robot_devices/motors/dynamixel.py

@@ -23,7 +23,10 @@ import numpy as np
 import tqdm
 
 from lerobot.common.robot_devices.motors.configs import DynamixelMotorsBusConfig
-from lerobot.common.robot_devices.utils import RobotDeviceAlreadyConnectedError, RobotDeviceNotConnectedError
+from lerobot.common.robot_devices.utils import (
+    RobotDeviceAlreadyConnectedError,
+    RobotDeviceNotConnectedError,
+)
 from lerobot.common.utils.utils import capture_timestamp_utc
 
 PROTOCOL_VERSION = 2.0
@@ -784,7 +787,12 @@ class DynamixelMotorsBus:
                 f"{self.packet_handler.getTxRxResult(comm)}"
             )
 
-    def write(self, data_name, values: int | float | np.ndarray, motor_names: str | list[str] | None = None):
+    def write(
+        self,
+        data_name,
+        values: int | float | np.ndarray,
+        motor_names: str | list[str] | None = None,
+    ):
         if not self.is_connected:
             raise RobotDeviceNotConnectedError(
                 f"DynamixelMotorsBus({self.port}) is not connected. You need to run `motors_bus.connect()`."

lerobot/common/robot_devices/motors/feetech.py

@@ -23,7 +23,10 @@ import numpy as np
 import tqdm
 
 from lerobot.common.robot_devices.motors.configs import FeetechMotorsBusConfig
-from lerobot.common.robot_devices.utils import RobotDeviceAlreadyConnectedError, RobotDeviceNotConnectedError
+from lerobot.common.robot_devices.utils import (
+    RobotDeviceAlreadyConnectedError,
+    RobotDeviceNotConnectedError,
+)
 from lerobot.common.utils.utils import capture_timestamp_utc
 
 PROTOCOL_VERSION = 0
@@ -809,7 +812,12 @@ class FeetechMotorsBus:
                 f"{self.packet_handler.getTxRxResult(comm)}"
             )
 
-    def write(self, data_name, values: int | float | np.ndarray, motor_names: str | list[str] | None = None):
+    def write(
+        self,
+        data_name,
+        values: int | float | np.ndarray,
+        motor_names: str | list[str] | None = None,
+    ):
         if not self.is_connected:
             raise RobotDeviceNotConnectedError(
                 f"FeetechMotorsBus({self.port}) is not connected. You need to run `motors_bus.connect()`."

lerobot/common/robot_devices/motors/realman.py

@@ -1,150 +0,0 @@
-import time
-from typing import Dict
-from lerobot.common.robot_devices.motors.configs import RealmanMotorsBusConfig
-from Robotic_Arm.rm_robot_interface import *
-
-
-class RealmanMotorsBus:
-    """
-        对Realman SDK的二次封装
-    """
-    def __init__(self, 
-                 config: RealmanMotorsBusConfig):
-        self.rmarm = RoboticArm(rm_thread_mode_e.RM_TRIPLE_MODE_E)
-        self.handle = self.rmarm.rm_create_robot_arm(config.ip, config.port)
-        self.motors = config.motors
-        self.init_joint_position = config.init_joint['joint'] # [6 joints + 1 gripper]
-        self.safe_disable_position = config.init_joint['joint']
-        self.rmarm.rm_movej(self.init_joint_position[:-1], 5, 0, 0, 1)
-        time.sleep(3)
-        ret = self.rmarm.rm_get_current_arm_state()
-        self.init_pose = ret[1]['pose']
-
-    @property
-    def motor_names(self) -> list[str]:
-        return list(self.motors.keys())
-
-    @property
-    def motor_models(self) -> list[str]:
-        return [model for _, model in self.motors.values()]
-
-    @property
-    def motor_indices(self) -> list[int]:
-        return [idx for idx, _ in self.motors.values()]
-
-
-    def connect(self, enable=True) -> bool:
-        '''
-        使能机械臂并检测使能状态,尝试5s,如果使能超时则退出程序
-        '''
-        enable_flag = False
-        loop_flag = False
-        # 设置超时时间（秒）
-        timeout = 5
-        # 记录进入循环前的时间
-        start_time = time.time()
-        elapsed_time_flag = False
-        
-        while not loop_flag:
-            elapsed_time = time.time() - start_time
-            print("--------------------")
-            
-            if enable:
-                # 获取机械臂状态
-                ret = self.rmarm.rm_get_current_arm_state()
-                if ret[0] == 0:  # 成功获取状态
-                    enable_flag = True
-            else:
-                enable_flag = False
-                # 断开所有连接，销毁线程
-                RoboticArm.rm_destory()
-            print("使能状态:", enable_flag)
-            print("--------------------")
-            if(enable_flag == enable):
-                loop_flag = True
-                enable_flag = True
-            else: 
-                loop_flag = False
-                enable_flag = False
-            # 检查是否超过超时时间
-            if elapsed_time > timeout:
-                print("超时....")
-                elapsed_time_flag = True
-                enable_flag = True
-                break
-            time.sleep(1)
-
-        resp = enable_flag
-        print(f"Returning response: {resp}")
-        return resp
-    
-    def motor_names(self):
-        return
-
-    def set_calibration(self):
-        return
-    
-    def revert_calibration(self):
-        return
-
-    def apply_calibration(self):
-        """
-            移动到初始位置
-        """
-        self.write(target_joint=self.init_joint_position)
-
-    def write(self, target_joint:list):
-        # self.rmarm.rm_movej(target_joint[:-1], 50, 0, 0, 1)
-        self.rmarm.rm_movej_follow(target_joint[:-1])
-        self.rmarm.rm_set_gripper_position(target_joint[-1], block=False, timeout=2)
-    
-    def write_endpose(self, target_endpose: list, gripper: int):
-        self.rmarm.rm_movej_p(target_endpose, 50, 0, 0, 1)
-        self.rmarm.rm_set_gripper_position(gripper, block=False, timeout=2)
-
-    def write_joint_slow(self, target_joint: list):
-        self.rmarm.rm_movej(target_joint, 5, 0, 0, 0)
-
-    def write_joint_canfd(self, target_joint: list):
-        self.rmarm.rm_movej_canfd(target_joint, False)
-
-    def write_endpose_canfd(self, target_pose: list):
-        self.rmarm.rm_movep_canfd(target_pose, False)
-
-    def write_gripper(self, gripper: int):
-        self.rmarm.rm_set_gripper_position(gripper, False, 2)
-
-    def read(self) -> Dict:
-        """
-            - 机械臂关节消息,单位1度;[-1, 1]
-            - 机械臂夹爪消息,[-1, 1]
-        """
-        joint_msg = self.rmarm.rm_get_current_arm_state()[1]
-        joint_state = joint_msg['joint']
-
-        gripper_msg = self.rmarm.rm_get_gripper_state()[1]
-        gripper_state = gripper_msg['actpos']
-        
-        return {
-            "joint_1": joint_state[0]/180,
-            "joint_2": joint_state[1]/180,
-            "joint_3": joint_state[2]/180,
-            "joint_4": joint_state[3]/180,
-            "joint_5": joint_state[4]/180,
-            "joint_6": joint_state[5]/180,
-            "gripper": (gripper_state-500)/500
-        }
-    
-    def read_current_arm_joint_state(self):
-        return self.rmarm.rm_get_current_arm_state()[1]['joint']
-    
-    def read_current_arm_endpose_state(self):
-        return self.rmarm.rm_get_current_arm_state()[1]['pose']
-
-    def safe_disconnect(self):
-        """ 
-            Move to safe disconnect position
-        """
-        self.write(target_joint=self.safe_disable_position)
-        # 断开所有连接，销毁线程
-        RoboticArm.rm_destory()

lerobot/common/robot_devices/motors/utils.py

@@ -30,7 +30,9 @@ class MotorsBus(Protocol):
     def write(self): ...
 
 
-def make_motors_buses_from_configs(motors_bus_configs: dict[str, MotorsBusConfig]) -> list[MotorsBus]:
+def make_motors_buses_from_configs(
+    motors_bus_configs: dict[str, MotorsBusConfig],
+) -> list[MotorsBus]:
     motors_buses = {}
 
     for key, cfg in motors_bus_configs.items():
@@ -44,11 +46,6 @@ def make_motors_buses_from_configs(motors_bus_configs: dict[str, MotorsBusConfig
 
             motors_buses[key] = FeetechMotorsBus(cfg)
 
-        elif cfg.type == "realman":
-            from lerobot.common.robot_devices.motors.realman import RealmanMotorsBus
-
-            motors_buses[key] = RealmanMotorsBus(cfg)
-
         else:
             raise ValueError(f"The motor type '{cfg.type}' is not valid.")
 
@@ -70,7 +67,3 @@ def make_motors_bus(motor_type: str, **kwargs) -> MotorsBus:
 
     else:
         raise ValueError(f"The motor type '{motor_type}' is not valid.")
-
-
-def get_motor_names(arm: dict[str, MotorsBus]) -> list:
-    return [f"{arm}_{motor}" for arm, bus in arm.items() for motor in bus.motors]

lerobot/common/robot_devices/robots/configs.py

@@ -27,7 +27,6 @@ from lerobot.common.robot_devices.motors.configs import (
     DynamixelMotorsBusConfig,
     FeetechMotorsBusConfig,
     MotorsBusConfig,
-    RealmanMotorsBusConfig
 )
 
 
@@ -432,69 +431,6 @@ class MossRobotConfig(ManipulatorRobotConfig):
     mock: bool = False
 
 
-@RobotConfig.register_subclass("so101")
-@dataclass
-class So101RobotConfig(ManipulatorRobotConfig):
-    calibration_dir: str = ".cache/calibration/so101"
-    # `max_relative_target` limits the magnitude of the relative positional target vector for safety purposes.
-    # Set this to a positive scalar to have the same value for all motors, or a list that is the same length as
-    # the number of motors in your follower arms.
-    max_relative_target: int | None = None
-
-    leader_arms: dict[str, MotorsBusConfig] = field(
-        default_factory=lambda: {
-            "main": FeetechMotorsBusConfig(
-                port="/dev/tty.usbmodem58760431091",
-                motors={
-                    # name: (index, model)
-                    "shoulder_pan": [1, "sts3215"],
-                    "shoulder_lift": [2, "sts3215"],
-                    "elbow_flex": [3, "sts3215"],
-                    "wrist_flex": [4, "sts3215"],
-                    "wrist_roll": [5, "sts3215"],
-                    "gripper": [6, "sts3215"],
-                },
-            ),
-        }
-    )
-
-    follower_arms: dict[str, MotorsBusConfig] = field(
-        default_factory=lambda: {
-            "main": FeetechMotorsBusConfig(
-                port="/dev/tty.usbmodem585A0076891",
-                motors={
-                    # name: (index, model)
-                    "shoulder_pan": [1, "sts3215"],
-                    "shoulder_lift": [2, "sts3215"],
-                    "elbow_flex": [3, "sts3215"],
-                    "wrist_flex": [4, "sts3215"],
-                    "wrist_roll": [5, "sts3215"],
-                    "gripper": [6, "sts3215"],
-                },
-            ),
-        }
-    )
-
-    cameras: dict[str, CameraConfig] = field(
-        default_factory=lambda: {
-            "laptop": OpenCVCameraConfig(
-                camera_index=0,
-                fps=30,
-                width=640,
-                height=480,
-            ),
-            "phone": OpenCVCameraConfig(
-                camera_index=1,
-                fps=30,
-                width=640,
-                height=480,
-            ),
-        }
-    )
-
-    mock: bool = False
-
-
 @RobotConfig.register_subclass("so100")
 @dataclass
 class So100RobotConfig(ManipulatorRobotConfig):
@@ -507,7 +443,7 @@ class So100RobotConfig(ManipulatorRobotConfig):
     leader_arms: dict[str, MotorsBusConfig] = field(
         default_factory=lambda: {
             "main": FeetechMotorsBusConfig(
-                port="/dev/tty.usbmodem58760431091",
+                port="/dev/tty.usbmodem58760433331",
                 motors={
                     # name: (index, model)
                     "shoulder_pan": [1, "sts3215"],
@@ -524,7 +460,7 @@ class So100RobotConfig(ManipulatorRobotConfig):
     follower_arms: dict[str, MotorsBusConfig] = field(
         default_factory=lambda: {
             "main": FeetechMotorsBusConfig(
-                port="/dev/tty.usbmodem585A0076891",
+                port="/dev/tty.usbmodem58760431631",
                 motors={
                     # name: (index, model)
                     "shoulder_pan": [1, "sts3215"],
@@ -675,91 +611,3 @@ class LeKiwiRobotConfig(RobotConfig):
     )
 
     mock: bool = False
-
-
-@RobotConfig.register_subclass("realman")
-@dataclass
-class RealmanRobotConfig(RobotConfig):
-    inference_time: bool = False
-    max_gripper: int = 990
-    min_gripper: int = 10
-    servo_config_file: str = "/home/maic/LYT/lerobot/lerobot/common/robot_devices/teleop/servo_arm.yaml"
-    
-
-    left_follower_arm: dict[str, MotorsBusConfig] = field(
-        default_factory=lambda: {
-            "main": RealmanMotorsBusConfig(
-                ip = "192.168.3.18",
-                port = 8080,
-                motors={
-                    # name: (index, model)
-                    "joint_1": [1, "realman"],
-                    "joint_2": [2, "realman"],
-                    "joint_3": [3, "realman"],
-                    "joint_4": [4, "realman"],
-                    "joint_5": [5, "realman"],
-                    "joint_6": [6, "realman"],
-                    "gripper": [7, "realman"],
-                },
-                init_joint = {'joint': [-90, 90, 90, 90, 90, -90, 10]}
-            )
-        }
-    )
-
-    cameras: dict[str, CameraConfig] = field(
-        default_factory=lambda: {
-            # "one": OpenCVCameraConfig(
-            #     camera_index=4,
-            #     fps=30,
-            #     width=640,
-            #     height=480,
-            # ),
-            "left": IntelRealSenseCameraConfig(
-                serial_number="153122077516",
-                fps=30,
-                width=640,
-                height=480,
-                use_depth=False
-            ),
-            # "right": IntelRealSenseCameraConfig(
-            #     serial_number="405622075165",
-            #     fps=30,
-            #     width=640,
-            #     height=480,
-            #     use_depth=False
-            # ),
-            "front": IntelRealSenseCameraConfig(
-                serial_number="145422072751",
-                fps=30,
-                width=640,
-                height=480,
-                use_depth=False
-            ),
-            "high": IntelRealSenseCameraConfig(
-                serial_number="145422072193",
-                fps=30,
-                width=640,
-                height=480,
-                use_depth=False
-            ),
-        }
-    )
-
-    # right_follower_arm: dict[str, MotorsBusConfig] = field(
-    #     default_factory=lambda: {
-    #         "main": RealmanMotorsBusConfig(
-    #             ip = "192.168.3.19",
-    #             port = 8080,
-    #             motors={
-    #                 # name: (index, model)
-    #                 "joint_1": [1, "realman"],
-    #                 "joint_2": [2, "realman"],
-    #                 "joint_3": [3, "realman"],
-    #                 "joint_4": [4, "realman"],
-    #                 "joint_5": [5, "realman"],
-    #                 "joint_6": [6, "realman"],
-    #                 "gripper": (7, "realman"),
-    #             },
-    #         )
-    #     }
-    # )

lerobot/common/robot_devices/robots/feetech_calibration.py

@@ -36,12 +36,6 @@ ZERO_POSITION_DEGREE = 0
 ROTATED_POSITION_DEGREE = 90
 
 
-def reset_middle_positions(arm: MotorsBus):
-    input("Please move the robot to the new middle position for calibration, then press Enter...")
-    # Write 128 to Torque_Enable for all motors.
-    arm.write("Torque_Enable", 128)
-
-
 def assert_drive_mode(drive_mode):
     # `drive_mode` is in [0,1] with 0 means original rotation direction for the motor, and 1 means inverted.
     if not np.all(np.isin(drive_mode, [0, 1])):
@@ -213,7 +207,10 @@ def run_arm_auto_calibration_so100(arm: MotorsBus, robot_type: str, arm_name: st
 
     print("Calibrate elbow_flex")
     calib["elbow_flex"] = move_to_calibrate(
-        arm, "elbow_flex", positive_first=False, in_between_move_hook=in_between_move_hook
+        arm,
+        "elbow_flex",
+        positive_first=False,
+        in_between_move_hook=in_between_move_hook,
     )
     calib["elbow_flex"] = apply_offset(calib["elbow_flex"], offset=80 - 1024)
 
@@ -245,7 +242,11 @@ def run_arm_auto_calibration_so100(arm: MotorsBus, robot_type: str, arm_name: st
         }
         arm.write("Goal_Position", list(positions.values()), list(positions.keys()))
 
-    arm.write("Goal_Position", round(calib["shoulder_lift"]["zero_pos"] - 1600), "shoulder_lift")
+    arm.write(
+        "Goal_Position",
+        round(calib["shoulder_lift"]["zero_pos"] - 1600),
+        "shoulder_lift",
+    )
     time.sleep(2)
     arm.write("Goal_Position", round(calib["elbow_flex"]["zero_pos"] + 1700), "elbow_flex")
     time.sleep(2)
@@ -256,7 +257,11 @@ def run_arm_auto_calibration_so100(arm: MotorsBus, robot_type: str, arm_name: st
 
     print("Calibrate wrist_roll")
     calib["wrist_roll"] = move_to_calibrate(
-        arm, "wrist_roll", invert_drive_mode=True, positive_first=False, while_move_hook=while_move_hook
+        arm,
+        "wrist_roll",
+        invert_drive_mode=True,
+        positive_first=False,
+        while_move_hook=while_move_hook,
     )
 
     arm.write("Goal_Position", calib["wrist_roll"]["zero_pos"], "wrist_roll")
@@ -445,8 +450,6 @@ def run_arm_manual_calibration(arm: MotorsBus, robot_type: str, arm_name: str, a
 
     print(f"\nRunning calibration of {robot_type} {arm_name} {arm_type}...")
 
-    reset_middle_positions(arm)
-
     print("\nMove arm to zero position")
     print("See: " + URL_TEMPLATE.format(robot=robot_type, arm=arm_type, position="zero"))
     input("Press Enter to continue...")

lerobot/common/robot_devices/robots/lekiwi_remote.py

@@ -61,7 +61,9 @@ def calibrate_follower_arm(motors_bus, calib_dir_str):
     calib_dir.mkdir(parents=True, exist_ok=True)
     calib_file = calib_dir / "main_follower.json"
     try:
-        from lerobot.common.robot_devices.robots.feetech_calibration import run_arm_manual_calibration
+        from lerobot.common.robot_devices.robots.feetech_calibration import (
+            run_arm_manual_calibration,
+        )
     except ImportError:
         print("[WARNING] Calibration function not available. Skipping calibration.")
         return
@@ -116,7 +118,14 @@ def run_lekiwi(robot_config):
     robot = LeKiwi(motors_bus)
 
     # Define the expected arm motor IDs.
-    arm_motor_ids = ["shoulder_pan", "shoulder_lift", "elbow_flex", "wrist_flex", "wrist_roll", "gripper"]
+    arm_motor_ids = [
+        "shoulder_pan",
+        "shoulder_lift",
+        "elbow_flex",
+        "wrist_flex",
+        "wrist_roll",
+        "gripper",
+    ]
 
     # Disable torque for each arm motor.
     for motor in arm_motor_ids:
@@ -130,7 +139,9 @@ def run_lekiwi(robot_config):
     images_lock = threading.Lock()
     stop_event = threading.Event()
     cam_thread = threading.Thread(
-        target=run_camera_capture, args=(cameras, images_lock, latest_images_dict, stop_event), daemon=True
+        target=run_camera_capture,
+        args=(cameras, images_lock, latest_images_dict, stop_event),
+        daemon=True,
     )
     cam_thread.start()
 

lerobot/common/robot_devices/robots/manipulator.py

@@ -28,14 +28,22 @@ import numpy as np
 import torch
 
 from lerobot.common.robot_devices.cameras.utils import make_cameras_from_configs
-from lerobot.common.robot_devices.motors.utils import MotorsBus, make_motors_buses_from_configs
+from lerobot.common.robot_devices.motors.utils import (
+    MotorsBus,
+    make_motors_buses_from_configs,
+)
 from lerobot.common.robot_devices.robots.configs import ManipulatorRobotConfig
 from lerobot.common.robot_devices.robots.utils import get_arm_id
-from lerobot.common.robot_devices.utils import RobotDeviceAlreadyConnectedError, RobotDeviceNotConnectedError
+from lerobot.common.robot_devices.utils import (
+    RobotDeviceAlreadyConnectedError,
+    RobotDeviceNotConnectedError,
+)
 
 
 def ensure_safe_goal_position(
-    goal_pos: torch.Tensor, present_pos: torch.Tensor, max_relative_target: float | list[float]
+    goal_pos: torch.Tensor,
+    present_pos: torch.Tensor,
+    max_relative_target: float | list[float],
 ):
     # Cap relative action target magnitude for safety.
     diff = goal_pos - present_pos
@@ -45,7 +53,7 @@ def ensure_safe_goal_position(
     safe_goal_pos = present_pos + safe_diff
 
     if not torch.allclose(goal_pos, safe_goal_pos):
-        logging.warning(
+        logging.debug(
             "Relative goal position magnitude had to be clamped to be safe.\n"
             f"  requested relative goal position target: {diff}\n"
             f"    clamped relative goal position target: {safe_diff}"
@@ -243,7 +251,7 @@ class ManipulatorRobot:
 
         if self.robot_type in ["koch", "koch_bimanual", "aloha"]:
             from lerobot.common.robot_devices.motors.dynamixel import TorqueMode
-        elif self.robot_type in ["so100", "so101", "moss", "lekiwi"]:
+        elif self.robot_type in ["so100", "moss", "lekiwi"]:
             from lerobot.common.robot_devices.motors.feetech import TorqueMode
 
         # We assume that at connection time, arms are in a rest position, and torque can
@@ -260,7 +268,7 @@ class ManipulatorRobot:
             self.set_koch_robot_preset()
         elif self.robot_type == "aloha":
             self.set_aloha_robot_preset()
-        elif self.robot_type in ["so100", "so101", "moss", "lekiwi"]:
+        elif self.robot_type in ["so100", "moss", "lekiwi"]:
             self.set_so100_robot_preset()
 
         # Enable torque on all motors of the follower arms
@@ -309,11 +317,13 @@ class ManipulatorRobot:
                 print(f"Missing calibration file '{arm_calib_path}'")
 
                 if self.robot_type in ["koch", "koch_bimanual", "aloha"]:
-                    from lerobot.common.robot_devices.robots.dynamixel_calibration import run_arm_calibration
+                    from lerobot.common.robot_devices.robots.dynamixel_calibration import (
+                        run_arm_calibration,
+                    )
 
                     calibration = run_arm_calibration(arm, self.robot_type, name, arm_type)
 
-                elif self.robot_type in ["so100", "so101", "moss", "lekiwi"]:
+                elif self.robot_type in ["so100", "moss", "lekiwi"]:
                     from lerobot.common.robot_devices.robots.feetech_calibration import (
                         run_arm_manual_calibration,
                     )
@@ -464,6 +474,14 @@ class ManipulatorRobot:
             before_fwrite_t = time.perf_counter()
             goal_pos = leader_pos[name]
 
+            # If specified, clip the goal positions within predefined bounds specified in the config of the robot
+            # if self.config.joint_position_relative_bounds is not None:
+            #     goal_pos = torch.clamp(
+            #         goal_pos,
+            #         self.config.joint_position_relative_bounds["min"],
+            #         self.config.joint_position_relative_bounds["max"],
+            #     )
+
             # Cap goal position when too far away from present position.
             # Slower fps expected due to reading from the follower.
             if self.config.max_relative_target is not None:
@@ -585,6 +603,14 @@ class ManipulatorRobot:
             goal_pos = action[from_idx:to_idx]
             from_idx = to_idx
 
+            # If specified, clip the goal positions within predefined bounds specified in the config of the robot
+            # if self.config.joint_position_relative_bounds is not None:
+            #     goal_pos = torch.clamp(
+            #         goal_pos,
+            #         self.config.joint_position_relative_bounds["min"],
+            #         self.config.joint_position_relative_bounds["max"],
+            #     )
+
             # Cap goal position when too far away from present position.
             # Slower fps expected due to reading from the follower.
             if self.config.max_relative_target is not None:

lerobot/common/robot_devices/robots/mobile_manipulator.py

@@ -25,9 +25,14 @@ import zmq
 
 from lerobot.common.robot_devices.cameras.utils import make_cameras_from_configs
 from lerobot.common.robot_devices.motors.feetech import TorqueMode
-from lerobot.common.robot_devices.motors.utils import MotorsBus, make_motors_buses_from_configs
+from lerobot.common.robot_devices.motors.utils import (
+    MotorsBus,
+    make_motors_buses_from_configs,
+)
 from lerobot.common.robot_devices.robots.configs import LeKiwiRobotConfig
-from lerobot.common.robot_devices.robots.feetech_calibration import run_arm_manual_calibration
+from lerobot.common.robot_devices.robots.feetech_calibration import (
+    run_arm_manual_calibration,
+)
 from lerobot.common.robot_devices.robots.utils import get_arm_id
 from lerobot.common.robot_devices.utils import RobotDeviceNotConnectedError
 
@@ -324,7 +329,11 @@ class MobileManipulator:
         socks = dict(poller.poll(15))
         if self.video_socket not in socks or socks[self.video_socket] != zmq.POLLIN:
             # No new data arrived → reuse ALL old data
-            return (self.last_frames, self.last_present_speed, self.last_remote_arm_state)
+            return (
+                self.last_frames,
+                self.last_present_speed,
+                self.last_remote_arm_state,
+            )
 
         # Drain all messages, keep only the last
         last_msg = None
@@ -337,7 +346,11 @@ class MobileManipulator:
 
         if not last_msg:
             # No new message → also reuse old
-            return (self.last_frames, self.last_present_speed, self.last_remote_arm_state)
+            return (
+                self.last_frames,
+                self.last_present_speed,
+                self.last_remote_arm_state,
+            )
 
         # Decode only the final message
         try:
@@ -375,7 +388,11 @@ class MobileManipulator:
         except Exception as e:
             print(f"[DEBUG] Error decoding video message: {e}")
             # If decode fails, fall back to old data
-            return (self.last_frames, self.last_present_speed, self.last_remote_arm_state)
+            return (
+                self.last_frames,
+                self.last_present_speed,
+                self.last_remote_arm_state,
+            )
 
         return frames, present_speed, remote_arm_state_tensor
 
@@ -461,7 +478,11 @@ class MobileManipulator:
 
         body_state = self.wheel_raw_to_body(present_speed)
 
-        body_state_mm = (body_state[0] * 1000.0, body_state[1] * 1000.0, body_state[2])  # Convert x,y to mm/s
+        body_state_mm = (
+            body_state[0] * 1000.0,
+            body_state[1] * 1000.0,
+            body_state[2],
+        )  # Convert x,y to mm/s
         wheel_state_tensor = torch.tensor(body_state_mm, dtype=torch.float32)
         combined_state_tensor = torch.cat((remote_arm_state_tensor, wheel_state_tensor), dim=0)
 
@@ -620,7 +641,11 @@ class MobileManipulator:
         # Convert each wheel’s angular speed (deg/s) to a raw integer.
         wheel_raw = [MobileManipulator.degps_to_raw(deg) for deg in wheel_degps]
 
-        return {"left_wheel": wheel_raw[0], "back_wheel": wheel_raw[1], "right_wheel": wheel_raw[2]}
+        return {
+            "left_wheel": wheel_raw[0],
+            "back_wheel": wheel_raw[1],
+            "right_wheel": wheel_raw[2],
+        }
 
     def wheel_raw_to_body(
         self, wheel_raw: dict, wheel_radius: float = 0.05, base_radius: float = 0.125

lerobot/common/robot_devices/robots/realman.py

@@ -1,292 +0,0 @@
-"""
-    Teleoperation Realman with a PS5 controller and 
-"""
-
-import time
-import torch
-import numpy as np
-from dataclasses import dataclass, field, replace
-from collections import deque
-from lerobot.common.robot_devices.teleop.gamepad import HybridController
-from lerobot.common.robot_devices.motors.utils import get_motor_names, make_motors_buses_from_configs
-from lerobot.common.robot_devices.cameras.utils import make_cameras_from_configs
-from lerobot.common.robot_devices.utils import RobotDeviceAlreadyConnectedError, RobotDeviceNotConnectedError
-from lerobot.common.robot_devices.robots.configs import RealmanRobotConfig
-
-
-class RealmanRobot:
-    def __init__(self, config: RealmanRobotConfig | None = None, **kwargs):
-        if config is None:
-            config = RealmanRobotConfig()
-        # Overwrite config arguments using kwargs
-        self.config = replace(config, **kwargs)
-        self.robot_type = self.config.type
-        self.inference_time = self.config.inference_time # if it is inference time
-        
-        # build cameras
-        self.cameras = make_cameras_from_configs(self.config.cameras)
-        
-        # build realman motors
-        self.piper_motors = make_motors_buses_from_configs(self.config.left_follower_arm)
-        self.arm = self.piper_motors['main']
-        
-        # build init teleop info
-        self.init_info = {
-            'init_joint': self.arm.init_joint_position,
-            'init_pose': self.arm.init_pose,
-            'max_gripper': config.max_gripper,
-            'min_gripper': config.min_gripper,
-            'servo_config_file': config.servo_config_file
-        }
-        
-        # build state-action cache
-        self.joint_queue = deque(maxlen=2)
-        self.last_endpose = self.arm.init_pose
-
-        # build gamepad teleop
-        if not self.inference_time:
-            self.teleop = HybridController(self.init_info)
-        else:
-            self.teleop = None
-        
-        self.logs = {}
-        self.is_connected = False
-
-    @property
-    def camera_features(self) -> dict:
-        cam_ft = {}
-        for cam_key, cam in self.cameras.items():
-            key = f"observation.images.{cam_key}"
-            cam_ft[key] = {
-                "shape": (cam.height, cam.width, cam.channels),
-                "names": ["height", "width", "channels"],
-                "info": None,
-            }
-        return cam_ft
-
-    
-    @property
-    def motor_features(self) -> dict:
-        action_names = get_motor_names(self.piper_motors)
-        state_names = get_motor_names(self.piper_motors)
-        return {
-            "action": {
-                "dtype": "float32",
-                "shape": (len(action_names),),
-                "names": action_names,
-            },
-            "observation.state": {
-                "dtype": "float32",
-                "shape": (len(state_names),),
-                "names": state_names,
-            },
-        }
-    
-    @property
-    def has_camera(self):
-        return len(self.cameras) > 0
-
-    @property
-    def num_cameras(self):
-        return len(self.cameras)
-
-
-    def connect(self) -> None:
-        """Connect RealmanArm and cameras"""
-        if self.is_connected:
-            raise RobotDeviceAlreadyConnectedError(
-                "RealmanArm is already connected. Do not run `robot.connect()` twice."
-            )
-        
-        # connect RealmanArm
-        self.arm.connect(enable=True)
-        print("RealmanArm conneted")
-
-        # connect cameras
-        for name in self.cameras:
-            self.cameras[name].connect()
-            self.is_connected = self.is_connected and self.cameras[name].is_connected
-            print(f"camera {name} conneted")
-        
-        print("All connected")
-        self.is_connected = True
-        
-        self.run_calibration()
-
-
-    def disconnect(self) -> None:
-        """move to home position, disenable piper and cameras"""
-        # move piper to home position, disable
-        if not self.inference_time:
-            self.teleop.stop()
-
-        # disconnect piper
-        self.arm.safe_disconnect()
-        print("RealmanArm disable after 5 seconds")
-        time.sleep(5)
-        self.arm.connect(enable=False)
-
-        # disconnect cameras
-        if len(self.cameras) > 0:
-            for cam in self.cameras.values():
-                cam.disconnect()
-
-        self.is_connected = False
-
-
-    def run_calibration(self):
-        """move piper to the home position"""
-        if not self.is_connected:
-            raise ConnectionError()
-        
-        self.arm.apply_calibration()
-        if not self.inference_time:
-            self.teleop.reset()
-
-
-    def teleop_step(
-        self, record_data=False
-    ) -> None | tuple[dict[str, torch.Tensor], dict[str, torch.Tensor]]:
-        if not self.is_connected:
-            raise ConnectionError()
-
-        if self.teleop is None and self.inference_time:
-            self.teleop = HybridController(self.init_info)
-
-        # read target pose state as 
-        before_read_t = time.perf_counter()
-        state = self.arm.read() # read current joint position from robot
-        action = self.teleop.get_action() # target joint position and pose end pos from gamepad
-        self.logs["read_pos_dt_s"] = time.perf_counter() - before_read_t
-
-        if action['control_mode'] == 'joint':
-            # 关节控制模式（主模式）
-            current_pose = self.arm.read_current_arm_endpose_state()
-            self.teleop.update_endpose_state(current_pose)
-
-            target_joints = action['joint_angles'][:-1]
-            self.arm.write_gripper(action['gripper'])
-            print(action['gripper'])
-            if action['master_controller_status']['infrared'] == 1:
-                if action['master_controller_status']['button'] == 1:
-                    self.arm.write_joint_canfd(target_joints)
-                else:
-                    self.arm.write_joint_slow(target_joints)
-                
-            # do action
-            before_write_t = time.perf_counter()
-            self.joint_queue.append(list(self.arm.read().values()))
-            self.logs["write_pos_dt_s"] = time.perf_counter() - before_write_t
-
-        else:
-            target_pose = list(action['end_pose'])
-            # do action
-            before_write_t = time.perf_counter()
-            if self.last_endpose != target_pose:
-                self.arm.write_endpose_canfd(target_pose)
-                self.last_endpose = target_pose
-            self.arm.write_gripper(action['gripper'])
-            
-            target_joints = self.arm.read_current_arm_joint_state()
-            self.joint_queue.append(list(self.arm.read().values()))
-            self.teleop.update_joint_state(target_joints)
-            self.logs["write_pos_dt_s"] = time.perf_counter() - before_write_t
-        
-        if not record_data:
-            return
-        
-        state = torch.as_tensor(list(self.joint_queue[0]), dtype=torch.float32)
-        action = torch.as_tensor(list(self.joint_queue[-1]), dtype=torch.float32)
-
-        # Capture images from cameras
-        images = {}
-        for name in self.cameras:
-            before_camread_t = time.perf_counter()
-            images[name] = self.cameras[name].async_read()
-            images[name] = torch.from_numpy(images[name])
-            self.logs[f"read_camera_{name}_dt_s"] = self.cameras[name].logs["delta_timestamp_s"]
-            self.logs[f"async_read_camera_{name}_dt_s"] = time.perf_counter() - before_camread_t
-
-        # Populate output dictionnaries
-        obs_dict, action_dict = {}, {}
-        obs_dict["observation.state"] = state
-        action_dict["action"] = action
-        for name in self.cameras:
-            obs_dict[f"observation.images.{name}"] = images[name]
-
-        return obs_dict, action_dict
-
-
-
-    def send_action(self, action: torch.Tensor) -> torch.Tensor:
-        """Write the predicted actions from policy to the motors"""
-        if not self.is_connected:
-            raise RobotDeviceNotConnectedError(
-                "Piper is not connected. You need to run `robot.connect()`."
-            )
-
-        # send to motors, torch to list
-        target_joints = action.tolist()
-        len_joint = len(target_joints) - 1
-        target_joints = [target_joints[i]*180 for i in range(len_joint)] + [target_joints[-1]]
-        target_joints[-1] = int(target_joints[-1]*500 + 500)
-        self.arm.write(target_joints)
-
-        return action
-
-
-
-    def capture_observation(self) -> dict:
-        """capture current images and joint positions"""
-        if not self.is_connected:
-            raise RobotDeviceNotConnectedError(
-                "Piper is not connected. You need to run `robot.connect()`."
-            )
-        
-        # read current joint positions
-        before_read_t = time.perf_counter()
-        state = self.arm.read()  # 6 joints + 1 gripper
-        self.logs["read_pos_dt_s"] = time.perf_counter() - before_read_t
-
-        state = torch.as_tensor(list(state.values()), dtype=torch.float32)
-
-        # read images from cameras
-        images = {}
-        for name in self.cameras:
-            before_camread_t = time.perf_counter()
-            images[name] = self.cameras[name].async_read()
-            images[name] = torch.from_numpy(images[name])
-            self.logs[f"read_camera_{name}_dt_s"] = self.cameras[name].logs["delta_timestamp_s"]
-            self.logs[f"async_read_camera_{name}_dt_s"] = time.perf_counter() - before_camread_t
-
-        # Populate output dictionnaries and format to pytorch
-        obs_dict = {}
-        obs_dict["observation.state"] = state
-        for name in self.cameras:
-            obs_dict[f"observation.images.{name}"] = images[name]
-        return obs_dict
-    
-    def teleop_safety_stop(self):
-        """ move to home position after record one episode """
-        self.run_calibration()
-
-    
-    def __del__(self):
-        if self.is_connected:
-            self.disconnect()
-            if not self.inference_time:
-                self.teleop.stop()
-
-
-if __name__ == '__main__':
-    robot = RealmanRobot()
-    robot.connect()
-    # robot.run_calibration()
-    while True:
-        robot.teleop_step(record_data=True)
-    
-    robot.capture_observation()
-    dummy_action = torch.Tensor([-0.40586111280653214, 0.5522833506266276, 0.4998166826036241, -0.3539944542778863, -0.524433347913954, 0.9064999898274739, 0.482])
-    robot.send_action(dummy_action)
-    robot.disconnect()
-    print('ok')

lerobot/common/robot_devices/robots/utils.py

@@ -23,9 +23,7 @@ from lerobot.common.robot_devices.robots.configs import (
     MossRobotConfig,
     RobotConfig,
     So100RobotConfig,
-    So101RobotConfig,
     StretchRobotConfig,
-    RealmanRobotConfig
 )
 
 
@@ -60,15 +58,10 @@ def make_robot_config(robot_type: str, **kwargs) -> RobotConfig:
         return MossRobotConfig(**kwargs)
     elif robot_type == "so100":
         return So100RobotConfig(**kwargs)
-    elif robot_type == "so101":
-        return So101RobotConfig(**kwargs)
     elif robot_type == "stretch":
         return StretchRobotConfig(**kwargs)
     elif robot_type == "lekiwi":
         return LeKiwiRobotConfig(**kwargs)
-    elif robot_type == 'realman':
-        return RealmanRobotConfig(**kwargs)
-
     else:
         raise ValueError(f"Robot type '{robot_type}' is not available.")
 
@@ -79,15 +72,11 @@ def make_robot_from_config(config: RobotConfig):
 
         return ManipulatorRobot(config)
     elif isinstance(config, LeKiwiRobotConfig):
-        from lerobot.common.robot_devices.robots.mobile_manipulator import MobileManipulator
+        from lerobot.common.robot_devices.robots.mobile_manipulator import (
+            MobileManipulator,
+        )
 
         return MobileManipulator(config)
-    
-    elif isinstance(config, RealmanRobotConfig):
-        from lerobot.common.robot_devices.robots.realman import RealmanRobot
-
-        return RealmanRobot(config)
-
     else:
         from lerobot.common.robot_devices.robots.stretch import StretchRobot
 

lerobot/common/robot_devices/teleop/gamepad.py

@@ -1,466 +0,0 @@
-import pygame
-import threading
-import time
-import serial
-import binascii
-import logging
-import yaml
-from typing import Dict
-from Robotic_Arm.rm_robot_interface import *
-
-
-
-class ServoArm:
-    def __init__(self, config_file="config.yaml"):
-        """初始化机械臂的串口连接并发送初始数据。
-
-        Args:
-            config_file (str): 配置文件的路径。
-        """
-        self.config = self._load_config(config_file)
-        self.port = self.config["port"]
-        self.baudrate = self.config["baudrate"]
-        self.joint_hex_data = self.config["joint_hex_data"]
-        self.control_hex_data = self.config["control_hex_data"]
-        self.arm_axis = self.config.get("arm_axis", 7)
-
-        try:
-            self.serial_conn = serial.Serial(self.port, self.baudrate, timeout=0)
-            self.bytes_to_send = binascii.unhexlify(self.joint_hex_data.replace(" ", ""))
-            self.serial_conn.write(self.bytes_to_send)
-            time.sleep(1)
-            self.connected = True
-            logging.info(f"串口连接成功: {self.port}")
-        except Exception as e:
-            logging.error(f"串口连接失败: {e}")
-            self.connected = False
-
-    def _load_config(self, config_file):
-        """加载配置文件。
-
-        Args:
-            config_file (str): 配置文件的路径。
-
-        Returns:
-            dict: 配置文件内容。
-        """
-        try:
-            with open(config_file, "r") as file:
-                config = yaml.safe_load(file)
-            return config
-        except Exception as e:
-            logging.error(f"配置文件加载失败: {e}")
-            # 返回默认配置
-            return {
-                "port": "/dev/ttyUSB0",
-                "baudrate": 460800,
-                "joint_hex_data": "55 AA 02 00 00 67",
-                "control_hex_data": "55 AA 08 00 00 B9",
-                "arm_axis": 6
-            }
-
-    def _bytes_to_signed_int(self, byte_data):
-        """将字节数据转换为有符号整数。
-
-        Args:
-            byte_data (bytes): 字节数据。
-
-        Returns:
-            int: 有符号整数。
-        """
-        return int.from_bytes(byte_data, byteorder="little", signed=True)
-
-    def _parse_joint_data(self, hex_received):
-        """解析接收到的十六进制数据并提取关节数据。
-
-        Args:
-            hex_received (str): 接收到的十六进制字符串数据。
-
-        Returns:
-            dict: 解析后的关节数据。
-        """
-        logging.debug(f"hex_received: {hex_received}")
-        joints = {}
-        for i in range(self.arm_axis):
-            start = 14 + i * 10
-            end = start + 8
-            joint_hex = hex_received[start:end]
-            joint_byte_data = bytearray.fromhex(joint_hex)
-            joint_value = self._bytes_to_signed_int(joint_byte_data) / 10000.0
-            joints[f"joint_{i+1}"] = joint_value
-        grasp_start = 14 + self.arm_axis*10
-        grasp_hex = hex_received[grasp_start:grasp_start+8]
-        grasp_byte_data = bytearray.fromhex(grasp_hex)
-        # 夹爪进行归一化处理
-        grasp_value = self._bytes_to_signed_int(grasp_byte_data)/1000
-
-        joints["grasp"] = grasp_value
-        return joints
-    
-    def _parse_controller_data(self, hex_received):
-        status = {
-            'infrared': 0,
-            'button': 0
-        }
-        if len(hex_received) == 18:
-            status['infrared'] = self._bytes_to_signed_int(bytearray.fromhex(hex_received[12:14]))
-            status['button'] = self._bytes_to_signed_int(bytearray.fromhex(hex_received[14:16]))
-            # print(infrared)
-        return status
-
-    def get_joint_actions(self):
-        """从串口读取数据并解析关节动作。
-
-        Returns:
-            dict: 包含关节数据的字典。
-        """
-        if not self.connected:
-            return {}
-        
-        try:
-            self.serial_conn.write(self.bytes_to_send)
-            time.sleep(0.02)
-            bytes_received = self.serial_conn.read(self.serial_conn.inWaiting())
-            if len(bytes_received) == 0:
-                return {}
-            
-            hex_received = binascii.hexlify(bytes_received).decode("utf-8").upper()
-            actions = self._parse_joint_data(hex_received)
-            return actions
-        except Exception as e:
-            logging.error(f"读取串口数据错误: {e}")
-            return {}
-
-    def get_controller_status(self):
-        bytes_to_send = binascii.unhexlify(self.control_hex_data.replace(" ", ""))
-        self.serial_conn.write(bytes_to_send)
-        time.sleep(0.02)
-        bytes_received = self.serial_conn.read(self.serial_conn.inWaiting())
-        hex_received = binascii.hexlify(bytes_received).decode("utf-8").upper()
-        # print("control status:", hex_received)
-        status = self._parse_controller_data(hex_received)
-        return status
-
-    def close(self):
-        """关闭串口连接"""
-        if self.connected and hasattr(self, 'serial_conn'):
-            self.serial_conn.close()
-            self.connected = False
-            logging.info("串口连接已关闭")
-
-
-class HybridController:
-    def __init__(self, init_info):
-        # 初始化pygame和手柄
-        pygame.init()
-        pygame.joystick.init()
-        
-        # 检查是否有连接的手柄
-        if pygame.joystick.get_count() == 0:
-            raise Exception("未检测到手柄")
-        
-        # 初始化手柄
-        self.joystick = pygame.joystick.Joystick(0)
-        self.joystick.init()
-        # 摇杆死区
-        self.deadzone = 0.15
-        # 控制模式: True为关节控制（主模式），False为末端控制
-        self.joint_control_mode = True
-        # 精细控制模式
-        self.fine_control_mode = False
-        
-        # 初始化末端姿态和关节角度
-        self.init_joint = init_info['init_joint']
-        self.init_pose = init_info.get('init_pose', [0]*6)
-        self.max_gripper = init_info['max_gripper']
-        self.min_gripper = init_info['min_gripper']
-        servo_config_file = init_info['servo_config_file']
-        self.joint = self.init_joint.copy()
-        self.pose = self.init_pose.copy()
-        self.pose_speeds = [0.0] * 6
-        self.joint_speeds = [0.0] * 6
-        self.tozero = False
-        
-        # 主臂关节状态
-        self.master_joint_actions = {}
-        self.master_controller_status = {}
-        self.use_master_arm = False
-        
-        # 末端位姿限制
-        self.pose_limits = [
-            (-0.800, 0.800),     # X (m)
-            (-0.800, 0.800),     # Y (m)
-            (-0.800, 0.800),     # Z (m)
-            (-3.14, 3.14),       # RX (rad)
-            (-3.14, 3.14),       # RY (rad)
-            (-3.14, 3.14)        # RZ (rad)
-        ]
-        
-        # 关节角度限制 (度)
-        self.joint_limits = [
-            (-180, 180),  # joint 1
-            (-180, 180),  # joint 2
-            (-180, 180),  # joint 3
-            (-180, 180),  # joint 4
-            (-180, 180),  # joint 5
-            (-180, 180)   # joint 6
-        ]
-        
-        # 控制参数
-        self.linear_step = 0.002      # 线性移动步长(m)
-        self.angular_step = 0.01      # 角度步长(rad)
-        
-        # 夹爪状态和速度
-        self.gripper_speed = 10
-        self.gripper = self.min_gripper
-        
-        # 初始化串口通信（主臂关节状态获取）
-        self.servo_arm = None
-        if servo_config_file:
-            try:
-                self.servo_arm = ServoArm(servo_config_file)
-                self.use_master_arm = True
-                logging.info("串口主臂连接成功，启用主从控制模式")
-            except Exception as e:
-                logging.error(f"串口主臂连接失败: {e}")
-                self.use_master_arm = False
-        
-        # 启动更新线程
-        self.running = True
-        self.thread = threading.Thread(target=self.update_controller)
-        self.thread.start()
-        
-        print("混合控制器已启动")
-        print("主控制模式: 关节控制")
-        if self.use_master_arm:
-            print("主从控制: 启用")
-        print("Back按钮: 切换控制模式(关节/末端)")
-        print("L3按钮: 切换精细控制模式")
-        print("Start按钮: 重置到初始位置")
-    
-    def _apply_nonlinear_mapping(self, value):
-        """应用非线性映射以提高控制精度"""
-        sign = 1 if value >= 0 else -1
-        return sign * (abs(value) ** 2)
-    
-    def _normalize_angle(self, angle):
-        """将角度归一化到[-π, π]范围内"""
-        import math
-        while angle > math.pi:
-            angle -= 2 * math.pi
-        while angle < -math.pi:
-            angle += 2 * math.pi
-        return angle
-    
-    def update_controller(self):
-        while self.running:
-            try:
-                pygame.event.pump()
-            except Exception as e:
-                print(f"控制器错误: {e}")
-                self.stop()
-                continue
-            
-            # 检查控制模式切换 (Back按钮)
-            if self.joystick.get_button(6):  # Back按钮
-                self.joint_control_mode = not self.joint_control_mode
-                mode_str = "关节控制" if self.joint_control_mode else "末端位姿控制"
-                print(f"切换到{mode_str}模式")
-                time.sleep(0.3)  # 防止多次触发
-            
-            # 检查精细控制模式切换 (L3按钮)
-            if self.joystick.get_button(10):  # L3按钮
-                self.fine_control_mode = not self.fine_control_mode
-                print(f"切换到{'精细' if self.fine_control_mode else '普通'}控制模式")
-                time.sleep(0.3)  # 防止多次触发
-            
-            # 检查重置按钮 (Start按钮)
-            if self.joystick.get_button(7):  # Start按钮
-                print("重置机械臂到初始位置...")
-                # print("init_joint", self.init_joint.copy())
-                self.tozero = True
-                self.joint = self.init_joint.copy()
-                self.pose = self.init_pose.copy()
-                self.pose_speeds = [0.0] * 6
-                self.joint_speeds = [0.0] * 6
-                self.gripper_speed = 10
-                self.gripper = self.min_gripper
-                print("机械臂已重置到初始位置")
-                time.sleep(0.3)  # 防止多次触发
-
-            # 从串口获取主臂关节状态
-            if self.servo_arm and self.servo_arm.connected:
-                try:
-                    self.master_joint_actions = self.servo_arm.get_joint_actions()
-                    self.master_controller_status = self.servo_arm.get_controller_status()
-                    if self.master_joint_actions:
-                        logging.debug(f"主臂关节状态: {self.master_joint_actions}")
-                    
-                except Exception as e:
-                    logging.error(f"获取主臂状态错误: {e}")
-                    self.master_joint_actions = {}
-            # print(self.master_joint_actions)
-
-            # 根据控制模式更新相应的控制逻辑
-            if self.joint_control_mode:
-                # 关节控制模式下，优先使用主臂数据，Xbox作为辅助
-                self.update_joint_control()
-            else:
-                # 末端控制模式，使用Xbox控制
-                self.update_end_pose()
-            time.sleep(0.02)
-            # print('gripper:', self.gripper)
-    
-    def update_joint_control(self):
-        """更新关节角度控制 - 优先使用主臂数据"""
-        if self.use_master_arm and self.master_joint_actions:
-            # 主从控制模式：直接使用主臂的关节角度
-            try:
-                # 将主臂关节角度映射到从臂
-                for i in range(6):  # 假设只有6个关节需要控制
-                    joint_key = f"joint_{i+1}"
-                    if joint_key in self.master_joint_actions:
-                        # 直接使用主臂的关节角度（已经是度数）
-                        self.joint[i] = self.master_joint_actions[joint_key]
-                        
-                        # 应用关节限制
-                        min_val, max_val = self.joint_limits[i]
-                        self.joint[i] = max(min_val, min(max_val, self.joint[i]))
-                
-                # print(self.joint)
-                logging.debug(f"主臂关节映射到从臂: {self.joint[:6]}")
-                
-            except Exception as e:
-                logging.error(f"主臂数据映射错误: {e}")
-        
-        # 如果有主臂夹爪数据，使用主臂夹爪状态
-        if self.use_master_arm and "grasp" in self.master_joint_actions:
-            self.gripper = self.master_joint_actions["grasp"] * 1000
-            self.joint[-1] = self.gripper
-    
-
-    def update_end_pose(self):
-        """更新末端位姿控制"""
-        # 根据控制模式调整步长
-        current_linear_step = self.linear_step * (0.1 if self.fine_control_mode else 1.0)
-        current_angular_step = self.angular_step * (0.1 if self.fine_control_mode else 1.0)
-        
-        # 方向键控制XY
-        hat = self.joystick.get_hat(0)
-        hat_up = hat[1] == 1     # Y+
-        hat_down = hat[1] == -1  # Y-
-        hat_left = hat[0] == -1  # X-
-        hat_right = hat[0] == 1  # X+
-        
-        # 右摇杆控制Z
-        right_y_raw = -self.joystick.get_axis(4)
-        # 左摇杆控制RZ
-        left_y_raw = -self.joystick.get_axis(1)
-        
-        # 应用死区
-        right_y = 0.0 if abs(right_y_raw) < self.deadzone else right_y_raw
-        left_y = 0.0 if abs(left_y_raw) < self.deadzone else left_y_raw
-
-        # 计算各轴速度
-        self.pose_speeds[1] = current_linear_step if hat_up else (-current_linear_step if hat_down else 0.0)  # Y
-        self.pose_speeds[0] = -current_linear_step if hat_left else (current_linear_step if hat_right else 0.0)  # X
-        
-        # 设置Z速度（右摇杆Y轴控制）
-        z_mapping = self._apply_nonlinear_mapping(right_y)
-        self.pose_speeds[2] = z_mapping * current_linear_step  # Z
-        
-        # L1/R1控制RX旋转
-        LB = self.joystick.get_button(4)  # RX-
-        RB = self.joystick.get_button(5)  # RX+
-        self.pose_speeds[3] = (-current_angular_step if LB else (current_angular_step if RB else 0.0))
-        
-        # △/□控制RY旋转
-        triangle = self.joystick.get_button(2)  # RY+
-        square = self.joystick.get_button(3)    # RY-
-        self.pose_speeds[4] = (current_angular_step if triangle else (-current_angular_step if square else 0.0))
-        
-        # 左摇杆Y轴控制RZ旋转
-        rz_mapping = self._apply_nonlinear_mapping(left_y)
-        self.pose_speeds[5] = rz_mapping * current_angular_step * 2  # RZ
-        
-        # 夹爪控制（圈/叉）
-        circle = self.joystick.get_button(1)  # 夹爪开
-        cross = self.joystick.get_button(0)   # 夹爪关
-        if circle:
-            self.gripper = min(self.max_gripper, self.gripper + self.gripper_speed)
-        elif cross:
-            self.gripper = max(self.min_gripper, self.gripper - self.gripper_speed)
-
-        # 更新末端位姿
-        for i in range(6):
-            self.pose[i] += self.pose_speeds[i]
-        
-        # 角度归一化处理
-        for i in range(3, 6):
-            self.pose[i] = self._normalize_angle(self.pose[i])
-
-    def update_joint_state(self, joint):
-        self.joint = joint
-        # self.tozero = False
-
-    def update_endpose_state(self, end_pose):
-        self.pose = end_pose
-        # self.tozero = False
-    
-    def update_tozero_state(self, tozero):
-        self.tozero = tozero
-
-            
-    def get_action(self) -> Dict:
-        """获取当前控制命令"""
-        return {
-            'control_mode': 'joint' if self.joint_control_mode else 'end_pose',
-            'use_master_arm': self.use_master_arm,
-            'master_joint_actions': self.master_joint_actions,
-            'master_controller_status': self.master_controller_status,
-            'end_pose': self.pose,
-            'joint_angles': self.joint,
-            'gripper': int(self.gripper),
-            'tozero': self.tozero
-        }
-    
-    def stop(self):
-        """停止控制器"""
-        self.running = False
-        if self.thread.is_alive():
-            self.thread.join()
-        if self.servo_arm:
-            self.servo_arm.close()
-        pygame.quit()
-        print("混合控制器已退出")
-    
-    def reset(self):
-        """重置到初始状态"""
-        self.joint = self.init_joint.copy()
-        self.pose = self.init_pose.copy()
-        self.pose_speeds = [0.0] * 6
-        self.joint_speeds = [0.0] * 6
-        self.gripper_speed = 10
-        self.gripper = self.min_gripper
-        print("已重置到初始状态")
-
-
-# 使用示例
-if __name__ == "__main__":
-    # 初始化睿尔曼机械臂
-    arm = RoboticArm(rm_thread_mode_e.RM_TRIPLE_MODE_E)
-    # 创建机械臂连接
-    handle = arm.rm_create_robot_arm("192.168.3.18", 8080)
-    print(f"机械臂连接ID: {handle.id}")
-    init_pose = arm.rm_get_current_arm_state()[1]['pose']
-
-    with open('/home/maic/LYT/lerobot/lerobot/common/robot_devices/teleop/realman_mix.yaml', "r") as file:
-        config = yaml.safe_load(file)
-    config['init_pose'] = init_pose
-    arm_controller = HybridController(config)
-    try:
-        while True:
-            print(arm_controller.get_action())
-            time.sleep(0.1)
-    except KeyboardInterrupt:
-        arm_controller.stop()

lerobot/common/robot_devices/teleop/realman_mix.yaml

@@ -1,4 +0,0 @@
-init_joint: [-90, 90, 90, -90, -90, 90]
-max_gripper: 990
-min_gripper: 10
-servo_config_file: "/home/maic/LYT/lerobot/lerobot/common/robot_devices/teleop/servo_arm.yaml"

lerobot/common/robot_devices/teleop/servo_arm.yaml

@@ -1,6 +0,0 @@
-port: /dev/ttyUSB0
-right_port: /dev/ttyUSB1
-baudrate: 460800
-joint_hex_data: "55 AA 02 00 00 67"
-control_hex_data: "55 AA 08 00 00 B9"
-arm_axis: 6

lerobot/common/robot_devices/utils.py

@@ -48,7 +48,8 @@ class RobotDeviceNotConnectedError(Exception):
     """Exception raised when the robot device is not connected."""
 
     def __init__(
-        self, message="This robot device is not connected. Try calling `robot_device.connect()` first."
+        self,
+        message="This robot device is not connected. Try calling `robot_device.connect()` first.",
     ):
         self.message = message
         super().__init__(self.message)

lerobot/common/utils/io_utils.py

@@ -28,7 +28,9 @@ def write_video(video_path, stacked_frames, fps):
     # Filter out DeprecationWarnings raised from pkg_resources
     with warnings.catch_warnings():
         warnings.filterwarnings(
-            "ignore", "pkg_resources is deprecated as an API", category=DeprecationWarning
+            "ignore",
+            "pkg_resources is deprecated as an API",
+            category=DeprecationWarning,
         )
         imageio.mimsave(video_path, stacked_frames, fps=fps)
 

lerobot/common/utils/logging_utils.py

@@ -94,7 +94,7 @@ class MetricsTracker:
         metrics: dict[str, AverageMeter],
         initial_step: int = 0,
     ):
-        self.__dict__.update(dict.fromkeys(self.__keys__))
+        self.__dict__.update({k: None for k in self.__keys__})
         self._batch_size = batch_size
         self._num_frames = num_frames
         self._avg_samples_per_ep = num_frames / num_episodes