fix caching and dataset stats is optional

for more information, see https://pre-commit.ci

.gitignore

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

.pre-commit-config.yaml

@@ -36,8 +36,8 @@ repos:
       - id: end-of-file-fixer
       - id: trailing-whitespace
 
-  - repo: https://github.com/adhtruong/mirrors-typos
-    rev: v1.31.1
+  - repo: https://github.com/crate-ci/typos
+    rev: v1.30.2
     hooks:
       - id: typos
         args: [--force-exclude]
@@ -46,9 +46,9 @@ repos:
     rev: v3.19.1
     hooks:
     -   id: pyupgrade
-
+        exclude: '^(.*_pb2_grpc\.py|.*_pb2\.py$)'
   - repo: https://github.com/astral-sh/ruff-pre-commit
-    rev: v0.11.5
+    rev: v0.9.10
     hooks:
       - id: ruff
         args: [--fix]
@@ -57,12 +57,12 @@ repos:
 
   ##### Security #####
   - repo: https://github.com/gitleaks/gitleaks
-    rev: v8.24.3
+    rev: v8.24.0
     hooks:
       - id: gitleaks
 
   - repo: https://github.com/woodruffw/zizmor-pre-commit
-    rev: v1.5.2
+    rev: v1.4.1
     hooks:
       - id: zizmor
 

README.md

@@ -98,25 +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)
@@ -125,7 +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
@@ -198,7 +191,6 @@ Under the hood, the `LeRobotDataset` format makes use of several ways to seriali
 Here are the important details and internal structure organization of a typical `LeRobotDataset` instantiated with `dataset = LeRobotDataset("lerobot/aloha_static_coffee")`. The exact features will change from dataset to dataset but not the main aspects:
 
 ```
-TODO: IMPROVE
 dataset attributes:
   ├ hf_dataset: a Hugging Face dataset (backed by Arrow/parquet). Typical features example:
   │  ├ observation.images.cam_high (VideoFrame):
@@ -211,7 +203,7 @@ dataset attributes:
   │  ├ timestamp (float32): timestamp in the 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
-  ├ meta: contains 2 tensors with the start and end indices of each episode
+  ├ 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
   │  └ to: (1D int64 tensor): last frame index for each episode — shape (num episodes,)
   ├ stats: a dictionary of statistics (max, mean, min, std) for each feature in the dataset, for instance

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
@@ -104,12 +112,14 @@ 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:
-    episode_index = 0
-    ep_num_images = dataset.meta.episodes["length"][episode_index]
+    ep_num_images = dataset.episode_data_index["to"][0].item()
     if imgs_dir.exists() and len(list(imgs_dir.glob("frame_*.png"))) == ep_num_images:
         return
 
@@ -121,7 +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)
@@ -266,8 +280,7 @@ def benchmark_encoding_decoding(
             overwrite=True,
         )
 
-    episode_index = 0
-    ep_num_images = dataset.meta.episodes["length"][episode_index]
+    ep_num_images = dataset.episode_data_index["to"][0].item()
     width, height = tuple(dataset[0][dataset.meta.camera_keys[0]].shape[-2:])
     num_pixels = width * height
     video_size_bytes = video_path.stat().st_size
@@ -277,7 +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(

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:

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:.
@@ -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 \

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:.
@@ -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) |
 | ---------- | ------------------ | ---------------------- |

examples/11_use_moss.md

@@ -31,15 +31,9 @@ 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
@@ -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 \

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/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:")
@@ -92,11 +95,11 @@ print(dataset.hf_dataset)
 # LeRobot datasets also subclasses PyTorch datasets so you can do everything you know and love from working
 # with the latter, like iterating through the dataset.
 # The __getitem__ iterates over the frames of the dataset. Since our datasets are also structured by
-# episodes, you can access the frame indices of any episode using dataset.meta.episodes. Here, we access
+# episodes, you can access the frame indices of any episode using the episode_data_index. Here, we access
 # frame indices associated to the first episode:
 episode_index = 0
-from_idx = dataset.meta.episodes["dataset_from_index"][episode_index]
-to_idx = dataset.meta.episodes["dataset_to_index"][episode_index]
+from_idx = dataset.episode_data_index["from"][episode_index].item()
+to_idx = dataset.episode_data_index["to"][episode_index].item()
 
 # Then we grab all the image frames from the first camera:
 camera_key = dataset.meta.camera_keys[0]
@@ -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

@@ -18,7 +18,7 @@ training outputs directory. In the latter case, you might want to run examples/3
 
 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]"`
 ```
 """
 

examples/3_train_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

@@ -4,7 +4,7 @@ This tutorial will explain the training script, how to use it, and particularly
 
 ## 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.
@@ -21,7 +21,7 @@ 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 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.)
 
@@ -50,7 +50,7 @@ By default, every field takes its default value specified in the dataclass. If a
 
 ## 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 \

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 \
@@ -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
 ```
@@ -102,8 +97,6 @@ This is equivalent to running `stretch_robot_home.py`
 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 \

examples/advanced/1_add_image_transforms.py

@@ -31,7 +31,7 @@ dataset = LeRobotDataset(dataset_repo_id, episodes=[0])
 # This is equivalent to `dataset = LeRobotDataset(dataset_repo_id, image_transforms=None)`
 
 # Get the index of the first observation in the first episode
-first_idx = dataset.meta.episodes["dataset_from_index"][0]
+first_idx = dataset.episode_data_index["from"][0].item()
 
 # Get the frame corresponding to the first camera
 frame = dataset[first_idx][dataset.meta.camera_keys[0]]

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.

examples/port_datasets/droid_rlds/README.md

@@ -1,144 +0,0 @@
-# Port DROID 1.0.1 dataset to LeRobotDataset
-
-## Download
-
-TODO
-
-It will take 2 TB in your local disk.
-
-## Port on a single computer
-
-First, install tensorflow dataset utilities to read from raw files:
-```bash
-pip install tensorflow
-pip install tensorflow_datasets
-```
-
-Then run this script to start porting the dataset:
-```bash
-python examples/port_datasets/droid_rlds/port_droid.py \
-    --raw-dir /your/data/droid/1.0.1 \
-    --repo-id your_id/droid_1.0.1 \
-    --push-to-hub
-```
-
-It will take 400GB in your local disk.
-
-As usual, your LeRobotDataset will be stored in your huggingface/lerobot cache folder.
-
-WARNING: it will take 7 days for porting the dataset locally and 3 days to upload, so we will need to parallelize over multiple nodes on a slurm cluster.
-
-NOTE: For development, run this script to start porting a shard:
-```bash
-python examples/port_datasets/droid_rlds/port.py \
-    --raw-dir /your/data/droid/1.0.1 \
-    --repo-id your_id/droid_1.0.1 \
-    --num-shards 2048 \
-    --shard-index 0
-```
-
-## Port over SLURM
-
-Install slurm utilities from Hugging Face:
-```bash
-pip install datatrove
-```
-
-
-### 1. Port one shard per job
-
-Run this script to start porting shards of the dataset:
-```bash
-python examples/port_datasets/droid_rlds/slurm_port_shards.py \
-    --raw-dir /your/data/droid/1.0.1 \
-    --repo-id your_id/droid_1.0.1 \
-    --logs-dir /your/logs \
-    --job-name port_droid \
-    --partition your_partition \
-    --workers 2048 \
-    --cpus-per-task 8 \
-    --mem-per-cpu 1950M
-```
-
-**Note on how to set your command line arguments**
-
-Regarding `--partition`, find yours by running:
-```bash
-info --format="%R"`
-```
-and select the CPU partition if you have one. No GPU needed.
-
-Regarding `--workers`, it is the number of slurm jobs you will launch in parallel. 2048 is the maximum number, since there is 2048 shards in Droid. This big number will certainly max-out your cluster.
-
-Regarding `--cpus-per-task` and `--mem-per-cpu`, by default it will use ~16GB of RAM (8*1950M) which is recommended to load the raw frames and 8 CPUs which can be useful to parallelize the encoding of the frames.
-
-Find the number of CPUs and Memory of the nodes of your partition by running:
-```bash
-sinfo -N -p your_partition -h -o "%N cpus=%c mem=%m"
-```
-
-**Useful commands to check progress and debug**
-
-Check if your jobs are running:
-```bash
-squeue -u $USER`
-```
-
-You should see a list with job indices like `15125385_155` where `15125385` is the index of the run and `155` is the worker index. The output/print of this worker is written in real time in `/your/logs/job_name/slurm_jobs/15125385_155.out`. For instance, you can inspect the content of this file by running `less /your/logs/job_name/slurm_jobs/15125385_155.out`.
-
-Check the progression of your jobs by running:
-```bash
-jobs_status /your/logs
-```
-
-If it's not 100% and no more slurm job is running, it means that some of them failed. Inspect the logs by running:
-```bash
-failed_logs /your/logs/job_name
-```
-
-If there is an issue in the code, you can fix it in debug mode with `--slurm 0` which allows to set breakpoint:
-```bash
-python examples/port_datasets/droid_rlds/slurm_port_shards.py --slurm 0 ...
-```
-
-And you can relaunch the same command, which will skip the completed jobs:
-```bash
-python examples/port_datasets/droid_rlds/slurm_port_shards.py --slurm 1 ...
-```
-
-Once all jobs are completed, you will have one dataset per shard (e.g. `droid_1.0.1_world_2048_rank_1594`) saved on disk in your `/lerobot/home/dir/your_id` directory. You can find your `/lerobot/home/dir` by running:
-```bash
-python -c "from lerobot.common.constants import HF_LEROBOT_HOME;print(HF_LEROBOT_HOME)"
-```
-
-
-### 2. Aggregate all shards
-
-Run this script to start aggregation:
-```bash
-python examples/port_datasets/droid_rlds/slurm_aggregate_shards.py \
-    --repo-id your_id/droid_1.0.1 \
-    --logs-dir /your/logs \
-    --job-name aggr_droid \
-    --partition your_partition \
-    --workers 2048 \
-    --cpus-per-task 8 \
-    --mem-per-cpu 1950M
-```
-
-Once all jobs are completed, you will have one dataset your `/lerobot/home/dir/your_id/droid_1.0.1` directory.
-
-
-### 3. Upload dataset
-
-Run this script to start uploading:
-```bash
-python examples/port_datasets/droid_rlds/slurm_upload.py \
-    --repo-id your_id/droid_1.0.1 \
-    --logs-dir /your/logs \
-    --job-name upload_droid \
-    --partition your_partition \
-    --workers 50 \
-    --cpus-per-task 4 \
-    --mem-per-cpu 1950M
-```

examples/port_datasets/droid_rlds/port_droid.py

@@ -1,430 +0,0 @@
-#!/usr/bin/env python
-
-# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
-#
-# Licensed under the Apache License, Version 2.0 (the "License");
-# you may not use this file except in compliance with the License.
-# You may obtain a copy of the License at
-#
-#     http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing, software
-# distributed under the License is distributed on an "AS IS" BASIS,
-# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-# See the License for the specific language governing permissions and
-# limitations under the License.
-
-import argparse
-import logging
-import time
-from pathlib import Path
-
-import numpy as np
-import tensorflow_datasets as tfds
-
-from lerobot.common.datasets.lerobot_dataset import LeRobotDataset, LeRobotDatasetMetadata
-from lerobot.common.utils.utils import get_elapsed_time_in_days_hours_minutes_seconds
-
-DROID_SHARDS = 2048
-DROID_FPS = 15
-DROID_ROBOT_TYPE = "Franka"
-
-# Dataset schema slightly adapted from: https://droid-dataset.github.io/droid/the-droid-dataset.html#-dataset-schema
-DROID_FEATURES = {
-    # true on first step of the episode
-    "is_first": {
-        "dtype": "bool",
-        "shape": (1,),
-        "names": None,
-    },
-    # true on last step of the episode
-    "is_last": {
-        "dtype": "bool",
-        "shape": (1,),
-        "names": None,
-    },
-    # true on last step of the episode if it is a terminal step, True for demos
-    "is_terminal": {
-        "dtype": "bool",
-        "shape": (1,),
-        "names": None,
-    },
-    # language_instruction is also stored as "task" to follow LeRobot standard
-    "language_instruction": {
-        "dtype": "string",
-        "shape": (1,),
-        "names": None,
-    },
-    "language_instruction_2": {
-        "dtype": "string",
-        "shape": (1,),
-        "names": None,
-    },
-    "language_instruction_3": {
-        "dtype": "string",
-        "shape": (1,),
-        "names": None,
-    },
-    "observation.state.gripper_position": {
-        "dtype": "float32",
-        "shape": (1,),
-        "names": {
-            "axes": ["gripper"],
-        },
-    },
-    "observation.state.cartesian_position": {
-        "dtype": "float32",
-        "shape": (6,),
-        "names": {
-            "axes": ["x", "y", "z", "roll", "pitch", "yaw"],
-        },
-    },
-    "observation.state.joint_position": {
-        "dtype": "float32",
-        "shape": (7,),
-        "names": {
-            "axes": ["joint_0", "joint_1", "joint_2", "joint_3", "joint_4", "joint_5", "joint_6"],
-        },
-    },
-    # Add this new feature to follow LeRobot standard of using joint position + gripper
-    "observation.state": {
-        "dtype": "float32",
-        "shape": (8,),
-        "names": {
-            "axes": ["joint_0", "joint_1", "joint_2", "joint_3", "joint_4", "joint_5", "joint_6", "gripper"],
-        },
-    },
-    # Initially called wrist_image_left
-    "observation.images.wrist_left": {
-        "dtype": "video",
-        "shape": (180, 320, 3),
-        "names": [
-            "height",
-            "width",
-            "channels",
-        ],
-    },
-    # Initially called exterior_image_1_left
-    "observation.images.exterior_1_left": {
-        "dtype": "video",
-        "shape": (180, 320, 3),
-        "names": [
-            "height",
-            "width",
-            "channels",
-        ],
-    },
-    # Initially called exterior_image_2_left
-    "observation.images.exterior_2_left": {
-        "dtype": "video",
-        "shape": (180, 320, 3),
-        "names": [
-            "height",
-            "width",
-            "channels",
-        ],
-    },
-    "action.gripper_position": {
-        "dtype": "float32",
-        "shape": (1,),
-        "names": {
-            "axes": ["gripper"],
-        },
-    },
-    "action.gripper_velocity": {
-        "dtype": "float32",
-        "shape": (1,),
-        "names": {
-            "axes": ["gripper"],
-        },
-    },
-    "action.cartesian_position": {
-        "dtype": "float32",
-        "shape": (6,),
-        "names": {
-            "axes": ["x", "y", "z", "roll", "pitch", "yaw"],
-        },
-    },
-    "action.cartesian_velocity": {
-        "dtype": "float32",
-        "shape": (6,),
-        "names": {
-            "axes": ["x", "y", "z", "roll", "pitch", "yaw"],
-        },
-    },
-    "action.joint_position": {
-        "dtype": "float32",
-        "shape": (7,),
-        "names": {
-            "axes": ["joint_0", "joint_1", "joint_2", "joint_3", "joint_4", "joint_5", "joint_6"],
-        },
-    },
-    "action.joint_velocity": {
-        "dtype": "float32",
-        "shape": (7,),
-        "names": {
-            "axes": ["joint_0", "joint_1", "joint_2", "joint_3", "joint_4", "joint_5", "joint_6"],
-        },
-    },
-    # This feature was called "action" in RLDS dataset and consists of [6x joint velocities, 1x gripper position]
-    "action.original": {
-        "dtype": "float32",
-        "shape": (7,),
-        "names": {
-            "axes": ["x", "y", "z", "roll", "pitch", "yaw", "gripper"],
-        },
-    },
-    # Add this new feature to follow LeRobot standard of using joint position + gripper
-    "action": {
-        "dtype": "float32",
-        "shape": (8,),
-        "names": {
-            "axes": ["joint_0", "joint_1", "joint_2", "joint_3", "joint_4", "joint_5", "joint_6", "gripper"],
-        },
-    },
-    "discount": {
-        "dtype": "float32",
-        "shape": (1,),
-        "names": None,
-    },
-    "reward": {
-        "dtype": "float32",
-        "shape": (1,),
-        "names": None,
-    },
-    # Meta data that are the same for all frames in the episode
-    "task_category": {
-        "dtype": "string",
-        "shape": (1,),
-        "names": None,
-    },
-    "building": {
-        "dtype": "string",
-        "shape": (1,),
-        "names": None,
-    },
-    "collector_id": {
-        "dtype": "string",
-        "shape": (1,),
-        "names": None,
-    },
-    "date": {
-        "dtype": "string",
-        "shape": (1,),
-        "names": None,
-    },
-    "camera_extrinsics.wrist_left": {
-        "dtype": "float32",
-        "shape": (6,),
-        "names": {
-            "axes": ["x", "y", "z", "roll", "pitch", "yaw"],
-        },
-    },
-    "camera_extrinsics.exterior_1_left": {
-        "dtype": "float32",
-        "shape": (6,),
-        "names": {
-            "axes": ["x", "y", "z", "roll", "pitch", "yaw"],
-        },
-    },
-    "camera_extrinsics.exterior_2_left": {
-        "dtype": "float32",
-        "shape": (6,),
-        "names": {
-            "axes": ["x", "y", "z", "roll", "pitch", "yaw"],
-        },
-    },
-    "is_episode_successful": {
-        "dtype": "bool",
-        "shape": (1,),
-        "names": None,
-    },
-}
-
-
-def is_episode_successful(tf_episode_metadata):
-    # Adapted from: https://github.com/droid-dataset/droid_policy_learning/blob/dd1020eb20d981f90b5ff07dc80d80d5c0cb108b/robomimic/utils/rlds_utils.py#L8
-    return "/success/" in tf_episode_metadata["file_path"].numpy().decode()
-
-
-def generate_lerobot_frames(tf_episode):
-    m = tf_episode["episode_metadata"]
-    frame_meta = {
-        "task_category": m["building"].numpy().decode(),
-        "building": m["building"].numpy().decode(),
-        "collector_id": m["collector_id"].numpy().decode(),
-        "date": m["date"].numpy().decode(),
-        "camera_extrinsics.wrist_left": m["extrinsics_wrist_cam"].numpy(),
-        "camera_extrinsics.exterior_1_left": m["extrinsics_exterior_cam_1"].numpy(),
-        "camera_extrinsics.exterior_2_left": m["extrinsics_exterior_cam_2"].numpy(),
-        "is_episode_successful": np.array([is_episode_successful(m)]),
-    }
-    for f in tf_episode["steps"]:
-        # Dataset schema slightly adapted from: https://droid-dataset.github.io/droid/the-droid-dataset.html#-dataset-schema
-        frame = {
-            "is_first": np.array([f["is_first"].numpy()]),
-            "is_last": np.array([f["is_last"].numpy()]),
-            "is_terminal": np.array([f["is_terminal"].numpy()]),
-            "language_instruction": f["language_instruction"].numpy().decode(),
-            "language_instruction_2": f["language_instruction_2"].numpy().decode(),
-            "language_instruction_3": f["language_instruction_3"].numpy().decode(),
-            "observation.state.gripper_position": f["observation"]["gripper_position"].numpy(),
-            "observation.state.cartesian_position": f["observation"]["cartesian_position"].numpy(),
-            "observation.state.joint_position": f["observation"]["joint_position"].numpy(),
-            "observation.images.wrist_left": f["observation"]["wrist_image_left"].numpy(),
-            "observation.images.exterior_1_left": f["observation"]["exterior_image_1_left"].numpy(),
-            "observation.images.exterior_2_left": f["observation"]["exterior_image_2_left"].numpy(),
-            "action.gripper_position": f["action_dict"]["gripper_position"].numpy(),
-            "action.gripper_velocity": f["action_dict"]["gripper_velocity"].numpy(),
-            "action.cartesian_position": f["action_dict"]["cartesian_position"].numpy(),
-            "action.cartesian_velocity": f["action_dict"]["cartesian_velocity"].numpy(),
-            "action.joint_position": f["action_dict"]["joint_position"].numpy(),
-            "action.joint_velocity": f["action_dict"]["joint_velocity"].numpy(),
-            "discount": np.array([f["discount"].numpy()]),
-            "reward": np.array([f["reward"].numpy()]),
-            "action.original": f["action"].numpy(),
-        }
-
-        # language_instruction is also stored as "task" to follow LeRobot standard
-        frame["task"] = frame["language_instruction"]
-
-        # Add this new feature to follow LeRobot standard of using joint position + gripper
-        frame["observation.state"] = np.concatenate(
-            [frame["observation.state.joint_position"], frame["observation.state.gripper_position"]]
-        )
-        frame["action"] = np.concatenate([frame["action.joint_position"], frame["action.gripper_position"]])
-
-        # Meta data that are the same for all frames in the episode
-        frame.update(frame_meta)
-
-        # Cast fp64 to fp32
-        for key in frame:
-            if isinstance(frame[key], np.ndarray) and frame[key].dtype == np.float64:
-                frame[key] = frame[key].astype(np.float32)
-
-        yield frame
-
-
-def port_droid(
-    raw_dir: Path,
-    repo_id: str,
-    push_to_hub: bool = False,
-    num_shards: int | None = None,
-    shard_index: int | None = None,
-):
-    dataset_name = raw_dir.parent.name
-    version = raw_dir.name
-    data_dir = raw_dir.parent.parent
-
-    builder = tfds.builder(f"{dataset_name}/{version}", data_dir=data_dir, version="")
-
-    if num_shards is not None:
-        tfds_num_shards = builder.info.splits["train"].num_shards
-        if tfds_num_shards != DROID_SHARDS:
-            raise ValueError(
-                f"Number of shards of Droid dataset is expected to be {DROID_SHARDS} but is {tfds_num_shards}."
-            )
-        if num_shards != tfds_num_shards:
-            raise ValueError(
-                f"We only shard over the fixed number of shards provided by tensorflow dataset ({tfds_num_shards}), but {num_shards} shards provided instead."
-            )
-        if shard_index >= tfds_num_shards:
-            raise ValueError(
-                f"Shard index is greater than the num of shards ({shard_index} >= {num_shards})."
-            )
-
-        raw_dataset = builder.as_dataset(split=f"train[{shard_index}shard]")
-    else:
-        raw_dataset = builder.as_dataset(split="train")
-
-    lerobot_dataset = LeRobotDataset.create(
-        repo_id=repo_id,
-        robot_type=DROID_ROBOT_TYPE,
-        fps=DROID_FPS,
-        features=DROID_FEATURES,
-    )
-
-    start_time = time.time()
-    num_episodes = raw_dataset.cardinality().numpy().item()
-    logging.info(f"Number of episodes {num_episodes}")
-
-    for episode_index, episode in enumerate(raw_dataset):
-        elapsed_time = time.time() - start_time
-        d, h, m, s = get_elapsed_time_in_days_hours_minutes_seconds(elapsed_time)
-
-        logging.info(
-            f"{episode_index} / {num_episodes} episodes processed (after {d} days, {h} hours, {m} minutes, {s:.3f} seconds)"
-        )
-
-        for frame in generate_lerobot_frames(episode):
-            lerobot_dataset.add_frame(frame)
-
-        lerobot_dataset.save_episode()
-        logging.info("Save_episode")
-
-    if push_to_hub:
-        lerobot_dataset.push_to_hub(
-            # Add openx tag, since it belongs to the openx collection of datasets
-            tags=["openx"],
-            private=False,
-        )
-
-
-def validate_dataset(repo_id):
-    """Sanity check that ensure meta data can be loaded and all files are present."""
-    meta = LeRobotDatasetMetadata(repo_id)
-
-    if meta.total_episodes == 0:
-        raise ValueError("Number of episodes is 0.")
-
-    for ep_idx in range(meta.total_episodes):
-        data_path = meta.root / meta.get_data_file_path(ep_idx)
-
-        if not data_path.exists():
-            raise ValueError(f"Parquet file is missing in: {data_path}")
-
-        for vid_key in meta.video_keys:
-            vid_path = meta.root / meta.get_video_file_path(ep_idx, vid_key)
-            if not vid_path.exists():
-                raise ValueError(f"Video file is missing in: {vid_path}")
-
-
-def main():
-    parser = argparse.ArgumentParser()
-
-    parser.add_argument(
-        "--raw-dir",
-        type=Path,
-        required=True,
-        help="Directory containing input raw datasets (e.g. `path/to/dataset` or `path/to/dataset/version).",
-    )
-    parser.add_argument(
-        "--repo-id",
-        type=str,
-        help="Repositery identifier on Hugging Face: a community or a user name `/` the name of the dataset, required when push-to-hub is True",
-    )
-    parser.add_argument(
-        "--push-to-hub",
-        action="store_true",
-        help="Upload to hub.",
-    )
-    parser.add_argument(
-        "--num-shards",
-        type=int,
-        default=None,
-        help="Number of shards. Can be either None to load the full dataset, or 2048 to load one of the 2048 tensorflow dataset files.",
-    )
-    parser.add_argument(
-        "--shard-index",
-        type=int,
-        default=None,
-        help="Index of the shard. Can be either None to load the full dataset, or in [0,2047] to load one of the 2048 tensorflow dataset files.",
-    )
-
-    args = parser.parse_args()
-
-    port_droid(**vars(args))
-
-
-if __name__ == "__main__":
-    main()

examples/port_datasets/droid_rlds/slurm_aggregate_shards.py

@@ -1,293 +0,0 @@
-#!/usr/bin/env python
-
-# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
-#
-# Licensed under the Apache License, Version 2.0 (the "License");
-# you may not use this file except in compliance with the License.
-# You may obtain a copy of the License at
-#
-#     http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing, software
-# distributed under the License is distributed on an "AS IS" BASIS,
-# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-# See the License for the specific language governing permissions and
-# limitations under the License.
-
-import argparse
-import logging
-from pathlib import Path
-
-import tqdm
-from datatrove.executor import LocalPipelineExecutor
-from datatrove.executor.slurm import SlurmPipelineExecutor
-from datatrove.pipeline.base import PipelineStep
-
-from examples.port_datasets.droid_rlds.port_droid import DROID_SHARDS
-from lerobot.common.datasets.aggregate import validate_all_metadata
-from lerobot.common.datasets.lerobot_dataset import LeRobotDatasetMetadata
-from lerobot.common.datasets.utils import (
-    legacy_write_episode_stats,
-    legacy_write_task,
-    write_episode,
-    write_info,
-)
-from lerobot.common.utils.utils import init_logging
-
-
-class AggregateDatasets(PipelineStep):
-    def __init__(
-        self,
-        repo_ids: list[str],
-        aggregated_repo_id: str,
-    ):
-        super().__init__()
-        self.repo_ids = repo_ids
-        self.aggr_repo_id = aggregated_repo_id
-
-        self.create_aggr_dataset()
-
-    def create_aggr_dataset(self):
-        init_logging()
-
-        logging.info("Start aggregate_datasets")
-
-        all_metadata = [LeRobotDatasetMetadata(repo_id) for repo_id in self.repo_ids]
-
-        fps, robot_type, features = validate_all_metadata(all_metadata)
-
-        # Create resulting dataset folder
-        aggr_meta = LeRobotDatasetMetadata.create(
-            repo_id=self.aggr_repo_id,
-            fps=fps,
-            robot_type=robot_type,
-            features=features,
-        )
-
-        logging.info("Find all tasks")
-        # find all tasks, deduplicate them, create new task indices for each dataset
-        # indexed by dataset index
-        datasets_task_index_to_aggr_task_index = {}
-        aggr_task_index = 0
-        for dataset_index, meta in enumerate(tqdm.tqdm(all_metadata, desc="Find all tasks")):
-            task_index_to_aggr_task_index = {}
-
-            for task_index, task in meta.tasks.items():
-                if task not in aggr_meta.task_to_task_index:
-                    # add the task to aggr tasks mappings
-                    aggr_meta.tasks[aggr_task_index] = task
-                    aggr_meta.task_to_task_index[task] = aggr_task_index
-                    aggr_task_index += 1
-
-                # add task_index anyway
-                task_index_to_aggr_task_index[task_index] = aggr_meta.task_to_task_index[task]
-
-            datasets_task_index_to_aggr_task_index[dataset_index] = task_index_to_aggr_task_index
-
-        logging.info("Prepare copy data and videos")
-        datasets_ep_idx_to_aggr_ep_idx = {}
-        datasets_aggr_episode_index_shift = {}
-        aggr_episode_index_shift = 0
-        for dataset_index, meta in enumerate(tqdm.tqdm(all_metadata, desc="Prepare copy data and videos")):
-            ep_idx_to_aggr_ep_idx = {}
-
-            for episode_index in range(meta.total_episodes):
-                aggr_episode_index = episode_index + aggr_episode_index_shift
-                ep_idx_to_aggr_ep_idx[episode_index] = aggr_episode_index
-
-            datasets_ep_idx_to_aggr_ep_idx[dataset_index] = ep_idx_to_aggr_ep_idx
-            datasets_aggr_episode_index_shift[dataset_index] = aggr_episode_index_shift
-
-            # populate episodes
-            for episode_index, episode_dict in meta.episodes.items():
-                aggr_episode_index = episode_index + aggr_episode_index_shift
-                episode_dict["episode_index"] = aggr_episode_index
-                aggr_meta.episodes[aggr_episode_index] = episode_dict
-
-            # populate episodes_stats
-            for episode_index, episode_stats in meta.episodes_stats.items():
-                aggr_episode_index = episode_index + aggr_episode_index_shift
-                aggr_meta.episodes_stats[aggr_episode_index] = episode_stats
-
-            # populate info
-            aggr_meta.info["total_episodes"] += meta.total_episodes
-            aggr_meta.info["total_frames"] += meta.total_frames
-            aggr_meta.info["total_videos"] += len(aggr_meta.video_keys) * meta.total_episodes
-
-            aggr_episode_index_shift += meta.total_episodes
-
-        logging.info("Write meta data")
-        aggr_meta.info["total_tasks"] = len(aggr_meta.tasks)
-        aggr_meta.info["total_chunks"] = aggr_meta.get_episode_chunk(aggr_episode_index_shift - 1)
-        aggr_meta.info["splits"] = {"train": f"0:{aggr_meta.info['total_episodes']}"}
-
-        # create a new episodes jsonl with updated episode_index using write_episode
-        for episode_dict in tqdm.tqdm(aggr_meta.episodes.values(), desc="Write episodes"):
-            write_episode(episode_dict, aggr_meta.root)
-
-        # create a new episode_stats jsonl with updated episode_index using write_episode_stats
-        for episode_index, episode_stats in tqdm.tqdm(
-            aggr_meta.episodes_stats.items(), desc="Write episodes stats"
-        ):
-            legacy_write_episode_stats(episode_index, episode_stats, aggr_meta.root)
-
-        # create a new task jsonl with updated episode_index using write_task
-        for task_index, task in tqdm.tqdm(aggr_meta.tasks.items(), desc="Write tasks"):
-            legacy_write_task(task_index, task, aggr_meta.root)
-
-        write_info(aggr_meta.info, aggr_meta.root)
-
-        self.datasets_task_index_to_aggr_task_index = datasets_task_index_to_aggr_task_index
-        self.datasets_ep_idx_to_aggr_ep_idx = datasets_ep_idx_to_aggr_ep_idx
-        self.datasets_aggr_episode_index_shift = datasets_aggr_episode_index_shift
-
-        logging.info("Meta data done writing!")
-
-    def run(self, data=None, rank: int = 0, world_size: int = 1):
-        import logging
-        import shutil
-
-        import pandas as pd
-
-        from lerobot.common.datasets.aggregate import get_update_episode_and_task_func
-        from lerobot.common.datasets.lerobot_dataset import LeRobotDatasetMetadata
-        from lerobot.common.utils.utils import init_logging
-
-        init_logging()
-
-        aggr_meta = LeRobotDatasetMetadata(self.aggr_repo_id)
-        all_metadata = [LeRobotDatasetMetadata(repo_id) for repo_id in self.repo_ids]
-
-        if world_size != len(all_metadata):
-            raise ValueError()
-
-        dataset_index = rank
-        meta = all_metadata[dataset_index]
-        aggr_episode_index_shift = self.datasets_aggr_episode_index_shift[dataset_index]
-
-        logging.info("Copy data")
-        for episode_index in range(meta.total_episodes):
-            aggr_episode_index = self.datasets_ep_idx_to_aggr_ep_idx[dataset_index][episode_index]
-            data_path = meta.root / meta.get_data_file_path(episode_index)
-            aggr_data_path = aggr_meta.root / aggr_meta.get_data_file_path(aggr_episode_index)
-
-            # update episode_index and task_index
-            df = pd.read_parquet(data_path)
-            update_row_func = get_update_episode_and_task_func(
-                aggr_episode_index_shift, self.datasets_task_index_to_aggr_task_index[dataset_index]
-            )
-            df = df.apply(update_row_func, axis=1)
-
-            aggr_data_path.parent.mkdir(parents=True, exist_ok=True)
-            df.to_parquet(aggr_data_path)
-
-        logging.info("Copy videos")
-        for episode_index in range(meta.total_episodes):
-            aggr_episode_index = episode_index + aggr_episode_index_shift
-            for vid_key in meta.video_keys:
-                video_path = meta.root / meta.get_video_file_path(episode_index, vid_key)
-                aggr_video_path = aggr_meta.root / aggr_meta.get_video_file_path(aggr_episode_index, vid_key)
-                aggr_video_path.parent.mkdir(parents=True, exist_ok=True)
-                shutil.copy(video_path, aggr_video_path)
-
-                # copy_command = f"cp {video_path} {aggr_video_path} &"
-                # subprocess.Popen(copy_command, shell=True)
-
-        logging.info("Done!")
-
-
-def make_aggregate_executor(
-    repo_ids, repo_id, job_name, logs_dir, workers, partition, cpus_per_task, mem_per_cpu, slurm=True
-):
-    kwargs = {
-        "pipeline": [
-            AggregateDatasets(repo_ids, repo_id),
-        ],
-        "logging_dir": str(logs_dir / job_name),
-    }
-
-    if slurm:
-        kwargs.update(
-            {
-                "job_name": job_name,
-                "tasks": DROID_SHARDS,
-                "workers": workers,
-                "time": "08:00:00",
-                "partition": partition,
-                "cpus_per_task": cpus_per_task,
-                "sbatch_args": {"mem-per-cpu": mem_per_cpu},
-            }
-        )
-        executor = SlurmPipelineExecutor(**kwargs)
-    else:
-        kwargs.update(
-            {
-                "tasks": DROID_SHARDS,
-                "workers": 1,
-            }
-        )
-        executor = LocalPipelineExecutor(**kwargs)
-
-    return executor
-
-
-def main():
-    parser = argparse.ArgumentParser()
-
-    parser.add_argument(
-        "--repo-id",
-        type=str,
-        help="Repositery identifier on Hugging Face: a community or a user name `/` the name of the dataset, required when push-to-hub is True.",
-    )
-    parser.add_argument(
-        "--logs-dir",
-        type=Path,
-        help="Path to logs directory for `datatrove`.",
-    )
-    parser.add_argument(
-        "--job-name",
-        type=str,
-        default="aggr_droid",
-        help="Job name used in slurm, and name of the directory created inside the provided logs directory.",
-    )
-    parser.add_argument(
-        "--slurm",
-        type=int,
-        default=1,
-        help="Launch over slurm. Use `--slurm 0` to launch sequentially (useful to debug).",
-    )
-    parser.add_argument(
-        "--workers",
-        type=int,
-        default=2048,
-        help="Number of slurm workers. It should be less than the maximum number of shards.",
-    )
-    parser.add_argument(
-        "--partition",
-        type=str,
-        help="Slurm partition. Ideally a CPU partition. No need for GPU partition.",
-    )
-    parser.add_argument(
-        "--cpus-per-task",
-        type=int,
-        default=8,
-        help="Number of cpus that each slurm worker will use.",
-    )
-    parser.add_argument(
-        "--mem-per-cpu",
-        type=str,
-        default="1950M",
-        help="Memory per cpu that each worker will use.",
-    )
-
-    args = parser.parse_args()
-    kwargs = vars(args)
-    kwargs["slurm"] = kwargs.pop("slurm") == 1
-
-    repo_ids = [f"{args.repo_id}_world_{DROID_SHARDS}_rank_{rank}" for rank in range(DROID_SHARDS)]
-    aggregate_executor = make_aggregate_executor(repo_ids, **kwargs)
-    aggregate_executor.run()
-
-
-if __name__ == "__main__":
-    main()

examples/port_datasets/droid_rlds/slurm_port_shards.py

@@ -1,147 +0,0 @@
-import argparse
-from pathlib import Path
-
-from datatrove.executor import LocalPipelineExecutor
-from datatrove.executor.slurm import SlurmPipelineExecutor
-from datatrove.pipeline.base import PipelineStep
-
-from examples.port_datasets.droid_rlds.port_droid import DROID_SHARDS
-
-
-class PortDroidShards(PipelineStep):
-    def __init__(
-        self,
-        raw_dir: Path | str,
-        repo_id: str = None,
-    ):
-        super().__init__()
-        self.raw_dir = Path(raw_dir)
-        self.repo_id = repo_id
-
-    def run(self, data=None, rank: int = 0, world_size: int = 1):
-        from datasets.utils.tqdm import disable_progress_bars
-
-        from examples.port_datasets.droid_rlds.port_droid import port_droid, validate_dataset
-        from lerobot.common.utils.utils import init_logging
-
-        init_logging()
-        disable_progress_bars()
-
-        shard_repo_id = f"{self.repo_id}_world_{world_size}_rank_{rank}"
-
-        try:
-            validate_dataset(shard_repo_id)
-            return
-        except:
-            pass
-
-        port_droid(
-            self.raw_dir,
-            shard_repo_id,
-            push_to_hub=False,
-            num_shards=world_size,
-            shard_index=rank,
-        )
-
-        validate_dataset(shard_repo_id)
-
-
-def make_port_executor(
-    raw_dir, repo_id, job_name, logs_dir, workers, partition, cpus_per_task, mem_per_cpu, slurm=True
-):
-    kwargs = {
-        "pipeline": [
-            PortDroidShards(raw_dir, repo_id),
-        ],
-        "logging_dir": str(logs_dir / job_name),
-    }
-
-    if slurm:
-        kwargs.update(
-            {
-                "job_name": job_name,
-                "tasks": DROID_SHARDS,
-                "workers": workers,
-                "time": "08:00:00",
-                "partition": partition,
-                "cpus_per_task": cpus_per_task,
-                "sbatch_args": {"mem-per-cpu": mem_per_cpu},
-            }
-        )
-        executor = SlurmPipelineExecutor(**kwargs)
-    else:
-        kwargs.update(
-            {
-                "tasks": 1,
-                "workers": 1,
-            }
-        )
-        executor = LocalPipelineExecutor(**kwargs)
-
-    return executor
-
-
-def main():
-    parser = argparse.ArgumentParser()
-
-    parser.add_argument(
-        "--raw-dir",
-        type=Path,
-        required=True,
-        help="Directory containing input raw datasets (e.g. `path/to/dataset` or `path/to/dataset/version).",
-    )
-    parser.add_argument(
-        "--repo-id",
-        type=str,
-        help="Repositery identifier on Hugging Face: a community or a user name `/` the name of the dataset, required when push-to-hub is True.",
-    )
-    parser.add_argument(
-        "--logs-dir",
-        type=Path,
-        help="Path to logs directory for `datatrove`.",
-    )
-    parser.add_argument(
-        "--job-name",
-        type=str,
-        default="port_droid",
-        help="Job name used in slurm, and name of the directory created inside the provided logs directory.",
-    )
-    parser.add_argument(
-        "--slurm",
-        type=int,
-        default=1,
-        help="Launch over slurm. Use `--slurm 0` to launch sequentially (useful to debug).",
-    )
-    parser.add_argument(
-        "--workers",
-        type=int,
-        default=2048,
-        help="Number of slurm workers. It should be less than the maximum number of shards.",
-    )
-    parser.add_argument(
-        "--partition",
-        type=str,
-        help="Slurm partition. Ideally a CPU partition. No need for GPU partition.",
-    )
-    parser.add_argument(
-        "--cpus-per-task",
-        type=int,
-        default=8,
-        help="Number of cpus that each slurm worker will use.",
-    )
-    parser.add_argument(
-        "--mem-per-cpu",
-        type=str,
-        default="1950M",
-        help="Memory per cpu that each worker will use.",
-    )
-
-    args = parser.parse_args()
-    kwargs = vars(args)
-    kwargs["slurm"] = kwargs.pop("slurm") == 1
-    port_executor = make_port_executor(**kwargs)
-    port_executor.run()
-
-
-if __name__ == "__main__":
-    main()

examples/port_datasets/droid_rlds/slurm_upload.py

@@ -1,263 +0,0 @@
-import argparse
-import logging
-import os
-from pathlib import Path
-
-from datatrove.executor import LocalPipelineExecutor
-from datatrove.executor.slurm import SlurmPipelineExecutor
-from datatrove.pipeline.base import PipelineStep
-from huggingface_hub import HfApi
-from huggingface_hub.constants import REPOCARD_NAME
-
-from examples.port_datasets.droid_rlds.port_droid import DROID_SHARDS
-from lerobot.common.datasets.lerobot_dataset import CODEBASE_VERSION, LeRobotDatasetMetadata
-from lerobot.common.datasets.utils import create_lerobot_dataset_card
-from lerobot.common.utils.utils import init_logging
-
-
-class UploadDataset(PipelineStep):
-    def __init__(
-        self,
-        repo_id: str,
-        branch: str | None = None,
-        revision: str | None = None,
-        tags: list | None = None,
-        license: str | None = "apache-2.0",
-        private: bool = False,
-        distant_repo_id: str | None = None,
-        **card_kwargs,
-    ):
-        super().__init__()
-        self.repo_id = repo_id
-        self.distant_repo_id = self.repo_id if distant_repo_id is None else distant_repo_id
-        self.branch = branch
-        self.tags = tags
-        self.license = license
-        self.private = private
-        self.card_kwargs = card_kwargs
-        self.revision = revision if revision else CODEBASE_VERSION
-
-        if os.environ.get("HF_HUB_ENABLE_HF_TRANSFER", "0") != "1":
-            logging.warning(
-                'HF_HUB_ENABLE_HF_TRANSFER is not set to "1". Install hf_transfer and set the env '
-                "variable for faster uploads:\npip install hf-transfer\nexport HF_HUB_ENABLE_HF_TRANSFER=1"
-            )
-
-        self.create_repo()
-
-    def create_repo(self):
-        logging.info(f"Loading meta data from {self.repo_id}...")
-        meta = LeRobotDatasetMetadata(self.repo_id)
-
-        logging.info(f"Creating repo {self.distant_repo_id}...")
-        hub_api = HfApi()
-        hub_api.create_repo(
-            repo_id=self.distant_repo_id,
-            private=self.private,
-            repo_type="dataset",
-            exist_ok=True,
-        )
-        if self.branch:
-            hub_api.create_branch(
-                repo_id=self.distant_repo_id,
-                branch=self.branch,
-                revision=self.revision,
-                repo_type="dataset",
-                exist_ok=True,
-            )
-
-        if not hub_api.file_exists(
-            self.distant_repo_id, REPOCARD_NAME, repo_type="dataset", revision=self.branch
-        ):
-            card = create_lerobot_dataset_card(
-                tags=self.tags, dataset_info=meta.info, license=self.license, **self.card_kwargs
-            )
-            card.push_to_hub(repo_id=self.distant_repo_id, repo_type="dataset", revision=self.branch)
-
-        def list_files_recursively(directory):
-            base_path = Path(directory)
-            return [str(file.relative_to(base_path)) for file in base_path.rglob("*") if file.is_file()]
-
-        logging.info(f"Listing all local files from {self.repo_id}...")
-        self.file_paths = list_files_recursively(meta.root)
-        self.file_paths = sorted(self.file_paths)
-
-    def create_chunks(self, lst, n):
-        from itertools import islice
-
-        it = iter(lst)
-        return [list(islice(it, size)) for size in [len(lst) // n + (i < len(lst) % n) for i in range(n)]]
-
-    def create_commits(self, additions):
-        import logging
-        import math
-        import random
-        import time
-
-        from huggingface_hub import create_commit
-        from huggingface_hub.utils import HfHubHTTPError
-
-        FILES_BETWEEN_COMMITS = 10  # noqa: N806
-        BASE_DELAY = 0.1  # noqa: N806
-        MAX_RETRIES = 12  # noqa: N806
-
-        # Split the files into smaller chunks for faster commit
-        # and avoiding "A commit has happened since" error
-        num_chunks = math.ceil(len(additions) / FILES_BETWEEN_COMMITS)
-        chunks = self.create_chunks(additions, num_chunks)
-
-        for chunk in chunks:
-            retries = 0
-            while True:
-                try:
-                    create_commit(
-                        self.distant_repo_id,
-                        repo_type="dataset",
-                        operations=chunk,
-                        commit_message=f"DataTrove upload ({len(chunk)} files)",
-                        revision=self.branch,
-                    )
-                    # TODO: every 100 chunks super_squach_commits()
-                    logging.info("create_commit completed!")
-                    break
-                except HfHubHTTPError as e:
-                    if "A commit has happened since" in e.server_message:
-                        if retries >= MAX_RETRIES:
-                            logging.error(f"Failed to create commit after {MAX_RETRIES=}. Giving up.")
-                            raise e
-                        logging.info("Commit creation race condition issue. Waiting...")
-                        time.sleep(BASE_DELAY * 2**retries + random.uniform(0, 2))
-                        retries += 1
-                    else:
-                        raise e
-
-    def run(self, data=None, rank: int = 0, world_size: int = 1):
-        import logging
-
-        from datasets.utils.tqdm import disable_progress_bars
-        from huggingface_hub import CommitOperationAdd, preupload_lfs_files
-
-        from lerobot.common.datasets.lerobot_dataset import LeRobotDatasetMetadata
-        from lerobot.common.utils.utils import init_logging
-
-        init_logging()
-        disable_progress_bars()
-
-        chunks = self.create_chunks(self.file_paths, world_size)
-        file_paths = chunks[rank]
-
-        if len(file_paths) == 0:
-            raise ValueError(file_paths)
-
-        logging.info("Pre-uploading LFS files...")
-        for i, path in enumerate(file_paths):
-            logging.info(f"{i}: {path}")
-
-        meta = LeRobotDatasetMetadata(self.repo_id)
-        additions = [
-            CommitOperationAdd(path_in_repo=path, path_or_fileobj=meta.root / path) for path in file_paths
-        ]
-        preupload_lfs_files(
-            repo_id=self.distant_repo_id, repo_type="dataset", additions=additions, revision=self.branch
-        )
-
-        logging.info("Creating commits...")
-        self.create_commits(additions)
-        logging.info("Done!")
-
-
-def make_upload_executor(
-    repo_id, job_name, logs_dir, workers, partition, cpus_per_task, mem_per_cpu, slurm=True
-):
-    kwargs = {
-        "pipeline": [
-            UploadDataset(repo_id),
-        ],
-        "logging_dir": str(logs_dir / job_name),
-    }
-
-    if slurm:
-        kwargs.update(
-            {
-                "job_name": job_name,
-                "tasks": DROID_SHARDS,
-                "workers": workers,
-                "time": "08:00:00",
-                "partition": partition,
-                "cpus_per_task": cpus_per_task,
-                "sbatch_args": {"mem-per-cpu": mem_per_cpu},
-            }
-        )
-        executor = SlurmPipelineExecutor(**kwargs)
-    else:
-        kwargs.update(
-            {
-                "tasks": DROID_SHARDS,
-                "workers": 1,
-            }
-        )
-        executor = LocalPipelineExecutor(**kwargs)
-
-    return executor
-
-
-def main():
-    parser = argparse.ArgumentParser()
-
-    parser.add_argument(
-        "--repo-id",
-        type=str,
-        help="Repositery identifier on Hugging Face: a community or a user name `/` the name of the dataset, required when push-to-hub is True.",
-    )
-    parser.add_argument(
-        "--logs-dir",
-        type=Path,
-        help="Path to logs directory for `datatrove`.",
-    )
-    parser.add_argument(
-        "--job-name",
-        type=str,
-        default="upload_droid",
-        help="Job name used in slurm, and name of the directory created inside the provided logs directory.",
-    )
-    parser.add_argument(
-        "--slurm",
-        type=int,
-        default=1,
-        help="Launch over slurm. Use `--slurm 0` to launch sequentially (useful to debug).",
-    )
-    parser.add_argument(
-        "--workers",
-        type=int,
-        default=50,
-        help="Number of slurm workers. It should be less than the maximum number of shards.",
-    )
-    parser.add_argument(
-        "--partition",
-        type=str,
-        help="Slurm partition. Ideally a CPU partition. No need for GPU partition.",
-    )
-    parser.add_argument(
-        "--cpus-per-task",
-        type=int,
-        default=8,
-        help="Number of cpus that each slurm worker will use.",
-    )
-    parser.add_argument(
-        "--mem-per-cpu",
-        type=str,
-        default="1950M",
-        help="Memory per cpu that each worker will use.",
-    )
-
-    init_logging()
-
-    args = parser.parse_args()
-    kwargs = vars(args)
-    kwargs["slurm"] = kwargs.pop("slurm") == 1
-    upload_executor = make_upload_executor(**kwargs)
-    upload_executor.run()
-
-
-if __name__ == "__main__":
-    main()

lerobot/common/datasets/aggregate.py

@@ -1,413 +0,0 @@
-import logging
-import shutil
-from pathlib import Path
-
-import pandas as pd
-import tqdm
-
-from lerobot.common.constants import HF_LEROBOT_HOME
-from lerobot.common.datasets.compute_stats import aggregate_stats
-from lerobot.common.datasets.lerobot_dataset import LeRobotDataset, LeRobotDatasetMetadata
-from lerobot.common.datasets.utils import (
-    DEFAULT_CHUNK_SIZE,
-    DEFAULT_DATA_FILE_SIZE_IN_MB,
-    DEFAULT_DATA_PATH,
-    DEFAULT_EPISODES_PATH,
-    DEFAULT_VIDEO_FILE_SIZE_IN_MB,
-    DEFAULT_VIDEO_PATH,
-    concat_video_files,
-    get_parquet_file_size_in_mb,
-    get_video_size_in_mb,
-    safe_write_dataframe_to_parquet,
-    update_chunk_file_indices,
-    write_info,
-    write_stats,
-    write_tasks,
-)
-from lerobot.common.utils.utils import init_logging
-
-
-def validate_all_metadata(all_metadata: list[LeRobotDatasetMetadata]):
-    # validate same fps, robot_type, features
-
-    fps = all_metadata[0].fps
-    robot_type = all_metadata[0].robot_type
-    features = all_metadata[0].features
-
-    for meta in tqdm.tqdm(all_metadata, desc="Validate all meta data"):
-        if fps != meta.fps:
-            raise ValueError(f"Same fps is expected, but got fps={meta.fps} instead of {fps}.")
-        if robot_type != meta.robot_type:
-            raise ValueError(
-                f"Same robot_type is expected, but got robot_type={meta.robot_type} instead of {robot_type}."
-            )
-        if features != meta.features:
-            raise ValueError(
-                f"Same features is expected, but got features={meta.features} instead of {features}."
-            )
-
-    return fps, robot_type, features
-
-
-def update_data_df(df, src_meta, dst_meta):
-    def _update(row):
-        row["episode_index"] = row["episode_index"] + dst_meta["total_episodes"]
-        row["index"] = row["index"] + dst_meta["total_frames"]
-        task = src_meta.tasks.iloc[row["task_index"]].name
-        row["task_index"] = dst_meta.tasks.loc[task].task_index.item()
-        return row
-
-    return df.apply(_update, axis=1)
-
-
-def update_meta_data(
-    df,
-    dst_meta,
-    meta_idx,
-    data_idx,
-    videos_idx,
-):
-    def _update(row):
-        row["meta/episodes/chunk_index"] = row["meta/episodes/chunk_index"] + meta_idx["chunk_index"]
-        row["meta/episodes/file_index"] = row["meta/episodes/file_index"] + meta_idx["file_index"]
-        row["data/chunk_index"] = row["data/chunk_index"] + data_idx["chunk_index"]
-        row["data/file_index"] = row["data/file_index"] + data_idx["file_index"]
-        for key, video_idx in videos_idx.items():
-            row[f"videos/{key}/chunk_index"] = row[f"videos/{key}/chunk_index"] + video_idx["chunk_index"]
-            row[f"videos/{key}/file_index"] = row[f"videos/{key}/file_index"] + video_idx["file_index"]
-            row[f"videos/{key}/from_timestamp"] = (
-                row[f"videos/{key}/from_timestamp"] + video_idx["latest_duration"]
-            )
-            row[f"videos/{key}/to_timestamp"] = (
-                row[f"videos/{key}/to_timestamp"] + video_idx["latest_duration"]
-            )
-        row["dataset_from_index"] = row["dataset_from_index"] + dst_meta.info["total_frames"]
-        row["dataset_to_index"] = row["dataset_to_index"] + dst_meta.info["total_frames"]
-        row["episode_index"] = row["episode_index"] + dst_meta.info["total_episodes"]
-        return row
-
-    return df.apply(_update, axis=1)
-
-
-def aggregate_datasets(repo_ids: list[str], aggr_repo_id: str, roots: list[Path] = None, aggr_root=None):
-    logging.info("Start aggregate_datasets")
-
-    # Load metadata
-    all_metadata = (
-        [LeRobotDatasetMetadata(repo_id) for repo_id in repo_ids]
-        if roots is None
-        else [
-            LeRobotDatasetMetadata(repo_id, root=root) for repo_id, root in zip(repo_ids, roots, strict=False)
-        ]
-    )
-    fps, robot_type, features = validate_all_metadata(all_metadata)
-    video_keys = [key for key in features if features[key]["dtype"] == "video"]
-    image_keys = [key for key in features if features[key]["dtype"] == "image"]
-
-    # Initialize output dataset metadata
-    dst_meta = LeRobotDatasetMetadata.create(
-        repo_id=aggr_repo_id,
-        fps=fps,
-        robot_type=robot_type,
-        features=features,
-        root=aggr_root,
-    )
-
-    # Aggregate task info
-    logging.info("Find all tasks")
-    unique_tasks = pd.concat([m.tasks for m in all_metadata]).index.unique()
-    dst_meta.tasks = pd.DataFrame({"task_index": range(len(unique_tasks))}, index=unique_tasks)
-
-    # Track counters and indices
-    meta_idx = {"chunk": 0, "file": 0}
-    data_idx = {"chunk": 0, "file": 0}
-    videos_idx = {
-        key: {"chunk": 0, "file": 0, "latest_duration": 0, "episode_duration": 0} for key in video_keys
-    }
-
-    # Process each dataset
-    for src_meta in tqdm.tqdm(all_metadata, desc="Copy data and videos"):
-        videos_idx = aggregate_videos(src_meta, dst_meta, videos_idx)
-        data_idx = aggregate_data(src_meta, dst_meta, data_idx)
-        meta_idx = aggregate_metadata(src_meta, dst_meta, meta_idx, data_idx, videos_idx, video_keys)
-
-        dst_meta.info["total_episodes"] += src_meta.total_episodes
-        dst_meta.info["total_frames"] += src_meta.total_frames
-
-    finalize_aggregation(aggr_meta, all_metadata)
-    logging.info("Aggregation complete.")
-
-
-# -------------------------------
-# Helper Functions
-# -------------------------------
-
-
-def aggregate_videos(src_meta, dst_meta, videos_idx):
-    """
-    Aggregates video chunks from a dataset into the aggregated dataset folder.
-    """
-    for key, video_idx in videos_idx.items():
-        # Get unique (chunk, file) combinations
-        unique_chunk_file_pairs = {
-            (chunk, file)
-            for chunk, file in zip(
-                src_meta.episodes[f"videos/{key}/chunk_index"],
-                src_meta.episodes[f"videos/{key}/file_index"],
-                strict=False,
-            )
-        }
-
-        # Current target chunk/file index
-        chunk_idx = video_idx["chunk_idx"]
-        file_idx = video_idx["file_idx"]
-
-        for src_chunk_idx, src_file_idx in unique_chunk_file_pairs:
-            src_path = src_meta.root / DEFAULT_VIDEO_PATH.format(
-                video_key=key,
-                chunk_index=src_chunk_idx,
-                file_index=src_file_idx,
-            )
-
-            dst_path = dst_meta.root / DEFAULT_VIDEO_PATH.format(
-                video_key=key,
-                chunk_index=chunk_idx,
-                file_index=file_idx,
-            )
-
-            if not dst_path.exists():
-                # First write to this destination file
-                dst_path.parent.mkdir(parents=True, exist_ok=True)
-                shutil.copy(str(src_path), str(dst_path))
-                continue
-
-            # Check file sizes before appending
-            src_size = get_video_size_in_mb(src_path)
-            dst_size = get_video_size_in_mb(dst_path)
-
-            if dst_size + src_size >= DEFAULT_VIDEO_FILE_SIZE_IN_MB:
-                # Rotate to a new chunk/file
-                chunk_idx, file_idx = update_chunk_file_indices(chunk_idx, file_idx, DEFAULT_CHUNK_SIZE)
-                dst_path = dst_meta.root / DEFAULT_VIDEO_PATH.format(
-                    video_key=key,
-                    chunk_index=chunk_idx,
-                    file_index=file_idx,
-                )
-                dst_path.parent.mkdir(parents=True, exist_ok=True)
-                shutil.copy(str(src_path), str(dst_path))
-            else:
-                # Append to existing video file
-                concat_video_files(
-                    [dst_path, src_path],
-                    dst_meta.root,
-                    key,
-                    chunk_idx,
-                    file_idx,
-                )
-
-                if aggr_size_in_mb + size_in_mb >= DEFAULT_DATA_FILE_SIZE_IN_MB:
-                    # Size limit is reached, prepare new parquet file
-                    aggr_data_chunk_idx, aggr_data_file_idx = update_chunk_file_indices(
-                        aggr_data_chunk_idx, aggr_data_file_idx, DEFAULT_CHUNK_SIZE
-                    )
-                    aggr_path = aggr_root / DEFAULT_DATA_PATH.format(
-                        chunk_index=aggr_data_chunk_idx, file_index=aggr_data_file_idx
-                    )
-                    aggr_path.parent.mkdir(parents=True, exist_ok=True)
-                    df.to_parquet(aggr_path)
-                else:
-                    # Update the existing parquet file with new rows
-                    aggr_df = pd.read_parquet(aggr_path)
-                    df = pd.concat([aggr_df, df], ignore_index=True)
-                    safe_write_dataframe_to_parquet(df, aggr_path, image_keys)
-
-        return videos_idx
-
-
-def aggregate_data(src_meta, dst_meta, data_idx):
-    unique_chunk_file_ids = {
-        (c, f)
-        for c, f in zip(
-            src_meta.episodes["data/chunk_index"], src_meta.episodes["data/file_index"], strict=False
-        )
-    }
-    for src_chunk_idx, src_file_idx in unique_chunk_file_ids:
-        src_path = src_meta.root / DEFAULT_DATA_PATH.format(
-            chunk_index=src_chunk_idx, file_index=src_file_idx
-        )
-        df = pd.read_parquet(src_path)
-        df = update_data_df(df, src_meta, dst_meta)
-
-        dst_path = aggr_root / DEFAULT_DATA_PATH.format(
-            chunk_index=data_idx["chunk"], file_index=data_idx["file"]
-        )
-        data_idx = write_parquet_safely(
-            df,
-            src_path,
-            dst_path,
-            data_idx,
-            DEFAULT_DATA_FILE_SIZE_IN_MB,
-            DEFAULT_CHUNK_SIZE,
-            DEFAULT_DATA_PATH,
-        )
-
-    return data_idx
-
-
-def aggregate_metadata(src_meta, dst_meta, meta_idx, data_idx, videos_idx):
-    chunk_file_ids = {
-        (c, f)
-        for c, f in zip(
-            src_meta.episodes["meta/episodes/chunk_index"],
-            src_meta.episodes["meta/episodes/file_index"],
-            strict=False,
-        )
-    }
-
-    for chunk_idx, file_idx in chunk_file_ids:
-        src_path = src_meta.root / DEFAULT_EPISODES_PATH.format(chunk_index=chunk_idx, file_index=file_idx)
-        df = pd.read_parquet(src_path)
-        df = update_meta_data(
-            df,
-            dst_meta,
-            meta_idx,
-            data_idx,
-            videos_idx,
-        )
-
-        # for k in video_keys:
-        #     video_idx[k]["latest_duration"] += video_idx[k]["episode_duration"]
-
-        dst_path = dst_meta.root / DEFAULT_EPISODES_PATH.format(
-            chunk_index=meta_idx["chunk"], file_index=meta_idx["file"]
-        )
-        write_parquet_safely(
-            df,
-            src_path,
-            dst_path,
-            meta_idx,
-            DEFAULT_DATA_FILE_SIZE_IN_MB,
-            DEFAULT_CHUNK_SIZE,
-            DEFAULT_EPISODES_PATH,
-        )
-
-    return meta_idx
-
-
-def write_parquet_safely(
-    df: pd.DataFrame,
-    src_path: Path,
-    dst_path: Path,
-    idx: dict[str, int],
-    max_mb: float,
-    chunk_size: int,
-    default_path: str,
-):
-    """
-    Safely appends or creates a Parquet file at dst_path based on size constraints.
-
-    Parameters:
-        df (pd.DataFrame): Data to write.
-        src_path (Path): Path to source file (used to get size).
-        dst_path (Path): Target path for writing.
-        idx (dict): Dictionary containing 'chunk' and 'file' indices.
-        max_mb (float): Maximum allowed file size in MB.
-        chunk_size (int): Maximum number of files per chunk.
-        default_path (str): Format string for generating a new file path.
-
-    Returns:
-        dict: Updated index dictionary.
-    """
-
-    # If destination file doesn't exist, just write the new one
-    if not dst_path.exists():
-        dst_path.parent.mkdir(parents=True, exist_ok=True)
-        df.to_parquet(dst_path)
-        return idx
-
-    # Otherwise, check if we exceed the size limit
-    src_size = get_parquet_file_size_in_mb(src_path)
-    dst_size = get_parquet_file_size_in_mb(dst_path)
-
-    if dst_size + src_size >= max_mb:
-        # File is too large, move to a new one
-        idx["chunk"], idx["file"] = update_chunk_file_indices(idx["chunk"], idx["file"], chunk_size)
-        new_path = dst_path.parent / default_path.format(chunk_index=idx["chunk"], file_index=idx["file"])
-        new_path.parent.mkdir(parents=True, exist_ok=True)
-        df.to_parquet(new_path)
-    else:
-        # Append to existing file
-        existing_df = pd.read_parquet(dst_path)
-        combined_df = pd.concat([existing_df, df], ignore_index=True)
-        combined_df.to_parquet(dst_path)
-
-    return idx
-
-
-def finalize_aggregation(aggr_meta, all_metadata):
-    logging.info("write tasks")
-    write_tasks(aggr_meta.tasks, aggr_meta.root)
-
-    logging.info("write info")
-    aggr_meta.info.update(
-        {
-            "total_tasks": len(aggr_meta.tasks),
-            "total_episodes": sum(m.total_episodes for m in all_metadata),
-            "total_frames": sum(m.total_frames for m in all_metadata),
-            "splits": {"train": f"0:{sum(m.total_episodes for m in all_metadata)}"},
-        }
-    )
-    write_info(aggr_meta.info, aggr_meta.root)
-
-    logging.info("write stats")
-    aggr_meta.stats = aggregate_stats([m.stats for m in all_metadata])
-    write_stats(aggr_meta.stats, aggr_meta.root)
-
-
-if __name__ == "__main__":
-    init_logging()
-
-    num_shards = 2048
-    repo_id = "cadene/droid_1.0.1_v30"
-    aggr_repo_id = f"{repo_id}_compact_6"
-    tags = ["openx"]
-
-    # num_shards = 210
-    # repo_id = "cadene/agibot_alpha_v30"
-    # aggr_repo_id = f"{repo_id}"
-    # tags = None
-
-    # aggr_root = Path(f"/tmp/{aggr_repo_id}")
-    aggr_root = HF_LEROBOT_HOME / aggr_repo_id
-    if aggr_root.exists():
-        shutil.rmtree(aggr_root)
-
-    repo_ids = []
-    roots = []
-    for rank in range(num_shards):
-        shard_repo_id = f"{repo_id}_world_{num_shards}_rank_{rank}"
-        shard_root = HF_LEROBOT_HOME / shard_repo_id
-        try:
-            meta = LeRobotDatasetMetadata(shard_repo_id, root=shard_root)
-            if len(meta.video_keys) == 0:
-                continue
-            repo_ids.append(shard_repo_id)
-            roots.append(shard_root)
-        except:
-            pass
-
-        if rank == 1:
-            break
-
-    aggregate_datasets(
-        repo_ids,
-        aggr_repo_id,
-        roots=roots,
-        aggr_root=aggr_root,
-    )
-
-    aggr_dataset = LeRobotDataset(repo_id=aggr_repo_id, root=aggr_root)
-    # for i in tqdm.tqdm(range(len(aggr_dataset))):
-    #     aggr_dataset[i]
-    #     pass
-    aggr_dataset.push_to_hub(tags=tags, upload_large_folder=True)

lerobot/common/datasets/backward_compatibility.py

@@ -47,18 +47,6 @@ If you encounter a problem, contact LeRobot maintainers on [Discord](https://dis
 or open an [issue on GitHub](https://github.com/huggingface/lerobot/issues/new/choose).
 """
 
-V30_MESSAGE = """
-The dataset you requested ({repo_id}) is in {version} format.
-While current version of LeRobot is backward-compatible with it, the version of your dataset still uses global
-stats instead of per-episode stats. Update your dataset stats to the new format using this command:
-```
-python lerobot/common/datasets/v30/convert_dataset_v21_to_v30.py --repo-id={repo_id}
-```
-
-If you encounter a problem, contact LeRobot maintainers on [Discord](https://discord.com/invite/s3KuuzsPFb)
-or open an [issue on GitHub](https://github.com/huggingface/lerobot/issues/new/choose).
-"""
-
 FUTURE_MESSAGE = """
 The dataset you requested ({repo_id}) is only available in {version} format.
 As we cannot ensure forward compatibility with it, please update your current version of lerobot.
@@ -70,14 +58,7 @@ class CompatibilityError(Exception): ...
 
 class BackwardCompatibilityError(CompatibilityError):
     def __init__(self, repo_id: str, version: packaging.version.Version):
-        if version.major == 3:
-            message = V30_MESSAGE.format(repo_id=repo_id, version=version)
-        elif version.major == 2:
-            message = V2_MESSAGE.format(repo_id=repo_id, version=version)
-        else:
-            raise NotImplementedError(
-                "Contact the maintainer on [Discord](https://discord.com/invite/s3KuuzsPFb)."
-            )
+        message = V2_MESSAGE.format(repo_id=repo_id, version=version)
         super().__init__(message)
 
 

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/lerobot_dataset.py

@@ -16,18 +16,16 @@
 import contextlib
 import logging
 import shutil
-import tempfile
 from pathlib import Path
 from typing import Callable
 
 import datasets
 import numpy as np
 import packaging.version
-import pandas as pd
 import PIL.Image
 import torch
 import torch.utils
-from datasets import Dataset
+from datasets import concatenate_datasets, load_dataset
 from huggingface_hub import HfApi, snapshot_download
 from huggingface_hub.constants import REPOCARD_NAME
 from huggingface_hub.errors import RevisionNotFoundError
@@ -36,41 +34,36 @@ from lerobot.common.constants import HF_LEROBOT_HOME
 from lerobot.common.datasets.compute_stats import aggregate_stats, compute_episode_stats
 from lerobot.common.datasets.image_writer import AsyncImageWriter, write_image
 from lerobot.common.datasets.utils import (
-    DEFAULT_EPISODES_PATH,
     DEFAULT_FEATURES,
     DEFAULT_IMAGE_PATH,
     INFO_PATH,
+    TASKS_PATH,
+    append_jsonlines,
+    backward_compatible_episodes_stats,
     check_delta_timestamps,
+    check_timestamps_sync,
     check_version_compatibility,
-    concat_video_files,
     create_empty_dataset_info,
     create_lerobot_dataset_card,
     embed_images,
-    flatten_dict,
     get_delta_indices,
+    get_episode_data_index,
     get_features_from_robot,
-    get_hf_dataset_size_in_mb,
     get_hf_features_from_features,
-    get_parquet_file_size_in_mb,
-    get_parquet_num_frames,
     get_safe_version,
-    get_video_duration_in_s,
-    get_video_size_in_mb,
     hf_transform_to_torch,
     is_valid_version,
     load_episodes,
+    load_episodes_stats,
     load_info,
-    load_nested_dataset,
     load_stats,
     load_tasks,
-    safe_write_dataframe_to_parquet,
-    update_chunk_file_indices,
     validate_episode_buffer,
     validate_frame,
+    write_episode,
+    write_episode_stats,
     write_info,
     write_json,
-    write_stats,
-    write_tasks,
 )
 from lerobot.common.datasets.video_utils import (
     VideoFrame,
@@ -81,7 +74,7 @@ from lerobot.common.datasets.video_utils import (
 )
 from lerobot.common.robot_devices.robots.utils import Robot
 
-CODEBASE_VERSION = "v3.0"
+CODEBASE_VERSION = "v2.1"
 
 
 class LeRobotDatasetMetadata:
@@ -105,18 +98,20 @@ class LeRobotDatasetMetadata:
                 self.revision = get_safe_version(self.repo_id, self.revision)
 
             (self.root / "meta").mkdir(exist_ok=True, parents=True)
-            # TODO(rcadene): instead of downloading all episodes metadata files,
-            # download only the ones associated to the requested episodes. This would
-            # require adding `episodes: list[int]` as argument.
             self.pull_from_repo(allow_patterns="meta/")
             self.load_metadata()
 
     def load_metadata(self):
         self.info = load_info(self.root)
         check_version_compatibility(self.repo_id, self._version, CODEBASE_VERSION)
-        self.tasks = load_tasks(self.root)
+        self.tasks, self.task_to_task_index = load_tasks(self.root)
         self.episodes = load_episodes(self.root)
-        self.stats = load_stats(self.root)
+        if self._version < packaging.version.parse("v2.1"):
+            self.stats = load_stats(self.root)
+            self.episodes_stats = backward_compatible_episodes_stats(self.stats, self.episodes)
+        else:
+            self.episodes_stats = load_episodes_stats(self.root)
+            self.stats = aggregate_stats(list(self.episodes_stats.values()))
 
     def pull_from_repo(
         self,
@@ -138,22 +133,17 @@ class LeRobotDatasetMetadata:
         return packaging.version.parse(self.info["codebase_version"])
 
     def get_data_file_path(self, ep_index: int) -> Path:
-        ep = self.episodes[ep_index]
-        chunk_idx = ep["data/chunk_index"]
-        file_idx = ep["data/file_index"]
-        fpath = self.data_path.format(chunk_index=chunk_idx, file_index=file_idx)
+        ep_chunk = self.get_episode_chunk(ep_index)
+        fpath = self.data_path.format(episode_chunk=ep_chunk, episode_index=ep_index)
         return Path(fpath)
 
     def get_video_file_path(self, ep_index: int, vid_key: str) -> Path:
-        ep = self.episodes[ep_index]
-        chunk_idx = ep[f"videos/{vid_key}/chunk_index"]
-        file_idx = ep[f"videos/{vid_key}/file_index"]
-        fpath = self.video_path.format(video_key=vid_key, chunk_index=chunk_idx, file_index=file_idx)
+        ep_chunk = self.get_episode_chunk(ep_index)
+        fpath = self.video_path.format(episode_chunk=ep_chunk, video_key=vid_key, episode_index=ep_index)
         return Path(fpath)
 
-    @property
-    def url_root(self) -> str:
-        return f"https://huggingface.co/datasets/{self.repo_id}/resolve/main"
+    def get_episode_chunk(self, ep_index: int) -> int:
+        return ep_index // self.chunks_size
 
     @property
     def data_path(self) -> str:
@@ -220,109 +210,40 @@ class LeRobotDatasetMetadata:
         """Total number of different tasks performed in this dataset."""
         return self.info["total_tasks"]
 
+    @property
+    def total_chunks(self) -> int:
+        """Total number of chunks (groups of episodes)."""
+        return self.info["total_chunks"]
+
     @property
     def chunks_size(self) -> int:
-        """Max number of files per chunk."""
+        """Max number of episodes per chunk."""
         return self.info["chunks_size"]
 
-    @property
-    def data_files_size_in_mb(self) -> int:
-        """Max size of data file in mega bytes."""
-        return self.info["data_files_size_in_mb"]
-
-    @property
-    def video_files_size_in_mb(self) -> int:
-        """Max size of video file in mega bytes."""
-        return self.info["video_files_size_in_mb"]
-
     def get_task_index(self, task: str) -> int | None:
         """
         Given a task in natural language, returns its task_index if the task already exists in the dataset,
         otherwise return None.
         """
-        if task in self.tasks.index:
-            return int(self.tasks.loc[task].task_index)
-        else:
-            return None
+        return self.task_to_task_index.get(task, None)
 
-    def save_episode_tasks(self, tasks: list[str]):
-        if len(set(tasks)) != len(tasks):
-            raise ValueError(f"Tasks are not unique: {tasks}")
-
-        if self.tasks is None:
-            new_tasks = tasks
-            task_indices = range(len(tasks))
-            self.tasks = pd.DataFrame({"task_index": task_indices}, index=tasks)
-        else:
-            new_tasks = [task for task in tasks if task not in self.tasks.index]
-            new_task_indices = range(len(self.tasks), len(self.tasks) + len(new_tasks))
-            for task_idx, task in zip(new_task_indices, new_tasks, strict=False):
-                self.tasks.loc[task] = task_idx
-
-        if len(new_tasks) > 0:
-            # Update on disk
-            write_tasks(self.tasks, self.root)
-
-    def _save_episode_metadata(self, episode_dict: dict) -> None:
-        """Save episode metadata to a parquet file and update the Hugging Face dataset of episodes metadata.
-
-        This function processes episodes metadata from a dictionary, converts it into a Hugging Face dataset,
-        and saves it as a parquet file. It handles both the creation of new parquet files and the
-        updating of existing ones based on size constraints. After saving the metadata, it reloads
-        the Hugging Face dataset to ensure it is up-to-date.
-
-        Notes: We both need to update parquet files and HF dataset:
-        - `pandas` loads parquet file in RAM
-        - `datasets` relies on a memory mapping from pyarrow (no RAM). It either converts parquet files to a pyarrow cache on disk,
-          or loads directly from pyarrow cache.
+    def add_task(self, task: str):
         """
-        # Convert buffer into HF Dataset
-        episode_dict = {key: [value] for key, value in episode_dict.items()}
-        ep_dataset = Dataset.from_dict(episode_dict)
-        ep_size_in_mb = get_hf_dataset_size_in_mb(ep_dataset)
-        df = pd.DataFrame(ep_dataset)
-        num_frames = episode_dict["length"][0]
+        Given a task in natural language, add it to the dictionary of tasks.
+        """
+        if task in self.task_to_task_index:
+            raise ValueError(f"The task '{task}' already exists and can't be added twice.")
 
-        if self.episodes is None:
-            # Initialize indices and frame count for a new dataset made of the first episode data
-            chunk_idx, file_idx = 0, 0
-            df["meta/episodes/chunk_index"] = [chunk_idx]
-            df["meta/episodes/file_index"] = [file_idx]
-            df["dataset_from_index"] = [0]
-            df["dataset_to_index"] = [num_frames]
-        else:
-            # Retrieve information from the latest parquet file
-            latest_ep = self.episodes[-1]
-            chunk_idx = latest_ep["meta/episodes/chunk_index"]
-            file_idx = latest_ep["meta/episodes/file_index"]
+        task_index = self.info["total_tasks"]
+        self.task_to_task_index[task] = task_index
+        self.tasks[task_index] = task
+        self.info["total_tasks"] += 1
 
-            latest_path = self.root / DEFAULT_EPISODES_PATH.format(chunk_index=chunk_idx, file_index=file_idx)
-            latest_size_in_mb = get_parquet_file_size_in_mb(latest_path)
-
-            if latest_size_in_mb + ep_size_in_mb >= self.data_files_size_in_mb:
-                # Size limit is reached, prepare new parquet file
-                chunk_idx, file_idx = update_chunk_file_indices(chunk_idx, file_idx, self.chunks_size)
-
-            # Update the existing pandas dataframe with new row
-            df["meta/episodes/chunk_index"] = [chunk_idx]
-            df["meta/episodes/file_index"] = [file_idx]
-            df["dataset_from_index"] = [latest_ep["dataset_to_index"]]
-            df["dataset_to_index"] = [latest_ep["dataset_to_index"] + num_frames]
-
-            if latest_size_in_mb + ep_size_in_mb < self.data_files_size_in_mb:
-                # Size limit wasnt reached, concatenate latest dataframe with new one
-                latest_df = pd.read_parquet(latest_path)
-                df = pd.concat([latest_df, df], ignore_index=True)
-
-        # Write the resulting dataframe from RAM to disk
-        path = self.root / DEFAULT_EPISODES_PATH.format(chunk_index=chunk_idx, file_index=file_idx)
-        path.parent.mkdir(parents=True, exist_ok=True)
-        df.to_parquet(path, index=False)
-
-        # Update the Hugging Face dataset by reloading it.
-        # This process should be fast because only the latest Parquet file has been modified.
-        # Therefore, only this file needs to be converted to PyArrow; the rest is loaded from the PyArrow memory-mapped cache.
-        self.episodes = load_episodes(self.root)
+        task_dict = {
+            "task_index": task_index,
+            "task": task,
+        }
+        append_jsonlines(task_dict, self.root / TASKS_PATH)
 
     def save_episode(
         self,
@@ -330,28 +251,32 @@ class LeRobotDatasetMetadata:
         episode_length: int,
         episode_tasks: list[str],
         episode_stats: dict[str, dict],
-        episode_metadata: dict,
     ) -> None:
+        self.info["total_episodes"] += 1
+        self.info["total_frames"] += episode_length
+
+        chunk = self.get_episode_chunk(episode_index)
+        if chunk >= self.total_chunks:
+            self.info["total_chunks"] += 1
+
+        self.info["splits"] = {"train": f"0:{self.info['total_episodes']}"}
+        self.info["total_videos"] += len(self.video_keys)
+        if len(self.video_keys) > 0:
+            self.update_video_info()
+
+        write_info(self.info, self.root)
+
         episode_dict = {
             "episode_index": episode_index,
             "tasks": episode_tasks,
             "length": episode_length,
         }
-        episode_dict.update(episode_metadata)
-        episode_dict.update(flatten_dict({"stats": episode_stats}))
-        self._save_episode_metadata(episode_dict)
+        self.episodes[episode_index] = episode_dict
+        write_episode(episode_dict, self.root)
 
-        # Update info
-        self.info["total_episodes"] += 1
-        self.info["total_frames"] += episode_length
-        self.info["total_tasks"] = len(self.tasks)
-        self.info["splits"] = {"train": f"0:{self.info['total_episodes']}"}
-        if len(self.video_keys) > 0:
-            self.update_video_info()
-        write_info(self.info, self.root)
-
-        self.stats = aggregate_stats([self.stats, episode_stats]) if self.stats is not None else episode_stats
-        write_stats(self.stats, self.root)
+        self.episodes_stats[episode_index] = episode_stats
+        self.stats = aggregate_stats([self.stats, episode_stats]) if self.stats else episode_stats
+        write_episode_stats(episode_index, episode_stats, self.root)
 
     def update_video_info(self) -> None:
         """
@@ -393,7 +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(
@@ -416,9 +341,8 @@ class LeRobotDatasetMetadata:
 
             features = {**features, **DEFAULT_FEATURES}
 
-        obj.tasks = None
-        obj.episodes = None
-        obj.stats = None
+        obj.tasks, obj.task_to_task_index = {}, {}
+        obj.episodes_stats, obj.stats, obj.episodes = {}, {}, {}
         obj.info = create_empty_dataset_info(CODEBASE_VERSION, fps, robot_type, features, use_videos)
         if len(obj.video_keys) > 0 and not use_videos:
             raise ValueError()
@@ -563,17 +487,29 @@ class LeRobotDataset(torch.utils.data.Dataset):
         self.meta = LeRobotDatasetMetadata(
             self.repo_id, self.root, self.revision, force_cache_sync=force_cache_sync
         )
+        if self.episodes is not None and self.meta._version >= packaging.version.parse("v2.1"):
+            episodes_stats = [self.meta.episodes_stats[ep_idx] for ep_idx in self.episodes]
+            self.stats = aggregate_stats(episodes_stats)
 
         # Load actual data
         try:
             if force_cache_sync:
                 raise FileNotFoundError
+            assert all((self.root / fpath).is_file() for fpath in self.get_episodes_file_paths())
             self.hf_dataset = self.load_hf_dataset()
         except (AssertionError, FileNotFoundError, NotADirectoryError):
             self.revision = get_safe_version(self.repo_id, self.revision)
-            self.download(download_videos)
+            self.download_episodes(download_videos)
             self.hf_dataset = self.load_hf_dataset()
 
+        self.episode_data_index = get_episode_data_index(self.meta.episodes, self.episodes)
+
+        # Check timestamps
+        timestamps = torch.stack(self.hf_dataset["timestamp"]).numpy()
+        episode_indices = torch.stack(self.hf_dataset["episode_index"]).numpy()
+        ep_data_index_np = {k: t.numpy() for k, t in self.episode_data_index.items()}
+        check_timestamps_sync(timestamps, episode_indices, ep_data_index_np, self.fps, self.tolerance_s)
+
         # Setup delta_indices
         if self.delta_timestamps is not None:
             check_delta_timestamps(self.delta_timestamps, self.fps, self.tolerance_s)
@@ -649,7 +585,7 @@ class LeRobotDataset(torch.utils.data.Dataset):
             ignore_patterns=ignore_patterns,
         )
 
-    def download(self, download_videos: bool = True) -> None:
+    def download_episodes(self, download_videos: bool = True) -> None:
         """Downloads the dataset from the given 'repo_id' at the provided version. If 'episodes' is given, this
         will only download those episodes (selected by their episode_index). If 'episodes' is None, the whole
         dataset will be downloaded. Thanks to the behavior of snapshot_download, if the files are already present
@@ -657,10 +593,11 @@ class LeRobotDataset(torch.utils.data.Dataset):
         """
         # TODO(rcadene, aliberts): implement faster transfer
         # https://huggingface.co/docs/huggingface_hub/en/guides/download#faster-downloads
-        ignore_patterns = None if download_videos else "videos/"
         files = None
+        ignore_patterns = None if download_videos else "videos/"
         if self.episodes is not None:
             files = self.get_episodes_file_paths()
+
         self.pull_from_repo(allow_patterns=files, ignore_patterns=ignore_patterns)
 
     def get_episodes_file_paths(self) -> list[Path]:
@@ -673,13 +610,19 @@ class LeRobotDataset(torch.utils.data.Dataset):
                 for ep_idx in episodes
             ]
             fpaths += video_files
-        # episodes are stored in the same files, so we return unique paths only
-        fpaths = list(set(fpaths))
+
         return fpaths
 
     def load_hf_dataset(self) -> datasets.Dataset:
         """hf_dataset contains all the observations, states, actions, rewards, etc."""
-        hf_dataset = load_nested_dataset(self.root / "data")
+        if self.episodes is None:
+            path = str(self.root / "data")
+            hf_dataset = load_dataset("parquet", data_dir=path, split="train")
+        else:
+            files = [str(self.root / self.meta.get_data_file_path(ep_idx)) for ep_idx in self.episodes]
+            hf_dataset = load_dataset("parquet", data_files=files, split="train")
+
+        # TODO(aliberts): hf_dataset.set_format("torch")
         hf_dataset.set_transform(hf_transform_to_torch)
         return hf_dataset
 
@@ -687,6 +630,8 @@ class LeRobotDataset(torch.utils.data.Dataset):
         features = get_hf_features_from_features(self.features)
         ft_dict = {col: [] for col in features}
         hf_dataset = datasets.Dataset.from_dict(ft_dict, features=features, split="train")
+
+        # TODO(aliberts): hf_dataset.set_format("torch")
         hf_dataset.set_transform(hf_transform_to_torch)
         return hf_dataset
 
@@ -718,16 +663,15 @@ class LeRobotDataset(torch.utils.data.Dataset):
             return get_hf_features_from_features(self.features)
 
     def _get_query_indices(self, idx: int, ep_idx: int) -> tuple[dict[str, list[int | bool]]]:
-        ep = self.meta.episodes[ep_idx]
-        ep_start = ep["dataset_from_index"]
-        ep_end = ep["dataset_to_index"]
+        ep_start = self.episode_data_index["from"][ep_idx]
+        ep_end = self.episode_data_index["to"][ep_idx]
         query_indices = {
-            key: [max(ep_start, min(ep_end - 1, idx + delta)) for delta in delta_idx]
+            key: [max(ep_start.item(), min(ep_end.item() - 1, idx + delta)) for delta in delta_idx]
             for key, delta_idx in self.delta_indices.items()
         }
         padding = {  # Pad values outside of current episode range
             f"{key}_is_pad": torch.BoolTensor(
-                [(idx + delta < ep_start) | (idx + delta >= ep_end) for delta in delta_idx]
+                [(idx + delta < ep_start.item()) | (idx + delta >= ep_end.item()) for delta in delta_idx]
             )
             for key, delta_idx in self.delta_indices.items()
         }
@@ -741,7 +685,7 @@ class LeRobotDataset(torch.utils.data.Dataset):
         query_timestamps = {}
         for key in self.meta.video_keys:
             if query_indices is not None and key in query_indices:
-                timestamps = self.hf_dataset[query_indices[key]]["timestamp"]
+                timestamps = self.hf_dataset.select(query_indices[key])["timestamp"]
                 query_timestamps[key] = torch.stack(timestamps).tolist()
             else:
                 query_timestamps[key] = [current_ts]
@@ -750,7 +694,7 @@ class LeRobotDataset(torch.utils.data.Dataset):
 
     def _query_hf_dataset(self, query_indices: dict[str, list[int]]) -> dict:
         return {
-            key: torch.stack(self.hf_dataset[q_idx][key])
+            key: torch.stack(self.hf_dataset.select(q_idx)[key])
             for key, q_idx in query_indices.items()
             if key not in self.meta.video_keys
         }
@@ -761,17 +705,10 @@ class LeRobotDataset(torch.utils.data.Dataset):
         Segmentation Fault. This probably happens because a memory reference to the video loader is created in
         the main process and a subprocess fails to access it.
         """
-        ep = self.meta.episodes[ep_idx]
         item = {}
         for vid_key, query_ts in query_timestamps.items():
-            # Episodes are stored sequentially on a single mp4 to reduce the number of files.
-            # Thus we load the start timestamp of the episode on this mp4 and
-            # shift the query timestamp accordingly.
-            from_timestamp = ep[f"videos/{vid_key}/from_timestamp"]
-            shifted_query_ts = [from_timestamp + ts for ts in query_ts]
-
             video_path = self.root / self.meta.get_video_file_path(ep_idx, vid_key)
-            frames = decode_video_frames(video_path, shifted_query_ts, self.tolerance_s, self.video_backend)
+            frames = decode_video_frames(video_path, query_ts, self.tolerance_s, self.video_backend)
             item[vid_key] = frames.squeeze(0)
 
         return item
@@ -809,7 +746,8 @@ class LeRobotDataset(torch.utils.data.Dataset):
 
         # Add task as a string
         task_idx = item["task_index"].item()
-        item["task"] = self.meta.tasks.iloc[task_idx].name
+        item["task"] = self.meta.tasks[task_idx]
+
         return item
 
     def __repr__(self):
@@ -839,9 +777,6 @@ class LeRobotDataset(torch.utils.data.Dataset):
         )
         return self.root / fpath
 
-    def _get_image_file_dir(self, episode_index: int, image_key: str) -> Path:
-        return self._get_image_file_path(episode_index, image_key, frame_index=0).parent
-
     def _save_image(self, image: torch.Tensor | np.ndarray | PIL.Image.Image, fpath: Path) -> None:
         if self.image_writer is None:
             if isinstance(image, torch.Tensor):
@@ -886,7 +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)
@@ -920,8 +857,11 @@ class LeRobotDataset(torch.utils.data.Dataset):
         episode_buffer["index"] = np.arange(self.meta.total_frames, self.meta.total_frames + episode_length)
         episode_buffer["episode_index"] = np.full((episode_length,), episode_index)
 
-        # Update tasks and task indices with new tasks if any
-        self.meta.save_episode_tasks(episode_tasks)
+        # Add new tasks to the tasks dictionary
+        for task in episode_tasks:
+            task_index = self.meta.get_task_index(task)
+            if task_index is None:
+                self.meta.add_task(task)
 
         # Given tasks in natural language, find their corresponding task indices
         episode_buffer["task_index"] = np.array([self.meta.get_task_index(task) for task in tasks])
@@ -929,155 +869,58 @@ 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])
 
-        # Wait for image writer to end, so that episode stats over images can be computed
         self._wait_image_writer()
+        self._save_episode_table(episode_buffer, episode_index)
         ep_stats = compute_episode_stats(episode_buffer, self.features)
 
-        ep_metadata = self._save_episode_data(episode_buffer)
-        for video_key in self.meta.video_keys:
-            ep_metadata.update(self._save_episode_video(video_key, episode_index))
+        if len(self.meta.video_keys) > 0:
+            video_paths = self.encode_episode_videos(episode_index)
+            for key in self.meta.video_keys:
+                episode_buffer[key] = video_paths[key]
 
-        # `meta.save_episode` need to be executed after encoding the videos
-        self.meta.save_episode(episode_index, episode_length, episode_tasks, ep_stats, ep_metadata)
+        # `meta.save_episode` be executed after encoding the videos
+        self.meta.save_episode(episode_index, episode_length, episode_tasks, ep_stats)
 
-        # TODO(rcadene): remove? there is only one episode in the episode buffer, no need for ep_data_index
-        # ep_data_index = get_episode_data_index(self.meta.episodes, [episode_index])
-        # ep_data_index_np = {k: t.numpy() for k, t in ep_data_index.items()}
-        # check_timestamps_sync(
-        #     episode_buffer["timestamp"],
-        #     episode_buffer["episode_index"],
-        #     ep_data_index_np,
-        #     self.fps,
-        #     self.tolerance_s,
-        # )
+        ep_data_index = get_episode_data_index(self.meta.episodes, [episode_index])
+        ep_data_index_np = {k: t.numpy() for k, t in ep_data_index.items()}
+        check_timestamps_sync(
+            episode_buffer["timestamp"],
+            episode_buffer["episode_index"],
+            ep_data_index_np,
+            self.fps,
+            self.tolerance_s,
+        )
+
+        video_files = list(self.root.rglob("*.mp4"))
+        assert len(video_files) == self.num_episodes * len(self.meta.video_keys)
+
+        parquet_files = list(self.root.rglob("*.parquet"))
+        assert len(parquet_files) == self.num_episodes
 
-        # TODO(rcadene): images are also deleted in clear_episode_buffer
         # delete images
         img_dir = self.root / "images"
         if img_dir.is_dir():
             shutil.rmtree(self.root / "images")
 
-        if not episode_data:
-            # Reset episode buffer
+        if not episode_data:  # Reset the buffer
             self.episode_buffer = self.create_episode_buffer()
 
-    def _save_episode_data(self, episode_buffer: dict) -> dict:
-        """Save episode data to a parquet file and update the Hugging Face dataset of frames data.
-
-        This function processes episodes data from a buffer, converts it into a Hugging Face dataset,
-        and saves it as a parquet file. It handles both the creation of new parquet files and the
-        updating of existing ones based on size constraints. After saving the data, it reloads
-        the Hugging Face dataset to ensure it is up-to-date.
-
-        Notes: We both need to update parquet files and HF dataset:
-        - `pandas` loads parquet file in RAM
-        - `datasets` relies on a memory mapping from pyarrow (no RAM). It either converts parquet files to a pyarrow cache on disk,
-          or loads directly from pyarrow cache.
-        """
-        # Convert buffer into HF Dataset
-        ep_dict = {key: episode_buffer[key] for key in self.hf_features}
-        ep_dataset = datasets.Dataset.from_dict(ep_dict, features=self.hf_features, split="train")
+    def _save_episode_table(self, episode_buffer: dict, episode_index: int) -> None:
+        episode_dict = {key: episode_buffer[key] for key in self.hf_features}
+        ep_dataset = datasets.Dataset.from_dict(episode_dict, features=self.hf_features, split="train")
         ep_dataset = embed_images(ep_dataset)
-        ep_size_in_mb = get_hf_dataset_size_in_mb(ep_dataset)
-        ep_num_frames = len(ep_dataset)
-        df = pd.DataFrame(ep_dataset)
-
-        if self.meta.episodes is None:
-            # Initialize indices and frame count for a new dataset made of the first episode data
-            chunk_idx, file_idx = 0, 0
-            latest_num_frames = 0
-        else:
-            # Retrieve information from the latest parquet file
-            latest_ep = self.meta.episodes[-1]
-            chunk_idx = latest_ep["data/chunk_index"]
-            file_idx = latest_ep["data/file_index"]
-
-            latest_path = self.root / self.meta.data_path.format(chunk_index=chunk_idx, file_index=file_idx)
-            latest_size_in_mb = get_parquet_file_size_in_mb(latest_path)
-            latest_num_frames = get_parquet_num_frames(latest_path)
-
-            # Determine if a new parquet file is needed
-            if latest_size_in_mb + ep_size_in_mb >= self.meta.data_files_size_in_mb:
-                # Size limit is reached, prepare new parquet file
-                chunk_idx, file_idx = update_chunk_file_indices(chunk_idx, file_idx, self.meta.chunks_size)
-                latest_num_frames = 0
-            else:
-                # Update the existing parquet file with new rows
-                latest_df = pd.read_parquet(latest_path)
-                df = pd.concat([latest_df, df], ignore_index=True)
-
-        # Write the resulting dataframe from RAM to disk
-        path = self.root / self.meta.data_path.format(chunk_index=chunk_idx, file_index=file_idx)
-        path.parent.mkdir(parents=True, exist_ok=True)
-        safe_write_dataframe_to_parquet(df, path, self.meta.image_keys)
-
-        # Update the Hugging Face dataset by reloading it.
-        # This process should be fast because only the latest Parquet file has been modified.
-        # Therefore, only this file needs to be converted to PyArrow; the rest is loaded from the PyArrow memory-mapped cache.
-        self.hf_dataset = self.load_hf_dataset()
-
-        metadata = {
-            "data/chunk_index": chunk_idx,
-            "data/file_index": file_idx,
-            "dataset_from_index": latest_num_frames,
-            "dataset_to_index": latest_num_frames + ep_num_frames,
-        }
-        return metadata
-
-    def _save_episode_video(self, video_key: str, episode_index: int):
-        # Encode episode frames into a temporary video
-        ep_path = self._encode_temporary_episode_video(video_key, episode_index)
-        ep_size_in_mb = get_video_size_in_mb(ep_path)
-        ep_duration_in_s = get_video_duration_in_s(ep_path)
-
-        if self.meta.episodes is None:
-            # Initialize indices for a new dataset made of the first episode data
-            chunk_idx, file_idx = 0, 0
-            latest_duration_in_s = 0
-            new_path = self.root / self.meta.video_path.format(
-                video_key=video_key, chunk_index=chunk_idx, file_index=file_idx
-            )
-            new_path.parent.mkdir(parents=True, exist_ok=True)
-            shutil.move(str(ep_path), str(new_path))
-        else:
-            # Retrieve information from the latest video file
-            latest_ep = self.meta.episodes[-1]
-            chunk_idx = latest_ep[f"videos/{video_key}/chunk_index"]
-            file_idx = latest_ep[f"videos/{video_key}/file_index"]
-
-            latest_path = self.root / self.meta.video_path.format(
-                video_key=video_key, chunk_index=chunk_idx, file_index=file_idx
-            )
-            latest_size_in_mb = get_video_size_in_mb(latest_path)
-            latest_duration_in_s = get_video_duration_in_s(latest_path)
-
-            if latest_size_in_mb + ep_size_in_mb >= self.meta.video_files_size_in_mb:
-                # Move temporary episode video to a new video file in the dataset
-                chunk_idx, file_idx = update_chunk_file_indices(chunk_idx, file_idx, self.meta.chunks_size)
-                new_path = self.root / self.meta.video_path.format(
-                    video_key=video_key, chunk_index=chunk_idx, file_index=file_idx
-                )
-                new_path.parent.mkdir(parents=True, exist_ok=True)
-                shutil.move(str(ep_path), str(new_path))
-            else:
-                # Update latest video file
-                concat_video_files([latest_path, ep_path], self.root, video_key, chunk_idx, file_idx)
-
-        # Remove temporary directory
-        shutil.rmtree(str(ep_path.parent))
-
-        metadata = {
-            "episode_index": episode_index,
-            f"videos/{video_key}/chunk_index": chunk_idx,
-            f"videos/{video_key}/file_index": file_idx,
-            f"videos/{video_key}/from_timestamp": latest_duration_in_s,
-            f"videos/{video_key}/to_timestamp": latest_duration_in_s + ep_duration_in_s,
-        }
-        return metadata
+        self.hf_dataset = concatenate_datasets([self.hf_dataset, ep_dataset])
+        self.hf_dataset.set_transform(hf_transform_to_torch)
+        ep_data_path = self.root / self.meta.get_data_file_path(ep_index=episode_index)
+        ep_data_path.parent.mkdir(parents=True, exist_ok=True)
+        ep_dataset.to_parquet(ep_data_path)
 
     def clear_episode_buffer(self) -> None:
         episode_index = self.episode_buffer["episode_index"]
@@ -1117,16 +960,34 @@ class LeRobotDataset(torch.utils.data.Dataset):
         if self.image_writer is not None:
             self.image_writer.wait_until_done()
 
-    def _encode_temporary_episode_video(self, video_key: str, episode_index: int) -> dict:
+    def encode_videos(self) -> None:
         """
         Use ffmpeg to convert frames stored as png into mp4 videos.
         Note: `encode_video_frames` is a blocking call. Making it asynchronous shouldn't speedup encoding,
         since video encoding with ffmpeg is already using multithreading.
         """
-        temp_path = Path(tempfile.mkdtemp(dir=self.root)) / f"{video_key}_{episode_index:03d}.mp4"
-        img_dir = self._get_image_file_dir(episode_index, video_key)
-        encode_video_frames(img_dir, temp_path, self.fps, overwrite=True)
-        return temp_path
+        for ep_idx in range(self.meta.total_episodes):
+            self.encode_episode_videos(ep_idx)
+
+    def encode_episode_videos(self, episode_index: int) -> dict:
+        """
+        Use ffmpeg to convert frames stored as png into mp4 videos.
+        Note: `encode_video_frames` is a blocking call. Making it asynchronous shouldn't speedup encoding,
+        since video encoding with ffmpeg is already using multithreading.
+        """
+        video_paths = {}
+        for key in self.meta.video_keys:
+            video_path = self.root / self.meta.get_video_file_path(episode_index, key)
+            video_paths[key] = str(video_path)
+            if video_path.is_file():
+                # Skip if video is already encoded. Could be the case when resuming data recording.
+                continue
+            img_dir = self._get_image_file_path(
+                episode_index=episode_index, image_key=key, frame_index=0
+            ).parent
+            encode_video_frames(img_dir, video_path, self.fps, overwrite=True)
+
+        return video_paths
 
     @classmethod
     def create(
@@ -1171,6 +1032,7 @@ class LeRobotDataset(torch.utils.data.Dataset):
         obj.image_transforms = None
         obj.delta_timestamps = None
         obj.delta_indices = None
+        obj.episode_data_index = None
         obj.video_backend = video_backend if video_backend is not None else get_safe_default_codec()
         return obj
 
@@ -1196,7 +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]):
@@ -337,11 +355,13 @@ def compute_sampler_weights(
     if len(offline_dataset) > 0:
         offline_data_mask_indices = []
         for start_index, end_index in zip(
-            offline_dataset.meta.episodes["dataset_from_index"],
-            offline_dataset.meta.episodes["dataset_to_index"],
+            offline_dataset.episode_data_index["from"],
+            offline_dataset.episode_data_index["to"],
             strict=True,
         ):
-            offline_data_mask_indices.extend(range(start_index, end_index - offline_drop_n_last_frames))
+            offline_data_mask_indices.extend(
+                range(start_index.item(), end_index.item() - offline_drop_n_last_frames)
+            )
         offline_data_mask = torch.zeros(len(offline_dataset), dtype=torch.bool)
         offline_data_mask[torch.tensor(offline_data_mask_indices)] = True
         weights.append(

lerobot/common/datasets/profile_streaming_dataset.py

@@ -1,295 +0,0 @@
-#!/usr/bin/env python
-"""
-Script to profile the StreamingLeRobotDataset iteration speed.
-Run with: python -m lerobot.common.datasets.profile_streaming_dataset
-"""
-
-import argparse
-import time
-
-import numpy as np
-from line_profiler import LineProfiler
-from tqdm import tqdm
-
-from lerobot.common.datasets.streaming_dataset import StreamingLeRobotDataset
-
-
-def timing_stats(times):
-    return {
-        "mean": np.mean(times),
-        "std": np.std(times),
-        "min": np.min(times),
-        "max": np.max(times),
-        "median": np.median(times),
-    }
-
-
-def measure_iteration_times(dataset, num_samples=100, num_runs=5, warmup_iters=2):
-    """
-    Measure individual iteration times and compute statistics.
-
-    Args:
-        dataset: The dataset to iterate over.
-        num_samples (int): Number of samples to iterate through per run.
-        num_runs (int): Number of timing runs to perform.
-        warmup_iters (int): Number of warmup iterations before timing.
-
-    Returns:
-        dict: Statistics including mean, std, min, max times per sample.
-    """
-    print(f"Measuring iteration times over {num_runs} runs of {num_samples} samples each...")
-    print(f"Using {warmup_iters} warmup iterations per run")
-
-    all_sample_times = []
-    run_times = []
-
-    for run in range(num_runs):
-        print(f"Run {run + 1}/{num_runs}...")
-        run_start = time.time()
-
-        # Warmup phase
-        print(f"  Performing {warmup_iters} warmup iterations...")
-        warmup_iterator = iter(dataset)
-        warmup_times = []
-        try:
-            for _ in range(warmup_iters):
-                start_time = time.time()
-                next(warmup_iterator)
-                end_time = time.time()
-                elapsed_ms = (end_time - start_time) * 1000
-                warmup_times.append(elapsed_ms)
-        except StopIteration:
-            print("  Warning: Iterator exhausted during warmup")
-
-        # timing phase
-        iterator = iter(dataset)
-        sample_times = []
-
-        for _ in tqdm(range(num_samples)):
-            start_time = time.time()
-            next(iterator)
-            end_time = time.time()
-            elapsed_ms = (end_time - start_time) * 1000
-            sample_times.append(elapsed_ms)
-
-        run_end = time.time()
-        run_times.append(run_end - run_start)
-        all_sample_times.extend(sample_times)
-
-    # Compute statistics
-    sample_times_array = np.array(all_sample_times)
-    warmup_times_array = np.array(warmup_times)
-    run_times_array = np.array(run_times)
-
-    stats = {
-        "sample_times_ms": timing_stats(sample_times_array),
-        "warmup_times_ms": timing_stats(warmup_times_array),
-        "run_times_s": timing_stats(run_times_array),
-        "samples_per_second": num_samples / np.mean(run_times_array),
-    }
-
-    return stats
-
-
-def print_timing_stats(stats):
-    """Print timing statistics in a readable format."""
-    print("\n" + "=" * 60)
-    print("TIMING STATISTICS")
-    print("=" * 60)
-
-    warmup_stats = stats["warmup_times_ms"]
-    print("Warmup timing (ms):")
-    print(f"  Mean: {warmup_stats['mean']:.2f} ± {warmup_stats['std']:.2f}")
-    print(f"  Median: {warmup_stats['median']:.2f}")
-    print(f"  Range: [{warmup_stats['min']:.2f}, {warmup_stats['max']:.2f}]")
-
-    sample_stats = stats["sample_times_ms"]
-    print("\nPer-sample timing (ms):")
-    print(f"  Mean: {sample_stats['mean']:.2f} ± {sample_stats['std']:.2f}")
-    print(f"  Median: {sample_stats['median']:.2f}")
-    print(f"  Range: [{sample_stats['min']:.2f}, {sample_stats['max']:.2f}]")
-
-    run_stats = stats["run_times_s"]
-    print("\nPer-run timing (seconds):")
-    print(f"  Mean: {run_stats['mean']:.2f} ± {run_stats['std']:.2f}")
-    print(f"  Range: [{run_stats['min']:.2f}, {run_stats['max']:.2f}]")
-
-    print("\nThroughput:")
-    print(f"  Samples/second: {stats['samples_per_second']:.2f}")
-    print("=" * 60)
-
-
-def _time_iterations(dataset, num_samples, num_runs, warmup_iters, stats_file_path):
-    # Measure iteration times with statistics
-    timing_stats = measure_iteration_times(dataset, num_samples, num_runs, warmup_iters)
-    print_timing_stats(timing_stats)
-
-    # Save results to a file
-    with open(stats_file_path, "w") as f:
-        f.write("TIMING STATISTICS\n")
-        f.write("=" * 60 + "\n")
-        warmup_stats = timing_stats["warmup_times_ms"]
-        f.write("Warmup timing (ms):\n")
-        f.write(f"  Mean: {warmup_stats['mean']:.2f} ± {warmup_stats['std']:.2f}\n")
-        f.write(f"  Median: {warmup_stats['median']:.2f}\n")
-        f.write(f"  Range: [{warmup_stats['min']:.2f}, {warmup_stats['max']:.2f}]\n\n")
-
-        sample_stats = timing_stats["sample_times_ms"]
-        f.write("Per-sample timing (ms):\n")
-        f.write(f"  Mean: {sample_stats['mean']:.2f} ± {sample_stats['std']:.2f}\n")
-        f.write(f"  Median: {sample_stats['median']:.2f}\n")
-        f.write(f"  Range: [{sample_stats['min']:.2f}, {sample_stats['max']:.2f}]\n\n")
-
-        run_stats = timing_stats["run_times_s"]
-        f.write("Per-run timing (seconds):\n")
-        f.write(f"  Mean: {run_stats['mean']:.2f} ± {run_stats['std']:.2f}\n")
-        f.write(f"  Range: [{run_stats['min']:.2f}, {run_stats['max']:.2f}]\n\n")
-
-        throughput_stats = timing_stats["samples_per_second"]
-        f.write("Throughput:\n")
-        f.write(f"  Samples/second: {throughput_stats:.2f}\n")
-        f.write("=" * 60 + "\n\n")
-
-        f.write("DETAILED LINE PROFILING RESULTS\n")
-        f.write("=" * 60 + "\n")
-
-    print(f"\nDetailed profiling results saved to {stats_file_path}")
-
-
-def _profile_iteration(dataset, num_samples, stats_file_path):
-    # Create a line profiler instance for detailed profiling
-    profiler = LineProfiler()
-
-    # Add functions to profile
-    profiler.add_function(dataset.__iter__)
-    profiler.add_function(dataset.make_frame)
-    profiler.add_function(dataset._make_iterable_dataset)
-
-    # Profile the iteration
-
-    # Define the function to profile
-    def iterate_dataset(ds, n):
-        # Iterating without warmup for line profiling
-        iterator = iter(ds)
-        start_time = time.time()
-        for _ in range(n):
-            next(iterator)
-        end_time = time.time()
-        return end_time - start_time
-
-    # Add the function to the profiler
-    profiler.add_function(iterate_dataset)
-
-    # Run the profiled function
-    profiler.runcall(iterate_dataset, dataset, num_samples)
-
-    with open(stats_file_path, "a") as f:
-        profiler.print_stats(stream=f)
-
-
-def _analyze_randomness(dataset, num_samples, stats_file_path):
-    """
-    Analyze the randomness of dataset iteration by checking correlation between
-    iteration index and frame index.
-
-    Args:
-        dataset: The dataset to analyze.
-        num_samples: Number of samples to use for analysis.
-        stats_file_path: Path to save the analysis results.
-    """
-    print("\nAnalyzing randomness of dataset iteration...")
-
-    # Collect iteration index and frame index pairs
-    points = []
-    iterator = iter(dataset)
-
-    for i in tqdm(range(num_samples)):
-        try:
-            frame = next(iterator)
-            points.append(np.array([i, frame["frame_index"]]))
-        except (StopIteration, KeyError) as e:
-            if isinstance(e, StopIteration):
-                print(f"  Warning: Iterator exhausted after {i} samples")
-            else:
-                print(f"  Warning: frame_index not found in sample {i}")
-            break
-
-    # Compute correlation between iteration index and frame index
-    points_array = np.array(points)
-    correlation = np.corrcoef(points_array[:, 0], points_array[:, 1])[0, 1]
-
-    # Save results to file
-    with open(stats_file_path, "a") as f:
-        f.write("\nRANDOMNESS ANALYSIS\n")
-        f.write("=" * 60 + "\n")
-        f.write(f"Correlation between iteration index and frame index: {correlation:.4f}\n")
-        f.write("(Correlation close to 0 indicates more random access pattern)\n")
-        f.write("(Correlation close to 1 indicates sequential access pattern)\n")
-        f.write("=" * 60 + "\n")
-
-    print(f"Correlation between iteration index and frame index: {correlation:.4f}")
-    print("(Correlation close to 0 indicates more random access pattern)")
-    print("(Correlation close to 1 indicates sequential access pattern)")
-
-
-def profile_dataset(
-    repo_id, num_samples=100, buffer_size=1000, max_num_shards=16, seed=42, num_runs=3, warmup_iters=10
-):
-    """
-    Profile the streaming dataset iteration speed.
-
-    Args:
-        repo_id (str): HuggingFace repository ID for the dataset.
-        num_samples (int): Number of samples to iterate through.
-        buffer_size (int): Buffer size for the dataset.
-        max_num_shards (int): Number of shards to use.
-        seed (int): Random seed for reproducibility.
-        num_runs (int): Number of timing runs to perform.
-        warmup_iters (int): Number of warmup iterations before timing.
-    """
-    stats_file_path = "streaming_dataset_profile.txt"
-
-    print(f"Creating dataset from {repo_id} with buffer_size={buffer_size}, max_num_shards={max_num_shards}")
-    dataset = StreamingLeRobotDataset(
-        repo_id=repo_id, buffer_size=buffer_size, max_num_shards=max_num_shards, seed=seed
-    )
-
-    _time_iterations(dataset, num_samples, num_runs, warmup_iters, stats_file_path)
-    _profile_iteration(dataset, num_samples, stats_file_path)
-    _analyze_randomness(dataset, num_samples, stats_file_path)
-
-
-def main():
-    parser = argparse.ArgumentParser(description="Profile StreamingLeRobotDataset iteration speed")
-    parser.add_argument(
-        "--repo-id",
-        type=str,
-        default="lerobot/aloha_mobile_cabinet",
-        help="HuggingFace repository ID for the dataset",
-    )
-    parser.add_argument("--num-samples", type=int, default=2_000, help="Number of samples to iterate through")
-    parser.add_argument("--buffer-size", type=int, default=1000, help="Buffer size for the dataset")
-    parser.add_argument("--max-num-shards", type=int, default=1, help="Number of shards to use")
-    parser.add_argument("--seed", type=int, default=42, help="Random seed for reproducibility")
-    parser.add_argument(
-        "--num-runs", type=int, default=10, help="Number of timing runs to perform for statistics"
-    )
-    parser.add_argument(
-        "--warmup-iters", type=int, default=1, help="Number of warmup iterations before timing"
-    )
-
-    args = parser.parse_args()
-
-    profile_dataset(
-        repo_id=args.repo_id,
-        num_samples=args.num_samples,
-        buffer_size=args.buffer_size,
-        max_num_shards=args.max_num_shards,
-        seed=args.seed,
-        num_runs=args.num_runs,
-        warmup_iters=args.warmup_iters,
-    )
-
-
-if __name__ == "__main__":
-    main()

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

@@ -21,8 +21,7 @@ import torch
 class EpisodeAwareSampler:
     def __init__(
         self,
-        dataset_from_indices: list[int],
-        dataset_to_indices: list[int],
+        episode_data_index: dict,
         episode_indices_to_use: Union[list, None] = None,
         drop_n_first_frames: int = 0,
         drop_n_last_frames: int = 0,
@@ -31,8 +30,7 @@ class EpisodeAwareSampler:
         """Sampler that optionally incorporates episode boundary information.
 
         Args:
-            dataset_from_indices: List of indices containing the start of each episode in the dataset.
-            dataset_to_indices: List of indices containing the end of each episode in the dataset.
+            episode_data_index: Dictionary with keys 'from' and 'to' containing the start and end indices of each episode.
             episode_indices_to_use: List of episode indices to use. If None, all episodes are used.
                                     Assumes that episodes are indexed from 0 to N-1.
             drop_n_first_frames: Number of frames to drop from the start of each episode.
@@ -41,10 +39,15 @@ class EpisodeAwareSampler:
         """
         indices = []
         for episode_idx, (start_index, end_index) in enumerate(
-            zip(dataset_from_indices, dataset_to_indices, strict=True)
+            zip(episode_data_index["from"], episode_data_index["to"], strict=True)
         ):
             if episode_indices_to_use is None or episode_idx in episode_indices_to_use:
-                indices.extend(range(start_index + drop_n_first_frames, end_index - drop_n_last_frames))
+                indices.extend(
+                    range(
+                        start_index.item() + drop_n_first_frames,
+                        end_index.item() - drop_n_last_frames,
+                    )
+                )
 
         self.indices = indices
         self.shuffle = shuffle

lerobot/common/datasets/streaming_dataset.py

@@ -1,252 +0,0 @@
-import random
-from pathlib import Path
-from typing import Callable, Dict, Generator, Iterator
-
-import datasets
-import numpy as np
-import torch
-from datasets import load_dataset
-from line_profiler import profile
-
-from lerobot.common.constants import HF_LEROBOT_HOME
-from lerobot.common.datasets.lerobot_dataset import CODEBASE_VERSION, LeRobotDatasetMetadata
-from lerobot.common.datasets.utils import (
-    check_version_compatibility,
-    item_to_torch,
-)
-from lerobot.common.datasets.video_utils import (
-    VideoDecoderCache,
-    decode_video_frames_torchcodec,
-    get_safe_default_codec,
-)
-
-
-class StreamingLeRobotDataset(torch.utils.data.IterableDataset):
-    """LeRobotDataset with streaming capabilities.
-
-    This class extends LeRobotDataset to add streaming functionality, allowing data to be streamed
-    rather than loaded entirely into memory. This is especially useful for large datasets that may
-    not fit in memory or when you want to quickly explore a dataset without downloading it completely.
-
-    The key innovation is using a Backtrackable iterator that maintains a bounded buffer of recent
-    items, allowing us to access previous frames for delta timestamps without loading the entire
-    dataset into memory.
-
-    Example:
-        Basic usage:
-        ```python
-        from lerobot.common.datasets.streaming_dataset import StreamingLeRobotDataset
-
-        # Create a streaming dataset with delta timestamps
-        delta_timestamps = {
-            "observation.image": [-1.0, -0.5, 0.0],  # 1 sec ago, 0.5 sec ago, current
-            "action": [0.0, 0.1, 0.2],  # current, 0.1 sec future, 0.2 sec future
-        }
-
-        dataset = StreamingLeRobotDataset(
-            repo_id="your-dataset-repo-id",
-            delta_timestamps=delta_timestamps,
-            streaming=True,
-            buffer_size=1000,
-        )
-
-        # Iterate over the dataset
-        for i, item in enumerate(dataset):
-            print(f"Sample {i}: Episode {item['episode_index']} Frame {item['frame_index']}")
-            # item will contain stacked frames according to delta_timestamps
-            if i >= 10:
-                break
-        ```
-    """
-
-    def __init__(
-        self,
-        repo_id: str,
-        root: str | Path | None = None,
-        episodes: list[int] | None = None,
-        image_transforms: Callable | None = None,
-        tolerance_s: float = 1e-4,
-        revision: str | None = None,
-        force_cache_sync: bool = False,
-        video_backend: str | None = "torchcodec",
-        streaming: bool = True,
-        buffer_size: int = 1000,
-        max_num_shards: int = 16,
-        seed: int = 42,
-        rng: np.random.Generator | None = None,
-    ):
-        """Initialize a StreamingLeRobotDataset.
-
-        Args:
-            repo_id (str): This is the repo id that will be used to fetch the dataset.
-            root (Path | None, optional): Local directory to use for downloading/writing files.
-            episodes (list[int] | None, optional): If specified, this will only load episodes specified by
-                their episode_index in this list.
-            image_transforms (Callable | None, optional): Transform to apply to image data.
-            tolerance_s (float, optional): Tolerance in seconds for timestamp matching.
-            revision (str, optional): Git revision id (branch name, tag, or commit hash).
-            force_cache_sync (bool, optional): Flag to sync and refresh local files first.
-            video_backend (str | None, optional): Video backend to use for decoding videos. Uses "torchcodec" by default.
-            streaming (bool, optional): Whether to stream the dataset or load it all. Defaults to True.
-            buffer_size (int, optional): Buffer size for shuffling when streaming. Defaults to 1000.
-            max_num_shards (int, optional): Number of shards to re-shard the input dataset into. Defaults to 16.
-            seed (int, optional): Reproducibility random seed.
-            rng (np.random.Generator | None, optional): Random number generator.
-        """
-        super().__init__()
-        self.repo_id = repo_id
-        self.root = Path(root) if root else HF_LEROBOT_HOME / repo_id
-        self.image_transforms = image_transforms
-        self.episodes = episodes
-        self.tolerance_s = tolerance_s
-        self.revision = revision if revision else CODEBASE_VERSION
-        self.video_backend = video_backend if video_backend else get_safe_default_codec()
-        self.seed = seed
-        self.rng = rng if rng is not None else np.random.default_rng(seed)
-
-        self.streaming = streaming
-        self.buffer_size = buffer_size
-
-        # We cache the video decoders to avoid re-initializing them at each frame (avoiding a ~10x slowdown)
-        self.video_decoder_cache = VideoDecoderCache()
-
-        # Unused attributes
-        self.image_writer = None
-        self.episode_buffer = None
-
-        self.root.mkdir(exist_ok=True, parents=True)
-
-        # Load metadata
-        self.meta = LeRobotDatasetMetadata(
-            self.repo_id, self.root, self.revision, force_cache_sync=force_cache_sync
-        )
-        # Check version
-        check_version_compatibility(self.repo_id, self.meta._version, CODEBASE_VERSION)
-
-        self.hf_dataset = self.load_hf_dataset()
-        self.num_shards = min(self.hf_dataset.num_shards, max_num_shards)
-
-    @property
-    def fps(self):
-        return self.meta.fps
-
-    @staticmethod
-    def _iter_random_indices(
-        rng: np.random.Generator, buffer_size: int, random_batch_size=1000
-    ) -> Iterator[int]:
-        while True:
-            yield from (int(i) for i in rng.integers(0, buffer_size, size=random_batch_size))
-
-    @staticmethod
-    def _infinite_generator_over_elements(elements: list[int]) -> Iterator[int]:
-        return (random.choice(list(elements)) for _ in iter(int, 1))
-
-    def load_hf_dataset(self) -> datasets.IterableDataset:
-        dataset = load_dataset(self.repo_id, split="train", streaming=self.streaming)
-        self.streaming_from_local = False
-
-        # TODO(fracapuano): Add support for streaming from a local folder and not only from HF Hub
-        return dataset
-
-    def __iter__(self) -> Iterator[Dict[str, torch.Tensor]]:
-        buffer_indices_generator = self._iter_random_indices(self.rng, self.buffer_size)
-
-        # This buffer is populated while iterating on the dataset's shards
-        frames_buffer = []
-        idx_to_iterable_dataset = {
-            idx: self._make_iterable_dataset(self.hf_dataset.shard(self.num_shards, index=idx))
-            for idx in range(self.num_shards)
-        }
-
-        try:
-            while available_shards := list(idx_to_iterable_dataset.keys()):
-                shard_key = next(self._infinite_generator_over_elements(available_shards))
-                dataset = idx_to_iterable_dataset[shard_key]  # selects which shard to iterate on
-                for frame in self.make_frame(dataset):
-                    if len(frames_buffer) == self.buffer_size:
-                        i = next(buffer_indices_generator)
-                        yield frames_buffer[i]
-                        frames_buffer[i] = frame
-                    else:
-                        frames_buffer.append(frame)
-                    break  # random shard sampled, switch shard
-
-        except (
-            RuntimeError,
-            StopIteration,
-        ):  # NOTE: StopIteration inside a generator throws a RuntimeError since 3.7
-            # Remove exhausted shard
-            del idx_to_iterable_dataset[shard_key]
-
-        # Once shards are all exhausted, shuffle the buffer and yield the remaining frames
-        self.rng.shuffle(frames_buffer)
-        yield from frames_buffer
-
-    def _make_iterable_dataset(self, dataset: datasets.IterableDataset) -> Iterator:
-        return iter(dataset)
-
-    @profile
-    def make_frame(self, dataset_iterator: datasets.IterableDataset) -> Generator:
-        """Makes a frame starting from a dataset iterator"""
-        item = next(dataset_iterator)
-        item = item_to_torch(item)
-
-        # Get episode index from the item
-        ep_idx = item["episode_index"]
-
-        # Load video frames, when needed
-        if len(self.meta.video_keys) > 0:
-            current_ts = item["timestamp"]
-            query_timestamps = self._get_query_timestamps(current_ts, None)
-            video_frames = self._query_videos(query_timestamps, ep_idx)
-            item = {**video_frames, **item}
-
-        # Add task as a string
-        task_idx = item["task_index"]
-        item["task"] = self.meta.tasks.iloc[task_idx].name
-
-        yield item
-
-    def _get_query_timestamps(
-        self,
-        current_ts: float,
-        query_indices: dict[str, list[int]] | None = None,
-    ) -> dict[str, list[float]]:
-        query_timestamps = {}
-        for key in self.meta.video_keys:
-            if query_indices is not None and key in query_indices:
-                timestamps = self.hf_dataset.select(query_indices[key])["timestamp"]
-                query_timestamps[key] = torch.stack(timestamps).tolist()
-            else:
-                query_timestamps[key] = [current_ts]
-
-        return query_timestamps
-
-    def _query_videos(self, query_timestamps: dict[str, list[float]], ep_idx: int) -> dict:
-        """Note: When using data workers (e.g. DataLoader with num_workers>0), do not call this function
-        in the main process (e.g. by using a second Dataloader with num_workers=0). It will result in a
-        Segmentation Fault. This probably happens because a memory reference to the video loader is created in
-        the main process and a subprocess fails to access it.
-        """
-        item = {}
-        for vid_key, query_ts in query_timestamps.items():
-            root = self.meta.url_root if self.streaming and not self.streaming_from_local else self.root
-            video_path = f"{root}/{self.meta.get_video_file_path(ep_idx, vid_key)}"
-            frames = decode_video_frames_torchcodec(
-                video_path, query_ts, self.tolerance_s, decoder_cache=self.video_decoder_cache
-            )
-            item[vid_key] = frames.squeeze(0)
-
-        return item
-
-
-# Example usage
-if __name__ == "__main__":
-    repo_id = "lerobot/aloha_mobile_cabinet"
-    dataset = StreamingLeRobotDataset(repo_id)
-
-    for i, frame in enumerate(dataset):
-        print(frame)
-
-        if i > 10:  # only stream first 10 frames
-            break

lerobot/common/datasets/utils.py

@@ -17,23 +17,18 @@ import contextlib
 import importlib.resources
 import json
 import logging
-import shutil
-import subprocess
-import tempfile
 from collections.abc import Iterator
+from itertools import accumulate
 from pathlib import Path
 from pprint import pformat
 from types import SimpleNamespace
-from typing import Any, TypeVar
+from typing import Any
 
 import datasets
+import jsonlines
 import numpy as np
 import packaging.version
-import pandas
-import pandas as pd
-import pyarrow.parquet as pq
 import torch
-from datasets import Dataset, concatenate_datasets
 from datasets.table import embed_table_storage
 from huggingface_hub import DatasetCard, DatasetCardData, HfApi
 from huggingface_hub.errors import RevisionNotFoundError
@@ -47,25 +42,19 @@ from lerobot.common.datasets.backward_compatibility import (
 )
 from lerobot.common.robot_devices.robots.utils import Robot
 from lerobot.common.utils.utils import is_valid_numpy_dtype_string
-from lerobot.configs.types import FeatureType, PolicyFeature
+from lerobot.configs.types import DictLike, FeatureType, PolicyFeature
 
-DEFAULT_CHUNK_SIZE = 1000  # Max number of files per chunk
-DEFAULT_DATA_FILE_SIZE_IN_MB = 100  # Max size per file
-DEFAULT_VIDEO_FILE_SIZE_IN_MB = 500  # Max size per file
+DEFAULT_CHUNK_SIZE = 1000  # Max number of episodes per chunk
 
 INFO_PATH = "meta/info.json"
+EPISODES_PATH = "meta/episodes.jsonl"
 STATS_PATH = "meta/stats.json"
+EPISODES_STATS_PATH = "meta/episodes_stats.jsonl"
+TASKS_PATH = "meta/tasks.jsonl"
 
-EPISODES_DIR = "meta/episodes"
-DATA_DIR = "data"
-VIDEO_DIR = "videos"
-
-CHUNK_FILE_PATTERN = "chunk-{chunk_index:03d}/file-{file_index:03d}"
-DEFAULT_TASKS_PATH = "meta/tasks.parquet"
-DEFAULT_EPISODES_PATH = EPISODES_DIR + "/" + CHUNK_FILE_PATTERN + ".parquet"
-DEFAULT_DATA_PATH = DATA_DIR + "/" + CHUNK_FILE_PATTERN + ".parquet"
-DEFAULT_VIDEO_PATH = VIDEO_DIR + "/{video_key}/" + CHUNK_FILE_PATTERN + ".mp4"
-DEFAULT_IMAGE_PATH = "images/{image_key}/episode-{episode_index:06d}/frame-{frame_index:06d}.png"
+DEFAULT_VIDEO_PATH = "videos/chunk-{episode_chunk:03d}/{video_key}/episode_{episode_index:06d}.mp4"
+DEFAULT_PARQUET_PATH = "data/chunk-{episode_chunk:03d}/episode_{episode_index:06d}.parquet"
+DEFAULT_IMAGE_PATH = "images/{image_key}/episode_{episode_index:06d}/frame_{frame_index:06d}.png"
 
 DATASET_CARD_TEMPLATE = """
 ---
@@ -85,117 +74,6 @@ DEFAULT_FEATURES = {
     "task_index": {"dtype": "int64", "shape": (1,), "names": None},
 }
 
-T = TypeVar("T")
-
-
-def get_parquet_file_size_in_mb(parquet_path):
-    metadata = pq.read_metadata(parquet_path)
-    total_uncompressed_size = 0
-    for row_group in range(metadata.num_row_groups):
-        rg_metadata = metadata.row_group(row_group)
-        for column in range(rg_metadata.num_columns):
-            col_metadata = rg_metadata.column(column)
-            total_uncompressed_size += col_metadata.total_uncompressed_size
-    return total_uncompressed_size / (1024**2)
-
-
-def get_hf_dataset_size_in_mb(hf_ds: Dataset) -> int:
-    return hf_ds.data.nbytes / (1024**2)
-
-
-def get_pd_dataframe_size_in_mb(df: pandas.DataFrame) -> int:
-    # TODO(rcadene): unused?
-    memory_usage_bytes = df.memory_usage(deep=True).sum()
-    return memory_usage_bytes / (1024**2)
-
-
-def update_chunk_file_indices(chunk_idx: int, file_idx: int, chunks_size: int):
-    if file_idx == chunks_size - 1:
-        file_idx = 0
-        chunk_idx += 1
-    else:
-        file_idx += 1
-    return chunk_idx, file_idx
-
-
-def load_nested_dataset(pq_dir: Path) -> Dataset:
-    """Find parquet files in provided directory {pq_dir}/chunk-xxx/file-xxx.parquet
-    Convert parquet files to pyarrow memory mapped in a cache folder for efficient RAM usage
-    Concatenate all pyarrow references to return HF Dataset format
-    """
-    paths = sorted(pq_dir.glob("*/*.parquet"))
-    if len(paths) == 0:
-        raise FileNotFoundError(f"Provided directory does not contain any parquet file: {pq_dir}")
-
-    # TODO(rcadene): set num_proc to accelerate conversion to pyarrow
-    datasets = [Dataset.from_parquet(str(path)) for path in paths]
-    return concatenate_datasets(datasets)
-
-
-def get_parquet_num_frames(parquet_path):
-    metadata = pq.read_metadata(parquet_path)
-    return metadata.num_rows
-
-
-def get_video_size_in_mb(mp4_path: Path):
-    file_size_bytes = mp4_path.stat().st_size
-    file_size_mb = file_size_bytes / (1024**2)
-    return file_size_mb
-
-
-def concat_video_files(paths_to_cat: list[Path], root: Path, video_key: str, chunk_idx: int, file_idx: int):
-    # TODO(rcadene): move to video_utils.py
-    # TODO(rcadene): add docstring
-    tmp_dir = Path(tempfile.mkdtemp(dir=root))
-    # Create a text file with the list of files to concatenate
-    path_concat_video_files = tmp_dir / "concat_video_files.txt"
-    with open(path_concat_video_files, "w") as f:
-        for ep_path in paths_to_cat:
-            f.write(f"file '{str(ep_path)}'\n")
-
-    path_tmp_output = tmp_dir / "tmp_output.mp4"
-    command = [
-        "ffmpeg",
-        "-y",
-        "-f",
-        "concat",
-        "-safe",
-        "0",
-        "-i",
-        str(path_concat_video_files),
-        "-c",
-        "copy",
-        str(path_tmp_output),
-    ]
-    subprocess.run(command, check=True)
-
-    output_path = root / DEFAULT_VIDEO_PATH.format(
-        video_key=video_key, chunk_index=chunk_idx, file_index=file_idx
-    )
-    output_path.parent.mkdir(parents=True, exist_ok=True)
-    shutil.move(str(path_tmp_output), str(output_path))
-    shutil.rmtree(str(tmp_dir))
-
-
-def get_video_duration_in_s(mp4_file: Path):
-    # TODO(rcadene): move to video_utils.py
-    command = [
-        "ffprobe",
-        "-v",
-        "error",
-        "-show_entries",
-        "format=duration",
-        "-of",
-        "default=noprint_wrappers=1:nokey=1",
-        str(mp4_file),
-    ]
-    result = subprocess.run(
-        command,
-        stdout=subprocess.PIPE,
-        stderr=subprocess.STDOUT,
-    )
-    return float(result.stdout)
-
 
 def flatten_dict(d: dict, parent_key: str = "", sep: str = "/") -> dict:
     """Flatten a nested dictionary structure by collapsing nested keys into one key with a separator.
@@ -229,13 +107,23 @@ def unflatten_dict(d: dict, sep: str = "/") -> dict:
     return outdict
 
 
+def get_nested_item(obj: DictLike, flattened_key: str, sep: str = "/") -> Any:
+    split_keys = flattened_key.split(sep)
+    getter = obj[split_keys[0]]
+    if len(split_keys) == 1:
+        return getter
+
+    for key in split_keys[1:]:
+        getter = getter[key]
+
+    return getter
+
+
 def serialize_dict(stats: dict[str, torch.Tensor | np.ndarray | dict]) -> dict:
     serialized_dict = {}
     for key, value in flatten_dict(stats).items():
         if isinstance(value, (torch.Tensor, np.ndarray)):
             serialized_dict[key] = value.tolist()
-        elif isinstance(value, list) and isinstance(value[0], (int, float, list)):
-            serialized_dict[key] = value
         elif isinstance(value, np.generic):
             serialized_dict[key] = value.item()
         elif isinstance(value, (int, float)):
@@ -265,6 +153,23 @@ def write_json(data: dict, fpath: Path) -> None:
         json.dump(data, f, indent=4, ensure_ascii=False)
 
 
+def load_jsonlines(fpath: Path) -> list[Any]:
+    with jsonlines.open(fpath, "r") as reader:
+        return list(reader)
+
+
+def write_jsonlines(data: dict, fpath: Path) -> None:
+    fpath.parent.mkdir(exist_ok=True, parents=True)
+    with jsonlines.open(fpath, "w") as writer:
+        writer.write_all(data)
+
+
+def append_jsonlines(data: dict, fpath: Path) -> None:
+    fpath.parent.mkdir(exist_ok=True, parents=True)
+    with jsonlines.open(fpath, "a") as writer:
+        writer.write(data)
+
+
 def write_info(info: dict, local_dir: Path):
     write_json(info, local_dir / INFO_PATH)
 
@@ -293,48 +198,52 @@ def load_stats(local_dir: Path) -> dict[str, dict[str, np.ndarray]]:
     return cast_stats_to_numpy(stats)
 
 
-def write_hf_dataset(hf_dataset: Dataset, local_dir: Path):
-    if get_hf_dataset_size_in_mb(hf_dataset) > DEFAULT_DATA_FILE_SIZE_IN_MB:
-        raise NotImplementedError("Contact a maintainer.")
-
-    path = local_dir / DEFAULT_DATA_PATH.format(chunk_index=0, file_index=0)
-    path.parent.mkdir(parents=True, exist_ok=True)
-    hf_dataset.to_parquet(path)
+def write_task(task_index: int, task: dict, local_dir: Path):
+    task_dict = {
+        "task_index": task_index,
+        "task": task,
+    }
+    append_jsonlines(task_dict, local_dir / TASKS_PATH)
 
 
-def write_tasks(tasks: pandas.DataFrame, local_dir: Path):
-    path = local_dir / DEFAULT_TASKS_PATH
-    path.parent.mkdir(parents=True, exist_ok=True)
-    tasks.to_parquet(path)
+def load_tasks(local_dir: Path) -> tuple[dict, dict]:
+    tasks = load_jsonlines(local_dir / TASKS_PATH)
+    tasks = {item["task_index"]: item["task"] for item in sorted(tasks, key=lambda x: x["task_index"])}
+    task_to_task_index = {task: task_index for task_index, task in tasks.items()}
+    return tasks, task_to_task_index
 
 
-def load_tasks(local_dir: Path):
-    tasks = pd.read_parquet(local_dir / DEFAULT_TASKS_PATH)
-    return tasks
+def write_episode(episode: dict, local_dir: Path):
+    append_jsonlines(episode, local_dir / EPISODES_PATH)
 
 
-def write_episodes(episodes: Dataset, local_dir: Path):
-    if get_hf_dataset_size_in_mb(episodes) > DEFAULT_DATA_FILE_SIZE_IN_MB:
-        raise NotImplementedError("Contact a maintainer.")
-
-    fpath = local_dir / DEFAULT_EPISODES_PATH.format(chunk_index=0, file_index=0)
-    fpath.parent.mkdir(parents=True, exist_ok=True)
-    episodes.to_parquet(fpath)
+def load_episodes(local_dir: Path) -> dict:
+    episodes = load_jsonlines(local_dir / EPISODES_PATH)
+    return {item["episode_index"]: item for item in sorted(episodes, key=lambda x: x["episode_index"])}
 
 
-def load_episodes(local_dir: Path) -> datasets.Dataset:
-    episodes = load_nested_dataset(local_dir / EPISODES_DIR)
-    # Select episode features/columns containing references to episode data and videos
-    # (e.g. tasks, dataset_from_index, dataset_to_index, data/chunk_index, data/file_index, etc.)
-    # This is to speedup access to these data, instead of having to load episode stats.
-    episodes = episodes.select_columns([key for key in episodes.features if not key.startswith("stats/")])
-    return episodes
+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),
+    }
+    append_jsonlines(episode_stats, local_dir / EPISODES_STATS_PATH)
+
+
+def load_episodes_stats(local_dir: Path) -> dict:
+    episodes_stats = load_jsonlines(local_dir / EPISODES_STATS_PATH)
+    return {
+        item["episode_index"]: cast_stats_to_numpy(item["stats"])
+        for item in sorted(episodes_stats, key=lambda x: x["episode_index"])
+    }
 
 
 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(
@@ -482,7 +391,6 @@ def get_hf_features_from_features(features: dict) -> datasets.Features:
 
 
 def get_features_from_robot(robot: Robot, use_videos: bool = True) -> dict:
-    # TODO(rcadene): add fps for each feature
     camera_ft = {}
     if robot.cameras:
         camera_ft = {
@@ -504,7 +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
@@ -536,17 +444,31 @@ def create_empty_dataset_info(
         "total_episodes": 0,
         "total_frames": 0,
         "total_tasks": 0,
+        "total_videos": 0,
+        "total_chunks": 0,
         "chunks_size": DEFAULT_CHUNK_SIZE,
-        "data_files_size_in_mb": DEFAULT_DATA_FILE_SIZE_IN_MB,
-        "video_files_size_in_mb": DEFAULT_VIDEO_FILE_SIZE_IN_MB,
         "fps": fps,
         "splits": {},
-        "data_path": DEFAULT_DATA_PATH,
+        "data_path": DEFAULT_PARQUET_PATH,
         "video_path": DEFAULT_VIDEO_PATH if use_videos else None,
         "features": features,
     }
 
 
+def get_episode_data_index(
+    episode_dicts: dict[dict], episodes: list[int] | None = None
+) -> dict[str, torch.Tensor]:
+    episode_lengths = {ep_idx: ep_dict["length"] for ep_idx, ep_dict in episode_dicts.items()}
+    if episodes is not None:
+        episode_lengths = {ep_idx: episode_lengths[ep_idx] for ep_idx in episodes}
+
+    cumulative_lengths = list(accumulate(episode_lengths.values()))
+    return {
+        "from": torch.LongTensor([0] + cumulative_lengths[:-1]),
+        "to": torch.LongTensor(cumulative_lengths),
+    }
+
+
 def check_timestamps_sync(
     timestamps: np.ndarray,
     episode_indices: np.ndarray,
@@ -621,7 +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
@@ -892,32 +817,3 @@ def validate_episode_buffer(episode_buffer: dict, total_episodes: int, features:
             f"In episode_buffer not in features: {buffer_keys - set(features)}"
             f"In features not in episode_buffer: {set(features) - buffer_keys}"
         )
-
-
-def safe_write_dataframe_to_parquet(df: pandas.DataFrame, path: Path, image_keys: list[str]):
-    if len(image_keys) > 0:
-        # TODO(qlhoest): replace this weird synthax by `df.to_parquet(path)` only
-        datasets.Dataset.from_dict(df.to_dict(orient="list")).to_parquet(path)
-    else:
-        df.to_parquet(path)
-
-
-def item_to_torch(item: dict) -> dict:
-    """Convert all items in a dictionary to PyTorch tensors where appropriate.
-
-    This function is used to convert an item from a streaming dataset to PyTorch tensors.
-
-    Args:
-        item (dict): Dictionary of items from a dataset.
-
-    Returns:
-        dict: Dictionary with all tensor-like items converted to torch.Tensor.
-    """
-    import numpy as np
-    import torch
-
-    for key, val in item.items():
-        if isinstance(val, (np.ndarray, list)) and key not in ["task"]:
-            # Convert numpy arrays and lists to torch tensors
-            item[key] = torch.tensor(val)
-    return item

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

@@ -121,12 +121,12 @@ from safetensors.torch import load_file
 
 from lerobot.common.datasets.utils import (
     DEFAULT_CHUNK_SIZE,
-    DEFAULT_DATA_PATH,
+    DEFAULT_PARQUET_PATH,
     DEFAULT_VIDEO_PATH,
+    EPISODES_PATH,
     INFO_PATH,
-    LEGACY_EPISODES_PATH,
-    LEGACY_TASKS_PATH,
     STATS_PATH,
+    TASKS_PATH,
     create_branch,
     create_lerobot_dataset_card,
     flatten_dict,
@@ -291,12 +291,14 @@ def split_parquet_by_episodes(
     for ep_chunk in range(total_chunks):
         ep_chunk_start = DEFAULT_CHUNK_SIZE * ep_chunk
         ep_chunk_end = min(DEFAULT_CHUNK_SIZE * (ep_chunk + 1), total_episodes)
-        chunk_dir = "/".join(DEFAULT_DATA_PATH.split("/")[:-1]).format(episode_chunk=ep_chunk)
+        chunk_dir = "/".join(DEFAULT_PARQUET_PATH.split("/")[:-1]).format(episode_chunk=ep_chunk)
         (output_dir / chunk_dir).mkdir(parents=True, exist_ok=True)
         for ep_idx in range(ep_chunk_start, ep_chunk_end):
             ep_table = table.filter(pc.equal(table["episode_index"], ep_idx))
             episode_lengths.insert(ep_idx, len(ep_table))
-            output_file = output_dir / DEFAULT_DATA_PATH.format(episode_chunk=ep_chunk, episode_index=ep_idx)
+            output_file = output_dir / DEFAULT_PARQUET_PATH.format(
+                episode_chunk=ep_chunk, episode_index=ep_idx
+            )
             pq.write_table(ep_table, output_file)
 
     return episode_lengths
@@ -377,7 +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)
@@ -400,7 +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,
     )
@@ -408,7 +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]
@@ -422,7 +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):
@@ -449,7 +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:
@@ -479,7 +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:
@@ -494,7 +523,7 @@ def convert_dataset(
 
     assert set(tasks) == {task for ep_tasks in tasks_by_episodes.values() for task in ep_tasks}
     tasks = [{"task_index": task_idx, "task": task} for task_idx, task in enumerate(tasks)]
-    write_jsonlines(tasks, v20_dir / LEGACY_TASKS_PATH)
+    write_jsonlines(tasks, v20_dir / TASKS_PATH)
     features["task_index"] = {
         "dtype": "int64",
         "shape": (1,),
@@ -507,12 +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:
@@ -541,10 +579,14 @@ 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 / LEGACY_EPISODES_PATH)
+    write_jsonlines(episodes, v20_dir / EPISODES_PATH)
 
     # Assemble metadata v2.0
     metadata_v2_0 = {
@@ -558,7 +600,7 @@ def convert_dataset(
         "chunks_size": DEFAULT_CHUNK_SIZE,
         "fps": metadata_v1["fps"],
         "splits": {"train": f"0:{total_episodes}"},
-        "data_path": DEFAULT_DATA_PATH,
+        "data_path": DEFAULT_PARQUET_PATH,
         "video_path": DEFAULT_VIDEO_PATH if video_keys else None,
         "features": features,
     }
@@ -570,7 +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 LEGACY_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"
@@ -61,8 +69,8 @@ def convert_dataset(
     with SuppressWarnings():
         dataset = LeRobotDataset(repo_id, revision=V20, force_cache_sync=True)
 
-    if (dataset.root / LEGACY_EPISODES_STATS_PATH).is_file():
-        (dataset.root / LEGACY_EPISODES_STATS_PATH).unlink()
+    if (dataset.root / EPISODES_STATS_PATH).is_file():
+        (dataset.root / EPISODES_STATS_PATH).unlink()
 
     convert_stats(dataset, num_workers=num_workers)
     ref_stats = load_stats(dataset.root)
@@ -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

@@ -17,9 +17,13 @@ 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 legacy_write_episode_stats
+from lerobot.common.datasets.utils import write_episode_stats
 
 
 def sample_episode_video_frames(dataset: LeRobotDataset, episode_index: int, ft_key: str) -> np.ndarray:
@@ -72,7 +76,7 @@ def convert_stats(dataset: LeRobotDataset, num_workers: int = 0):
             convert_episode_stats(dataset, ep_idx)
 
     for ep_idx in tqdm(range(total_episodes)):
-        legacy_write_episode_stats(ep_idx, dataset.meta.episodes_stats[ep_idx], dataset.root)
+        write_episode_stats(ep_idx, dataset.meta.episodes_stats[ep_idx], dataset.root)
 
 
 def check_aggregate_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/v30/convert_dataset_v21_to_v30.py

@@ -1,452 +0,0 @@
-"""
-This script will help you convert any LeRobot dataset already pushed to the hub from codebase version 2.1 to
-3.0. It will:
-
-- Generate per-episodes stats and writes them in `episodes_stats.jsonl`
-- Check consistency between these new stats and the old ones.
-- Remove the deprecated `stats.json`.
-- Update codebase_version in `info.json`.
-- Push this new version to the hub on the 'main' branch and tags it with "v2.1".
-
-Usage:
-
-```bash
-python lerobot/common/datasets/v30/convert_dataset_v21_to_v30.py \
-    --repo-id=lerobot/pusht
-```
-
-"""
-
-import argparse
-import shutil
-from pathlib import Path
-from typing import Any
-
-import jsonlines
-import pandas as pd
-import pyarrow as pa
-import tqdm
-from datasets import Dataset, Features, Image
-from huggingface_hub import HfApi, snapshot_download
-from requests import HTTPError
-
-from lerobot.common.constants import HF_LEROBOT_HOME
-from lerobot.common.datasets.compute_stats import aggregate_stats
-from lerobot.common.datasets.lerobot_dataset import CODEBASE_VERSION, LeRobotDataset
-from lerobot.common.datasets.utils import (
-    DEFAULT_CHUNK_SIZE,
-    DEFAULT_DATA_FILE_SIZE_IN_MB,
-    DEFAULT_DATA_PATH,
-    DEFAULT_VIDEO_FILE_SIZE_IN_MB,
-    DEFAULT_VIDEO_PATH,
-    cast_stats_to_numpy,
-    concat_video_files,
-    flatten_dict,
-    get_parquet_file_size_in_mb,
-    get_parquet_num_frames,
-    get_video_duration_in_s,
-    get_video_size_in_mb,
-    load_info,
-    update_chunk_file_indices,
-    write_episodes,
-    write_info,
-    write_stats,
-    write_tasks,
-)
-
-LEGACY_EPISODES_PATH = "meta/episodes.jsonl"
-LEGACY_EPISODES_STATS_PATH = "meta/episodes_stats.jsonl"
-LEGACY_TASKS_PATH = "meta/tasks.jsonl"
-LEGACY_DEFAULT_VIDEO_PATH = "videos/chunk-{episode_chunk:03d}/{video_key}/episode_{episode_index:06d}.mp4"
-LEGACY_DEFAULT_PARQUET_PATH = "data/chunk-{episode_chunk:03d}/episode_{episode_index:06d}.parquet"
-
-V21 = "v2.1"
-
-
-"""
--------------------------
-OLD
-data/chunk-000/episode_000000.parquet
-
-NEW
-data/chunk-000/file_000.parquet
--------------------------
-OLD
-videos/chunk-000/CAMERA/episode_000000.mp4
-
-NEW
-videos/chunk-000/file_000.mp4
--------------------------
-OLD
-episodes.jsonl
-{"episode_index": 1, "tasks": ["Put the blue block in the green bowl"], "length": 266}
-
-NEW
-meta/episodes/chunk-000/episodes_000.parquet
-episode_index | video_chunk_index | video_file_index | data_chunk_index | data_file_index | tasks | length
--------------------------
-OLD
-tasks.jsonl
-{"task_index": 1, "task": "Put the blue block in the green bowl"}
-
-NEW
-meta/tasks/chunk-000/file_000.parquet
-task_index | task
--------------------------
-OLD
-episodes_stats.jsonl
-
-NEW
-meta/episodes_stats/chunk-000/file_000.parquet
-episode_index | mean | std | min | max
--------------------------
-UPDATE
-meta/info.json
--------------------------
-"""
-
-
-def load_jsonlines(fpath: Path) -> list[Any]:
-    with jsonlines.open(fpath, "r") as reader:
-        return list(reader)
-
-
-def legacy_load_episodes(local_dir: Path) -> dict:
-    episodes = load_jsonlines(local_dir / LEGACY_EPISODES_PATH)
-    return {item["episode_index"]: item for item in sorted(episodes, key=lambda x: x["episode_index"])}
-
-
-def legacy_load_episodes_stats(local_dir: Path) -> dict:
-    episodes_stats = load_jsonlines(local_dir / LEGACY_EPISODES_STATS_PATH)
-    return {
-        item["episode_index"]: cast_stats_to_numpy(item["stats"])
-        for item in sorted(episodes_stats, key=lambda x: x["episode_index"])
-    }
-
-
-def legacy_load_tasks(local_dir: Path) -> tuple[dict, dict]:
-    tasks = load_jsonlines(local_dir / LEGACY_TASKS_PATH)
-    tasks = {item["task_index"]: item["task"] for item in sorted(tasks, key=lambda x: x["task_index"])}
-    task_to_task_index = {task: task_index for task_index, task in tasks.items()}
-    return tasks, task_to_task_index
-
-
-def convert_tasks(root, new_root):
-    tasks, _ = legacy_load_tasks(root)
-    task_indices = tasks.keys()
-    task_strings = tasks.values()
-    df_tasks = pd.DataFrame({"task_index": task_indices}, index=task_strings)
-    write_tasks(df_tasks, new_root)
-
-
-def concat_data_files(paths_to_cat, new_root, chunk_idx, file_idx, image_keys):
-    # TODO(rcadene): to save RAM use Dataset.from_parquet(file) and concatenate_datasets
-    dataframes = [pd.read_parquet(file) for file in paths_to_cat]
-    # Concatenate all DataFrames along rows
-    concatenated_df = pd.concat(dataframes, ignore_index=True)
-
-    path = new_root / DEFAULT_DATA_PATH.format(chunk_index=chunk_idx, file_index=file_idx)
-    path.parent.mkdir(parents=True, exist_ok=True)
-
-    if len(image_keys) > 0:
-        schema = pa.Schema.from_pandas(concatenated_df)
-        features = Features.from_arrow_schema(schema)
-        for key in image_keys:
-            features[key] = Image()
-        schema = features.arrow_schema
-    else:
-        schema = None
-
-    concatenated_df.to_parquet(path, index=False, schema=schema)
-
-
-def convert_data(root, new_root):
-    data_dir = root / "data"
-    ep_paths = sorted(data_dir.glob("*/*.parquet"))
-
-    image_keys = get_image_keys(root)
-
-    ep_idx = 0
-    chunk_idx = 0
-    file_idx = 0
-    size_in_mb = 0
-    num_frames = 0
-    paths_to_cat = []
-    episodes_metadata = []
-    for ep_path in ep_paths:
-        ep_size_in_mb = get_parquet_file_size_in_mb(ep_path)
-        ep_num_frames = get_parquet_num_frames(ep_path)
-        ep_metadata = {
-            "episode_index": ep_idx,
-            "data/chunk_index": chunk_idx,
-            "data/file_index": file_idx,
-            "dataset_from_index": num_frames,
-            "dataset_to_index": num_frames + ep_num_frames,
-        }
-        size_in_mb += ep_size_in_mb
-        num_frames += ep_num_frames
-        episodes_metadata.append(ep_metadata)
-        ep_idx += 1
-
-        if size_in_mb < DEFAULT_DATA_FILE_SIZE_IN_MB:
-            paths_to_cat.append(ep_path)
-            continue
-
-        concat_data_files(paths_to_cat, new_root, chunk_idx, file_idx, image_keys)
-
-        # Reset for the next file
-        size_in_mb = ep_size_in_mb
-        num_frames = ep_num_frames
-        paths_to_cat = [ep_path]
-
-        chunk_idx, file_idx = update_chunk_file_indices(chunk_idx, file_idx, DEFAULT_CHUNK_SIZE)
-
-    # Write remaining data if any
-    if paths_to_cat:
-        concat_data_files(paths_to_cat, new_root, chunk_idx, file_idx, image_keys)
-
-    return episodes_metadata
-
-
-def get_video_keys(root):
-    info = load_info(root)
-    features = info["features"]
-    video_keys = [key for key, ft in features.items() if ft["dtype"] == "video"]
-    return video_keys
-
-
-def get_image_keys(root):
-    info = load_info(root)
-    features = info["features"]
-    image_keys = [key for key, ft in features.items() if ft["dtype"] == "image"]
-    return image_keys
-
-
-def convert_videos(root: Path, new_root: Path):
-    video_keys = get_video_keys(root)
-    if len(video_keys) == 0:
-        return None
-
-    video_keys = sorted(video_keys)
-
-    eps_metadata_per_cam = []
-    for camera in video_keys:
-        eps_metadata = convert_videos_of_camera(root, new_root, camera)
-        eps_metadata_per_cam.append(eps_metadata)
-
-    num_eps_per_cam = [len(eps_cam_map) for eps_cam_map in eps_metadata_per_cam]
-    if len(set(num_eps_per_cam)) != 1:
-        raise ValueError(f"All cams dont have same number of episodes ({num_eps_per_cam}).")
-
-    episods_metadata = []
-    num_cameras = len(video_keys)
-    num_episodes = num_eps_per_cam[0]
-    for ep_idx in range(num_episodes):
-        # Sanity check
-        ep_ids = [eps_metadata_per_cam[cam_idx][ep_idx]["episode_index"] for cam_idx in range(num_cameras)]
-        ep_ids += [ep_idx]
-        if len(set(ep_ids)) != 1:
-            raise ValueError(f"All episode indices need to match ({ep_ids}).")
-
-        ep_dict = {}
-        for cam_idx in range(num_cameras):
-            ep_dict.update(eps_metadata_per_cam[cam_idx][ep_idx])
-        episods_metadata.append(ep_dict)
-
-    return episods_metadata
-
-
-def convert_videos_of_camera(root: Path, new_root: Path, video_key):
-    # Access old paths to mp4
-    videos_dir = root / "videos"
-    ep_paths = sorted(videos_dir.glob(f"*/{video_key}/*.mp4"))
-
-    ep_idx = 0
-    chunk_idx = 0
-    file_idx = 0
-    size_in_mb = 0
-    duration_in_s = 0.0
-    paths_to_cat = []
-    episodes_metadata = []
-    for ep_path in tqdm.tqdm(ep_paths, desc=f"convert videos of {video_key}"):
-        ep_size_in_mb = get_video_size_in_mb(ep_path)
-        ep_duration_in_s = get_video_duration_in_s(ep_path)
-        ep_metadata = {
-            "episode_index": ep_idx,
-            f"videos/{video_key}/chunk_index": chunk_idx,
-            f"videos/{video_key}/file_index": file_idx,
-            f"videos/{video_key}/from_timestamp": duration_in_s,
-            f"videos/{video_key}/to_timestamp": duration_in_s + ep_duration_in_s,
-        }
-        size_in_mb += ep_size_in_mb
-        duration_in_s += ep_duration_in_s
-        episodes_metadata.append(ep_metadata)
-        ep_idx += 1
-
-        if size_in_mb < DEFAULT_VIDEO_FILE_SIZE_IN_MB:
-            paths_to_cat.append(ep_path)
-            continue
-
-        concat_video_files(paths_to_cat, new_root, video_key, chunk_idx, file_idx)
-
-        # Reset for the next file
-        size_in_mb = ep_size_in_mb
-        duration_in_s = ep_duration_in_s
-        paths_to_cat = [ep_path]
-
-        chunk_idx, file_idx = update_chunk_file_indices(chunk_idx, file_idx, DEFAULT_CHUNK_SIZE)
-
-    # Write remaining videos if any
-    if paths_to_cat:
-        concat_video_files(paths_to_cat, new_root, video_key, chunk_idx, file_idx)
-
-    return episodes_metadata
-
-
-def generate_episode_metadata_dict(
-    episodes_legacy_metadata, episodes_metadata, episodes_stats, episodes_videos=None
-):
-    num_episodes = len(episodes_metadata)
-    episodes_legacy_metadata_vals = list(episodes_legacy_metadata.values())
-    episodes_stats_vals = list(episodes_stats.values())
-    episodes_stats_keys = list(episodes_stats.keys())
-
-    for i in range(num_episodes):
-        ep_legacy_metadata = episodes_legacy_metadata_vals[i]
-        ep_metadata = episodes_metadata[i]
-        ep_stats = episodes_stats_vals[i]
-
-        ep_ids_set = {
-            ep_legacy_metadata["episode_index"],
-            ep_metadata["episode_index"],
-            episodes_stats_keys[i],
-        }
-
-        if episodes_videos is None:
-            ep_video = {}
-        else:
-            ep_video = episodes_videos[i]
-            ep_ids_set.add(ep_video["episode_index"])
-
-        if len(ep_ids_set) != 1:
-            raise ValueError(f"Number of episodes is not the same ({ep_ids_set}).")
-
-        ep_dict = {**ep_metadata, **ep_video, **ep_legacy_metadata, **flatten_dict({"stats": ep_stats})}
-        ep_dict["meta/episodes/chunk_index"] = 0
-        ep_dict["meta/episodes/file_index"] = 0
-        yield ep_dict
-
-
-def convert_episodes_metadata(root, new_root, episodes_metadata, episodes_video_metadata=None):
-    episodes_legacy_metadata = legacy_load_episodes(root)
-    episodes_stats = legacy_load_episodes_stats(root)
-
-    num_eps_set = {len(episodes_legacy_metadata), len(episodes_metadata)}
-    if episodes_video_metadata is not None:
-        num_eps_set.add(len(episodes_video_metadata))
-
-    if len(num_eps_set) != 1:
-        raise ValueError(f"Number of episodes is not the same ({num_eps_set}).")
-
-    ds_episodes = Dataset.from_generator(
-        lambda: generate_episode_metadata_dict(
-            episodes_legacy_metadata, episodes_metadata, episodes_stats, episodes_video_metadata
-        )
-    )
-    write_episodes(ds_episodes, new_root)
-
-    stats = aggregate_stats(list(episodes_stats.values()))
-    write_stats(stats, new_root)
-
-
-def convert_info(root, new_root):
-    info = load_info(root)
-    info["codebase_version"] = "v3.0"
-    del info["total_chunks"]
-    del info["total_videos"]
-    info["data_files_size_in_mb"] = DEFAULT_DATA_FILE_SIZE_IN_MB
-    info["video_files_size_in_mb"] = DEFAULT_VIDEO_FILE_SIZE_IN_MB
-    info["data_path"] = DEFAULT_DATA_PATH
-    info["video_path"] = DEFAULT_VIDEO_PATH
-    info["fps"] = float(info["fps"])
-    for key in info["features"]:
-        if info["features"][key]["dtype"] == "video":
-            # already has fps in video_info
-            continue
-        info["features"][key]["fps"] = info["fps"]
-    write_info(info, new_root)
-
-
-def convert_dataset(
-    repo_id: str,
-    branch: str | None = None,
-    num_workers: int = 4,
-):
-    root = HF_LEROBOT_HOME / repo_id
-    old_root = HF_LEROBOT_HOME / f"{repo_id}_old"
-    new_root = HF_LEROBOT_HOME / f"{repo_id}_v30"
-
-    if old_root.is_dir() and root.is_dir():
-        shutil.rmtree(str(root))
-        shutil.move(str(old_root), str(root))
-
-    if new_root.is_dir():
-        shutil.rmtree(new_root)
-
-    snapshot_download(
-        repo_id,
-        repo_type="dataset",
-        revision=V21,
-        local_dir=root,
-    )
-
-    convert_info(root, new_root)
-    convert_tasks(root, new_root)
-    episodes_metadata = convert_data(root, new_root)
-    episodes_videos_metadata = convert_videos(root, new_root)
-    convert_episodes_metadata(root, new_root, episodes_metadata, episodes_videos_metadata)
-
-    shutil.move(str(root), str(old_root))
-    shutil.move(str(new_root), str(root))
-
-    hub_api = HfApi()
-    try:
-        hub_api.delete_tag(repo_id, tag=CODEBASE_VERSION, repo_type="dataset")
-    except HTTPError as e:
-        print(f"tag={CODEBASE_VERSION} probably doesn't exist. Skipping exception ({e})")
-        pass
-    hub_api.delete_files(
-        delete_patterns=["data/chunk*/episode_*", "meta/*.jsonl", "videos/chunk*"],
-        repo_id=repo_id,
-        revision=branch,
-        repo_type="dataset",
-    )
-    hub_api.create_tag(repo_id, tag=CODEBASE_VERSION, revision=branch, repo_type="dataset")
-
-    LeRobotDataset(repo_id).push_to_hub()
-
-
-if __name__ == "__main__":
-    parser = argparse.ArgumentParser()
-    parser.add_argument(
-        "--repo-id",
-        type=str,
-        required=True,
-        help="Repository identifier on Hugging Face: a community or a user name `/` the name of the dataset "
-        "(e.g. `lerobot/pusht`, `cadene/aloha_sim_insertion_human`).",
-    )
-    parser.add_argument(
-        "--branch",
-        type=str,
-        default=None,
-        help="Repo branch to push your dataset. Defaults to the main branch.",
-    )
-    parser.add_argument(
-        "--num-workers",
-        type=int,
-        default=4,
-        help="Number of workers for parallelizing stats compute. Defaults to 4.",
-    )
-
-    args = parser.parse_args()
-    convert_dataset(**vars(args))

lerobot/common/datasets/video_utils.py

@@ -21,15 +21,12 @@ import warnings
 from collections import OrderedDict
 from dataclasses import dataclass, field
 from pathlib import Path
-from threading import Lock
-from typing import Any, ClassVar, Dict, Literal, Optional
+from typing import Any, ClassVar
 
-import fsspec
 import pyarrow as pa
 import torch
 import torchvision
 from datasets.features.features import register_feature
-from line_profiler import profile
 from PIL import Image
 
 
@@ -77,7 +74,7 @@ def decode_video_frames_torchvision(
     video_path: Path | str,
     timestamps: list[float],
     tolerance_s: float,
-    backend: Literal["pyav", "video_reader"] = "pyav",
+    backend: str = "pyav",
     log_loaded_timestamps: bool = False,
 ) -> torch.Tensor:
     """Loads frames associated to the requested timestamps of a video
@@ -158,7 +155,6 @@ def decode_video_frames_torchvision(
     )
 
     # get closest frames to the query timestamps
-    # TODO(rcadene): remove torch.stack
     closest_frames = torch.stack([loaded_frames[idx] for idx in argmin_])
     closest_ts = loaded_ts[argmin_]
 
@@ -172,62 +168,15 @@ def decode_video_frames_torchvision(
     return closest_frames
 
 
-class VideoDecoderCache:
-    """Thread-safe cache for video decoders to avoid expensive re-initialization."""
-
-    def __init__(self):
-        self._cache: Dict[str, Any] = {}
-        self._lock = Lock()
-
-    def get_decoder(self, video_path: str):
-        """Get a cached decoder or create a new one."""
-        if importlib.util.find_spec("torchcodec"):
-            from torchcodec.decoders import VideoDecoder
-        else:
-            raise ImportError("torchcodec is required but not available.")
-
-        video_path = str(video_path)
-
-        with self._lock:
-            if video_path not in self._cache:
-                file_handle = fsspec.open(video_path, client_kwargs={"trust_env": True}).__enter__()
-                decoder = VideoDecoder(file_handle, seek_mode="approximate")
-                self._cache[video_path] = decoder
-
-            return self._cache[video_path]
-
-    def clear(self):
-        """Clear the cache."""
-        with self._lock:
-            self._cache.clear()
-
-    def size(self) -> int:
-        """Return the number of cached decoders."""
-        with self._lock:
-            return len(self._cache)
-
-
-# Global instance
-_default_decoder_cache = VideoDecoderCache()
-
-
-@profile
 def decode_video_frames_torchcodec(
     video_path: Path | str,
     timestamps: list[float],
     tolerance_s: float,
+    device: str = "cpu",
     log_loaded_timestamps: bool = False,
-    decoder_cache: Optional[VideoDecoderCache] = None,
 ) -> torch.Tensor:
     """Loads frames associated with the requested timestamps of a video using torchcodec.
 
-    Args:
-        video_path: Path to the video file.
-        timestamps: List of timestamps to extract frames.
-        tolerance_s: Allowed deviation in seconds for frame retrieval.
-        log_loaded_timestamps: Whether to log loaded timestamps.
-        decoder_cache: Optional decoder cache instance. Uses default if None.
-
     Note: Setting device="cuda" outside the main process, e.g. in data loader workers, will lead to CUDA initialization errors.
 
     Note: Video benefits from inter-frame compression. Instead of storing every frame individually,
@@ -236,20 +185,23 @@ def decode_video_frames_torchcodec(
     and all subsequent frames until reaching the requested frame. The number of key frames in a video
     can be adjusted during encoding to take into account decoding time and video size in bytes.
     """
-    if decoder_cache is None:
-        decoder_cache = _default_decoder_cache
 
-    # Use cached decoder instead of creating new one each time
-    decoder = decoder_cache.get_decoder(str(video_path))
+    if importlib.util.find_spec("torchcodec"):
+        from torchcodec.decoders import VideoDecoder
+    else:
+        raise ImportError("torchcodec is required but not available.")
 
-    loaded_ts = []
+    # initialize video decoder
+    decoder = VideoDecoder(video_path, device=device, seek_mode="approximate")
     loaded_frames = []
-
+    loaded_ts = []
     # get metadata for frame information
     metadata = decoder.metadata
     average_fps = metadata.average_fps
+
     # convert timestamps to frame indices
     frame_indices = [round(ts * average_fps) for ts in timestamps]
+
     # retrieve frames based on indices
     frames_batch = decoder.get_frames_at(indices=frame_indices)
 
@@ -300,7 +252,7 @@ def encode_video_frames(
     g: int | None = 2,
     crf: int | None = 30,
     fast_decode: int = 0,
-    log_level: str | None = "quiet",
+    log_level: str | None = "error",
     overwrite: bool = False,
 ) -> None:
     """More info on ffmpeg arguments tuning on `benchmark/video/README.md`"""
@@ -312,7 +264,7 @@ def encode_video_frames(
         [
             ("-f", "image2"),
             ("-r", str(fps)),
-            ("-i", str(imgs_dir / "frame-%06d.png")),
+            ("-i", str(imgs_dir / "frame_%06d.png")),
             ("-vcodec", vcodec),
             ("-pix_fmt", pix_fmt),
         ]

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 +156,136 @@ 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."""
+
+    delta_action: float | None = None
+    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
+    delta_action: float = 0.1
+    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 = 0.8
+    gripper_penalty: float = 0.0
+    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/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

@@ -24,8 +24,8 @@ 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.tdmpc.configuration_tdmpc import TDMPCConfig
 from lerobot.common.policies.vqbet.configuration_vqbet import VQBeTConfig
@@ -55,10 +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
+        return SACPolicy
+    elif name == "hilserl_classifier":
+        from lerobot.common.policies.hilserl.classifier.modeling_classifier import Classifier
+
+        return Classifier
     else:
         raise NotImplementedError(f"Policy with name {name} is not implemented.")
 
@@ -74,8 +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 == "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:
@@ -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 expacted 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 +1,227 @@
+#!/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: int = -5
+    log_std_max: int = 2
+    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.
+        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
+
+    # 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/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
@@ -128,14 +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(
@@ -160,6 +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}")
 
@@ -174,13 +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,
@@ -192,7 +208,7 @@ def record_episode(
     dataset,
     events,
     episode_time_s,
-    display_data,
+    display_cameras,
     policy,
     fps,
     single_task,
@@ -200,11 +216,12 @@ 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,
         fps=fps,
+        # record_delta_actions=record_delta_actions,
         teleoperate=policy is None,
         single_task=single_task,
     )
@@ -215,7 +232,7 @@ 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,
@@ -246,14 +263,21 @@ def control_loop(
     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)
+            # if record_delta_actions:
+            #     action["action"] = action["action"] - current_joint_positions
         else:
             observation = robot.capture_observation()
 
             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.
@@ -264,15 +288,14 @@ def control_loop(
             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")):
-            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
@@ -301,11 +324,25 @@ 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):
@@ -327,12 +364,20 @@ 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/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/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():

lerobot/common/robot_devices/robots/configs.py

@@ -443,7 +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"],
@@ -460,7 +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"],

lerobot/common/robot_devices/robots/feetech_calibration.py

@@ -207,7 +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)
 
@@ -239,7 +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)
@@ -250,7 +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")

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}"
@@ -309,7 +317,9 @@ 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)
 
@@ -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/utils.py

@@ -72,7 +72,9 @@ 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)
     else:

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

lerobot/common/utils/random_utils.py

@@ -42,7 +42,11 @@ def deserialize_python_rng_state(rng_state_dict: dict[str, torch.Tensor]) -> Non
     """
     Restores the rng state for `random` from a dictionary produced by `serialize_python_rng_state()`.
     """
-    py_state = (rng_state_dict["py_rng_version"].item(), tuple(rng_state_dict["py_rng_state"].tolist()), None)
+    py_state = (
+        rng_state_dict["py_rng_version"].item(),
+        tuple(rng_state_dict["py_rng_state"].tolist()),
+        None,
+    )
     random.setstate(py_state)
 
 

lerobot/common/utils/utils.py

@@ -18,6 +18,7 @@ import os
 import os.path as osp
 import platform
 import subprocess
+import time
 from copy import copy
 from datetime import datetime, timezone
 from pathlib import Path
@@ -107,11 +108,11 @@ def is_amp_available(device: str):
         raise ValueError(f"Unknown device '{device}.")
 
 
-def init_logging():
+def init_logging(log_file=None):
     def custom_format(record):
         dt = datetime.now().strftime("%Y-%m-%d %H:%M:%S")
         fnameline = f"{record.pathname}:{record.lineno}"
-        message = f"{record.levelname} {dt} {fnameline[-15:]:>15} {record.msg}"
+        message = f"{record.levelname} [PID: {os.getpid()}] {dt} {fnameline[-15:]:>15} {record.msg}"
         return message
 
     logging.basicConfig(level=logging.INFO)
@@ -125,6 +126,12 @@ def init_logging():
     console_handler.setFormatter(formatter)
     logging.getLogger().addHandler(console_handler)
 
+    if log_file is not None:
+        # File handler
+        file_handler = logging.FileHandler(log_file)
+        file_handler.setFormatter(formatter)
+        logging.getLogger().addHandler(file_handler)
+
 
 def format_big_number(num, precision=0):
     suffixes = ["", "K", "M", "B", "T", "Q"]
@@ -230,11 +237,31 @@ def is_valid_numpy_dtype_string(dtype_str: str) -> bool:
         return False
 
 
-def get_elapsed_time_in_days_hours_minutes_seconds(elapsed_time_s: float):
-    days = int(elapsed_time_s // (24 * 3600))
-    elapsed_time_s %= 24 * 3600
-    hours = int(elapsed_time_s // 3600)
-    elapsed_time_s %= 3600
-    minutes = int(elapsed_time_s // 60)
-    seconds = elapsed_time_s % 60
-    return days, hours, minutes, seconds
+class TimerManager:
+    def __init__(
+        self,
+        elapsed_time_list: list[float] | None = None,
+        label="Elapsed time",
+        log=True,
+    ):
+        self.label = label
+        self.elapsed_time_list = elapsed_time_list
+        self.log = log
+        self.elapsed = 0.0
+
+    def __enter__(self):
+        self.start = time.perf_counter()
+        return self
+
+    def __exit__(self, exc_type, exc_value, traceback):
+        self.elapsed: float = time.perf_counter() - self.start
+
+        if self.elapsed_time_list is not None:
+            self.elapsed_time_list.append(self.elapsed)
+
+        if self.log:
+            print(f"{self.label}: {self.elapsed:.6f} seconds")
+
+    @property
+    def elapsed_seconds(self):
+        return self.elapsed

lerobot/common/utils/wandb_utils.py

@@ -30,9 +30,10 @@ def cfg_to_group(cfg: TrainPipelineConfig, return_list: bool = False) -> list[st
     """Return a group name for logging. Optionally returns group name as list."""
     lst = [
         f"policy:{cfg.policy.type}",
-        f"dataset:{cfg.dataset.repo_id}",
         f"seed:{cfg.seed}",
     ]
+    if cfg.dataset is not None:
+        lst.append(f"dataset:{cfg.dataset.repo_id}")
     if cfg.env is not None:
         lst.append(f"env:{cfg.env.type}")
     return lst if return_list else "-".join(lst)
@@ -92,6 +93,12 @@ class WandBLogger:
             resume="must" if cfg.resume else None,
             mode=self.cfg.mode if self.cfg.mode in ["online", "offline", "disabled"] else "online",
         )
+        run_id = wandb.run.id
+        # NOTE: We will override the cfg.wandb.run_id with the wandb run id.
+        # This is because we want to be able to resume the run from the wandb run id.
+        cfg.wandb.run_id = run_id
+        # Handle custom step key for rl asynchronous training.
+        self._wandb_custom_step_key: set[str] | None = None
         print(colored("Logs will be synced with wandb.", "blue", attrs=["bold"]))
         logging.info(f"Track this run --> {colored(wandb.run.get_url(), 'yellow', attrs=['bold'])}")
         self._wandb = wandb
@@ -108,9 +115,26 @@ class WandBLogger:
         artifact.add_file(checkpoint_dir / PRETRAINED_MODEL_DIR / SAFETENSORS_SINGLE_FILE)
         self._wandb.log_artifact(artifact)
 
-    def log_dict(self, d: dict, step: int, mode: str = "train"):
+    def log_dict(
+        self, d: dict, step: int | None = None, mode: str = "train", custom_step_key: str | None = None
+    ):
         if mode not in {"train", "eval"}:
             raise ValueError(mode)
+        if step is None and custom_step_key is None:
+            raise ValueError("Either step or custom_step_key must be provided.")
+
+        # NOTE: This is not simple. Wandb step is it must always monotonically increase and it
+        # increases with each wandb.log call, but in the case of asynchronous RL for example,
+        # multiple time steps is possible for example, the interaction step with the environment,
+        # the training step, the evaluation step, etc. So we need to define a custom step key
+        # to log the correct step for each metric.
+        if custom_step_key is not None:
+            if self._wandb_custom_step_key is None:
+                self._wandb_custom_step_key = set()
+            new_custom_key = f"{mode}/{custom_step_key}"
+            if new_custom_key not in self._wandb_custom_step_key:
+                self._wandb_custom_step_key.add(new_custom_key)
+                self._wandb.define_metric(new_custom_key, hidden=True)
 
         for k, v in d.items():
             if not isinstance(v, (int, float, str)):
@@ -118,7 +142,22 @@ class WandBLogger:
                     f'WandB logging of key "{k}" was ignored as its type is not handled by this wrapper.'
                 )
                 continue
-            self._wandb.log({f"{mode}/{k}": v}, step=step)
+
+            # Do not log the custom step key itself.
+            if self._wandb_custom_step_key is not None and k in self._wandb_custom_step_key:
+                continue
+
+            if custom_step_key is not None:
+                value_custom_step = d[custom_step_key]
+                self._wandb.log(
+                    {
+                        f"{mode}/{k}": v,
+                        f"{mode}/{custom_step_key}": value_custom_step,
+                    }
+                )
+                continue
+
+            self._wandb.log(data={f"{mode}/{k}": v}, step=step)
 
     def log_video(self, video_path: str, step: int, mode: str = "train"):
         if mode not in {"train", "eval"}:

lerobot/configs/parser.py

@@ -42,7 +42,10 @@ def get_cli_overrides(field_name: str, args: Sequence[str] | None = None) -> lis
         args = sys.argv[1:]
     attr_level_args = []
     detect_string = f"--{field_name}."
-    exclude_strings = (f"--{field_name}.{draccus.CHOICE_TYPE_KEY}=", f"--{field_name}.{PATH_KEY}=")
+    exclude_strings = (
+        f"--{field_name}.{draccus.CHOICE_TYPE_KEY}=",
+        f"--{field_name}.{PATH_KEY}=",
+    )
     for arg in args:
         if arg.startswith(detect_string) and not arg.startswith(exclude_strings):
             denested_arg = f"--{arg.removeprefix(detect_string)}"

lerobot/configs/policies.py

@@ -26,7 +26,11 @@ from huggingface_hub.errors import HfHubHTTPError
 from lerobot.common.optim.optimizers import OptimizerConfig
 from lerobot.common.optim.schedulers import LRSchedulerConfig
 from lerobot.common.utils.hub import HubMixin
-from lerobot.common.utils.utils import auto_select_torch_device, is_amp_available, is_torch_device_available
+from lerobot.common.utils.utils import (
+    auto_select_torch_device,
+    is_amp_available,
+    is_torch_device_available,
+)
 from lerobot.configs.types import FeatureType, NormalizationMode, PolicyFeature
 
 # Generic variable that is either PreTrainedConfig or a subclass thereof

lerobot/configs/train.py

@@ -34,11 +34,10 @@ TRAIN_CONFIG_NAME = "train_config.json"
 
 @dataclass
 class TrainPipelineConfig(HubMixin):
-    dataset: DatasetConfig
+    dataset: DatasetConfig | None = None  # NOTE: In RL, we don't need a dataset
     env: envs.EnvConfig | None = None
     policy: PreTrainedConfig | None = None
-    # Set `dir` to where you would like to save all of the run outputs. If you run another training session
-    # with the same value for `dir` its contents will be overwritten unless you set `resume` to true.
+    # Set `dir` to where you would like to save all of the run outputs. If you run another training session # with the same value for `dir` its contents will be overwritten unless you set `resume` to true.
     output_dir: Path | None = None
     job_name: str | None = None
     # Set `resume` to true to resume a previous run. In order for this to work, you will need to make sure
@@ -107,8 +106,9 @@ class TrainPipelineConfig(HubMixin):
             train_dir = f"{now:%Y-%m-%d}/{now:%H-%M-%S}_{self.job_name}"
             self.output_dir = Path("outputs/train") / train_dir
 
-        if isinstance(self.dataset.repo_id, list):
-            raise NotImplementedError("LeRobotMultiDataset is not currently implemented.")
+        if self.dataset is not None:
+            if isinstance(self.dataset.repo_id, list):
+                raise NotImplementedError("LeRobotMultiDataset is not currently implemented.")
 
         if not self.use_policy_training_preset and (self.optimizer is None or self.scheduler is None):
             raise ValueError("Optimizer and Scheduler must be set when the policy presets are not used.")
@@ -125,7 +125,10 @@ class TrainPipelineConfig(HubMixin):
         return draccus.encode(self)
 
     def _save_pretrained(self, save_directory: Path) -> None:
-        with open(save_directory / TRAIN_CONFIG_NAME, "w") as f, draccus.config_type("json"):
+        with (
+            open(save_directory / TRAIN_CONFIG_NAME, "w") as f,
+            draccus.config_type("json"),
+        ):
             draccus.dump(self, f, indent=4)
 
     @classmethod

lerobot/configs/types.py

@@ -23,6 +23,7 @@ class FeatureType(str, Enum):
     VISUAL = "VISUAL"
     ENV = "ENV"
     ACTION = "ACTION"
+    REWARD = "REWARD"
 
 
 class NormalizationMode(str, Enum):

lerobot/scripts/configure_motor.py

@@ -38,7 +38,12 @@ def get_motor_bus_cls(brand: str) -> tuple:
             FeetechMotorsBus,
         )
 
-        return FeetechMotorsBusConfig, FeetechMotorsBus, MODEL_BAUDRATE_TABLE, SCS_SERIES_BAUDRATE_TABLE
+        return (
+            FeetechMotorsBusConfig,
+            FeetechMotorsBus,
+            MODEL_BAUDRATE_TABLE,
+            SCS_SERIES_BAUDRATE_TABLE,
+        )
 
     elif brand == "dynamixel":
         from lerobot.common.robot_devices.motors.configs import DynamixelMotorsBusConfig
@@ -48,7 +53,12 @@ def get_motor_bus_cls(brand: str) -> tuple:
             DynamixelMotorsBus,
         )
 
-        return DynamixelMotorsBusConfig, DynamixelMotorsBus, MODEL_BAUDRATE_TABLE, X_SERIES_BAUDRATE_TABLE
+        return (
+            DynamixelMotorsBusConfig,
+            DynamixelMotorsBus,
+            MODEL_BAUDRATE_TABLE,
+            X_SERIES_BAUDRATE_TABLE,
+        )
 
     else:
         raise ValueError(
@@ -164,12 +174,25 @@ def configure_motor(port, brand, model, motor_idx_des, baudrate_des):
 
 if __name__ == "__main__":
     parser = argparse.ArgumentParser()
-    parser.add_argument("--port", type=str, required=True, help="Motors bus port (e.g. dynamixel,feetech)")
+    parser.add_argument(
+        "--port",
+        type=str,
+        required=True,
+        help="Motors bus port (e.g. dynamixel,feetech)",
+    )
     parser.add_argument("--brand", type=str, required=True, help="Motor brand (e.g. dynamixel,feetech)")
     parser.add_argument("--model", type=str, required=True, help="Motor model (e.g. xl330-m077,sts3215)")
-    parser.add_argument("--ID", type=int, required=True, help="Desired ID of the current motor (e.g. 1,2,3)")
     parser.add_argument(
-        "--baudrate", type=int, default=1000000, help="Desired baudrate for the motor (default: 1000000)"
+        "--ID",
+        type=int,
+        required=True,
+        help="Desired ID of the current motor (e.g. 1,2,3)",
+    )
+    parser.add_argument(
+        "--baudrate",
+        type=int,
+        default=1000000,
+        help="Desired baudrate for the motor (default: 1000000)",
     )
     args = parser.parse_args()
 

lerobot/scripts/control_robot.py

@@ -135,19 +135,15 @@ python lerobot/scripts/control_robot.py \
 """
 
 import logging
-import os
 import time
 from dataclasses import asdict
 from pprint import pformat
 
-import rerun as rr
-
 # from safetensors.torch import load_file, save_file
 from lerobot.common.datasets.lerobot_dataset import LeRobotDataset
 from lerobot.common.policies.factory import make_policy
 from lerobot.common.robot_devices.control_configs import (
     CalibrateControlConfig,
-    ControlConfig,
     ControlPipelineConfig,
     RecordControlConfig,
     RemoteRobotConfig,
@@ -157,7 +153,6 @@ from lerobot.common.robot_devices.control_configs import (
 from lerobot.common.robot_devices.control_utils import (
     control_loop,
     init_keyboard_listener,
-    is_headless,
     log_control_info,
     record_episode,
     reset_environment,
@@ -237,7 +232,7 @@ def teleoperate(robot: Robot, cfg: TeleoperateControlConfig):
         control_time_s=cfg.teleop_time_s,
         fps=cfg.fps,
         teleoperate=True,
-        display_data=cfg.display_data,
+        display_cameras=cfg.display_cameras,
     )
 
 
@@ -274,10 +269,12 @@ def record(
     # Load pretrained policy
     policy = None if cfg.policy is None else make_policy(cfg.policy, ds_meta=dataset.meta)
 
+    # Load pretrained policy
+    policy = None if cfg.policy is None else make_policy(cfg.policy, ds_meta=dataset.meta)
+
     if not robot.is_connected:
         robot.connect()
-
-    listener, events = init_keyboard_listener()
+    listener, events = init_keyboard_listener(assign_rewards=cfg.assign_rewards)
 
     # Execute a few seconds without recording to:
     # 1. teleoperate the robot to move it in starting position if no policy provided,
@@ -285,7 +282,14 @@ def record(
     # 3. place the cameras windows on screen
     enable_teleoperation = policy is None
     log_say("Warmup record", cfg.play_sounds)
-    warmup_record(robot, events, enable_teleoperation, cfg.warmup_time_s, cfg.display_data, cfg.fps)
+    warmup_record(
+        robot,
+        events,
+        enable_teleoperation,
+        cfg.warmup_time_s,
+        cfg.display_cameras,
+        cfg.fps,
+    )
 
     if has_method(robot, "teleop_safety_stop"):
         robot.teleop_safety_stop()
@@ -301,7 +305,7 @@ def record(
             dataset=dataset,
             events=events,
             episode_time_s=cfg.episode_time_s,
-            display_data=cfg.display_data,
+            display_cameras=cfg.display_cameras,
             policy=policy,
             fps=cfg.fps,
             single_task=cfg.single_task,
@@ -331,7 +335,7 @@ def record(
             break
 
     log_say("Stop recording", cfg.play_sounds, blocking=True)
-    stop_recording(robot, listener, cfg.display_data)
+    stop_recording(robot, listener, cfg.display_cameras)
 
     if cfg.push_to_hub:
         dataset.push_to_hub(tags=cfg.tags, private=cfg.private)
@@ -350,15 +354,17 @@ def replay(
 
     dataset = LeRobotDataset(cfg.repo_id, root=cfg.root, episodes=[cfg.episode])
     actions = dataset.hf_dataset.select_columns("action")
-
     if not robot.is_connected:
         robot.connect()
 
     log_say("Replaying episode", cfg.play_sounds, blocking=True)
     for idx in range(dataset.num_frames):
+        current_joint_positions = robot.follower_arms["main"].read("Present_Position")
         start_episode_t = time.perf_counter()
 
         action = actions[idx]["action"]
+        # if replay_delta_actions:
+        #     action = action + current_joint_positions
         robot.send_action(action)
 
         dt_s = time.perf_counter() - start_episode_t
@@ -368,40 +374,6 @@ def replay(
         log_control_info(robot, dt_s, fps=cfg.fps)
 
 
-def _init_rerun(control_config: ControlConfig, session_name: str = "lerobot_control_loop") -> None:
-    """Initializes the Rerun SDK for visualizing the control loop.
-
-    Args:
-        control_config: Configuration determining data display and robot type.
-        session_name: Rerun session name. Defaults to "lerobot_control_loop".
-
-    Raises:
-        ValueError: If viewer IP is missing for non-remote configurations with display enabled.
-    """
-    if (control_config.display_data and not is_headless()) or (
-        control_config.display_data and isinstance(control_config, RemoteRobotConfig)
-    ):
-        # Configure Rerun flush batch size default to 8KB if not set
-        batch_size = os.getenv("RERUN_FLUSH_NUM_BYTES", "8000")
-        os.environ["RERUN_FLUSH_NUM_BYTES"] = batch_size
-
-        # Initialize Rerun based on configuration
-        rr.init(session_name)
-        if isinstance(control_config, RemoteRobotConfig):
-            viewer_ip = control_config.viewer_ip
-            viewer_port = control_config.viewer_port
-            if not viewer_ip or not viewer_port:
-                raise ValueError(
-                    "Viewer IP & Port are required for remote config. Set via config file/CLI or disable control_config.display_data."
-                )
-            logging.info(f"Connecting to viewer at {viewer_ip}:{viewer_port}")
-            rr.connect_tcp(f"{viewer_ip}:{viewer_port}")
-        else:
-            # Get memory limit for rerun viewer parameters
-            memory_limit = os.getenv("LEROBOT_RERUN_MEMORY_LIMIT", "10%")
-            rr.spawn(memory_limit=memory_limit)
-
-
 @parser.wrap()
 def control_robot(cfg: ControlPipelineConfig):
     init_logging()
@@ -409,22 +381,17 @@ def control_robot(cfg: ControlPipelineConfig):
 
     robot = make_robot_from_config(cfg.robot)
 
-    # TODO(Steven): Blueprint for fixed window size
-
     if isinstance(cfg.control, CalibrateControlConfig):
         calibrate(robot, cfg.control)
     elif isinstance(cfg.control, TeleoperateControlConfig):
-        _init_rerun(control_config=cfg.control, session_name="lerobot_control_loop_teleop")
         teleoperate(robot, cfg.control)
     elif isinstance(cfg.control, RecordControlConfig):
-        _init_rerun(control_config=cfg.control, session_name="lerobot_control_loop_record")
         record(robot, cfg.control)
     elif isinstance(cfg.control, ReplayControlConfig):
         replay(robot, cfg.control)
     elif isinstance(cfg.control, RemoteRobotConfig):
         from lerobot.common.robot_devices.robots.lekiwi_remote import run_lekiwi
 
-        _init_rerun(control_config=cfg.control, session_name="lerobot_control_loop_remote")
         run_lekiwi(cfg.robot)
 
     if robot.is_connected:

lerobot/scripts/control_sim_robot.py

@@ -149,7 +149,11 @@ def init_sim_calibration(robot, cfg):
     axis_directions = np.array(cfg.get("axis_directions", [1]))
     offsets = np.array(cfg.get("offsets", [0])) * np.pi
 
-    return {"start_pos": start_pos, "axis_directions": axis_directions, "offsets": offsets}
+    return {
+        "start_pos": start_pos,
+        "axis_directions": axis_directions,
+        "offsets": offsets,
+    }
 
 
 def real_positions_to_sim(real_positions, axis_directions, start_pos, offsets):
@@ -197,8 +201,14 @@ def record(
     resume: bool = False,
     local_files_only: bool = False,
     run_compute_stats: bool = True,
+    assign_rewards: bool = False,
 ) -> LeRobotDataset:
     # Load pretrained policy
+
+    extra_features = (
+        {"next.reward": {"dtype": "int64", "shape": (1,), "names": None}} if assign_rewards else None
+    )
+
     policy = None
     if pretrained_policy_name_or_path is not None:
         policy, policy_fps, device, use_amp = init_policy(pretrained_policy_name_or_path, policy_overrides)
@@ -211,7 +221,7 @@ def record(
         raise ValueError("Either policy or process_action_fn has to be set to enable control in sim.")
 
     # initialize listener before sim env
-    listener, events = init_keyboard_listener()
+    listener, events = init_keyboard_listener(assign_rewards=assign_rewards)
 
     # create sim env
     env = env()
@@ -241,7 +251,11 @@ def record(
             shape = env.observation_space[key].shape
             if not key.startswith("observation.image."):
                 key = "observation.image." + key
-            features[key] = {"dtype": "video", "names": ["channels", "height", "width"], "shape": shape}
+            features[key] = {
+                "dtype": "video",
+                "names": ["channels", "height", "width"],
+                "shape": shape,
+            }
 
         for key, obs_key in state_keys_dict.items():
             features[key] = {
@@ -250,7 +264,12 @@ def record(
                 "shape": env.observation_space[obs_key].shape,
             }
 
-        features["action"] = {"dtype": "float32", "shape": env.action_space.shape, "names": None}
+        features["action"] = {
+            "dtype": "float32",
+            "shape": env.action_space.shape,
+            "names": None,
+        }
+        features = {**features, **extra_features}
 
         # Create empty dataset or load existing saved episodes
         sanity_check_dataset_name(repo_id, policy)
@@ -302,6 +321,13 @@ def record(
                 "timestamp": env_timestamp,
             }
 
+            # Overwrite environment reward with manually assigned reward
+            if assign_rewards:
+                frame["next.reward"] = events["next.reward"]
+
+                # Should success always be false to match what we do in control_utils?
+                frame["next.success"] = False
+
             for key in image_keys:
                 if not key.startswith("observation.image"):
                     frame["observation.image." + key] = observation[key]
@@ -361,7 +387,12 @@ def record(
 
 
 def replay(
-    env, root: Path, repo_id: str, episode: int, fps: int | None = None, local_files_only: bool = True
+    env,
+    root: Path,
+    repo_id: str,
+    episode: int,
+    fps: int | None = None,
+    local_files_only: bool = True,
 ):
     env = env()
 
@@ -408,7 +439,10 @@ if __name__ == "__main__":
 
     parser_record = subparsers.add_parser("record", parents=[base_parser])
     parser_record.add_argument(
-        "--fps", type=none_or_int, default=None, help="Frames per second (set to None to disable)"
+        "--fps",
+        type=none_or_int,
+        default=None,
+        help="Frames per second (set to None to disable)",
     )
     parser_record.add_argument(
         "--root",
@@ -486,9 +520,19 @@ if __name__ == "__main__":
         default=0,
         help="Resume recording on an existing dataset.",
     )
+    parser_record.add_argument(
+        "--assign-rewards",
+        type=int,
+        default=0,
+        help="Enables the assignation of rewards to frames (by default no assignation). When enabled, assign a 0 reward to frames until the space bar is pressed which assign a 1 reward. Press the space bar a second time to assign a 0 reward. The reward assigned is reset to 0 when the episode ends.",
+    )
+
     parser_replay = subparsers.add_parser("replay", parents=[base_parser])
     parser_replay.add_argument(
-        "--fps", type=none_or_int, default=None, help="Frames per second (set to None to disable)"
+        "--fps",
+        type=none_or_int,
+        default=None,
+        help="Frames per second (set to None to disable)",
     )
     parser_replay.add_argument(
         "--root",

lerobot/scripts/eval.py

@@ -66,7 +66,7 @@ from torch import Tensor, nn
 from tqdm import trange
 
 from lerobot.common.envs.factory import make_env
-from lerobot.common.envs.utils import add_envs_task, check_env_attributes_and_types, preprocess_observation
+from lerobot.common.envs.utils import preprocess_observation
 from lerobot.common.policies.factory import make_policy
 from lerobot.common.policies.pretrained import PreTrainedPolicy
 from lerobot.common.policies.utils import get_device_from_parameters
@@ -124,6 +124,7 @@ def rollout(
 
     # Reset the policy and environments.
     policy.reset()
+
     observation, info = env.reset(seed=seeds)
     if render_callback is not None:
         render_callback(env)
@@ -144,7 +145,6 @@ def rollout(
         disable=inside_slurm(),  # we dont want progress bar when we use slurm, since it clutters the logs
         leave=False,
     )
-    check_env_attributes_and_types(env)
     while not np.all(done):
         # Numpy array to tensor and changing dictionary keys to LeRobot policy format.
         observation = preprocess_observation(observation)
@@ -155,10 +155,6 @@ def rollout(
             key: observation[key].to(device, non_blocking=device.type == "cuda") for key in observation
         }
 
-        # Infer "task" from attributes of environments.
-        # TODO: works with SyncVectorEnv but not AsyncVectorEnv
-        observation = add_envs_task(env, observation)
-
         with torch.inference_mode():
             action = policy.select_action(observation)
 
@@ -293,7 +289,8 @@ def eval_policy(
             seeds = None
         else:
             seeds = range(
-                start_seed + (batch_ix * env.num_envs), start_seed + ((batch_ix + 1) * env.num_envs)
+                start_seed + (batch_ix * env.num_envs),
+                start_seed + ((batch_ix + 1) * env.num_envs),
             )
         rollout_data = rollout(
             env,
@@ -413,7 +410,11 @@ def eval_policy(
 
 
 def _compile_episode_data(
-    rollout_data: dict, done_indices: Tensor, start_episode_index: int, start_data_index: int, fps: float
+    rollout_data: dict,
+    done_indices: Tensor,
+    start_episode_index: int,
+    start_data_index: int,
+    fps: float,
 ) -> dict:
     """Convenience function for `eval_policy(return_episode_data=True)`
 
@@ -481,7 +482,10 @@ def eval_main(cfg: EvalPipelineConfig):
     )
     policy.eval()
 
-    with torch.no_grad(), torch.autocast(device_type=device.type) if cfg.policy.use_amp else nullcontext():
+    with (
+        torch.no_grad(),
+        torch.autocast(device_type=device.type) if cfg.policy.use_amp else nullcontext(),
+    ):
         info = eval_policy(
             env,
             policy,

lerobot/scripts/eval_on_robot.py

@@ -0,0 +1,412 @@
+#!/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.
+"""Evaluate a policy by running rollouts on the real robot and computing metrics.
+
+Usage examples: evaluate a checkpoint from the LeRobot training script for 10 episodes.
+
+```
+python lerobot/scripts/eval_on_robot.py \
+    -p outputs/train/model/checkpoints/005000/pretrained_model \
+    eval.n_episodes=10
+```
+
+Test reward classifier with teleoperation (you need to press space to take over)
+```
+python lerobot/scripts/eval_on_robot.py \
+    --robot-path lerobot/configs/robot/so100.yaml \
+    --reward-classifier-pretrained-path outputs/classifier/checkpoints/best/pretrained_model \
+    --reward-classifier-config-file lerobot/configs/policy/hilserl_classifier.yaml \
+    --display-cameras 1
+```
+
+**NOTE** (michel-aractingi): This script is incomplete and it is being prepared
+for running training on the real robot.
+"""
+
+import argparse
+import logging
+import time
+
+import cv2
+import numpy as np
+import torch
+from tqdm import trange
+
+from lerobot.common.policies.policy_protocol import Policy
+from lerobot.common.robot_devices.control_utils import (
+    busy_wait,
+    is_headless,
+    reset_follower_position,
+)
+from lerobot.common.robot_devices.robots.factory import Robot, make_robot
+from lerobot.common.utils.utils import (
+    init_hydra_config,
+    init_logging,
+    log_say,
+)
+
+
+def get_classifier(pretrained_path, config_path):
+    if pretrained_path is None or config_path is None:
+        return
+
+    from lerobot.common.policies.factory import _policy_cfg_from_hydra_cfg
+    from lerobot.common.policies.hilserl.classifier.configuration_classifier import (
+        ClassifierConfig,
+    )
+    from lerobot.common.policies.hilserl.classifier.modeling_classifier import (
+        Classifier,
+    )
+
+    cfg = init_hydra_config(config_path)
+
+    classifier_config = _policy_cfg_from_hydra_cfg(ClassifierConfig, cfg)
+    classifier_config.num_cameras = len(cfg.training.image_keys)  # TODO automate these paths
+    model = Classifier(classifier_config)
+    model.load_state_dict(Classifier.from_pretrained(pretrained_path).state_dict())
+    model = model.to("mps")
+    return model
+
+
+def rollout(
+    robot: Robot,
+    policy: Policy,
+    reward_classifier,
+    fps: int,
+    control_time_s: float = 20,
+    use_amp: bool = True,
+    display_cameras: bool = False,
+) -> dict:
+    """Run a batched policy rollout on the real robot.
+
+    The return dictionary contains:
+        "robot": A a dictionary of (batch, sequence + 1, *) tensors mapped to observation
+            keys. NOTE the that this has an extra sequence element relative to the other keys in the
+            dictionary. This is because an extra observation is included for after the environment is
+            terminated or truncated.
+        "action": A (batch, sequence, action_dim) tensor of actions applied based on the observations (not
+            including the last observations).
+        "reward": A (batch, sequence) tensor of rewards received for applying the actions.
+        "success": A (batch, sequence) tensor of success conditions (the only time this can be True is upon
+            environment termination/truncation).
+        "done": A (batch, sequence) tensor of **cumulative** done conditions. For any given batch element,
+            the first True is followed by True's all the way till the end. This can be used for masking
+            extraneous elements from the sequences above.
+
+    Args:
+        robot: The robot class that defines the interface with the real robot.
+        policy: The policy. Must be a PyTorch nn module.
+
+    Returns:
+        The dictionary described above.
+    """
+    # TODO (michel-aractingi): Infer the device from policy parameters when policy is added
+    # assert isinstance(policy, nn.Module), "Policy must be a PyTorch nn module."
+    # device = get_device_from_parameters(policy)
+
+    # define keyboard listener
+    listener, events = init_keyboard_listener()
+
+    # Reset the policy. TODO (michel-aractingi) add real policy evaluation once the code is ready.
+    # policy.reset()
+
+    # NOTE: sorting to make sure the key sequence is the same during training and testing.
+    observation = robot.capture_observation()
+    image_keys = [key for key in observation if "image" in key]
+    image_keys.sort()
+
+    all_actions = []
+    all_rewards = []
+    all_successes = []
+
+    start_episode_t = time.perf_counter()
+    init_pos = robot.follower_arms["main"].read("Present_Position")
+    timestamp = 0.0
+    while timestamp < control_time_s:
+        start_loop_t = time.perf_counter()
+
+        # Apply the next action.
+        while events["pause_policy"] and not events["human_intervention_step"]:
+            busy_wait(0.5)
+
+        if events["human_intervention_step"]:
+            # take over the robot's actions
+            observation, action = robot.teleop_step(record_data=True)
+            action = action["action"]  # teleop step returns torch tensors but in a dict
+        else:
+            # explore with policy
+            with torch.inference_mode():
+                # TODO (michel-aractingi) replace this part with policy (predict_action)
+                action = robot.follower_arms["main"].read("Present_Position")
+                action = torch.from_numpy(action)
+                robot.send_action(action)
+                # action = predict_action(observation, policy, device, use_amp)
+
+        observation = robot.capture_observation()
+        images = []
+        for key in image_keys:
+            if display_cameras:
+                cv2.imshow(key, cv2.cvtColor(observation[key].numpy(), cv2.COLOR_RGB2BGR))
+                cv2.waitKey(1)
+            images.append(observation[key].to("mps"))
+
+        reward = reward_classifier.predict_reward(images) if reward_classifier is not None else 0.0
+        all_rewards.append(reward)
+
+        # print("REWARD : ", reward)
+
+        all_actions.append(action)
+        all_successes.append(torch.tensor([False]))
+
+        dt_s = time.perf_counter() - start_loop_t
+        busy_wait(1 / fps - dt_s)
+        timestamp = time.perf_counter() - start_episode_t
+        if events["exit_early"]:
+            events["exit_early"] = False
+            events["human_intervention_step"] = False
+            events["pause_policy"] = False
+            break
+
+    reset_follower_position(robot, target_position=init_pos)
+
+    dones = torch.tensor([False] * len(all_actions))
+    dones[-1] = True
+    # Stack the sequence along the first dimension so that we have (batch, sequence, *) tensors.
+    ret = {
+        "action": torch.stack(all_actions, dim=1),
+        "next.reward": torch.stack(all_rewards, dim=1),
+        "next.success": torch.stack(all_successes, dim=1),
+        "done": dones,
+    }
+
+    listener.stop()
+
+    return ret
+
+
+def eval_policy(
+    robot: Robot,
+    policy: torch.nn.Module,
+    fps: float,
+    n_episodes: int,
+    control_time_s: int = 20,
+    use_amp: bool = True,
+    display_cameras: bool = False,
+    reward_classifier_pretrained_path: str | None = None,
+    reward_classifier_config_file: str | None = None,
+) -> dict:
+    """
+    Args:
+        env: The batch of environments.
+        policy: The policy.
+        n_episodes: The number of episodes to evaluate.
+    Returns:
+        Dictionary with metrics and data regarding the rollouts.
+    """
+    # TODO (michel-aractingi) comment this out for testing with a fixed policy
+    # assert isinstance(policy, Policy)
+    # policy.eval()
+
+    sum_rewards = []
+    max_rewards = []
+    successes = []
+    rollouts = []
+
+    start_eval = time.perf_counter()
+    progbar = trange(n_episodes, desc="Evaluating policy on real robot")
+    reward_classifier = get_classifier(reward_classifier_pretrained_path, reward_classifier_config_file)
+
+    for _ in progbar:
+        rollout_data = rollout(
+            robot,
+            policy,
+            reward_classifier,
+            fps,
+            control_time_s,
+            use_amp,
+            display_cameras,
+        )
+
+        rollouts.append(rollout_data)
+        sum_rewards.append(sum(rollout_data["next.reward"]))
+        max_rewards.append(max(rollout_data["next.reward"]))
+        successes.append(rollout_data["next.success"][-1])
+
+    info = {
+        "per_episode": [
+            {
+                "episode_ix": i,
+                "sum_reward": sum_reward,
+                "max_reward": max_reward,
+                "pc_success": success * 100,
+            }
+            for i, (sum_reward, max_reward, success) in enumerate(
+                zip(
+                    sum_rewards[:n_episodes],
+                    max_rewards[:n_episodes],
+                    successes[:n_episodes],
+                    strict=False,
+                )
+            )
+        ],
+        "aggregated": {
+            "avg_sum_reward": float(np.nanmean(torch.cat(sum_rewards[:n_episodes]))),
+            "avg_max_reward": float(np.nanmean(torch.cat(max_rewards[:n_episodes]))),
+            "pc_success": float(np.nanmean(torch.cat(successes[:n_episodes])) * 100),
+            "eval_s": time.time() - start_eval,
+            "eval_ep_s": (time.time() - start_eval) / n_episodes,
+        },
+    }
+
+    if robot.is_connected:
+        robot.disconnect()
+
+    return info
+
+
+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.
+    events = {}
+    events["exit_early"] = False
+    events["rerecord_episode"] = False
+    events["pause_policy"] = False
+    events["human_intervention_step"] = False
+
+    if is_headless():
+        logging.warning(
+            "Headless environment detected. On-screen cameras display and keyboard inputs will not be available."
+        )
+        listener = None
+        return listener, events
+
+    # Only import pynput if not in a headless environment
+    from pynput import keyboard
+
+    def on_press(key):
+        try:
+            if key == keyboard.Key.right:
+                print("Right arrow key pressed. Exiting loop...")
+                events["exit_early"] = True
+            elif key == keyboard.Key.left:
+                print("Left arrow key pressed. Exiting loop and rerecord the last episode...")
+                events["rerecord_episode"] = True
+                events["exit_early"] = True
+            elif key == keyboard.Key.space:
+                # check if first space press then pause the policy for the user to get ready
+                # if second space press then the user is ready to start intervention
+                if not events["pause_policy"]:
+                    print(
+                        "Space key pressed. Human intervention required.\n"
+                        "Place the leader in similar pose to the follower and press space again."
+                    )
+                    events["pause_policy"] = True
+                    log_say(
+                        "Human intervention stage. Get ready to take over.",
+                        play_sounds=True,
+                    )
+                else:
+                    events["human_intervention_step"] = True
+                    print("Space key pressed. Human intervention starting.")
+                    log_say("Starting human intervention.", play_sounds=True)
+
+        except Exception as e:
+            print(f"Error handling key press: {e}")
+
+    listener = keyboard.Listener(on_press=on_press)
+    listener.start()
+
+    return listener, events
+
+
+if __name__ == "__main__":
+    init_logging()
+
+    parser = argparse.ArgumentParser(
+        description=__doc__, formatter_class=argparse.RawDescriptionHelpFormatter
+    )
+    group = parser.add_mutually_exclusive_group(required=True)
+    group.add_argument(
+        "--robot-path",
+        type=str,
+        default="lerobot/configs/robot/koch.yaml",
+        help="Path to robot yaml file used to instantiate the robot using `make_robot` factory function.",
+    )
+    group.add_argument(
+        "--robot-overrides",
+        type=str,
+        nargs="*",
+        help="Any key=value arguments to override config values (use dots for.nested=overrides)",
+    )
+    group.add_argument(
+        "-p",
+        "--pretrained-policy-name-or-path",
+        help=(
+            "Either the repo ID of a model hosted on the Hub or a path to a directory containing weights "
+            "saved using `Policy.save_pretrained`. If not provided, the policy is initialized from scratch "
+            "(useful for debugging). This argument is mutually exclusive with `--config`."
+        ),
+    )
+    group.add_argument(
+        "--config",
+        help=(
+            "Path to a yaml config you want to use for initializing a policy from scratch (useful for "
+            "debugging). This argument is mutually exclusive with `--pretrained-policy-name-or-path` (`-p`)."
+        ),
+    )
+    parser.add_argument("--revision", help="Optionally provide the Hugging Face Hub revision ID.")
+    parser.add_argument(
+        "--out-dir",
+        help=(
+            "Where to save the evaluation outputs. If not provided, outputs are saved in "
+            "outputs/eval/{timestamp}_{env_name}_{policy_name}"
+        ),
+    )
+    parser.add_argument(
+        "--display-cameras",
+        help=("Whether to display the camera feed while the rollout is happening"),
+    )
+    parser.add_argument(
+        "--reward-classifier-pretrained-path",
+        type=str,
+        default=None,
+        help="Path to the pretrained classifier weights.",
+    )
+    parser.add_argument(
+        "--reward-classifier-config-file",
+        type=str,
+        default=None,
+        help="Path to a yaml config file that is necessary to build the reward classifier model.",
+    )
+
+    args = parser.parse_args()
+
+    robot_cfg = init_hydra_config(args.robot_path, args.robot_overrides)
+    robot = make_robot(robot_cfg)
+    if not robot.is_connected:
+        robot.connect()
+
+    eval_policy(
+        robot,
+        None,
+        fps=40,
+        n_episodes=2,
+        control_time_s=100,
+        display_cameras=args.display_cameras,
+        reward_classifier_config_file=args.reward_classifier_config_file,
+        reward_classifier_pretrained_path=args.reward_classifier_pretrained_path,
+    )

lerobot/scripts/server/actor_server.py

@@ -0,0 +1,675 @@
+#!/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 logging
+import os
+import time
+from functools import lru_cache
+from queue import Empty
+from statistics import mean, quantiles
+
+# from lerobot.scripts.eval import eval_policy
+import grpc
+import torch
+from torch import nn
+from torch.multiprocessing import Event, Queue
+
+# TODO: Remove the import of maniskill
+from lerobot.common.policies.factory import make_policy
+from lerobot.common.policies.sac.modeling_sac import SACPolicy
+from lerobot.common.robot_devices.utils import busy_wait
+from lerobot.common.utils.random_utils import set_seed
+from lerobot.common.utils.utils import (
+    TimerManager,
+    get_safe_torch_device,
+    init_logging,
+)
+from lerobot.configs import parser
+from lerobot.configs.train import TrainPipelineConfig
+from lerobot.scripts.server import hilserl_pb2, hilserl_pb2_grpc, learner_service
+from lerobot.scripts.server.buffer import (
+    Transition,
+    bytes_to_state_dict,
+    move_state_dict_to_device,
+    move_transition_to_device,
+    python_object_to_bytes,
+    transitions_to_bytes,
+)
+from lerobot.scripts.server.gym_manipulator import make_robot_env
+from lerobot.scripts.server.network_utils import (
+    receive_bytes_in_chunks,
+    send_bytes_in_chunks,
+)
+from lerobot.scripts.server.utils import get_last_item_from_queue, setup_process_handlers
+
+ACTOR_SHUTDOWN_TIMEOUT = 30
+
+
+#################################################
+# Main entry point #
+#################################################
+
+
+@parser.wrap()
+def actor_cli(cfg: TrainPipelineConfig):
+    cfg.validate()
+    if not use_threads(cfg):
+        import torch.multiprocessing as mp
+
+        mp.set_start_method("spawn")
+
+    # Create logs directory to ensure it exists
+    log_dir = os.path.join(cfg.output_dir, "logs")
+    os.makedirs(log_dir, exist_ok=True)
+    log_file = os.path.join(log_dir, f"actor_{cfg.job_name}.log")
+
+    # Initialize logging with explicit log file
+    init_logging(log_file=log_file)
+    logging.info(f"Actor logging initialized, writing to {log_file}")
+
+    shutdown_event = setup_process_handlers(use_threads(cfg))
+
+    learner_client, grpc_channel = learner_service_client(
+        host=cfg.policy.actor_learner_config.learner_host,
+        port=cfg.policy.actor_learner_config.learner_port,
+    )
+
+    logging.info("[ACTOR] Establishing connection with Learner")
+    if not establish_learner_connection(learner_client, shutdown_event):
+        logging.error("[ACTOR] Failed to establish connection with Learner")
+        return
+
+    if not use_threads(cfg):
+        # If we use multithreading, we can reuse the channel
+        grpc_channel.close()
+        grpc_channel = None
+
+    logging.info("[ACTOR] Connection with Learner established")
+
+    parameters_queue = Queue()
+    transitions_queue = Queue()
+    interactions_queue = Queue()
+
+    concurrency_entity = None
+    if use_threads(cfg):
+        from threading import Thread
+
+        concurrency_entity = Thread
+    else:
+        from multiprocessing import Process
+
+        concurrency_entity = Process
+
+    receive_policy_process = concurrency_entity(
+        target=receive_policy,
+        args=(cfg, parameters_queue, shutdown_event, grpc_channel),
+        daemon=True,
+    )
+
+    transitions_process = concurrency_entity(
+        target=send_transitions,
+        args=(cfg, transitions_queue, shutdown_event, grpc_channel),
+        daemon=True,
+    )
+
+    interactions_process = concurrency_entity(
+        target=send_interactions,
+        args=(cfg, interactions_queue, shutdown_event, grpc_channel),
+        daemon=True,
+    )
+
+    transitions_process.start()
+    interactions_process.start()
+    receive_policy_process.start()
+
+    # HACK: FOR MANISKILL we do not have a reward classifier
+    # TODO: Remove this once we merge into main
+    reward_classifier = None
+    # if (
+    #     cfg.env.reward_classifier["pretrained_path"] is not None
+    #     and cfg.env.reward_classifier["config_path"] is not None
+    # ):
+    #     reward_classifier = get_classifier(
+    #         pretrained_path=cfg.env.reward_classifier["pretrained_path"],
+    #         config_path=cfg.env.reward_classifier["config_path"],
+    #     )
+
+    act_with_policy(
+        cfg=cfg,
+        reward_classifier=reward_classifier,
+        shutdown_event=shutdown_event,
+        parameters_queue=parameters_queue,
+        transitions_queue=transitions_queue,
+        interactions_queue=interactions_queue,
+    )
+    logging.info("[ACTOR] Policy process joined")
+
+    logging.info("[ACTOR] Closing queues")
+    transitions_queue.close()
+    interactions_queue.close()
+    parameters_queue.close()
+
+    transitions_process.join()
+    logging.info("[ACTOR] Transitions process joined")
+    interactions_process.join()
+    logging.info("[ACTOR] Interactions process joined")
+    receive_policy_process.join()
+    logging.info("[ACTOR] Receive policy process joined")
+
+    logging.info("[ACTOR] join queues")
+    transitions_queue.cancel_join_thread()
+    interactions_queue.cancel_join_thread()
+    parameters_queue.cancel_join_thread()
+
+    logging.info("[ACTOR] queues closed")
+
+
+#################################################
+# Core algorithm functions #
+#################################################
+
+
+def act_with_policy(
+    cfg: TrainPipelineConfig,
+    reward_classifier: nn.Module,
+    shutdown_event: any,  # Event,
+    parameters_queue: Queue,
+    transitions_queue: Queue,
+    interactions_queue: Queue,
+):
+    """
+    Executes policy interaction within the environment.
+
+    This function rolls out the policy in the environment, collecting interaction data and pushing it to a queue for streaming to the learner.
+    Once an episode is completed, updated network parameters received from the learner are retrieved from a queue and loaded into the network.
+
+    Args:
+        cfg: Configuration settings for the interaction process.
+        reward_classifier: Reward classifier to use for the interaction process.
+        shutdown_event: Event to check if the process should shutdown.
+        parameters_queue: Queue to receive updated network parameters from the learner.
+        transitions_queue: Queue to send transitions to the learner.
+        interactions_queue: Queue to send interactions to the learner.
+    """
+    # Initialize logging for multiprocessing
+    if not use_threads(cfg):
+        log_dir = os.path.join(cfg.output_dir, "logs")
+        os.makedirs(log_dir, exist_ok=True)
+        log_file = os.path.join(log_dir, f"actor_policy_{os.getpid()}.log")
+        init_logging(log_file=log_file)
+        logging.info("Actor policy process logging initialized")
+
+    logging.info("make_env online")
+
+    online_env = make_robot_env(cfg=cfg.env)
+
+    set_seed(cfg.seed)
+    device = get_safe_torch_device(cfg.policy.device, log=True)
+
+    torch.backends.cudnn.benchmark = True
+    torch.backends.cuda.matmul.allow_tf32 = True
+
+    logging.info("make_policy")
+
+    ### Instantiate the policy in both the actor and learner processes
+    ### To avoid sending a SACPolicy object through the port, we create a policy intance
+    ### on both sides, the learner sends the updated parameters every n steps to update the actor's parameters
+    # TODO: At some point we should just need make sac policy
+    policy: SACPolicy = make_policy(
+        cfg=cfg.policy,
+        env_cfg=cfg.env,
+    )
+    assert isinstance(policy, nn.Module)
+
+    obs, info = online_env.reset()
+
+    # NOTE: For the moment we will solely handle the case of a single environment
+    sum_reward_episode = 0
+    list_transition_to_send_to_learner = []
+    list_policy_time = []
+    episode_intervention = False
+    # Add counters for intervention rate calculation
+    episode_intervention_steps = 0
+    episode_total_steps = 0
+
+    for interaction_step in range(cfg.policy.online_steps):
+        start_time = time.perf_counter()
+        if shutdown_event.is_set():
+            logging.info("[ACTOR] Shutting down act_with_policy")
+            return
+
+        if interaction_step >= cfg.policy.online_step_before_learning:
+            # Time policy inference and check if it meets FPS requirement
+            with TimerManager(
+                elapsed_time_list=list_policy_time,
+                label="Policy inference time",
+                log=False,
+            ) as timer:  # noqa: F841
+                action = policy.select_action(batch=obs)
+            policy_fps = 1.0 / (list_policy_time[-1] + 1e-9)
+
+            log_policy_frequency_issue(policy_fps=policy_fps, cfg=cfg, interaction_step=interaction_step)
+
+            next_obs, reward, done, truncated, info = online_env.step(action.squeeze(dim=0).cpu().numpy())
+        else:
+            # TODO (azouitine): Make a custom space for torch tensor
+            action = online_env.action_space.sample()
+            next_obs, reward, done, truncated, info = online_env.step(action)
+
+            # HACK: We have only one env but we want to batch it, it will be resolved with the torch box
+            action = (
+                torch.from_numpy(action[0]).to(device, non_blocking=device.type == "cuda").unsqueeze(dim=0)
+            )
+
+        sum_reward_episode += float(reward)
+        # Increment total steps counter for intervention rate
+        episode_total_steps += 1
+
+        # NOTE: We overide the action if the intervention is True, because the action applied is the intervention action
+        if "is_intervention" in info and info["is_intervention"]:
+            # TODO: Check the shape
+            # NOTE: The action space for demonstration before hand is with the full action space
+            # but sometimes for example we want to deactivate the gripper
+            action = info["action_intervention"]
+            episode_intervention = True
+            # Increment intervention steps counter
+            episode_intervention_steps += 1
+
+        # Check for NaN values in observations
+        for key, tensor in obs.items():
+            if torch.isnan(tensor).any():
+                logging.error(f"[ACTOR] NaN values found in obs[{key}] at step {interaction_step}")
+
+        list_transition_to_send_to_learner.append(
+            Transition(
+                state=obs,
+                action=action,
+                reward=reward,
+                next_state=next_obs,
+                done=done,
+                truncated=truncated,  # TODO: (azouitine) Handle truncation properly
+                complementary_info=info,  # TODO Handle information for the transition, is_demonstraction: bool
+            )
+        )
+        # assign obs to the next obs and continue the rollout
+        obs = next_obs
+
+        # HACK: We have only one env but we want to batch it, it will be resolved with the torch box
+        # Because we are using a single environment we can index at zero
+        if done or truncated:
+            # TODO: Handle logging for episode information
+            logging.info(f"[ACTOR] Global step {interaction_step}: Episode reward: {sum_reward_episode}")
+
+            update_policy_parameters(policy=policy.actor, parameters_queue=parameters_queue, device=device)
+
+            if len(list_transition_to_send_to_learner) > 0:
+                push_transitions_to_transport_queue(
+                    transitions=list_transition_to_send_to_learner,
+                    transitions_queue=transitions_queue,
+                )
+                list_transition_to_send_to_learner = []
+
+            stats = get_frequency_stats(list_policy_time)
+            list_policy_time.clear()
+
+            # Calculate intervention rate
+            intervention_rate = 0.0
+            if episode_total_steps > 0:
+                intervention_rate = episode_intervention_steps / episode_total_steps
+
+            # Send episodic reward to the learner
+            interactions_queue.put(
+                python_object_to_bytes(
+                    {
+                        "Episodic reward": sum_reward_episode,
+                        "Interaction step": interaction_step,
+                        "Episode intervention": int(episode_intervention),
+                        "Intervention rate": intervention_rate,
+                        **stats,
+                    }
+                )
+            )
+            sum_reward_episode = 0.0
+            episode_intervention = False
+            # Reset intervention counters
+            episode_intervention_steps = 0
+            episode_total_steps = 0
+            obs, info = online_env.reset()
+
+        if cfg.env.fps is not None:
+            dt_time = time.perf_counter() - start_time
+            busy_wait(1 / cfg.env.fps - dt_time)
+
+
+#################################################
+#  Communication Functions - Group all gRPC/messaging functions  #
+#################################################
+
+
+def establish_learner_connection(
+    stub,
+    shutdown_event: any,  # Event,
+    attempts=30,
+):
+    for _ in range(attempts):
+        if shutdown_event.is_set():
+            logging.info("[ACTOR] Shutting down establish_learner_connection")
+            return False
+
+        # Force a connection attempt and check state
+        try:
+            logging.info("[ACTOR] Send ready message to Learner")
+            if stub.Ready(hilserl_pb2.Empty()) == hilserl_pb2.Empty():
+                return True
+        except grpc.RpcError as e:
+            logging.error(f"[ACTOR] Waiting for Learner to be ready... {e}")
+            time.sleep(2)
+    return False
+
+
+@lru_cache(maxsize=1)
+def learner_service_client(
+    host="127.0.0.1", port=50051
+) -> tuple[hilserl_pb2_grpc.LearnerServiceStub, grpc.Channel]:
+    import json
+
+    """
+    Returns a client for the learner service.
+
+    GRPC uses HTTP/2, which is a binary protocol and multiplexes requests over a single connection.
+    So we need to create only one client and reuse it.
+    """
+
+    service_config = {
+        "methodConfig": [
+            {
+                "name": [{}],  # Applies to ALL methods in ALL services
+                "retryPolicy": {
+                    "maxAttempts": 5,  # Max retries (total attempts = 5)
+                    "initialBackoff": "0.1s",  # First retry after 0.1s
+                    "maxBackoff": "2s",  # Max wait time between retries
+                    "backoffMultiplier": 2,  # Exponential backoff factor
+                    "retryableStatusCodes": [
+                        "UNAVAILABLE",
+                        "DEADLINE_EXCEEDED",
+                    ],  # Retries on network failures
+                },
+            }
+        ]
+    }
+
+    service_config_json = json.dumps(service_config)
+
+    channel = grpc.insecure_channel(
+        f"{host}:{port}",
+        options=[
+            ("grpc.max_receive_message_length", learner_service.MAX_MESSAGE_SIZE),
+            ("grpc.max_send_message_length", learner_service.MAX_MESSAGE_SIZE),
+            ("grpc.enable_retries", 1),
+            ("grpc.service_config", service_config_json),
+        ],
+    )
+    stub = hilserl_pb2_grpc.LearnerServiceStub(channel)
+    logging.info("[ACTOR] Learner service client created")
+    return stub, channel
+
+
+def receive_policy(
+    cfg: TrainPipelineConfig,
+    parameters_queue: Queue,
+    shutdown_event: any,  # Event,
+    learner_client: hilserl_pb2_grpc.LearnerServiceStub | None = None,
+    grpc_channel: grpc.Channel | None = None,
+):
+    logging.info("[ACTOR] Start receiving parameters from the Learner")
+
+    if not use_threads(cfg):
+        # Create a process-specific log file
+        log_dir = os.path.join(cfg.output_dir, "logs")
+        os.makedirs(log_dir, exist_ok=True)
+        log_file = os.path.join(log_dir, f"actor_receive_policy_{os.getpid()}.log")
+
+        # Initialize logging with explicit log file
+        init_logging(log_file=log_file)
+        logging.info("Actor receive policy process logging initialized")
+
+        # Setup process handlers to handle shutdown signal
+        # But use shutdown event from the main process
+        setup_process_handlers(use_threads=False)
+
+    if grpc_channel is None or learner_client is None:
+        learner_client, grpc_channel = learner_service_client(
+            host=cfg.policy.actor_learner_config.learner_host,
+            port=cfg.policy.actor_learner_config.learner_port,
+        )
+
+    try:
+        iterator = learner_client.StreamParameters(hilserl_pb2.Empty())
+        receive_bytes_in_chunks(
+            iterator,
+            parameters_queue,
+            shutdown_event,
+            log_prefix="[ACTOR] parameters",
+        )
+
+    except grpc.RpcError as e:
+        logging.error(f"[ACTOR] gRPC error: {e}")
+
+    if not use_threads(cfg):
+        grpc_channel.close()
+    logging.info("[ACTOR] Received policy loop stopped")
+
+
+def send_transitions(
+    cfg: TrainPipelineConfig,
+    transitions_queue: Queue,
+    shutdown_event: any,  # Event,
+    learner_client: hilserl_pb2_grpc.LearnerServiceStub | None = None,
+    grpc_channel: grpc.Channel | None = None,
+) -> hilserl_pb2.Empty:
+    """
+    Sends transitions to the learner.
+
+    This function continuously retrieves messages from the queue and processes:
+
+    - **Transition Data:**
+        - A batch of transitions (observation, action, reward, next observation) is collected.
+        - Transitions are moved to the CPU and serialized using PyTorch.
+        - The serialized data is wrapped in a `hilserl_pb2.Transition` message and sent to the learner.
+    """
+
+    if not use_threads(cfg):
+        # Create a process-specific log file
+        log_dir = os.path.join(cfg.output_dir, "logs")
+        os.makedirs(log_dir, exist_ok=True)
+        log_file = os.path.join(log_dir, f"actor_transitions_{os.getpid()}.log")
+
+        # Initialize logging with explicit log file
+        init_logging(log_file=log_file)
+        logging.info("Actor transitions process logging initialized")
+
+        # Setup process handlers to handle shutdown signal
+        # But use shutdown event from the main process
+        setup_process_handlers(False)
+
+    if grpc_channel is None or learner_client is None:
+        learner_client, grpc_channel = learner_service_client(
+            host=cfg.policy.actor_learner_config.learner_host,
+            port=cfg.policy.actor_learner_config.learner_port,
+        )
+
+    try:
+        learner_client.SendTransitions(transitions_stream(shutdown_event, transitions_queue))
+    except grpc.RpcError as e:
+        logging.error(f"[ACTOR] gRPC error: {e}")
+
+    logging.info("[ACTOR] Finished streaming transitions")
+
+    if not use_threads(cfg):
+        grpc_channel.close()
+    logging.info("[ACTOR] Transitions process stopped")
+
+
+def send_interactions(
+    cfg: TrainPipelineConfig,
+    interactions_queue: Queue,
+    shutdown_event: any,  # Event,
+    learner_client: hilserl_pb2_grpc.LearnerServiceStub | None = None,
+    grpc_channel: grpc.Channel | None = None,
+) -> hilserl_pb2.Empty:
+    """
+    Sends interactions to the learner.
+
+    This function continuously retrieves messages from the queue and processes:
+
+    - **Interaction Messages:**
+        - Contains useful statistics about episodic rewards and policy timings.
+        - The message is serialized using `pickle` and sent to the learner.
+    """
+
+    if not use_threads(cfg):
+        # Create a process-specific log file
+        log_dir = os.path.join(cfg.output_dir, "logs")
+        os.makedirs(log_dir, exist_ok=True)
+        log_file = os.path.join(log_dir, f"actor_interactions_{os.getpid()}.log")
+
+        # Initialize logging with explicit log file
+        init_logging(log_file=log_file)
+        logging.info("Actor interactions process logging initialized")
+
+        # Setup process handlers to handle shutdown signal
+        # But use shutdown event from the main process
+        setup_process_handlers(False)
+
+    if grpc_channel is None or learner_client is None:
+        learner_client, grpc_channel = learner_service_client(
+            host=cfg.policy.actor_learner_config.learner_host,
+            port=cfg.policy.actor_learner_config.learner_port,
+        )
+
+    try:
+        learner_client.SendInteractions(interactions_stream(shutdown_event, interactions_queue))
+    except grpc.RpcError as e:
+        logging.error(f"[ACTOR] gRPC error: {e}")
+
+    logging.info("[ACTOR] Finished streaming interactions")
+
+    if not use_threads(cfg):
+        grpc_channel.close()
+    logging.info("[ACTOR] Interactions process stopped")
+
+
+def transitions_stream(shutdown_event: Event, transitions_queue: Queue) -> hilserl_pb2.Empty:
+    while not shutdown_event.is_set():
+        try:
+            message = transitions_queue.get(block=True, timeout=5)
+        except Empty:
+            logging.debug("[ACTOR] Transition queue is empty")
+            continue
+
+        yield from send_bytes_in_chunks(
+            message, hilserl_pb2.Transition, log_prefix="[ACTOR] Send transitions"
+        )
+
+    return hilserl_pb2.Empty()
+
+
+def interactions_stream(
+    shutdown_event: any,  # Event,
+    interactions_queue: Queue,
+) -> hilserl_pb2.Empty:
+    while not shutdown_event.is_set():
+        try:
+            message = interactions_queue.get(block=True, timeout=5)
+        except Empty:
+            logging.debug("[ACTOR] Interaction queue is empty")
+            continue
+
+        yield from send_bytes_in_chunks(
+            message,
+            hilserl_pb2.InteractionMessage,
+            log_prefix="[ACTOR] Send interactions",
+        )
+
+    return hilserl_pb2.Empty()
+
+
+#################################################
+#  Policy functions #
+#################################################
+
+
+def update_policy_parameters(policy: SACPolicy, parameters_queue: Queue, device):
+    if not parameters_queue.empty():
+        logging.info("[ACTOR] Load new parameters from Learner.")
+        bytes_state_dict = get_last_item_from_queue(parameters_queue)
+        state_dict = bytes_to_state_dict(bytes_state_dict)
+        state_dict = move_state_dict_to_device(state_dict, device=device)
+        policy.load_state_dict(state_dict)
+
+
+#################################################
+#  Utilities functions #
+#################################################
+
+
+def push_transitions_to_transport_queue(transitions: list, transitions_queue):
+    """Send transitions to learner in smaller chunks to avoid network issues.
+
+    Args:
+        transitions: List of transitions to send
+        message_queue: Queue to send messages to learner
+        chunk_size: Size of each chunk to send
+    """
+    transition_to_send_to_learner = []
+    for transition in transitions:
+        tr = move_transition_to_device(transition=transition, device="cpu")
+        for key, value in tr["state"].items():
+            if torch.isnan(value).any():
+                logging.warning(f"Found NaN values in transition {key}")
+
+        transition_to_send_to_learner.append(tr)
+
+    transitions_queue.put(transitions_to_bytes(transition_to_send_to_learner))
+
+
+def get_frequency_stats(list_policy_time: list[float]) -> dict[str, float]:
+    stats = {}
+    list_policy_fps = [1.0 / t for t in list_policy_time]
+    if len(list_policy_fps) > 1:
+        policy_fps = mean(list_policy_fps)
+        quantiles_90 = quantiles(list_policy_fps, n=10)[-1]
+        logging.debug(f"[ACTOR] Average policy frame rate: {policy_fps}")
+        logging.debug(f"[ACTOR] Policy frame rate 90th percentile: {quantiles_90}")
+        stats = {
+            "Policy frequency [Hz]": policy_fps,
+            "Policy frequency 90th-p [Hz]": quantiles_90,
+        }
+    return stats
+
+
+def log_policy_frequency_issue(policy_fps: float, cfg: TrainPipelineConfig, interaction_step: int):
+    if policy_fps < cfg.env.fps:
+        logging.warning(
+            f"[ACTOR] Policy FPS {policy_fps:.1f} below required {cfg.env.fps} at step {interaction_step}"
+        )
+
+
+def use_threads(cfg: TrainPipelineConfig) -> bool:
+    return cfg.policy.concurrency.actor == "threads"
+
+
+if __name__ == "__main__":
+    actor_cli()

lerobot/scripts/server/crop_dataset_roi.py

@@ -0,0 +1,282 @@
+import argparse  # noqa: I001
+import json
+from copy import deepcopy
+from typing import Dict, Tuple
+from pathlib import Path
+import cv2
+
+# import torch.nn.functional as F  # noqa: N812
+import torchvision.transforms.functional as F  # type: ignore  # noqa: N812
+from tqdm import tqdm  # type: ignore
+
+from lerobot.common.datasets.lerobot_dataset import LeRobotDataset
+
+
+def select_rect_roi(img):
+    """
+    Allows the user to draw a rectangular ROI on the image.
+
+    The user must click and drag to draw the rectangle.
+    - While dragging, the rectangle is dynamically drawn.
+    - On mouse button release, the rectangle is fixed.
+    - Press 'c' to confirm the selection.
+    - Press 'r' to reset the selection.
+    - Press ESC to cancel.
+
+    Returns:
+        A tuple (top, left, height, width) representing the rectangular ROI,
+        or None if no valid ROI is selected.
+    """
+    # Create a working copy of the image
+    clone = img.copy()
+    working_img = clone.copy()
+
+    roi = None  # Will store the final ROI as (top, left, height, width)
+    drawing = False
+    ix, iy = -1, -1  # Initial click coordinates
+
+    def mouse_callback(event, x, y, flags, param):
+        nonlocal ix, iy, drawing, roi, working_img
+
+        if event == cv2.EVENT_LBUTTONDOWN:
+            # Start drawing: record starting coordinates
+            drawing = True
+            ix, iy = x, y
+
+        elif event == cv2.EVENT_MOUSEMOVE:
+            if drawing:
+                # Compute the top-left and bottom-right corners regardless of drag direction
+                top = min(iy, y)
+                left = min(ix, x)
+                bottom = max(iy, y)
+                right = max(ix, x)
+                # Show a temporary image with the current rectangle drawn
+                temp = working_img.copy()
+                cv2.rectangle(temp, (left, top), (right, bottom), (0, 255, 0), 2)
+                cv2.imshow("Select ROI", temp)
+
+        elif event == cv2.EVENT_LBUTTONUP:
+            # Finish drawing
+            drawing = False
+            top = min(iy, y)
+            left = min(ix, x)
+            bottom = max(iy, y)
+            right = max(ix, x)
+            height = bottom - top
+            width = right - left
+            roi = (top, left, height, width)  # (top, left, height, width)
+            # Draw the final rectangle on the working image and display it
+            working_img = clone.copy()
+            cv2.rectangle(working_img, (left, top), (right, bottom), (0, 255, 0), 2)
+            cv2.imshow("Select ROI", working_img)
+
+    # Create the window and set the callback
+    cv2.namedWindow("Select ROI")
+    cv2.setMouseCallback("Select ROI", mouse_callback)
+    cv2.imshow("Select ROI", working_img)
+
+    print("Instructions for ROI selection:")
+    print("  - Click and drag to draw a rectangular ROI.")
+    print("  - Press 'c' to confirm the selection.")
+    print("  - Press 'r' to reset and draw again.")
+    print("  - Press ESC to cancel the selection.")
+
+    # Wait until the user confirms with 'c', resets with 'r', or cancels with ESC
+    while True:
+        key = cv2.waitKey(1) & 0xFF
+        # Confirm ROI if one has been drawn
+        if key == ord("c") and roi is not None:
+            break
+        # Reset: clear the ROI and restore the original image
+        elif key == ord("r"):
+            working_img = clone.copy()
+            roi = None
+            cv2.imshow("Select ROI", working_img)
+        # Cancel selection for this image
+        elif key == 27:  # ESC key
+            roi = None
+            break
+
+    cv2.destroyWindow("Select ROI")
+    return roi
+
+
+def select_square_roi_for_images(images: dict) -> dict:
+    """
+    For each image in the provided dictionary, open a window to allow the user
+    to select a rectangular ROI. Returns a dictionary mapping each key to a tuple
+    (top, left, height, width) representing the ROI.
+
+    Parameters:
+        images (dict): Dictionary where keys are identifiers and values are OpenCV images.
+
+    Returns:
+        dict: Mapping of image keys to the selected rectangular ROI.
+    """
+    selected_rois = {}
+
+    for key, img in images.items():
+        if img is None:
+            print(f"Image for key '{key}' is None, skipping.")
+            continue
+
+        print(f"\nSelect rectangular ROI for image with key: '{key}'")
+        roi = select_rect_roi(img)
+
+        if roi is None:
+            print(f"No valid ROI selected for '{key}'.")
+        else:
+            selected_rois[key] = roi
+            print(f"ROI for '{key}': {roi}")
+
+    return selected_rois
+
+
+def get_image_from_lerobot_dataset(dataset: LeRobotDataset):
+    """
+    Find the first row in the dataset and extract the image in order to be used for the crop.
+    """
+    row = dataset[0]
+    image_dict = {}
+    for k in row:
+        if "image" in k:
+            image_dict[k] = deepcopy(row[k])
+    return image_dict
+
+
+def convert_lerobot_dataset_to_cropper_lerobot_dataset(
+    original_dataset: LeRobotDataset,
+    crop_params_dict: Dict[str, Tuple[int, int, int, int]],
+    new_repo_id: str,
+    new_dataset_root: str,
+    resize_size: Tuple[int, int] = (128, 128),
+) -> LeRobotDataset:
+    """
+    Converts an existing LeRobotDataset by iterating over its episodes and frames,
+    applying cropping and resizing to image observations, and saving a new dataset
+    with the transformed data.
+
+    Args:
+        original_dataset (LeRobotDataset): The source dataset.
+        crop_params_dict (Dict[str, Tuple[int, int, int, int]]):
+            A dictionary mapping observation keys to crop parameters (top, left, height, width).
+        new_repo_id (str): Repository id for the new dataset.
+        new_dataset_root (str): The root directory where the new dataset will be written.
+        resize_size (Tuple[int, int], optional): The target size (height, width) after cropping.
+            Defaults to (128, 128).
+
+    Returns:
+        LeRobotDataset: A new LeRobotDataset where the specified image observations have been cropped
+                        and resized.
+    """
+    # 1. Create a new (empty) LeRobotDataset for writing.
+    new_dataset = LeRobotDataset.create(
+        repo_id=new_repo_id,
+        fps=original_dataset.fps,
+        root=new_dataset_root,
+        robot_type=original_dataset.meta.robot_type,
+        features=original_dataset.meta.info["features"],
+        use_videos=len(original_dataset.meta.video_keys) > 0,
+    )
+
+    # Update the metadata for every image key that will be cropped:
+    # (Here we simply set the shape to be the final resize_size.)
+    for key in crop_params_dict:
+        if key in new_dataset.meta.info["features"]:
+            new_dataset.meta.info["features"][key]["shape"] = list(resize_size)
+
+    # 2. Process each episode in the original dataset.
+    episodes_info = original_dataset.meta.episodes
+    # (Sort episodes by episode_index for consistency.)
+
+    episodes_info = sorted(episodes_info, key=lambda x: x["episode_index"])
+    # Use the first task from the episode metadata (or "unknown" if not provided)
+    task = episodes_info[0]["tasks"][0] if episodes_info[0].get("tasks") else "unknown"
+
+    last_episode_index = 0
+    for sample in tqdm(original_dataset):
+        episode_index = sample.pop("episode_index")
+        if episode_index != last_episode_index:
+            new_dataset.save_episode(task, encode_videos=True)
+            last_episode_index = episode_index
+        sample.pop("frame_index")
+        # Make a shallow copy of the sample (the values—e.g. torch tensors—are assumed immutable)
+        new_sample = sample.copy()
+        # Loop over each observation key that should be cropped/resized.
+        for key, params in crop_params_dict.items():
+            if key in new_sample:
+                top, left, height, width = params
+                # Apply crop then resize.
+                cropped = F.crop(new_sample[key], top, left, height, width)
+                resized = F.resize(cropped, resize_size)
+                new_sample[key] = resized
+        # Add the transformed frame to the new dataset.
+        new_dataset.add_frame(new_sample)
+
+    # save last episode
+    new_dataset.save_episode(task, encode_videos=True)
+
+    # Optionally, consolidate the new dataset to compute statistics and update video info.
+    new_dataset.consolidate(run_compute_stats=True, keep_image_files=True)
+
+    new_dataset.push_to_hub(tags=None)
+
+    return new_dataset
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser(description="Crop rectangular ROIs from a LeRobot dataset.")
+    parser.add_argument(
+        "--repo-id",
+        type=str,
+        default="lerobot",
+        help="The repository id of the LeRobot dataset to process.",
+    )
+    parser.add_argument(
+        "--root",
+        type=str,
+        default=None,
+        help="The root directory of the LeRobot dataset.",
+    )
+    parser.add_argument(
+        "--crop-params-path",
+        type=str,
+        default=None,
+        help="The path to the JSON file containing the ROIs.",
+    )
+    args = parser.parse_args()
+
+    local_files_only = args.root is not None
+    dataset = LeRobotDataset(repo_id=args.repo_id, root=args.root, local_files_only=local_files_only)
+
+    images = get_image_from_lerobot_dataset(dataset)
+    images = {k: v.cpu().permute(1, 2, 0).numpy() for k, v in images.items()}
+    images = {k: (v * 255).astype("uint8") for k, v in images.items()}
+
+    if args.crop_params_path is None:
+        rois = select_square_roi_for_images(images)
+    else:
+        with open(args.crop_params_path) as f:
+            rois = json.load(f)
+
+    # Print the selected rectangular ROIs
+    print("\nSelected Rectangular Regions of Interest (top, left, height, width):")
+    for key, roi in rois.items():
+        print(f"{key}: {roi}")
+
+    new_repo_id = args.repo_id + "_cropped_resized"
+    new_dataset_root = Path(str(dataset.root) + "_cropped_resized")
+
+    croped_resized_dataset = convert_lerobot_dataset_to_cropper_lerobot_dataset(
+        original_dataset=dataset,
+        crop_params_dict=rois,
+        new_repo_id=new_repo_id,
+        new_dataset_root=new_dataset_root,
+        resize_size=(128, 128),
+    )
+
+    meta_dir = new_dataset_root / "meta"
+    meta_dir.mkdir(exist_ok=True)
+
+    with open(meta_dir / "crop_params.json", "w") as f:
+        json.dump(rois, f, indent=4)

lerobot/scripts/server/end_effector_control_utils.py

@@ -0,0 +1,781 @@
+import argparse
+import logging
+import sys
+import time
+
+import numpy as np
+import torch
+
+from lerobot.common.robot_devices.utils import busy_wait
+from lerobot.scripts.server.kinematics import RobotKinematics
+
+logging.basicConfig(level=logging.INFO)
+
+
+class InputController:
+    """Base class for input controllers that generate motion deltas."""
+
+    def __init__(self, x_step_size=0.01, y_step_size=0.01, z_step_size=0.01):
+        """
+        Initialize the controller.
+
+        Args:
+            x_step_size: Base movement step size in meters
+            y_step_size: Base movement step size in meters
+            z_step_size: Base movement step size in meters
+        """
+        self.x_step_size = x_step_size
+        self.y_step_size = y_step_size
+        self.z_step_size = z_step_size
+        self.running = True
+        self.episode_end_status = None  # None, "success", or "failure"
+        self.intervention_flag = False
+        self.open_gripper_command = False
+        self.close_gripper_command = False
+
+    def start(self):
+        """Start the controller and initialize resources."""
+        pass
+
+    def stop(self):
+        """Stop the controller and release resources."""
+        pass
+
+    def get_deltas(self):
+        """Get the current movement deltas (dx, dy, dz) in meters."""
+        return 0.0, 0.0, 0.0
+
+    def should_quit(self):
+        """Return True if the user has requested to quit."""
+        return not self.running
+
+    def update(self):
+        """Update controller state - call this once per frame."""
+        pass
+
+    def __enter__(self):
+        """Support for use in 'with' statements."""
+        self.start()
+        return self
+
+    def __exit__(self, exc_type, exc_val, exc_tb):
+        """Ensure resources are released when exiting 'with' block."""
+        self.stop()
+
+    def get_episode_end_status(self):
+        """
+        Get the current episode end status.
+
+        Returns:
+            None if episode should continue, "success" or "failure" otherwise
+        """
+        status = self.episode_end_status
+        self.episode_end_status = None  # Reset after reading
+        return status
+
+    def should_intervene(self):
+        """Return True if intervention flag was set."""
+        return self.intervention_flag
+
+    def gripper_command(self):
+        """Return the current gripper command."""
+        if self.open_gripper_command == self.close_gripper_command:
+            return "no-op"
+        elif self.open_gripper_command:
+            return "open"
+        elif self.close_gripper_command:
+            return "close"
+
+
+class KeyboardController(InputController):
+    """Generate motion deltas from keyboard input."""
+
+    def __init__(self, x_step_size=0.01, y_step_size=0.01, z_step_size=0.01):
+        super().__init__(x_step_size, y_step_size, z_step_size)
+        self.key_states = {
+            "forward_x": False,
+            "backward_x": False,
+            "forward_y": False,
+            "backward_y": False,
+            "forward_z": False,
+            "backward_z": False,
+            "quit": False,
+            "success": False,
+            "failure": False,
+        }
+        self.listener = None
+
+    def start(self):
+        """Start the keyboard listener."""
+        from pynput import keyboard
+
+        def on_press(key):
+            try:
+                if key == keyboard.Key.up:
+                    self.key_states["forward_x"] = True
+                elif key == keyboard.Key.down:
+                    self.key_states["backward_x"] = True
+                elif key == keyboard.Key.left:
+                    self.key_states["forward_y"] = True
+                elif key == keyboard.Key.right:
+                    self.key_states["backward_y"] = True
+                elif key == keyboard.Key.shift:
+                    self.key_states["backward_z"] = True
+                elif key == keyboard.Key.shift_r:
+                    self.key_states["forward_z"] = True
+                elif key == keyboard.Key.esc:
+                    self.key_states["quit"] = True
+                    self.running = False
+                    return False
+                elif key == keyboard.Key.enter:
+                    self.key_states["success"] = True
+                    self.episode_end_status = "success"
+                elif key == keyboard.Key.backspace:
+                    self.key_states["failure"] = True
+                    self.episode_end_status = "failure"
+            except AttributeError:
+                pass
+
+        def on_release(key):
+            try:
+                if key == keyboard.Key.up:
+                    self.key_states["forward_x"] = False
+                elif key == keyboard.Key.down:
+                    self.key_states["backward_x"] = False
+                elif key == keyboard.Key.left:
+                    self.key_states["forward_y"] = False
+                elif key == keyboard.Key.right:
+                    self.key_states["backward_y"] = False
+                elif key == keyboard.Key.shift:
+                    self.key_states["backward_z"] = False
+                elif key == keyboard.Key.shift_r:
+                    self.key_states["forward_z"] = False
+                elif key == keyboard.Key.enter:
+                    self.key_states["success"] = False
+                elif key == keyboard.Key.backspace:
+                    self.key_states["failure"] = False
+            except AttributeError:
+                pass
+
+        self.listener = keyboard.Listener(on_press=on_press, on_release=on_release)
+        self.listener.start()
+
+        print("Keyboard controls:")
+        print("  Arrow keys: Move in X-Y plane")
+        print("  Shift and Shift_R: Move in Z axis")
+        print("  Enter: End episode with SUCCESS")
+        print("  Backspace: End episode with FAILURE")
+        print("  ESC: Exit")
+
+    def stop(self):
+        """Stop the keyboard listener."""
+        if self.listener and self.listener.is_alive():
+            self.listener.stop()
+
+    def get_deltas(self):
+        """Get the current movement deltas from keyboard state."""
+        delta_x = delta_y = delta_z = 0.0
+
+        if self.key_states["forward_x"]:
+            delta_x += self.x_step_size
+        if self.key_states["backward_x"]:
+            delta_x -= self.x_step_size
+        if self.key_states["forward_y"]:
+            delta_y += self.y_step_size
+        if self.key_states["backward_y"]:
+            delta_y -= self.y_step_size
+        if self.key_states["forward_z"]:
+            delta_z += self.z_step_size
+        if self.key_states["backward_z"]:
+            delta_z -= self.z_step_size
+
+        return delta_x, delta_y, delta_z
+
+    def should_quit(self):
+        """Return True if ESC was pressed."""
+        return self.key_states["quit"]
+
+    def should_save(self):
+        """Return True if Enter was pressed (save episode)."""
+        return self.key_states["success"] or self.key_states["failure"]
+
+
+class GamepadController(InputController):
+    """Generate motion deltas from gamepad input."""
+
+    def __init__(self, x_step_size=0.01, y_step_size=0.01, z_step_size=0.01, deadzone=0.1):
+        super().__init__(x_step_size, y_step_size, z_step_size)
+        self.deadzone = deadzone
+        self.joystick = None
+        self.intervention_flag = False
+
+    def start(self):
+        """Initialize pygame and the gamepad."""
+        import pygame
+
+        pygame.init()
+        pygame.joystick.init()
+
+        if pygame.joystick.get_count() == 0:
+            logging.error("No gamepad detected. Please connect a gamepad and try again.")
+            self.running = False
+            return
+
+        self.joystick = pygame.joystick.Joystick(0)
+        self.joystick.init()
+        logging.info(f"Initialized gamepad: {self.joystick.get_name()}")
+
+        print("Gamepad controls:")
+        print("  Left analog stick: Move in X-Y plane")
+        print("  Right analog stick (vertical): Move in Z axis")
+        print("  B/Circle button: Exit")
+        print("  Y/Triangle button: End episode with SUCCESS")
+        print("  A/Cross button: End episode with FAILURE")
+        print("  X/Square button: Rerecord episode")
+
+    def stop(self):
+        """Clean up pygame resources."""
+        import pygame
+
+        if pygame.joystick.get_init():
+            if self.joystick:
+                self.joystick.quit()
+            pygame.joystick.quit()
+        pygame.quit()
+
+    def update(self):
+        """Process pygame events to get fresh gamepad readings."""
+        import pygame
+
+        for event in pygame.event.get():
+            if event.type == pygame.JOYBUTTONDOWN:
+                if event.button == 3:
+                    self.episode_end_status = "success"
+                # A button (1) for failure
+                elif event.button == 1:
+                    self.episode_end_status = "failure"
+                # X button (0) for rerecord
+                elif event.button == 0:
+                    self.episode_end_status = "rerecord_episode"
+
+                # RB button (6) for opening gripper
+                elif event.button == 6:
+                    self.open_gripper_command = True
+
+                # LT button (7) for closing gripper
+                elif event.button == 7:
+                    self.close_gripper_command = True
+
+            # Reset episode status on button release
+            elif event.type == pygame.JOYBUTTONUP:
+                if event.button in [0, 2, 3]:
+                    self.episode_end_status = None
+
+                elif event.button == 6:
+                    self.open_gripper_command = False
+
+                elif event.button == 7:
+                    self.close_gripper_command = False
+
+            # Check for RB button (typically button 5) for intervention flag
+            if self.joystick.get_button(5):
+                self.intervention_flag = True
+            else:
+                self.intervention_flag = False
+
+    def get_deltas(self):
+        """Get the current movement deltas from gamepad state."""
+        import pygame
+
+        try:
+            # Read joystick axes
+            # Left stick X and Y (typically axes 0 and 1)
+            x_input = self.joystick.get_axis(0)  # Left/Right
+            y_input = self.joystick.get_axis(1)  # Up/Down (often inverted)
+
+            # Right stick Y (typically axis 3 or 4)
+            z_input = self.joystick.get_axis(3)  # Up/Down for Z
+
+            # Apply deadzone to avoid drift
+            x_input = 0 if abs(x_input) < self.deadzone else x_input
+            y_input = 0 if abs(y_input) < self.deadzone else y_input
+            z_input = 0 if abs(z_input) < self.deadzone else z_input
+
+            # Calculate deltas (note: may need to invert axes depending on controller)
+            delta_x = -y_input * self.y_step_size  # Forward/backward
+            delta_y = -x_input * self.x_step_size  # Left/right
+            delta_z = -z_input * self.z_step_size  # Up/down
+
+            return delta_x, delta_y, delta_z
+
+        except pygame.error:
+            logging.error("Error reading gamepad. Is it still connected?")
+            return 0.0, 0.0, 0.0
+
+
+class GamepadControllerHID(InputController):
+    """Generate motion deltas from gamepad input using HIDAPI."""
+
+    def __init__(
+        self,
+        x_step_size=0.01,
+        y_step_size=0.01,
+        z_step_size=0.01,
+        deadzone=0.1,
+        vendor_id=0x046D,
+        product_id=0xC219,
+    ):
+        """
+        Initialize the HID gamepad controller.
+
+        Args:
+            step_size: Base movement step size in meters
+            z_scale: Scaling factor for Z-axis movement
+            deadzone: Joystick deadzone to prevent drift
+            vendor_id: USB vendor ID of the gamepad (default: Logitech)
+            product_id: USB product ID of the gamepad (default: RumblePad 2)
+        """
+        super().__init__(x_step_size, y_step_size, z_step_size)
+        self.deadzone = deadzone
+        self.vendor_id = vendor_id
+        self.product_id = product_id
+        self.device = None
+        self.device_info = None
+
+        # Movement values (normalized from -1.0 to 1.0)
+        self.left_x = 0.0
+        self.left_y = 0.0
+        self.right_x = 0.0
+        self.right_y = 0.0
+
+        # Button states
+        self.buttons = {}
+        self.quit_requested = False
+        self.save_requested = False
+
+    def find_device(self):
+        """Look for the gamepad device by vendor and product ID."""
+        import hid
+
+        devices = hid.enumerate()
+        for device in devices:
+            if device["vendor_id"] == self.vendor_id and device["product_id"] == self.product_id:
+                logging.info(f"Found gamepad: {device.get('product_string', 'Unknown')}")
+                return device
+
+        logging.error(
+            f"No gamepad with vendor ID 0x{self.vendor_id:04X} and product ID 0x{self.product_id:04X} found"
+        )
+        return None
+
+    def start(self):
+        """Connect to the gamepad using HIDAPI."""
+        import hid
+
+        self.device_info = self.find_device()
+        if not self.device_info:
+            self.running = False
+            return
+
+        try:
+            logging.info(f"Connecting to gamepad at path: {self.device_info['path']}")
+            self.device = hid.device()
+            self.device.open_path(self.device_info["path"])
+            self.device.set_nonblocking(1)
+
+            manufacturer = self.device.get_manufacturer_string()
+            product = self.device.get_product_string()
+            logging.info(f"Connected to {manufacturer} {product}")
+
+            logging.info("Gamepad controls (HID mode):")
+            logging.info("  Left analog stick: Move in X-Y plane")
+            logging.info("  Right analog stick: Move in Z axis (vertical)")
+            logging.info("  Button 1/B/Circle: Exit")
+            logging.info("  Button 2/A/Cross: End episode with SUCCESS")
+            logging.info("  Button 3/X/Square: End episode with FAILURE")
+
+        except OSError as e:
+            logging.error(f"Error opening gamepad: {e}")
+            logging.error("You might need to run this with sudo/admin privileges on some systems")
+            self.running = False
+
+    def stop(self):
+        """Close the HID device connection."""
+        if self.device:
+            self.device.close()
+            self.device = None
+
+    def update(self):
+        """
+        Read and process the latest gamepad data.
+        Due to an issue with the HIDAPI, we need to read the read the device several times in order to get a stable reading
+        """
+        for _ in range(10):
+            self._update()
+
+    def _update(self):
+        """Read and process the latest gamepad data."""
+        if not self.device or not self.running:
+            return
+
+        try:
+            # Read data from the gamepad
+            data = self.device.read(64)
+            if data:
+                # Interpret gamepad data - this will vary by controller model
+                # These offsets are for the Logitech RumblePad 2
+                if len(data) >= 8:
+                    # Normalize joystick values from 0-255 to -1.0-1.0
+                    self.left_x = (data[1] - 128) / 128.0
+                    self.left_y = (data[2] - 128) / 128.0
+                    self.right_x = (data[3] - 128) / 128.0
+                    self.right_y = (data[4] - 128) / 128.0
+
+                    # Apply deadzone
+                    self.left_x = 0 if abs(self.left_x) < self.deadzone else self.left_x
+                    self.left_y = 0 if abs(self.left_y) < self.deadzone else self.left_y
+                    self.right_x = 0 if abs(self.right_x) < self.deadzone else self.right_x
+                    self.right_y = 0 if abs(self.right_y) < self.deadzone else self.right_y
+
+                    # Parse button states (byte 5 in the Logitech RumblePad 2)
+                    buttons = data[5]
+
+                    # Check if RB is pressed then the intervention flag should be set
+                    self.intervention_flag = data[6] in [2, 6, 10, 14]
+
+                    # Check if RT is pressed
+                    self.open_gripper_command = data[6] in [8, 10, 12]
+
+                    # Check if LT is pressed
+                    self.close_gripper_command = data[6] in [4, 6, 12]
+
+                    # Check if Y/Triangle button (bit 7) is pressed for saving
+                    # Check if X/Square button (bit 5) is pressed for failure
+                    # Check if A/Cross button (bit 4) is pressed for rerecording
+                    if buttons & 1 << 7:
+                        self.episode_end_status = "success"
+                    elif buttons & 1 << 5:
+                        self.episode_end_status = "failure"
+                    elif buttons & 1 << 4:
+                        self.episode_end_status = "rerecord_episode"
+                    else:
+                        self.episode_end_status = None
+
+        except OSError as e:
+            logging.error(f"Error reading from gamepad: {e}")
+
+    def get_deltas(self):
+        """Get the current movement deltas from gamepad state."""
+        # Calculate deltas - invert as needed based on controller orientation
+        delta_x = -self.left_y * self.x_step_size  # Forward/backward
+        delta_y = -self.left_x * self.y_step_size  # Left/right
+        delta_z = -self.right_y * self.z_step_size  # Up/down
+
+        return delta_x, delta_y, delta_z
+
+    def should_quit(self):
+        """Return True if quit button was pressed."""
+        return self.quit_requested
+
+    def should_save(self):
+        """Return True if save button was pressed."""
+        return self.save_requested
+
+
+def test_forward_kinematics(robot, fps=10):
+    logging.info("Testing Forward Kinematics")
+    timestep = time.perf_counter()
+    kinematics = RobotKinematics(robot.robot_type)
+    while time.perf_counter() - timestep < 60.0:
+        loop_start_time = time.perf_counter()
+        robot.teleop_step()
+        obs = robot.capture_observation()
+        joint_positions = obs["observation.state"].cpu().numpy()
+        ee_pos = kinematics.fk_gripper_tip(joint_positions)
+        logging.info(f"EE Position: {ee_pos[:3, 3]}")
+        busy_wait(1 / fps - (time.perf_counter() - loop_start_time))
+
+
+def test_inverse_kinematics(robot, fps=10):
+    logging.info("Testing Inverse Kinematics")
+    timestep = time.perf_counter()
+    while time.perf_counter() - timestep < 60.0:
+        loop_start_time = time.perf_counter()
+        obs = robot.capture_observation()
+        joint_positions = obs["observation.state"].cpu().numpy()
+        ee_pos = RobotKinematics.fk_gripper_tip(joint_positions)
+        desired_ee_pos = ee_pos
+        target_joint_state = RobotKinematics.ik(joint_positions, desired_ee_pos, position_only=True)
+        robot.send_action(torch.from_numpy(target_joint_state))
+        logging.info(f"Target Joint State: {target_joint_state}")
+        busy_wait(1 / fps - (time.perf_counter() - loop_start_time))
+
+
+def teleoperate_inverse_kinematics_with_leader(robot, fps=10):
+    logging.info("Testing Inverse Kinematics")
+    kinematics = RobotKinematics(robot.robot_type)
+    timestep = time.perf_counter()
+    while time.perf_counter() - timestep < 60.0:
+        loop_start_time = time.perf_counter()
+        obs = robot.capture_observation()
+        joint_positions = obs["observation.state"].cpu().numpy()
+        ee_pos = kinematics.fk_gripper_tip(joint_positions)
+
+        leader_joint_positions = robot.leader_arms["main"].read("Present_Position")
+        leader_ee = kinematics.fk_gripper_tip(leader_joint_positions)
+
+        desired_ee_pos = leader_ee
+        target_joint_state = kinematics.ik(joint_positions, desired_ee_pos, position_only=True)
+        robot.send_action(torch.from_numpy(target_joint_state))
+        logging.info(f"Leader EE: {leader_ee[:3, 3]}, Follower EE: {ee_pos[:3, 3]}")
+        busy_wait(1 / fps - (time.perf_counter() - loop_start_time))
+
+
+def teleoperate_delta_inverse_kinematics_with_leader(robot, fps=10):
+    logging.info("Testing Delta End-Effector Control")
+    timestep = time.perf_counter()
+
+    # Initial position capture
+    obs = robot.capture_observation()
+    joint_positions = obs["observation.state"].cpu().numpy()
+
+    kinematics = RobotKinematics(robot.robot_type)
+
+    leader_joint_positions = robot.leader_arms["main"].read("Present_Position")
+    initial_leader_ee = kinematics.fk_gripper_tip(leader_joint_positions)
+
+    desired_ee_pos = np.diag(np.ones(4))
+
+    while time.perf_counter() - timestep < 60.0:
+        loop_start_time = time.perf_counter()
+
+        # Get leader state for teleoperation
+        leader_joint_positions = robot.leader_arms["main"].read("Present_Position")
+        leader_ee = kinematics.fk_gripper_tip(leader_joint_positions)
+
+        # Get current state
+        # obs = robot.capture_observation()
+        # joint_positions = obs["observation.state"].cpu().numpy()
+        joint_positions = robot.follower_arms["main"].read("Present_Position")
+        current_ee_pos = kinematics.fk_gripper_tip(joint_positions)
+
+        # Calculate delta between leader and follower end-effectors
+        # Scaling factor can be adjusted for sensitivity
+        scaling_factor = 1.0
+        ee_delta = (leader_ee - initial_leader_ee) * scaling_factor
+
+        # Apply delta to current position
+        desired_ee_pos[0, 3] = current_ee_pos[0, 3] + ee_delta[0, 3]
+        desired_ee_pos[1, 3] = current_ee_pos[1, 3] + ee_delta[1, 3]
+        desired_ee_pos[2, 3] = current_ee_pos[2, 3] + ee_delta[2, 3]
+
+        if np.any(np.abs(ee_delta[:3, 3]) > 0.01):
+            # Compute joint targets via inverse kinematics
+            target_joint_state = kinematics.ik(joint_positions, desired_ee_pos, position_only=True)
+
+            initial_leader_ee = leader_ee.copy()
+
+            # Send command to robot
+            robot.send_action(torch.from_numpy(target_joint_state))
+
+            # Logging
+            logging.info(f"Current EE: {current_ee_pos[:3, 3]}, Desired EE: {desired_ee_pos[:3, 3]}")
+            logging.info(f"Delta EE: {ee_delta[:3, 3]}")
+
+        busy_wait(1 / fps - (time.perf_counter() - loop_start_time))
+
+
+def teleoperate_delta_inverse_kinematics(robot, controller, fps=10, bounds=None, fk_func=None):
+    """
+    Control a robot using delta end-effector movements from any input controller.
+
+    Args:
+        robot: Robot instance to control
+        controller: InputController instance (keyboard, gamepad, etc.)
+        fps: Control frequency in Hz
+        bounds: Optional position limits
+        fk_func: Forward kinematics function to use
+    """
+    if fk_func is None:
+        fk_func = RobotKinematics.fk_gripper_tip
+
+    logging.info(f"Testing Delta End-Effector Control with {controller.__class__.__name__}")
+
+    # Initial position capture
+    obs = robot.capture_observation()
+    joint_positions = obs["observation.state"].cpu().numpy()
+    kinematics = RobotKinematics(robot.robot_type)
+    current_ee_pos = kinematics.fk_gripper_tip(joint_positions)
+
+    # Initialize desired position with current position
+    desired_ee_pos = np.eye(4)  # Identity matrix
+
+    timestep = time.perf_counter()
+    with controller:
+        while not controller.should_quit() and time.perf_counter() - timestep < 60.0:
+            loop_start_time = time.perf_counter()
+
+            # Process input events
+            controller.update()
+
+            # Get currrent robot state
+            joint_positions = robot.follower_arms["main"].read("Present_Position")
+            current_ee_pos = kinematics.fk_gripper_tip(joint_positions)
+
+            # Get movement deltas from the controller
+            delta_x, delta_y, delta_z = controller.get_deltas()
+
+            # Update desired position
+            desired_ee_pos[0, 3] = current_ee_pos[0, 3] + delta_x
+            desired_ee_pos[1, 3] = current_ee_pos[1, 3] + delta_y
+            desired_ee_pos[2, 3] = current_ee_pos[2, 3] + delta_z
+
+            # Apply bounds if provided
+            if bounds is not None:
+                desired_ee_pos[:3, 3] = np.clip(desired_ee_pos[:3, 3], bounds["min"], bounds["max"])
+
+            # Only send commands if there's actual movement
+            if any([abs(v) > 0.001 for v in [delta_x, delta_y, delta_z]]):
+                # Compute joint targets via inverse kinematics
+                target_joint_state = kinematics.ik(joint_positions, desired_ee_pos, position_only=True)
+
+                # Send command to robot
+                robot.send_action(torch.from_numpy(target_joint_state))
+
+            busy_wait(1 / fps - (time.perf_counter() - loop_start_time))
+
+
+def teleoperate_gym_env(env, controller, fps: int = 30):
+    """
+    Control a robot through a gym environment using keyboard inputs.
+
+    Args:
+        env: A gym environment created with make_robot_env
+        fps: Target control frequency
+    """
+
+    logging.info("Testing Keyboard Control of Gym Environment")
+    print("Keyboard controls:")
+    print("  Arrow keys: Move in X-Y plane")
+    print("  Shift and Shift_R: Move in Z axis")
+    print("  ESC: Exit")
+
+    # Reset the environment to get initial observation
+    obs, info = env.reset()
+
+    try:
+        with controller:
+            while not controller.should_quit():
+                loop_start_time = time.perf_counter()
+
+                # Process input events
+                controller.update()
+
+                # Get movement deltas from the controller
+                delta_x, delta_y, delta_z = controller.get_deltas()
+
+                # Create the action vector
+                action = np.array([delta_x, delta_y, delta_z])
+
+                # Skip if no movement
+                if any([abs(v) > 0.001 for v in [delta_x, delta_y, delta_z]]):
+                    # Step the environment - pass action as a tensor with intervention flag
+                    action_tensor = torch.from_numpy(action.astype(np.float32))
+                    obs, reward, terminated, truncated, info = env.step((action_tensor, False))
+
+                    # Log information
+                    logging.info(f"Action: [{delta_x:.4f}, {delta_y:.4f}, {delta_z:.4f}]")
+                    logging.info(f"Reward: {reward}")
+
+                    # Reset if episode ended
+                    if terminated or truncated:
+                        logging.info("Episode ended, resetting environment")
+                        obs, info = env.reset()
+
+                # Maintain target frame rate
+                busy_wait(1 / fps - (time.perf_counter() - loop_start_time))
+
+    finally:
+        # Close the environment
+        env.close()
+
+
+if __name__ == "__main__":
+    from lerobot.common.envs.configs import EEActionSpaceConfig, EnvWrapperConfig, HILSerlRobotEnvConfig
+    from lerobot.common.robot_devices.robots.configs import RobotConfig
+    from lerobot.common.robot_devices.robots.utils import make_robot_from_config
+    from lerobot.scripts.server.gym_manipulator import make_robot_env
+
+    parser = argparse.ArgumentParser(description="Test end-effector control")
+    parser.add_argument(
+        "--mode",
+        type=str,
+        default="keyboard",
+        choices=[
+            "keyboard",
+            "gamepad",
+            "keyboard_gym",
+            "gamepad_gym",
+            "leader",
+            "leader_abs",
+        ],
+        help="Control mode to use",
+    )
+    parser.add_argument(
+        "--robot-type",
+        type=str,
+        default="so100",
+        help="Robot type (so100, koch, aloha, etc.)",
+    )
+
+    args = parser.parse_args()
+
+    robot_config = RobotConfig.get_choice_class(args.robot_type)(mock=False)
+    robot = make_robot_from_config(robot_config)
+
+    if not robot.is_connected:
+        robot.connect()
+
+    # Example bounds
+    bounds = {
+        "max": np.array([0.32170487, 0.201285, 0.10273342]),
+        "min": np.array([0.16631757, -0.08237468, 0.03364977]),
+    }
+
+    try:
+        # Determine controller type based on mode prefix
+        controller = None
+        if args.mode.startswith("keyboard"):
+            controller = KeyboardController(x_step_size=0.01, y_step_size=0.01, z_step_size=0.05)
+        elif args.mode.startswith("gamepad"):
+            if sys.platform == "darwin":
+                controller = GamepadControllerHID(x_step_size=0.01, y_step_size=0.01, z_step_size=0.05)
+            else:
+                controller = GamepadController(x_step_size=0.01, y_step_size=0.01, z_step_size=0.05)
+
+        # Handle mode categories
+        if args.mode in ["keyboard", "gamepad"]:
+            # Direct robot control modes
+            teleoperate_delta_inverse_kinematics(robot, controller, bounds=bounds, fps=10)
+
+        elif args.mode in ["keyboard_gym", "gamepad_gym"]:
+            # Gym environment control modes
+            cfg = HILSerlRobotEnvConfig(robot=robot_config, wrapper=EnvWrapperConfig())
+            cfg.wrapper.ee_action_space_params = EEActionSpaceConfig(
+                x_step_size=0.03, y_step_size=0.03, z_step_size=0.03, bounds=bounds
+            )
+            cfg.wrapper.ee_action_space_params.use_gamepad = False
+            cfg.device = "cpu"
+            env = make_robot_env(cfg, robot)
+            teleoperate_gym_env(env, controller, fps=cfg.fps)
+
+        elif args.mode == "leader":
+            # Leader-follower modes don't use controllers
+            teleoperate_delta_inverse_kinematics_with_leader(robot)
+
+        elif args.mode == "leader_abs":
+            teleoperate_inverse_kinematics_with_leader(robot)
+
+    finally:
+        if robot.is_connected:
+            robot.disconnect()

lerobot/scripts/server/find_joint_limits.py

@@ -0,0 +1,140 @@
+import argparse
+import time
+
+import cv2
+import numpy as np
+
+from lerobot.common.robot_devices.control_utils import is_headless
+from lerobot.common.robot_devices.robots.configs import RobotConfig
+from lerobot.common.robot_devices.robots.utils import make_robot_from_config
+from lerobot.configs import parser
+from lerobot.scripts.server.kinematics import RobotKinematics
+
+follower_port = "/dev/tty.usbmodem58760431631"
+leader_port = "/dev/tty.usbmodem58760433331"
+
+
+def find_joint_bounds(
+    robot,
+    control_time_s=30,
+    display_cameras=False,
+):
+    if not robot.is_connected:
+        robot.connect()
+
+    start_episode_t = time.perf_counter()
+    pos_list = []
+    while True:
+        observation, action = robot.teleop_step(record_data=True)
+
+        # Wait for 5 seconds to stabilize the robot initial position
+        if time.perf_counter() - start_episode_t < 5:
+            continue
+
+        pos_list.append(robot.follower_arms["main"].read("Present_Position"))
+
+        if display_cameras and not is_headless():
+            image_keys = [key for key in observation if "image" in key]
+            for key in image_keys:
+                cv2.imshow(key, cv2.cvtColor(observation[key].numpy(), cv2.COLOR_RGB2BGR))
+            cv2.waitKey(1)
+
+        if time.perf_counter() - start_episode_t > control_time_s:
+            max = np.max(np.stack(pos_list), 0)
+            min = np.min(np.stack(pos_list), 0)
+            print(f"Max angle position per joint {max}")
+            print(f"Min angle position per joint {min}")
+            break
+
+
+def find_ee_bounds(
+    robot,
+    control_time_s=30,
+    display_cameras=False,
+):
+    if not robot.is_connected:
+        robot.connect()
+
+    start_episode_t = time.perf_counter()
+    ee_list = []
+    while True:
+        observation, action = robot.teleop_step(record_data=True)
+
+        # Wait for 5 seconds to stabilize the robot initial position
+        if time.perf_counter() - start_episode_t < 5:
+            continue
+
+        kinematics = RobotKinematics(robot.robot_type)
+        joint_positions = robot.follower_arms["main"].read("Present_Position")
+        print(f"Joint positions: {joint_positions}")
+        ee_list.append(kinematics.fk_gripper_tip(joint_positions)[:3, 3])
+
+        if display_cameras and not is_headless():
+            image_keys = [key for key in observation if "image" in key]
+            for key in image_keys:
+                cv2.imshow(key, cv2.cvtColor(observation[key].numpy(), cv2.COLOR_RGB2BGR))
+            cv2.waitKey(1)
+
+        if time.perf_counter() - start_episode_t > control_time_s:
+            max = np.max(np.stack(ee_list), 0)
+            min = np.min(np.stack(ee_list), 0)
+            print(f"Max ee position {max}")
+            print(f"Min ee position {min}")
+            break
+
+
+def make_robot(robot_type="so100"):
+    """
+    Create a robot instance using the appropriate robot config class.
+
+    Args:
+        robot_type: Robot type string (e.g., "so100", "koch", "aloha")
+
+    Returns:
+        Robot instance
+    """
+
+    # Get the appropriate robot config class based on robot_type
+    robot_config = RobotConfig.get_choice_class(robot_type)(mock=False)
+    robot_config.leader_arms["main"].port = leader_port
+    robot_config.follower_arms["main"].port = follower_port
+
+    return make_robot_from_config(robot_config)
+
+
+if __name__ == "__main__":
+    # Create argparse for script-specific arguments
+    parser = argparse.ArgumentParser(add_help=False)  # Set add_help=False to avoid conflict
+    parser.add_argument(
+        "--mode",
+        type=str,
+        default="joint",
+        choices=["joint", "ee"],
+        help="Mode to run the script in. Can be 'joint' or 'ee'.",
+    )
+    parser.add_argument(
+        "--control-time-s",
+        type=int,
+        default=30,
+        help="Time step to use for control.",
+    )
+    parser.add_argument(
+        "--robot-type",
+        type=str,
+        default="so100",
+        help="Robot type (so100, koch, aloha, etc.)",
+    )
+
+    # Only parse known args, leaving robot config args for Hydra if used
+    args = parser.parse_args()
+
+    # Create robot with the appropriate config
+    robot = make_robot(args.robot_type)
+
+    if args.mode == "joint":
+        find_joint_bounds(robot, args.control_time_s)
+    elif args.mode == "ee":
+        find_ee_bounds(robot, args.control_time_s)
+
+    if robot.is_connected:
+        robot.disconnect()

lerobot/scripts/server/hilserl.proto

@@ -0,0 +1,55 @@
+// !/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.
+syntax = "proto3";
+
+package hil_serl;
+
+// LearnerService: the Actor calls this to push transitions.
+// The Learner implements this service.
+service LearnerService {
+  // Actor -> Learner to store transitions
+  rpc SendInteractionMessage(InteractionMessage) returns (Empty);
+  rpc StreamParameters(Empty) returns (stream Parameters);
+  rpc SendTransitions(stream Transition) returns (Empty);
+  rpc SendInteractions(stream InteractionMessage) returns (Empty);
+  rpc Ready(Empty) returns (Empty);
+}
+
+enum TransferState {
+    TRANSFER_UNKNOWN = 0;
+    TRANSFER_BEGIN = 1;
+    TRANSFER_MIDDLE = 2;
+    TRANSFER_END = 3;
+}
+
+// Messages
+message Transition {
+  TransferState transfer_state = 1;
+  bytes data = 2;
+}
+
+message Parameters {
+  TransferState transfer_state = 1;
+  bytes data = 2;
+}
+
+message InteractionMessage {
+  TransferState transfer_state = 1;
+  bytes data = 2;
+}
+
+message Empty {}