Support accelerate training and add test configs for SmolVLA

- 2-GPU SLURM job (distributed training) - 1-GPU local accelerate and direct training scripts - Accelerate configs for 1-GPU and 2-GPU setups
Lower limits by 50% for current and torque for gripper motor (#1809 )
2026-05-12 15:19:43 +00:00 · 2025-09-04 13:07:25 +00:00 · 2025-08-29 16:06:55 +02:00 · 2025-08-28 11:18:54 +02:00 · 2025-08-25 12:39:32 +02:00 · 2025-08-22 15:24:02 +02:00
48 changed files with 520 additions and 152 deletions
@@ -30,7 +30,7 @@ pytest -sx tests/test_stuff.py::test_something
 ```

 ```bash
-python -m lerobot.scripts.train --some.option=true
+lerobot-train --some.option=true
 ```

 ## SECTION TO REMOVE BEFORE SUBMITTING YOUR PR
@@ -29,8 +29,8 @@ on:
 env:
  UV_VERSION: "0.8.0"
  PYTHON_VERSION: "3.10"
-  DOCKER_IMAGE_NAME_CPU: huggingface/lerobot-gpu:latest
-  DOCKER_IMAGE_NAME_GPU: huggingface/lerobot-cpu:latest
+  DOCKER_IMAGE_NAME_CPU: huggingface/lerobot-cpu:latest
+  DOCKER_IMAGE_NAME_GPU: huggingface/lerobot-gpu:latest

 # Ensures that only the latest commit is built, canceling older runs.
 concurrency:
@@ -44,7 +44,7 @@ test-end-to-end:
 	${MAKE} DEVICE=$(DEVICE) test-smolvla-ete-eval

 test-act-ete-train:
-	python -m lerobot.scripts.train \
+	lerobot-train \
 		--policy.type=act \
 		--policy.dim_model=64 \
 		--policy.n_action_steps=20 \
@@ -68,12 +68,12 @@ test-act-ete-train:
 		--output_dir=tests/outputs/act/

 test-act-ete-train-resume:
-	python -m lerobot.scripts.train \
+	lerobot-train \
 		--config_path=tests/outputs/act/checkpoints/000002/pretrained_model/train_config.json \
 		--resume=true

 test-act-ete-eval:
-	python -m lerobot.scripts.eval \
+	lerobot-eval \
 		--policy.path=tests/outputs/act/checkpoints/000004/pretrained_model \
 		--policy.device=$(DEVICE) \
 		--env.type=aloha \
@@ -82,7 +82,7 @@ test-act-ete-eval:
 		--eval.batch_size=1

 test-diffusion-ete-train:
-	python -m lerobot.scripts.train \
+	lerobot-train \
 		--policy.type=diffusion \
 		--policy.down_dims='[64,128,256]' \
 		--policy.diffusion_step_embed_dim=32 \
@@ -106,7 +106,7 @@ test-diffusion-ete-train:
 		--output_dir=tests/outputs/diffusion/

 test-diffusion-ete-eval:
-	python -m lerobot.scripts.eval \
+	lerobot-eval \
 		--policy.path=tests/outputs/diffusion/checkpoints/000002/pretrained_model \
 		--policy.device=$(DEVICE) \
 		--env.type=pusht \
@@ -115,7 +115,7 @@ test-diffusion-ete-eval:
 		--eval.batch_size=1

 test-tdmpc-ete-train:
-	python -m lerobot.scripts.train \
+	lerobot-train \
 		--policy.type=tdmpc \
 		--policy.device=$(DEVICE) \
 		--policy.push_to_hub=false \
@@ -137,7 +137,7 @@ test-tdmpc-ete-train:
 		--output_dir=tests/outputs/tdmpc/

 test-tdmpc-ete-eval:
-	python -m lerobot.scripts.eval \
+	lerobot-eval \
 		--policy.path=tests/outputs/tdmpc/checkpoints/000002/pretrained_model \
 		--policy.device=$(DEVICE) \
 		--env.type=xarm \
@@ -148,7 +148,7 @@ test-tdmpc-ete-eval:


 test-smolvla-ete-train:
-	python -m lerobot.scripts.train \
+	lerobot-train \
 		--policy.type=smolvla \
 		--policy.n_action_steps=20 \
 		--policy.chunk_size=20 \
@@ -171,7 +171,7 @@ test-smolvla-ete-train:
 		--output_dir=tests/outputs/smolvla/

 test-smolvla-ete-eval:
-	python -m lerobot.scripts.eval \
+	lerobot-eval \
 		--policy.path=tests/outputs/smolvla/checkpoints/000004/pretrained_model \
 		--policy.device=$(DEVICE) \
 		--env.type=aloha \
@@ -6,7 +6,7 @@

 <div align="center">

-[![Tests](https://github.com/huggingface/lerobot/actions/workflows/nightly.yml/badge.svg?branch=main)](https://github.com/huggingface/lerobot/actions/workflows/nighty.yml?query=branch%3Amain)
+[![Tests](https://github.com/huggingface/lerobot/actions/workflows/nightly.yml/badge.svg?branch=main)](https://github.com/huggingface/lerobot/actions/workflows/nightly.yml?query=branch%3Amain)
 [![Python versions](https://img.shields.io/pypi/pyversions/lerobot)](https://www.python.org/downloads/)
 [![License](https://img.shields.io/badge/License-Apache%202.0-blue.svg)](https://github.com/huggingface/lerobot/blob/main/LICENSE)
 [![Status](https://img.shields.io/pypi/status/lerobot)](https://pypi.org/project/lerobot/)
@@ -276,7 +276,7 @@ Check out [example 2](https://github.com/huggingface/lerobot/blob/main/examples/
 We also provide a more capable script to parallelize the evaluation over multiple environments during the same rollout. Here is an example with a pretrained model hosted on [lerobot/diffusion_pusht](https://huggingface.co/lerobot/diffusion_pusht):

 ```bash
-python -m lerobot.scripts.eval \
+lerobot-eval \
    --policy.path=lerobot/diffusion_pusht \
    --env.type=pusht \
    --eval.batch_size=10 \
@@ -288,10 +288,10 @@ python -m lerobot.scripts.eval \
 Note: After training your own policy, you can re-evaluate the checkpoints with:

 ```bash
-python -m lerobot.scripts.eval --policy.path={OUTPUT_DIR}/checkpoints/last/pretrained_model
+lerobot-eval --policy.path={OUTPUT_DIR}/checkpoints/last/pretrained_model
 ```

-See `python -m lerobot.scripts.eval --help` for more instructions.
+See `lerobot-eval --help` for more instructions.

 ### Train your own policy

@@ -303,7 +303,7 @@ A link to the wandb logs for the run will also show up in yellow in your termina

 \<img src="https://raw.githubusercontent.com/huggingface/lerobot/main/media/wandb.png" alt="WandB logs example"\>

-Note: For efficiency, during training every checkpoint is evaluated on a low number of episodes. You may use `--eval.n_episodes=500` to evaluate on more episodes than the default. Or, after training, you may want to re-evaluate your best checkpoints on more episodes or change the evaluation settings. See `python -m lerobot.scripts.eval --help` for more instructions.
+Note: For efficiency, during training every checkpoint is evaluated on a low number of episodes. You may use `--eval.n_episodes=500` to evaluate on more episodes than the default. Or, after training, you may want to re-evaluate your best checkpoints on more episodes or change the evaluation settings. See `lerobot-eval --help` for more instructions.

 #### Reproduce state-of-the-art (SOTA)

@@ -311,7 +311,7 @@ We provide some pretrained policies on our [hub page](https://huggingface.co/ler
 You can reproduce their training by loading the config from their run. Simply running:

 ```bash
-python -m lerobot.scripts.train --config_path=lerobot/diffusion_pusht
+lerobot-train --config_path=lerobot/diffusion_pusht
 ```

 reproduces SOTA results for Diffusion Policy on the PushT task.
@@ -0,0 +1,11 @@
+compute_environment: LOCAL_MACHINE
+debug: false
+distributed_type: NO
+downcast_bf16: 'no'
+enable_cpu_affinity: false
+machine_rank: 0
+main_training_function: main
+mixed_precision: 'no'
+num_machines: 1
+num_processes: 1
+use_cpu: false
@@ -0,0 +1,18 @@
+compute_environment: LOCAL_MACHINE
+debug: false
+distributed_type: MULTI_GPU
+downcast_bf16: 'no'
+enable_cpu_affinity: false
+gpu_ids: all
+machine_rank: 0
+main_training_function: main
+mixed_precision: 'no'
+num_machines: 1
+num_processes: 2
+rdzv_backend: static
+same_network: true
+tpu_env: []
+tpu_use_cluster: false
+tpu_use_sudo: false
+use_cpu: false
+dynamo_backend: "no"
@@ -29,7 +29,7 @@ ENV DEBIAN_FRONTEND=noninteractive \

 # Install system dependencies and uv (as root)
 RUN apt-get update && apt-get install -y --no-install-recommends \
-    build-essential git curl libglib2.0-0 libegl1-mesa ffmpeg \
+    build-essential git curl libglib2.0-0 libegl1-mesa-dev ffmpeg \
    libusb-1.0-0-dev speech-dispatcher libgeos-dev portaudio19-dev \
    && curl -LsSf https://astral.sh/uv/install.sh | sh \
    && mv /root/.local/bin/uv /usr/local/bin/uv \
@@ -39,6 +39,8 @@
 - sections:
  - local: notebooks
    title: Notebooks
+  - local: feetech
+    title: Updating Feetech Firmware
  title: "Resources"
 - sections:
  - local: contributing
@@ -9,7 +9,7 @@ To instantiate a camera, you need a camera identifier. This identifier might cha
 To find the camera indices of the cameras plugged into your system, run the following script:

 ```bash
-python -m lerobot.find_cameras opencv # or realsense for Intel Realsense cameras
+lerobot-find-cameras opencv # or realsense for Intel Realsense cameras
 ```

 The output will look something like this if you have two cameras connected:
@@ -0,0 +1,71 @@
+# Feetech Motor Firmware Update
+
+This tutorial guides you through updating the firmware of Feetech motors using the official Feetech software.
+
+## Prerequisites
+
+- Windows computer (Feetech software is only available for Windows)
+- Feetech motor control board
+- USB cable to connect the control board to your computer
+- Feetech motors connected to the control board
+
+## Step 1: Download Feetech Software
+
+1. Visit the official Feetech software download page: [https://www.feetechrc.com/software.html](https://www.feetechrc.com/software.html)
+2. Download the latest version of the Feetech debugging software (FD)
+3. Install the software on your Windows computer
+
+## Step 2: Hardware Setup
+
+1. Connect your Feetech motors to the motor control board
+2. Connect the motor control board to your Windows computer via USB cable
+3. Ensure power is supplied to the motors
+
+## Step 3: Configure Connection
+
+1. Launch the Feetech debugging software
+2. Select the correct COM port from the port dropdown menu
+   - If unsure which port to use, check Windows Device Manager under "Ports (COM & LPT)"
+3. Set the appropriate baud rate (typically 1000000 for most Feetech motors)
+4. Click "Open" to establish communication with the control board
+
+## Step 4: Scan for Motors
+
+1. Once connected, click the "Search" button to detect all connected motors
+2. The software will automatically discover and list all motors on the bus
+3. Each motor will appear with its ID number
+
+## Step 5: Update Firmware
+
+For each motor you want to update:
+
+1. **Select the motor** from the list by clicking on it
+2. **Click on Upgrade tab**:
+3. **Click on Online button**:
+   - If an potential firmware update is found, it will be displayed in the box
+4. **Click on Upgrade button**:
+   - The update progress will be displayed
+
+## Step 6: Verify Update
+
+1. After the update completes, the software should automatically refresh the motor information
+2. Verify that the firmware version has been updated to the expected version
+
+## Important Notes
+
+⚠️ **Warning**: Do not disconnect power or USB during firmware updates, it will potentially brick the motor.
+
+## Bonus: Motor Debugging on Linux/macOS
+
+For debugging purposes only, you can use the open-source Feetech Debug Tool:
+
+- **Repository**: [FT_SCServo_Debug_Qt](https://github.com/CarolinePascal/FT_SCServo_Debug_Qt/tree/fix/port-search-timer)
+
+### Installation Instructions
+
+Follow the instructions in the repository to install the tool, for Ubuntu you can directly install it, for MacOS you need to build it from source.
+
+**Limitations:**
+
+- This tool is for debugging and parameter adjustment only
+- Firmware updates must still be done on Windows with official Feetech software
@@ -412,7 +412,7 @@ Example configuration for training the [reward classifier](https://huggingface.c
 To train the classifier, use the `train.py` script with your configuration:

 ```bash
-python -m lerobot.scripts.train --config_path path/to/reward_classifier_train_config.json
+lerobot-train --config_path path/to/reward_classifier_train_config.json
 ```

 **Deploying and Testing the Model**
@@ -458,7 +458,7 @@ The reward classifier will automatically provide rewards based on the visual inp
 3. **Train the classifier**:

   ```bash
-   python -m lerobot.scripts.train --config_path src/lerobot/configs/reward_classifier_train_config.json
+   lerobot-train --config_path src/lerobot/configs/reward_classifier_train_config.json
   ```

 4. **Test the classifier**:
@@ -19,7 +19,7 @@ pip install -e ".[hopejr]"
 Before starting calibration and operation, you need to identify the USB ports for each HopeJR component. Run this script to find the USB ports for the arm, hand, glove, and exoskeleton:

 ```bash
-python -m lerobot.find_port
+lerobot-find-port
 ```

 This will display the available USB ports and their associated devices. Make note of the port paths (e.g., `/dev/tty.usbmodem58760433331`, `/dev/tty.usbmodem11301`) as you'll need to specify them in the `--robot.port` and `--teleop.port` parameters when recording data, replaying episodes, or running teleoperation scripts.
@@ -31,7 +31,7 @@ Before performing teleoperation, HopeJR's limbs need to be calibrated. Calibrati
 ### 1.1 Calibrate Robot Hand

 ```bash
-python -m lerobot.calibrate \
+lerobot-calibrate \
    --robot.type=hope_jr_hand \
    --robot.port=/dev/tty.usbmodem58760432281 \
    --robot.id=blue \
@@ -81,7 +81,7 @@ Once you have set the appropriate boundaries for all joints, click "Save" to sav
 ### 1.2 Calibrate Teleoperator Glove

 ```bash
-python -m lerobot.calibrate \
+lerobot-calibrate \
    --teleop.type=homunculus_glove \
    --teleop.port=/dev/tty.usbmodem11201 \
    --teleop.id=red \
@@ -120,7 +120,7 @@ Once calibration is complete, the system will save the calibration to `/Users/yo
 ### 1.3 Calibrate Robot Arm

 ```bash
-python -m lerobot.calibrate \
+lerobot-calibrate \
    --robot.type=hope_jr_arm \
    --robot.port=/dev/tty.usbserial-1110 \
    --robot.id=white
@@ -146,7 +146,7 @@ Use the calibration interface to set the range boundaries for each joint. Move e
 ### 1.4 Calibrate Teleoperator Exoskeleton

 ```bash
-python -m lerobot.calibrate \
+lerobot-calibrate \
    --teleop.type=homunculus_arm \
    --teleop.port=/dev/tty.usbmodem11201 \
    --teleop.id=black
@@ -178,7 +178,7 @@ Due to global variable conflicts in the Feetech middleware, teleoperation for ar
 ### Hand

 ```bash
-python -m lerobot.teleoperate \
+lerobot-teleoperate \
    --robot.type=hope_jr_hand \
    --robot.port=/dev/tty.usbmodem58760432281 \
    --robot.id=blue \
@@ -194,7 +194,7 @@ python -m lerobot.teleoperate \
 ### Arm

 ```bash
-python -m lerobot.teleoperate \
+lerobot-teleoperate \
    --robot.type=hope_jr_arm \
    --robot.port=/dev/tty.usbserial-1110 \
    --robot.id=white \
@@ -214,7 +214,7 @@ Record, Replay and Train with Hope-JR is still experimental.
 This step records the dataset, which can be seen as an example [here](https://huggingface.co/datasets/nepyope/hand_record_test_with_video_data/settings).

 ```bash
-python -m lerobot.record \
+lerobot-record \
    --robot.type=hope_jr_hand \
    --robot.port=/dev/tty.usbmodem58760432281 \
    --robot.id=right \
@@ -236,7 +236,7 @@ python -m lerobot.record \
 ### Replay

 ```bash
-python -m lerobot.replay \
+lerobot-replay \
    --robot.type=hope_jr_hand \
    --robot.port=/dev/tty.usbmodem58760432281 \
    --robot.id=right \
@@ -248,7 +248,7 @@ python -m lerobot.replay \
 ### Train

 ```bash
-python -m lerobot.scripts.train \
+lerobot-train \
  --dataset.repo_id=nepyope/hand_record_test_with_video_data \
  --policy.type=act \
  --output_dir=outputs/train/hopejr_hand \
@@ -263,7 +263,7 @@ python -m lerobot.scripts.train \
 This training run can be viewed as an example [here](https://wandb.ai/tino/lerobot/runs/rp0k8zvw?nw=nwusertino).

 ```bash
-python -m lerobot.record \
+lerobot-record \
  --robot.type=hope_jr_hand \
  --robot.port=/dev/tty.usbmodem58760432281 \
  --robot.id=right \
@@ -45,7 +45,7 @@ Note that the `id` associated with a robot is used to store the calibration file
 <hfoptions id="teleoperate_so101">
 <hfoption id="Command">
 ```bash
-python -m lerobot.teleoperate \
+lerobot-teleoperate \
    --robot.type=so101_follower \
    --robot.port=/dev/tty.usbmodem58760431541 \
    --robot.id=my_awesome_follower_arm \
@@ -101,7 +101,7 @@ With `rerun`, you can teleoperate again while simultaneously visualizing the cam
 <hfoptions id="teleoperate_koch_camera">
 <hfoption id="Command">
 ```bash
-python -m lerobot.teleoperate \
+lerobot-teleoperate \
    --robot.type=koch_follower \
    --robot.port=/dev/tty.usbmodem58760431541 \
    --robot.id=my_awesome_follower_arm \
@@ -174,7 +174,7 @@ Now you can record a dataset. To record 5 episodes and upload your dataset to th
 <hfoptions id="record">
 <hfoption id="Command">
 ```bash
-python -m lerobot.record \
+lerobot-record \
    --robot.type=so101_follower \
    --robot.port=/dev/tty.usbmodem585A0076841 \
    --robot.id=my_awesome_follower_arm \
@@ -376,7 +376,7 @@ You can replay the first episode on your robot with either the command below or
 <hfoptions id="replay">
 <hfoption id="Command">
 ```bash
-python -m lerobot.replay \
+lerobot-replay \
    --robot.type=so101_follower \
    --robot.port=/dev/tty.usbmodem58760431541 \
    --robot.id=my_awesome_follower_arm \
@@ -428,10 +428,10 @@ Your robot should replicate movements similar to those you recorded. For example

 ## Train a policy

-To train a policy to control your robot, use the [`python -m lerobot.scripts.train`](https://github.com/huggingface/lerobot/blob/main/src/lerobot/scripts/train.py) script. A few arguments are required. Here is an example command:
+To train a policy to control your robot, use the [`lerobot-train`](https://github.com/huggingface/lerobot/blob/main/src/lerobot/scripts/train.py) script. A few arguments are required. Here is an example command:

 ```bash
-python -m lerobot.scripts.train \
+lerobot-train \
  --dataset.repo_id=${HF_USER}/so101_test \
  --policy.type=act \
  --output_dir=outputs/train/act_so101_test \
@@ -453,7 +453,7 @@ Training should take several hours. You will find checkpoints in `outputs/train/
 To resume training from a checkpoint, below is an example command to resume from `last` checkpoint of the `act_so101_test` policy:

 ```bash
-python -m lerobot.scripts.train \
+lerobot-train \
  --config_path=outputs/train/act_so101_test/checkpoints/last/pretrained_model/train_config.json \
  --resume=true
 ```
@@ -490,7 +490,7 @@ You can use the `record` script from [`lerobot/record.py`](https://github.com/hu
 <hfoptions id="eval">
 <hfoption id="Command">
 ```bash
-python -m lerobot.record  \
+lerobot-record  \
  --robot.type=so100_follower \
  --robot.port=/dev/ttyACM1 \
  --robot.cameras="{ up: {type: opencv, index_or_path: /dev/video10, width: 640, height: 480, fps: 30}, side: {type: intelrealsense, serial_number_or_name: 233522074606, width: 640, height: 480, fps: 30}}" \
@@ -96,10 +96,10 @@ If you uploaded your dataset to the hub you can [visualize your dataset online](

 ## Train a policy

-To train a policy to control your robot, use the [`python -m lerobot.scripts.train`](https://github.com/huggingface/lerobot/blob/main/src/lerobot/scripts/train.py) script. A few arguments are required. Here is an example command:
+To train a policy to control your robot, use the [`lerobot-train`](https://github.com/huggingface/lerobot/blob/main/src/lerobot/scripts/train.py) script. A few arguments are required. Here is an example command:

 ```bash
-python -m lerobot.scripts.train \
+lerobot-train \
  --dataset.repo_id=${HF_USER}/il_gym \
  --policy.type=act \
  --output_dir=outputs/train/il_sim_test \
@@ -31,7 +31,7 @@ pip install -e ".[dynamixel]"
 To find the port for each bus servo adapter, run this script:

 ```bash
-python -m lerobot.find_port
+lerobot-find-port
 ```

 <hfoptions id="example">
@@ -98,7 +98,7 @@ For a visual reference on how to set the motor ids please refer to [this video](
 <hfoption id="Command">

 ```bash
-python -m lerobot.setup_motors \
+lerobot-setup-motors \
    --robot.type=koch_follower \
    --robot.port=/dev/tty.usbmodem575E0031751 # <- paste here the port found at previous step
 ```
@@ -174,7 +174,7 @@ Do the same steps for the leader arm but modify the command or script accordingl
 <hfoption id="Command">

 ```bash
-python -m lerobot.setup_motors \
+lerobot-setup-motors \
    --teleop.type=koch_leader \
    --teleop.port=/dev/tty.usbmodem575E0031751 \  # <- paste here the port found at previous step
 ```
@@ -211,7 +211,7 @@ Run the following command or API example to calibrate the follower arm:
 <hfoption id="Command">

 ```bash
-python -m lerobot.calibrate \
+lerobot-calibrate \
    --robot.type=koch_follower \
    --robot.port=/dev/tty.usbmodem58760431551 \ # <- The port of your robot
    --robot.id=my_awesome_follower_arm  # <- Give the robot a unique name
@@ -249,7 +249,7 @@ Do the same steps to calibrate the leader arm, run the following command or API
 <hfoption id="Command">

 ```bash
-python -m lerobot.calibrate \
+lerobot-calibrate \
    --teleop.type=koch_leader \
    --teleop.port=/dev/tty.usbmodem58760431551 \ # <- The port of your robot
    --teleop.id=my_awesome_leader_arm  # <- Give the robot a unique name
@@ -60,7 +60,7 @@ First, we will assemble the two SO100/SO101 arms. One to attach to the mobile ba
 To find the port for each bus servo adapter, run this script:

 ```bash
-python -m lerobot.find_port
+lerobot-find-port
 ```

 <hfoptions id="example">
@@ -116,7 +116,7 @@ The instructions for configuring the motors can be found in the SO101 [docs](./s
 You can run this command to setup motors for LeKiwi. It will first setup the motors for arm (id 6..1) and then setup motors for wheels (9,8,7)

 ```bash
-python -m lerobot.setup_motors \
+lerobot-setup-motors \
    --robot.type=lekiwi \
    --robot.port=/dev/tty.usbmodem58760431551 # <- paste here the port found at previous step
 ```
@@ -174,7 +174,7 @@ The calibration process is very important because it allows a neural network tra
 Make sure the arm is connected to the Raspberry Pi and run this script or API example (on the Raspberry Pi via SSH) to launch calibration of the follower arm:

 ```bash
-python -m lerobot.calibrate \
+lerobot-calibrate \
    --robot.type=lekiwi \
    --robot.id=my_awesome_kiwi # <- Give the robot a unique name
 ```
@@ -193,7 +193,7 @@ Then, to calibrate the leader arm (which is attached to the laptop/pc). Run the
 <hfoption id="Command">

 ```bash
-python -m lerobot.calibrate \
+lerobot-calibrate \
    --teleop.type=so100_leader \
    --teleop.port=/dev/tty.usbmodem58760431551 \ # <- The port of your robot
    --teleop.id=my_awesome_leader_arm # <- Give the robot a unique name
@@ -54,7 +54,7 @@ If you don't have a gpu device, you can train using our notebook on [![Google Co
 Pass your dataset to the training script using `--dataset.repo_id`. If you want to test your installation, run the following command where we use one of the datasets we collected for the [SmolVLA Paper](https://huggingface.co/papers/2506.01844).

 ```bash
-cd lerobot && python -m lerobot.scripts.train \
+cd lerobot && lerobot-train \
  --policy.path=lerobot/smolvla_base \
  --dataset.repo_id=${HF_USER}/mydataset \
  --batch_size=64 \
@@ -73,7 +73,7 @@ cd lerobot && python -m lerobot.scripts.train \
 Fine-tuning is an art. For a complete overview of the options for finetuning, run

 ```bash
-python -m lerobot.scripts.train --help
+lerobot-train --help
 ```

 <p align="center">
@@ -97,7 +97,7 @@ Similarly for when recording an episode, it is recommended that you are logged i
 Once you are logged in, you can run inference in your setup by doing:

 ```bash
-python -m lerobot.record \
+lerobot-record \
  --robot.type=so101_follower \
  --robot.port=/dev/ttyACM0 \ # <- Use your port
  --robot.id=my_blue_follower_arm \ # <- Use your robot id
@@ -26,7 +26,7 @@ Unlike the SO-101, the motor connectors are not easily accessible once the arm i
 To find the port for each bus servo adapter, run this script:

 ```bash
-python -m lerobot.find_port
+lerobot-find-port
 ```

 <hfoptions id="example">
@@ -93,7 +93,7 @@ For a visual reference on how to set the motor ids please refer to [this video](
 <hfoption id="Command">

 ```bash
-python -m lerobot.setup_motors \
+lerobot-setup-motors \
    --robot.type=so100_follower \
    --robot.port=/dev/tty.usbmodem585A0076841  # <- paste here the port found at previous step
 ```
@@ -168,7 +168,7 @@ Do the same steps for the leader arm.
 <hfoptions id="setup_motors">
 <hfoption id="Command">
 ```bash
-python -m lerobot.setup_motors \
+lerobot-setup-motors \
    --teleop.type=so100_leader \
    --teleop.port=/dev/tty.usbmodem575E0031751  # <- paste here the port found at previous step
 ```
@@ -568,7 +568,7 @@ Run the following command or API example to calibrate the follower arm:
 <hfoption id="Command">

 ```bash
-python -m lerobot.calibrate \
+lerobot-calibrate \
    --robot.type=so100_follower \
    --robot.port=/dev/tty.usbmodem58760431551 \ # <- The port of your robot
    --robot.id=my_awesome_follower_arm # <- Give the robot a unique name
@@ -606,7 +606,7 @@ Do the same steps to calibrate the leader arm, run the following command or API
 <hfoption id="Command">

 ```bash
-python -m lerobot.calibrate \
+lerobot-calibrate \
    --teleop.type=so100_leader \
    --teleop.port=/dev/tty.usbmodem58760431551 \ # <- The port of your robot
    --teleop.id=my_awesome_leader_arm # <- Give the robot a unique name
@@ -162,7 +162,7 @@ It is advisable to install one 3-pin cable in the motor after placing them befor
 To find the port for each bus servo adapter, connect MotorBus to your computer via USB and power. Run the following script and disconnect the MotorBus when prompted:

 ```bash
-python -m lerobot.find_port
+lerobot-find-port
 ```

 <hfoptions id="example">
@@ -240,7 +240,7 @@ Connect the usb cable from your computer and the power supply to the follower ar
 <hfoption id="Command">

 ```bash
-python -m lerobot.setup_motors \
+lerobot-setup-motors \
    --robot.type=so101_follower \
    --robot.port=/dev/tty.usbmodem585A0076841  # <- paste here the port found at previous step
 ```
@@ -316,7 +316,7 @@ Do the same steps for the leader arm.
 <hfoption id="Command">

 ```bash
-python -m lerobot.setup_motors \
+lerobot-setup-motors \
    --teleop.type=so101_leader \
    --teleop.port=/dev/tty.usbmodem575E0031751  # <- paste here the port found at previous step
 ```
@@ -353,7 +353,7 @@ Run the following command or API example to calibrate the follower arm:
 <hfoption id="Command">

 ```bash
-python -m lerobot.calibrate \
+lerobot-calibrate \
    --robot.type=so101_follower \
    --robot.port=/dev/tty.usbmodem58760431551 \ # <- The port of your robot
    --robot.id=my_awesome_follower_arm # <- Give the robot a unique name
@@ -402,7 +402,7 @@ Do the same steps to calibrate the leader arm, run the following command or API
 <hfoption id="Command">

 ```bash
-python -m lerobot.calibrate \
+lerobot-calibrate \
    --teleop.type=so101_leader \
    --teleop.port=/dev/tty.usbmodem58760431551 \ # <- The port of your robot
    --teleop.id=my_awesome_leader_arm # <- Give the robot a unique name
@@ -62,7 +62,7 @@ By default, every field takes its default value specified in the dataclass. If a
 Let's say that we want to train [Diffusion Policy](../src/lerobot/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 -m lerobot.scripts.train \
+lerobot-train \
    --dataset.repo_id=lerobot/pusht \
    --policy.type=diffusion \
    --env.type=pusht
@@ -77,7 +77,7 @@ Let's break this down:
 Let's see another example. Let's say you've been training [ACT](../src/lerobot/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 -m lerobot.scripts.train \
+lerobot-train \
    --policy.type=act \
    --dataset.repo_id=lerobot/aloha_sim_insertion_human \
    --env.type=aloha \
@@ -90,7 +90,7 @@ We now want to train a different policy for aloha on another task. We'll change
 Looking at the [`AlohaEnv`](../src/lerobot/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 -m lerobot.scripts.train \
+lerobot-train \
    --policy.type=act \
    --dataset.repo_id=lerobot/aloha_sim_transfer_cube_human \
    --env.type=aloha \
@@ -127,7 +127,7 @@ Now, let's assume that we want to reproduce the run just above. That run has pro
 We can then simply load the config values from this file using:

 ```bash
-python -m lerobot.scripts.train \
+lerobot-train \
    --config_path=outputs/train/act_aloha_transfer/checkpoints/last/pretrained_model/ \
    --output_dir=outputs/train/act_aloha_transfer_2
 ```
@@ -137,7 +137,7 @@ python -m lerobot.scripts.train \
 Similarly to Hydra, we can still override some parameters in the CLI if we want to, e.g.:

 ```bash
-python -m lerobot.scripts.train \
+lerobot-train \
    --config_path=outputs/train/act_aloha_transfer/checkpoints/last/pretrained_model/ \
    --output_dir=outputs/train/act_aloha_transfer_2
    --policy.n_action_steps=80
@@ -148,7 +148,7 @@ python -m lerobot.scripts.train \
 `--config_path` can also accept the repo_id of a repo on the hub that contains a `train_config.json` file, e.g. running:

 ```bash
-python -m lerobot.scripts.train --config_path=lerobot/diffusion_pusht
+lerobot-train --config_path=lerobot/diffusion_pusht
 ```

 will start a training run with the same configuration used for training [lerobot/diffusion_pusht](https://huggingface.co/lerobot/diffusion_pusht)
@@ -160,7 +160,7 @@ Being able to resume a training run is important in case it crashed or aborted f
 Let's reuse the command from the previous run and add a few more options:

 ```bash
-python -m lerobot.scripts.train \
+lerobot-train \
    --policy.type=act \
    --dataset.repo_id=lerobot/aloha_sim_transfer_cube_human \
    --env.type=aloha \
@@ -179,7 +179,7 @@ INFO 2025-01-24 16:10:56 ts/train.py:263 Checkpoint policy after step 100
 Now let's simulate a crash by killing the process (hit `ctrl`+`c`). We can then simply resume this run from the last checkpoint available with:

 ```bash
-python -m lerobot.scripts.train \
+lerobot-train \
    --config_path=outputs/train/run_resumption/checkpoints/last/pretrained_model/ \
    --resume=true
 ```
@@ -190,7 +190,7 @@ Another reason for which you might want to resume a run is simply to extend trai
 You could double the number of steps of the previous run with:

 ```bash
-python -m lerobot.scripts.train \
+lerobot-train \
    --config_path=outputs/train/run_resumption/checkpoints/last/pretrained_model/ \
    --resume=true \
    --steps=200000
@@ -224,7 +224,7 @@ In addition to the features currently in Draccus, we've added a special `.path`
 For example, we could fine-tune a [policy pre-trained on the aloha transfer task](https://huggingface.co/lerobot/act_aloha_sim_transfer_cube_human) on the aloha insertion task. We can achieve this with:

 ```bash
-python -m lerobot.scripts.train \
+lerobot-train \
    --policy.path=lerobot/act_aloha_sim_transfer_cube_human \
    --dataset.repo_id=lerobot/aloha_sim_insertion_human \
    --env.type=aloha \
@@ -270,7 +270,7 @@ We'll summarize here the main use cases to remember from this tutorial.
 #### Train a policy from scratch – CLI

 ```bash
-python -m lerobot.scripts.train \
+lerobot-train \
    --policy.type=act \  # <- select 'act' policy
    --env.type=pusht \  # <- select 'pusht' environment
    --dataset.repo_id=lerobot/pusht  # <- train on this dataset
@@ -279,7 +279,7 @@ python -m lerobot.scripts.train \
 #### Train a policy from scratch - config file + CLI

 ```bash
-python -m lerobot.scripts.train \
+lerobot-train \
    --config_path=path/to/pretrained_model \  # <- can also be a repo_id
    --policy.n_action_steps=80  # <- you may still override values
 ```
@@ -287,7 +287,7 @@ python -m lerobot.scripts.train \
 #### Resume/continue a training run

 ```bash
-python -m lerobot.scripts.train \
+lerobot-train \
    --config_path=checkpoint/pretrained_model/ \
    --resume=true \
    --steps=200000  # <- you can change some training parameters
@@ -296,7 +296,7 @@ python -m lerobot.scripts.train \
 #### Fine-tuning

 ```bash
-python -m lerobot.scripts.train \
+lerobot-train \
    --policy.path=lerobot/act_aloha_sim_transfer_cube_human \  # <- can also be a local path to a checkpoint
    --dataset.repo_id=lerobot/aloha_sim_insertion_human \
    --env.type=aloha \
@@ -18,7 +18,7 @@ Replays the actions of an episode from a dataset on a robot.
 Example:

 ```shell
-python -m lerobot.replay \
+lerobot-replay \
    --robot.type=so100_follower \
    --robot.port=/dev/tty.usbmodem58760431541 \
    --robot.id=black \
@@ -18,7 +18,7 @@ Helper to recalibrate your device (robot or teleoperator).
 Example:

 ```shell
-python -m lerobot.calibrate \
+lerobot-calibrate \
    --teleop.type=so100_leader \
    --teleop.port=/dev/tty.usbmodem58760431551 \
    --teleop.id=blue
@@ -60,7 +60,7 @@ class OpenCVCamera(Camera):
    or port changes, especially on Linux. Use the provided utility script to find
    available camera indices or paths:
    ```bash
-    python -m lerobot.find_cameras opencv
+    lerobot-find-cameras opencv
    ```

    The camera's default settings (FPS, resolution, color mode) are used unless
@@ -165,8 +165,7 @@ class OpenCVCamera(Camera):
            self.videocapture.release()
            self.videocapture = None
            raise ConnectionError(
-                f"Failed to open {self}."
-                f"Run `python -m lerobot.find_cameras opencv` to find available cameras."
+                f"Failed to open {self}.Run `lerobot-find-cameras opencv` to find available cameras."
            )

        self._configure_capture_settings()
@@ -51,7 +51,7 @@ class RealSenseCamera(Camera):

    Use the provided utility script to find available camera indices and default profiles:
    ```bash
-    python -m lerobot.find_cameras realsense
+    lerobot-find-cameras realsense
    ```

    A `RealSenseCamera` instance requires a configuration object specifying the
@@ -176,8 +176,7 @@ class RealSenseCamera(Camera):
            self.rs_profile = None
            self.rs_pipeline = None
            raise ConnectionError(
-                f"Failed to open {self}."
-                "Run `python -m lerobot.find_cameras realsense` to find available cameras."
+                f"Failed to open {self}.Run `lerobot-find-cameras realsense` to find available cameras."
            ) from e

        self._configure_capture_settings()
@@ -825,6 +825,8 @@ class LeRobotDataset(torch.utils.data.Dataset):
        """
        if not episode_data:
            episode_buffer = self.episode_buffer
+        else:
+            episode_buffer = episode_data

        validate_episode_buffer(episode_buffer, self.meta.total_episodes, self.features)

@@ -20,7 +20,7 @@ Helper to find the camera devices available in your system.
 Example:

 ```shell
-python -m lerobot.find_cameras
+lerobot-find-cameras
 ```
 """

@@ -18,7 +18,7 @@ Helper to find the USB port associated with your MotorsBus.
 Example:

 ```shell
-python -m lerobot.find_port
+lerobot-find-port
 ```
 """

@@ -107,6 +107,8 @@ X_SERIES_ENCODINGS_TABLE = {
    "Goal_PWM": X_SERIES_CONTROL_TABLE["Goal_PWM"][1],
    "Goal_Current": X_SERIES_CONTROL_TABLE["Goal_Current"][1],
    "Goal_Velocity": X_SERIES_CONTROL_TABLE["Goal_Velocity"][1],
+    "Goal_Position": X_SERIES_CONTROL_TABLE["Goal_Position"][1],
+    "Present_Position": X_SERIES_CONTROL_TABLE["Present_Position"][1],
    "Present_PWM": X_SERIES_CONTROL_TABLE["Present_PWM"][1],
    "Present_Current": X_SERIES_CONTROL_TABLE["Present_Current"][1],
    "Present_Velocity": X_SERIES_CONTROL_TABLE["Present_Velocity"][1],
@@ -222,7 +222,7 @@ class MotorsBus(abc.ABC):
    A MotorsBus subclass instance requires a port (e.g. `FeetechMotorsBus(port="/dev/tty.usbmodem575E0031751"`)).
    To find the port, you can run our utility script:
    ```bash
-    python -m lerobot.find_port.py
+    lerobot-find-port.py
    >>> Finding all available ports for the MotorsBus.
    >>> ["/dev/tty.usbmodem575E0032081", "/dev/tty.usbmodem575E0031751"]
    >>> Remove the usb cable from your MotorsBus and press Enter when done.
@@ -446,7 +446,7 @@ class MotorsBus(abc.ABC):
        except (FileNotFoundError, OSError, serial.SerialException) as e:
            raise ConnectionError(
                f"\nCould not connect on port '{self.port}'. Make sure you are using the correct port."
-                "\nTry running `python -m lerobot.find_port`\n"
+                "\nTry running `lerobot-find-port`\n"
            ) from e

    @abc.abstractmethod
@@ -30,7 +30,7 @@ pip install -e ".[pi0]"

 Example of finetuning the pi0 pretrained model (`pi0_base` in `openpi`):
 ```bash
-python -m lerobot.scripts.train \
+lerobot-train \
 --policy.path=lerobot/pi0 \
 --dataset.repo_id=danaaubakirova/koch_test
 ```
@@ -38,7 +38,7 @@ python -m lerobot.scripts.train \
 Example of finetuning the pi0 neural network with PaliGemma and expert Gemma
 pretrained with VLM default parameters before pi0 finetuning:
 ```bash
-python -m lerobot.scripts.train \
+lerobot-train \
 --policy.type=pi0 \
 --dataset.repo_id=danaaubakirova/koch_test
 ```
@@ -25,14 +25,14 @@ Disclaimer: It is not expected to perform as well as the original implementation

 Example of finetuning the pi0+FAST pretrained model (`pi0_fast_base` in `openpi`):
 ```bash
-python -m lerobot.scripts.train \
+lerobot-train \
 --policy.path=lerobot/pi0fast_base \
 --dataset.repo_id=danaaubakirova/koch_test
 ```

 Example of training the pi0+FAST neural network with from scratch:
 ```bash
-python -m lerobot.scripts.train \
+lerobot-train \
 --policy.type=pi0fast \
 --dataset.repo_id=danaaubakirova/koch_test
 ```
@@ -28,7 +28,7 @@ pip install -e ".[smolvla]"

 Example of finetuning the smolvla pretrained model (`smolvla_base`):
 ```bash
-python -m lerobot.scripts.train \
+lerobot-train \
 --policy.path=lerobot/smolvla_base \
 --dataset.repo_id=danaaubakirova/svla_so100_task1_v3 \
 --batch_size=64 \
@@ -38,7 +38,7 @@ python -m lerobot.scripts.train \
 Example of finetuning a smolVLA. SmolVLA is composed of a pretrained VLM,
 and an action expert.
 ```bash
-python -m lerobot.scripts.train \
+lerobot-train \
 --policy.type=smolvla \
 --dataset.repo_id=danaaubakirova/svla_so100_task1_v3 \
 --batch_size=64 \
@@ -18,7 +18,7 @@ Records a dataset. Actions for the robot can be either generated by teleoperatio
 Example:

 ```shell
-python -m lerobot.record \
+lerobot-record \
    --robot.type=so100_follower \
    --robot.port=/dev/tty.usbmodem58760431541 \
    --robot.cameras="{laptop: {type: opencv, camera_index: 0, width: 640, height: 480}}" \
@@ -36,7 +36,7 @@ python -m lerobot.record \

 Example recording with bimanual so100:
 ```shell
-python -m lerobot.record \
+lerobot-record \
  --robot.type=bi_so100_follower \
  --robot.left_arm_port=/dev/tty.usbmodem5A460851411 \
  --robot.right_arm_port=/dev/tty.usbmodem5A460812391 \
@@ -18,7 +18,7 @@ Replays the actions of an episode from a dataset on a robot.
 Examples:

 ```shell
-python -m lerobot.replay \
+lerobot-replay \
    --robot.type=so100_follower \
    --robot.port=/dev/tty.usbmodem58760431541 \
    --robot.id=black \
@@ -28,7 +28,7 @@ python -m lerobot.replay \

 Example replay with bimanual so100:
 ```shell
-python -m lerobot.replay \
+lerobot-replay \
  --robot.type=bi_so100_follower \
  --robot.left_arm_port=/dev/tty.usbmodem5A460851411 \
  --robot.right_arm_port=/dev/tty.usbmodem5A460812391 \
@@ -161,6 +161,11 @@ class SO100Follower(Robot):
                self.bus.write("I_Coefficient", motor, 0)
                self.bus.write("D_Coefficient", motor, 32)

+                if motor == "gripper":
+                    self.bus.write("Max_Torque_Limit", motor, 500)  # 50% of max torque to avoid burnout
+                    self.bus.write("Protection_Current", motor, 250)  # 50% of max current to avoid burnout
+                    self.bus.write("Overload_Torque", motor, 25)  # 25% torque when overloaded
+
    def setup_motors(self) -> None:
        for motor in reversed(self.bus.motors):
            input(f"Connect the controller board to the '{motor}' motor only and press enter.")
@@ -157,6 +157,13 @@ class SO101Follower(Robot):
                self.bus.write("I_Coefficient", motor, 0)
                self.bus.write("D_Coefficient", motor, 32)

+                if motor == "gripper":
+                    self.bus.write(
+                        "Max_Torque_Limit", motor, 500
+                    )  # 50% of the max torque limit to avoid burnout
+                    self.bus.write("Protection_Current", motor, 250)  # 50% of max current to avoid burnout
+                    self.bus.write("Overload_Torque", motor, 25)  # 25% torque when overloaded
+
    def setup_motors(self) -> None:
        for motor in reversed(self.bus.motors):
            input(f"Connect the controller board to the '{motor}' motor only and press enter.")
@@ -141,10 +141,10 @@ python lerobot/scripts/control_robot.py \

 ## Train a policy

-To train a policy to control your robot, use the [`python -m lerobot.scripts.train`](../src/lerobot/scripts/train.py) script. A few arguments are required. Here is an example command:
+To train a policy to control your robot, use the [`lerobot-train`](../src/lerobot/scripts/train.py) script. A few arguments are required. Here is an example command:

 ```bash
-python -m lerobot.scripts.train \
+lerobot-train \
  --dataset.repo_id=${HF_USER}/aloha_test \
  --policy.type=act \
  --output_dir=outputs/train/act_aloha_test \
@@ -21,7 +21,7 @@ You want to evaluate a model from the hub (eg: https://huggingface.co/lerobot/di
 for 10 episodes.

 ```
-python -m lerobot.scripts.eval \
+lerobot-eval \
    --policy.path=lerobot/diffusion_pusht \
    --env.type=pusht \
    --eval.batch_size=10 \
@@ -32,7 +32,7 @@ python -m lerobot.scripts.eval \

 OR, you want to evaluate a model checkpoint from the LeRobot training script for 10 episodes.
 ```
-python -m lerobot.scripts.eval \
+lerobot-eval \
    --policy.path=outputs/train/diffusion_pusht/checkpoints/005000/pretrained_model \
    --env.type=pusht \
    --eval.batch_size=10 \
@@ -243,7 +243,11 @@ def eval_policy(
    if max_episodes_rendered > 0 and not videos_dir:
        raise ValueError("If max_episodes_rendered > 0, videos_dir must be provided.")

-    if not isinstance(policy, PreTrainedPolicy):
+    # Handle accelerate-wrapped models by unwrapping them
+    if hasattr(policy, 'module') and isinstance(policy.module, PreTrainedPolicy):
+        # This is likely an accelerate-wrapped model (DistributedDataParallel)
+        policy = policy.module
+    elif not isinstance(policy, PreTrainedPolicy):
        raise ValueError(
            f"Policy of type 'PreTrainedPolicy' is expected, but type '{type(policy)}' was provided."
        )
@@ -302,11 +302,6 @@ class RobotClient:

                    self.logger.debug(f"Current latest action: {latest_action}")

-                    # Get queue state before changes
-                    old_size, old_timesteps = self._inspect_action_queue()
-                    if not old_timesteps:
-                        old_timesteps = [latest_action]  # queue was empty
-
                    # Get queue state before changes
                    old_size, old_timesteps = self._inspect_action_queue()
                    if not old_timesteps:
@@ -16,6 +16,7 @@
 import logging
 import time
 from contextlib import nullcontext
+from functools import partial
 from pprint import pformat
 from typing import Any

@@ -23,6 +24,8 @@ import torch
 from termcolor import colored
 from torch.amp import GradScaler
 from torch.optim import Optimizer
+import os
+from datetime import timedelta

 from lerobot.configs import parser
 from lerobot.configs.train import TrainPipelineConfig
@@ -52,6 +55,8 @@ from lerobot.utils.utils import (
 )
 from lerobot.utils.wandb_utils import WandBLogger

+def is_launched_with_accelerate() -> bool:
+    return "ACCELERATE_MIXED_PRECISION" in os.environ

 def update_policy(
    train_metrics: MetricsTracker,
@@ -59,36 +64,65 @@ def update_policy(
    batch: Any,
    optimizer: Optimizer,
    grad_clip_norm: float,
-    grad_scaler: GradScaler,
+    grad_scaler: GradScaler | None,
    lr_scheduler=None,
    use_amp: bool = False,
    lock=None,
+    accelerator=None,
 ) -> tuple[MetricsTracker, dict]:
    start_time = time.perf_counter()
    device = get_device_from_parameters(policy)
    policy.train()
-    with torch.autocast(device_type=device.type) if use_amp else nullcontext():
-        loss, output_dict = policy.forward(batch)
+    
+    grad_norm = 0.0  # Initialize grad_norm to avoid undefined variable
+
+    if accelerator:
+        with accelerator.accumulate(policy):
+            with torch.autocast(device_type=device.type) if use_amp else nullcontext():
+                loss, output_dict = policy.forward(batch)
+            # TODO(rcadene): policy.unnormalize_outputs(out_dict)
+            accelerator.backward(loss)
+            if accelerator.sync_gradients:
+                grad_norm = torch.nn.utils.clip_grad_norm_(
+                    policy.parameters(),
+                    grad_clip_norm,
+                    error_if_nonfinite=False,
+                )
+            optimizer.step()
+            optimizer.zero_grad()
+    else:
+        # Standard training loop without accelerate
+        with torch.autocast(device_type=device.type) if use_amp else nullcontext():
+            loss, output_dict = policy.forward(batch)
        # TODO(rcadene): policy.unnormalize_outputs(out_dict)
-    grad_scaler.scale(loss).backward()
+        
+        if grad_scaler is not None:
+            grad_scaler.scale(loss).backward()
+            # Unscale the gradient of the optimizer's assigned params in-place **prior to gradient clipping**.
+            grad_scaler.unscale_(optimizer)
+            grad_norm = torch.nn.utils.clip_grad_norm_(
+                policy.parameters(),
+                grad_clip_norm,
+                error_if_nonfinite=False,
+            )
+            # Optimizer's gradients are already unscaled, so scaler.step does not unscale them,
+            # although it still skips optimizer.step() if the gradients contain infs or NaNs.
+            with lock if lock is not None else nullcontext():
+                grad_scaler.step(optimizer)
+            # Updates the scale for next iteration.
+            grad_scaler.update()
+        else:
+            # Without GradScaler (fallback)
+            loss.backward()
+            grad_norm = torch.nn.utils.clip_grad_norm_(
+                policy.parameters(),
+                grad_clip_norm,
+                error_if_nonfinite=False,
+            )
+            with lock if lock is not None else nullcontext():
+                optimizer.step()

-    # Unscale the gradient of the optimizer's assigned params in-place **prior to gradient clipping**.
-    grad_scaler.unscale_(optimizer)
-
-    grad_norm = torch.nn.utils.clip_grad_norm_(
-        policy.parameters(),
-        grad_clip_norm,
-        error_if_nonfinite=False,
-    )
-
-    # Optimizer's gradients are already unscaled, so scaler.step does not unscale them,
-    # although it still skips optimizer.step() if the gradients contain infs or NaNs.
-    with lock if lock is not None else nullcontext():
-        grad_scaler.step(optimizer)
-    # Updates the scale for next iteration.
-    grad_scaler.update()
-
-    optimizer.zero_grad()
+        optimizer.zero_grad()

    # Step through pytorch scheduler at every batch instead of epoch
    if lr_scheduler is not None:
@@ -99,7 +133,7 @@ def update_policy(
        policy.update()

    train_metrics.loss = loss.item()
-    train_metrics.grad_norm = grad_norm.item()
+    train_metrics.grad_norm = grad_norm.item() if isinstance(grad_norm, torch.Tensor) else grad_norm
    train_metrics.lr = optimizer.param_groups[0]["lr"]
    train_metrics.update_s = time.perf_counter() - start_time
    return train_metrics, output_dict
@@ -108,8 +142,33 @@ def update_policy(
@parser.wrap()
 def train(cfg: TrainPipelineConfig):
    cfg.validate()
+    
+    accelerator = None
+    if is_launched_with_accelerate():
+        import accelerate
+
+        # For example pi0 has unused params (last llm block)
+        from accelerate import DistributedDataParallelKwargs
+        ddp_kwargs = DistributedDataParallelKwargs(find_unused_parameters=True)
+        # accelerator = accelerate.Accelerator(step_scheduler_with_optimizer=False, kwargs_handlers=[ddp_kwargs])
+        from accelerate import InitProcessGroupKwargs
+        # Set NCCL timeout (default 30 minutes = 1800 seconds)
+        nccl_timeout = getattr(cfg, 'nccl_timeout', 1800)
+        ddp_init_kwargs = InitProcessGroupKwargs(timeout=timedelta(seconds=nccl_timeout)) # FIXME(mshukor): allow user to set timeout. This should be longer than the evaluation time
+        # Set gradient accumulation steps (default 1)
+        gradient_accumulation_steps = getattr(cfg, 'gradient_accumulation_steps', 1)
+        accelerator = accelerate.Accelerator(step_scheduler_with_optimizer=False, gradient_accumulation_steps=gradient_accumulation_steps, kwargs_handlers=[ddp_init_kwargs, ddp_kwargs])
+        if accelerator is not None and not accelerator.is_main_process:
+            # Disable duplicate logging on non-main processes
+            logging.info(f"Setting logging level on non-main process {accelerator.process_index} to WARNING.")
+            logging.getLogger().setLevel(logging.WARNING)
+
    logging.info(pformat(cfg.to_dict()))

+    if accelerator and not accelerator.is_main_process:
+            # Disable logging on non-main processes.
+            cfg.wandb.enable = False
+
    if cfg.wandb.enable and cfg.wandb.project:
        wandb_logger = WandBLogger(cfg)
    else:
@@ -143,7 +202,8 @@ def train(cfg: TrainPipelineConfig):

    logging.info("Creating optimizer and scheduler")
    optimizer, lr_scheduler = make_optimizer_and_scheduler(cfg, policy)
-    grad_scaler = GradScaler(device.type, enabled=cfg.policy.use_amp)
+    # Only use GradScaler when not using accelerate (accelerate handles mixed precision internally)
+    grad_scaler = None if accelerator else GradScaler(device.type, enabled=cfg.policy.use_amp)

    step = 0  # number of policy updates (forward + backward + optim)

@@ -185,6 +245,11 @@ def train(cfg: TrainPipelineConfig):
    )
    dl_iter = cycle(dataloader)

+    # Prepare models for accelerate if using multi-GPU
+    if accelerator:
+        policy, optimizer, dataloader = accelerator.prepare(policy, optimizer, dataloader)
+        dl_iter = cycle(dataloader)
+
    policy.train()

    train_metrics = {
@@ -205,9 +270,10 @@ def train(cfg: TrainPipelineConfig):
        batch = next(dl_iter)
        train_tracker.dataloading_s = time.perf_counter() - start_time

-        for key in batch:
-            if isinstance(batch[key], torch.Tensor):
-                batch[key] = batch[key].to(device, non_blocking=device.type == "cuda")
+        if not accelerator:
+            for key in batch:
+                if isinstance(batch[key], torch.Tensor):
+                    batch[key] = batch[key].to(device, non_blocking=device.type == "cuda")

        train_tracker, output_dict = update_policy(
            train_tracker,
@@ -218,6 +284,7 @@ def train(cfg: TrainPipelineConfig):
            grad_scaler=grad_scaler,
            lr_scheduler=lr_scheduler,
            use_amp=cfg.policy.use_amp,
+            accelerator=accelerator,
        )

        # Note: eval and checkpoint happens *after* the `step`th training update has completed, so we
@@ -237,15 +304,17 @@ def train(cfg: TrainPipelineConfig):
                wandb_logger.log_dict(wandb_log_dict, step)
            train_tracker.reset_averages()

-        if cfg.save_checkpoint and is_saving_step:
+        if cfg.save_checkpoint and is_saving_step and (not accelerator or accelerator.is_main_process):
            logging.info(f"Checkpoint policy after step {step}")
            checkpoint_dir = get_step_checkpoint_dir(cfg.output_dir, cfg.steps, step)
-            save_checkpoint(checkpoint_dir, step, cfg, policy, optimizer, lr_scheduler)
+            # Unwrap model for accelerate
+            policy_to_save = accelerator.unwrap_model(policy) if accelerator else policy
+            save_checkpoint(checkpoint_dir, step, cfg, policy_to_save, optimizer, lr_scheduler)
            update_last_checkpoint(checkpoint_dir)
            if wandb_logger:
                wandb_logger.log_policy(checkpoint_dir)

-        if cfg.env and is_eval_step:
+        if cfg.env and is_eval_step and (not accelerator or accelerator.is_main_process):
            step_id = get_step_identifier(step, cfg.steps)
            logging.info(f"Eval policy at step {step}")
            with (
@@ -254,7 +323,7 @@ def train(cfg: TrainPipelineConfig):
            ):
                eval_info = eval_policy(
                    eval_env,
-                    policy,
+                    accelerator.unwrap_model(policy) if accelerator else policy,
                    cfg.eval.n_episodes,
                    videos_dir=cfg.output_dir / "eval" / f"videos_step_{step_id}",
                    max_episodes_rendered=4,
@@ -18,7 +18,7 @@ Helper to set motor ids and baudrate.
 Example:

 ```shell
-python -m lerobot.setup_motors \
+lerobot-setup-motors \
    --teleop.type=so100_leader \
    --teleop.port=/dev/tty.usbmodem575E0031751
 ```
@@ -18,7 +18,7 @@ Simple script to control a robot from teleoperation.
 Example:

 ```shell
-python -m lerobot.teleoperate \
+lerobot-teleoperate \
    --robot.type=so101_follower \
    --robot.port=/dev/tty.usbmodem58760431541 \
    --robot.cameras="{ front: {type: opencv, index_or_path: 0, width: 1920, height: 1080, fps: 30}}" \
@@ -32,7 +32,7 @@ python -m lerobot.teleoperate \
 Example teleoperation with bimanual so100:

 ```shell
-python -m lerobot.teleoperate \
+lerobot-teleoperate \
  --robot.type=bi_so100_follower \
  --robot.left_arm_port=/dev/tty.usbmodem5A460851411 \
  --robot.right_arm_port=/dev/tty.usbmodem5A460812391 \
@@ -44,7 +44,7 @@ Below is the short version on how to train and run inference/eval:
 ### Train from scratch

 ```bash
-python -m lerobot.scripts.train \
+lerobot-train \
  --dataset.repo_id=${HF_USER}/<dataset> \
  --policy.type=act \
  --output_dir=outputs/train/<desired_policy_repo_id> \
@@ -59,7 +59,7 @@ _Writes checkpoints to `outputs/train/<desired_policy_repo_id>/checkpoints/`._
 ### Evaluate the policy/run inference

 ```bash
-python -m lerobot.record \
+lerobot-record \
  --robot.type=so100_follower \
  --dataset.repo_id=<hf_user>/eval_<dataset> \
  --policy.path=<hf_user>/<desired_policy_repo_id> \
@@ -17,10 +17,9 @@ import time


 def busy_wait(seconds):
-    if platform.system() == "Darwin":
-        # On Mac, `time.sleep` is not accurate and we need to use this while loop trick,
+    if platform.system() == "Darwin" or platform.system() == "Windows":
+        # On Mac and Windows, `time.sleep` is not accurate and we need to use this while loop trick,
        # but it consumes CPU cycles.
-        # TODO(rcadene): find an alternative: from python 11, time.sleep is precise
        end_time = time.perf_counter() + seconds
        while time.perf_counter() < end_time:
            pass
@@ -60,11 +60,39 @@ def load_training_step(save_dir: Path) -> int:


 def update_last_checkpoint(checkpoint_dir: Path) -> Path:
+    import fcntl
+    import tempfile
+    import os
+    
    last_checkpoint_dir = checkpoint_dir.parent / LAST_CHECKPOINT_LINK
-    if last_checkpoint_dir.is_symlink():
-        last_checkpoint_dir.unlink()
    relative_target = checkpoint_dir.relative_to(checkpoint_dir.parent)
-    last_checkpoint_dir.symlink_to(relative_target)
+    
+    # Use file locking to prevent race conditions in multi-GPU training
+    lock_file = checkpoint_dir.parent / ".symlink_lock"
+    
+    try:
+        with open(lock_file, 'w') as f:
+            # Get exclusive lock
+            fcntl.flock(f.fileno(), fcntl.LOCK_EX)
+            
+            # Update symlink atomically
+            if last_checkpoint_dir.exists() or last_checkpoint_dir.is_symlink():
+                last_checkpoint_dir.unlink()
+            last_checkpoint_dir.symlink_to(relative_target)
+            
+    except (OSError, FileExistsError) as e:
+        # Handle race conditions gracefully - another process may have already updated
+        if not last_checkpoint_dir.exists():
+            try:
+                last_checkpoint_dir.symlink_to(relative_target)
+            except FileExistsError:
+                pass  # Another process created it, that's fine
+    finally:
+        # Clean up lock file
+        try:
+            lock_file.unlink()
+        except FileNotFoundError:
+            pass


 def save_checkpoint(
@@ -0,0 +1,45 @@
+#!/bin/bash
+
+echo "=== Local 1-GPU Accelerate Training Test with SmolVLA ==="
+echo "Environment: multi"
+echo "GPU: 1"
+echo "Steps: 50 (quick local test)"
+echo ""
+
+# Activate conda environment
+source /fsx/dana_aubakirova/miniconda3/etc/profile.d/conda.sh
+conda activate multi
+
+# Set CUDA environment for 1 GPU
+export CUDA_VISIBLE_DEVICES=0
+export PYTORCH_CUDA_ALLOC_CONF=max_split_size_mb:128,expandable_segments:True
+export TORCH_DISTRIBUTED_DEBUG=OFF
+export CUDA_LAUNCH_BLOCKING=0
+export TRANSFORMERS_NO_ADVISORY_WARNINGS=1
+
+# Change to working directory
+cd /fsx/dana_aubakirova/vla/pr/lerobot
+
+# Set output directory with timestamp
+export OUTPUT_DIR="outputs/test_accelerate_1gpu_local_$(date +%Y%m%d_%H%M%S)"
+
+echo "Output directory: $OUTPUT_DIR"
+echo ""
+
+# Test accelerate training with 1 GPU
+accelerate launch --config_file accelerate_configs/1gpu_config.yaml -m lerobot.scripts.train \
+    --policy.path=lerobot/smolvla_base \
+    --policy.push_to_hub=false \
+    --dataset.repo_id=lerobot/svla_so100_sorting \
+    --dataset.video_backend=pyav \
+    --steps=50 \
+    --save_freq=25 \
+    --log_freq=5 \
+    --batch_size=1 \
+    --num_workers=0 \
+    --output_dir=$OUTPUT_DIR \
+    --wandb.enable=false
+
+echo ""
+echo "=== Training completed! ==="
+echo "Check outputs in: $OUTPUT_DIR"
@@ -0,0 +1,67 @@
+#!/bin/bash
+#SBATCH --job-name=test_accelerate
+#SBATCH --nodes=1
+#SBATCH --ntasks-per-node=1
+#SBATCH --cpus-per-task=16
+#SBATCH --gres=gpu:2
+#SBATCH --time=1:00:00
+#SBATCH --partition=hopper-prod
+#SBATCH --output=/fsx/dana_aubakirova/vla/logs/test_accelerate_%j.out
+#SBATCH --error=/fsx/dana_aubakirova/vla/logs/test_accelerate_%j.err
+
+# Create logs directory if it doesn't exist
+mkdir -p /fsx/dana_aubakirova/vla/pr/lerobot/logs
+
+# Activate conda environment
+source /fsx/dana_aubakirova/miniconda3/etc/profile.d/conda.sh
+conda activate multi
+
+# 2-GPU Test CUDA environment
+export CUDA_VISIBLE_DEVICES=0,1
+export PYTORCH_CUDA_ALLOC_CONF=max_split_size_mb:128,expandable_segments:True
+export TORCH_DISTRIBUTED_DEBUG=OFF
+export NCCL_DEBUG=INFO
+export CUDA_LAUNCH_BLOCKING=0
+export ACCELERATE_USE_FSDP=false
+export ACCELERATE_USE_DEEPSPEED=false
+export HF_ACCELERATE_DEVICE_MAP=false
+export TRANSFORMERS_NO_ADVISORY_WARNINGS=1
+export SAFETENSORS_FAST_GPU=1
+export HF_HUB_ENABLE_HF_TRANSFER=1
+export PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True
+export ACCELERATE_TORCH_DEVICE_MAP_AUTO=false
+
+# Change to working directory
+cd /fsx/dana_aubakirova/vla/pr/lerobot
+
+echo "=== Testing Accelerate Multi-GPU Training with SmolVLA ==="
+echo "Dataset: lerobot/svla_so100_sorting"
+echo "GPUs: 2"
+echo "Steps: 100 (for quick test)"
+echo "Job ID: $SLURM_JOB_ID"
+echo ""
+
+# Set output directory with job ID
+export OUTPUT_DIR="outputs/test_accelerate_2gpu_job_${SLURM_JOB_ID}"
+
+echo "Output directory: $OUTPUT_DIR"
+echo ""
+
+# Test accelerate training
+accelerate launch --config_file accelerate_configs/2gpu_config_safe.yaml -m lerobot.scripts.train \
+    --policy.type=smolvla \
+    --policy.push_to_hub=false \
+    --dataset.repo_id=lerobot/svla_so100_sorting \
+    --dataset.video_backend=pyav \
+    --steps=100 \
+    --save_freq=50 \
+    --log_freq=5 \
+    --batch_size=2 \
+    --num_workers=0 \
+    --output_dir=$OUTPUT_DIR \
+    --wandb.enable=false
+
+echo ""
+echo "=== Training completed! ==="
+echo "Check logs and outputs in: $OUTPUT_DIR"
+echo "Job ID: $SLURM_JOB_ID"
@@ -0,0 +1,45 @@
+#!/bin/bash
+
+echo "=== Direct 1-GPU Training Test with SmolVLA (no accelerate) ==="
+echo "Environment: multi"
+echo "GPU: 1"
+echo "Steps: 50 (quick local test)"
+echo ""
+
+# Activate conda environment
+source /fsx/dana_aubakirova/miniconda3/etc/profile.d/conda.sh
+conda activate multi
+
+# Set CUDA environment for 1 GPU
+export CUDA_VISIBLE_DEVICES=0
+export PYTORCH_CUDA_ALLOC_CONF=max_split_size_mb:128,expandable_segments:True
+export TORCH_DISTRIBUTED_DEBUG=OFF
+export CUDA_LAUNCH_BLOCKING=0
+export TRANSFORMERS_NO_ADVISORY_WARNINGS=1
+
+# Change to working directory
+cd /fsx/dana_aubakirova/vla/pr/lerobot
+
+# Set output directory with timestamp
+export OUTPUT_DIR="outputs/test_direct_1gpu_local_$(date +%Y%m%d_%H%M%S)"
+
+echo "Output directory: $OUTPUT_DIR"
+echo ""
+
+# Test direct training with 1 GPU (no accelerate)
+python -m lerobot.scripts.train \
+    --policy.path=lerobot/smolvla_base \
+    --policy.push_to_hub=false \
+    --dataset.repo_id=lerobot/svla_so100_sorting \
+    --dataset.video_backend=pyav \
+    --steps=50 \
+    --save_freq=25 \
+    --log_freq=5 \
+    --batch_size=1 \
+    --num_workers=0 \
+    --output_dir=$OUTPUT_DIR \
+    --wandb.enable=false
+
+echo ""
+echo "=== Training completed! ==="
+echo "Check outputs in: $OUTPUT_DIR"
Author	SHA1	Message	Date
danaaubakirova	d148279921	Support accelerate training and add test configs for SmolVLA - 2-GPU SLURM job (distributed training) - 1-GPU local accelerate and direct training scripts - Accelerate configs for 1-GPU and 2-GPU setups	2025-09-04 13:07:25 +00:00
Pepijn	882c80d446	Lower limits by 50% for current and torque for gripper motor (#1809 ) Signed-off-by: Pepijn <138571049+pkooij@users.noreply.github.com>	2025-08-29 16:06:55 +02:00
Pepijn	61b0eeae4b	Add feetech firmware update docs (#1793 ) * Add feetech firmware update docs * add bonus * formatting * adapt text * feedback pr	2025-08-28 11:18:54 +02:00
mgiac-hexagon	577cd10974	Removed dupicate lines of code (#1709 )	2025-08-25 12:39:32 +02:00
lxk	b0923ab74b	fix(dataset): Use provided episode_data in save_episode (#1740 ) The 'episode_data' parameter was previously ignored, causing an error if provided. This change ensures it is correctly used, which allows for asynchronous episode saving by passing a copy of the episode buffer, preventing conflicts with the main data collection loop.	2025-08-22 15:24:02 +02:00
Jack Vial	7f70b78f32	Add missing encoding table entries for Koch arm (#1534 )	2025-08-20 17:24:05 +02:00
Steven Palma	55198de096	fix(ci): rename libegl1-mesa in deb13 trixie (#1735 )	2025-08-14 11:12:06 +02:00
Steven Palma	0878c6880f	fix(ci): inverted names (#1705 )	2025-08-09 00:21:42 +02:00
Caroline Pascal	11e6bd762a	fix(busy_wait): fix busy_wait implementation for Windows platforms and removing erronous TODO (#1695 )	2025-08-08 10:46:14 +02:00
Steven Palma	ce3b9f627e	chore(docs): prioritize use of entry points in docs + fix nightly badge (#1692 ) * chore(docs): fix typo in nightly badge * chore(docs): prioritize the use of entrypoints for consistency	2025-08-07 14:25:44 +02:00