LeRobot

0.5.1 · active · verified Thu Apr 16

LeRobot is a state-of-the-art open-source Python library for real-world robotics using PyTorch, developed by Hugging Face. It provides tools for creating robot environments, loading pre-trained policies, and managing robotics datasets. It is currently at version 0.5.1 and is under active development, with new releases often including significant updates and breaking changes.

Common errors

```
TypeError: DatasetFactory missing 1 required positional argument: 'root_dir'
```
cause Attempting to instantiate `DatasetFactory` without providing the required `root_dir` argument, which became mandatory in v0.5.0.

fix Pass a `root_dir` argument: `DatasetFactory('hf_dataset_id', root_dir='./local_data')`.
```
ValueError: The 'hf_hub_id' argument has been removed, please use 'repo_id' instead.
```
cause Using the old argument name `hf_hub_id` instead of `repo_id` for Hugging Face Hub identifiers.

fix Rename the argument: `load_policy(repo_id='your/model')`.
```
AttributeError: 'Policy' object has no attribute 'sample'
```
cause Using the deprecated `policy.sample()` method instead of the updated `policy.sample_actions(obs)`.

fix Change the method call to `action = policy.sample_actions(obs)` and ensure `obs` is the current observation.
```
ModuleNotFoundError: No module named 'lerobot'
```
cause The `lerobot` library is not installed in the current Python environment, or the environment is not activated.

fix Install the library using `pip install lerobot` or activate the correct virtual environment.

Warnings

breaking The `DatasetFactory` now requires the `root_dir` argument to be explicitly provided during instantiation. It was previously optional.
Fix: Ensure `DatasetFactory` is called with `DatasetFactory(root_dir='/path/to/data')`. For instance, `DatasetFactory("hf_dataset_id", root_dir="./data")`.
breaking The `hf_hub_id` argument in functions like `load_policy` has been renamed to `repo_id` for consistency with Hugging Face Hub terminology.
Fix: Replace `hf_hub_id='your/model'` with `repo_id='your/model'` when loading policies or datasets from the Hugging Face Hub.
breaking The signature and name of the policy sampling method have changed. `policy.sample()` is now `policy.sample_actions(obs)` and requires the current observation as input.
Fix: Update calls from `action = policy.sample()` to `action = policy.sample_actions(obs)`.
gotcha LeRobot requires Python 3.12+ and PyTorch >= 2.1. Incompatibilities can lead to installation issues or runtime errors.
Fix: Ensure your Python environment is `3.12` or newer, and PyTorch is `2.1` or newer. Use `pip install torch` to update PyTorch if needed.

Install

pip install lerobot Install core library
pip install 'lerobot[all]' Install with all optional dependencies (e.g. for mujoco)

Imports

load_policy

wrong:

from lerobot.common.policies.load import load_policy

correct:

from lerobot.common.policies.hf_hub.load import load_policy

The policy loading utility was moved to a submodule under `hf_hub`.

RobotEnvFactory

from lerobot.common.robot_env_factory import RobotEnvFactory

PolicyFactory

from lerobot.common.policies.factory import PolicyFactory

DatasetFactory

from lerobot.common.datasets.factory import DatasetFactory

Quickstart

This quickstart demonstrates how to load a pre-trained robotics policy from the Hugging Face Hub and use it to sample actions in a dummy environment. It showcases the core `load_policy` and `RobotEnvFactory` components.

from lerobot.common.policies.hf_hub.load import load_policy
from lerobot.common.robot_env_factory import RobotEnvFactory
import torch

# Load a pre-trained policy from the Hugging Face Hub
# This example uses a public model, so no explicit token is needed unless fine-tuning.
policy = load_policy("lerobot/diffusion_policy_v2_pen_push_multi_view_taco_play_sok_fixed_camera_v0")

# Create a dummy robot environment for demonstration
# Real environments require specific configurations like 'robomimic', 'taco_play', etc.
# e.g., env = RobotEnvFactory("taco_play", record_dir="./recordings")()
env = RobotEnvFactory("dummy")()

# Reset the environment to get initial observation
obs = env.reset()

# Sample actions from the policy
with torch.no_grad():
    action = policy.sample_actions(obs)

print(f"Initial observation keys: {obs.keys()}")
print(f"Sampled action shape: {action.shape}")
env.close()

view raw JSON →