Frequently Asked Questions

Find answers to common questions about LeRobot installation, usage, and troubleshooting.

Getting Started

What is LeRobot?

LeRobot is an open-source library that provides models, datasets, and tools for real-world robotics in PyTorch. It aims to lower the barrier to entry for robot learning by providing:

Hardware-agnostic robot control interface
Standardized dataset format hosted on Hugging Face Hub
State-of-the-art policies for imitation learning, RL, and VLA models
Tools for data collection, training, and evaluation

What Python version does LeRobot require?

LeRobot requires Python 3.12 or higher. We recommend using conda or uv to manage your Python environment.

# With conda
conda create -y -n lerobot python=3.12
conda activate lerobot

How do I install LeRobot?

You can install LeRobot via pip or from source:

# From PyPI (stable release)
pip install lerobot

# From source (latest development)
git clone https://github.com/huggingface/lerobot.git
cd lerobot
pip install -e .

See the Installation Guide for detailed instructions.

What robots does LeRobot support?

LeRobot supports a wide range of hardware:

Robot Arms: SO100, Koch, LeKiwi, HopeJR, OpenARM
Mobile Manipulators: OMX, EarthRover
Humanoids: Reachy2, Unitree G1
Teleoperation Devices: Gamepads, Keyboards, Phones

The library is designed to be extensible - you can implement the Robot interface for custom hardware.

Datasets

How do I load a dataset from the Hugging Face Hub?

Use the LeRobotDataset class:

from lerobot.datasets.lerobot_dataset import LeRobotDataset

# Load from Hub
dataset = LeRobotDataset("lerobot/aloha_mobile_cabinet")

# Access data
frame = dataset[0]
print(frame.keys())  # See available features

What format are LeRobot datasets?

LeRobotDataset uses:

Parquet files for state/action data (efficient columnar storage)
MP4 videos or images for visual observations
metadata.json for dataset information

This format enables efficient streaming, storage, and visualization of large robotic datasets.

Can I use my own dataset?

Yes! You can convert your dataset to LeRobotDataset format or load it directly. See:

Porting Datasets guide
examples/port_datasets/ for conversion examples
Dataset tools for merging and manipulating datasets

How do I visualize a dataset?

You can visualize datasets directly on the Hugging Face Hub or programmatically:

# View on Hub
# Visit: https://huggingface.co/datasets/lerobot/aloha_mobile_cabinet

# Visualize in Python
import matplotlib.pyplot as plt

frame = dataset[0]
plt.imshow(frame['observation.images.top'])
plt.title(f"Episode {frame['episode_index']}, Frame {frame['frame_index']}")
plt.show()

Training

How do I train a policy?

Use the lerobot-train command:

lerobot-train \
  --policy=act \
  --dataset.repo_id=lerobot/aloha_mobile_cabinet

Or use the Python API for more control:

from lerobot.policies.act.modeling_act import ACTPolicy
from lerobot.datasets.lerobot_dataset import LeRobotDataset

dataset = LeRobotDataset("lerobot/aloha_mobile_cabinet")
policy = ACTPolicy(config, dataset.meta)
# ... training loop ...

What policies are available?

LeRobot provides several state-of-the-art policies:Imitation Learning:

ACT (Action Chunking Transformer)
Diffusion Policy
VQ-BeT

Reinforcement Learning:

HIL-SERL
TDMPC

Vision-Language-Action:

Pi0Fast, Pi0.5
GR00T N1.5
SmolVLA
XVLA

How long does training take?

Training time varies significantly based on:

Policy type: Diffusion policies generally take longer than ACT
Dataset size: More episodes = longer training
Hardware: GPU type and number of GPUs
Hyperparameters: Batch size, number of steps

Example: Training ACT on PushT with a single GPU typically takes 2-4 hours for 100k steps.

Can I resume training from a checkpoint?

Yes! Use the --resume flag:

lerobot-train \
  --policy=act \
  --dataset.repo_id=lerobot/aloha_mobile_cabinet \
  --resume=outputs/train/my_checkpoint

How do I use multiple GPUs?

LeRobot supports distributed training via PyTorch DDP:

# Multi-GPU training
torchrun --nproc_per_node=4 -m lerobot.scripts.train \
  --policy=act \
  --dataset.repo_id=lerobot/aloha_mobile_cabinet

Inference and Deployment

How do I load a pretrained model?

Use the make_policy factory:

from lerobot.policies.factory import make_policy

# Load from Hub
policy = make_policy(pretrained="lerobot/act_aloha_sim_transfer_cube_human")

# Run inference
action = policy.select_action(observation)

How do I evaluate a policy?

Use the lerobot-eval command:

# Evaluate in simulation
lerobot-eval \
  --policy.path=lerobot/pi0_libero_finetuned \
  --env.type=libero \
  --env.task=libero_object \
  --eval.n_episodes=10

Can I deploy policies on real robots?

Yes! LeRobot is designed for real-world deployment:

Train your policy on real robot data or simulation
Load the policy and connect to your robot
Run the control loop

from lerobot.robots.myrobot import MyRobot
from lerobot.policies.factory import make_policy

robot = MyRobot()
policy = make_policy(pretrained="my_model")

while True:
    obs = robot.get_observation()
    action = policy.select_action(obs)
    robot.send_action(action)

What about latency in real-time control?

For low-latency control, consider:

Using async inference (see examples/tutorial/async-inf/)
Running on GPU for faster inference
Optimizing with TorchScript or ONNX
Using action chunking (ACT-style policies)

Hardware Integration

How do I integrate my custom robot?

Implement the Robot interface:

Create a new robot class inheriting from base Robot
Implement required methods: connect(), get_observation(), send_action()
Define your robot’s features and configuration

See the Robot API for details on implementing custom robots.

Do I need special hardware?

Not necessarily! You can:

Start with simulation environments (PushT, LIBERO)
Use low-cost hardware like SO-100 arms
Implement support for your existing robots

LeRobot works with a wide range of hardware from hobbyist to professional.

How do I calibrate my robot?

Calibration depends on your specific robot. Generally:

Follow your robot’s calibration procedure
Use LeRobot’s calibration tools if available
Record the calibration parameters in your robot config

For SO-100/101 robots, see the hardware-specific documentation.

Troubleshooting

I'm getting CUDA out of memory errors

Try these solutions:

Reduce batch size in your config
Use gradient accumulation
Enable mixed precision training (AMP)
Use a smaller model variant
Clear CUDA cache: torch.cuda.empty_cache()

See Troubleshooting for more details.

ffmpeg errors during dataset loading

Ensure you have ffmpeg with libsvtav1 support:

# Check encoders
ffmpeg -encoders | grep svt

# Reinstall if needed (with conda)
conda install ffmpeg=7.1.1 -c conda-forge

Where can I get help?

Discord: Join the LeRobot server for community support
GitHub Issues: Report bugs
Discussions: Ask questions
Documentation: Browse the full docs

Contributing

How can I contribute to LeRobot?

There are many ways to contribute:

Fix bugs or add features
Improve documentation
Share datasets on the Hub
Add support for new robots or policies
Help others in Discord

See the Contributing Guide to get started.

Do I need to be an expert to contribute?

Not at all! Contributions at all levels are welcome:

Documentation improvements
Bug reports
Example notebooks
Community support

Everyone started as a beginner - we’re here to help you learn!

Can't find your question?

Ask in the LeRobot Discord - the community is happy to help!

Community

FAQ

Frequently Asked Questions

Getting Started

Datasets

Training

Inference and Deployment

Hardware Integration

Troubleshooting

Contributing

Can't find your question?

Build docs developers (and LLMs) love

Community

FAQ

​Getting Started

​Datasets

​Training

​Inference and Deployment

​Hardware Integration

​Troubleshooting

​Contributing

Can't find your question?

Build docs developers (and LLMs) love

Getting Started

Datasets

Training

Inference and Deployment

Hardware Integration

Troubleshooting

Contributing