LeRobot

What it does

LeRobot takes in robot demonstrations and turns them into trained control policies. You teleoperate a robot with a leader arm, phone, gamepad or keyboard, and each run is recorded as a dataset of camera video plus movements. The library stores these in its LeRobotDataset format (Parquet plus MP4 or images) on the Hugging Face Hub. It then trains a policy and outputs a model that can drive the robot or be evaluated in simulation.

The documented workflow has four steps: teleoperate, record, train, deploy, with the same commands across supported robots. Command-line tools such as lerobot-record, lerobot-train, lerobot-eval and lerobot-rollout cover each step. Policies include imitation learning (ACT, Diffusion), reinforcement learning (HIL-SERL) and vision-language-action models such as SmolVLA and Pi0. A hardware-agnostic Robot interface and community plugins extend support to other arms, cameras and sensors.

LeRobot suits roboticists, students and hobbyists who are comfortable with Python and the command line. It can be used without a robot, through community datasets, notebooks and simulation. Installation is via pip and needs Python 3.12 or later with a recent PyTorch. Training needs a GPU, your own or cloud. Real-robot use needs supported hardware; the SO-101 arm is the suggested starting point. LeLab, a browser GUI, currently supports SO-101 only.

What it can do

• Teleoperate robots with leader arms, phones, gamepads or keyboards
• Record demonstrations as LeRobotDataset (video plus state/action data)
• Train imitation learning, reinforcement learning and vision-language-action policies
• Evaluate policies in simulation benchmarks such as LIBERO and MetaWorld
• Run trained policies on real robots
• Extend support to other robots, teleoperators and cameras through community plugins

Works with

• Hugging Face Hub
• PyTorch
• Weights & Biases

Example workflows

Teach an SO-101 arm a tabletop task

Inputs: an assembled, calibrated SO-101 arm, a leader arm and cameras. Steps: teleoperate the task roughly 50 times, record the runs as a dataset, train a policy on your GPU or in the cloud, then run it on the robot. Output: a trained policy controlling the arm.

Explore robot learning without hardware

Inputs: a community dataset from the Hugging Face Hub. Steps: play it back in the browser visualizer, train a policy in a free notebook, then test it in a simulation environment. Output: a trained policy evaluated in simulation.

Before you choose

• The base install is lightweight, so most workflows need extras for datasets, training, hardware or specific policies.
• Real-robot use depends on supported hardware, and training needs GPU compute you provide or rent separately.
• LeLab, the browser interface, currently supports SO-101 only.

Listed under

• Coding
• Developers
• Imitation learning
• Machine learning
• Open source
• Open-source
• Researchers

Get started

See the installation guide for extras and ffmpeg

1. Create a Python 3.12 environment, for example with conda, and install ffmpeg if your platform needs it.
2. Run pip install lerobot, adding extras such as 'lerobot[core_scripts,feetech]' for recording with an SO-101 arm or 'lerobot[training]' for training.
3. Assemble and calibrate your robot following its hardware guide, then record demonstrations with lerobot-record.

The source code is available under Apache 2.0. Compute, robot hardware and any hosted services are separate.

Source: github.com/huggingface/lerobot, checked 28 Oct 2026. Prices can change; check before you buy.

Sources for this overview

Product pages, documentation, and pricing checked 28 Oct 2026. Each source supports part of the overview; availability and prices can change.

• lerobot (LeRobot) ↗huggingface.co
• GitHub - huggingface/lerobot: 🤗 LeRobot: Making AI for Robotics more accessible with end-to-end learning · GitHub ↗github.com
• Pricing - Hugging Face ↗huggingface.co
• Installation · Hugging Face ↗huggingface.co

Company website

huggingface.co: product, pricing and policy pages

YC directory mirror

Batch and company name

Web search

Confirms the official site

Social accounts

Accounts the company links, and its posts on X

Human review

Checked by a person

knowledge_base

Last checked 28 Oct 2026. Every field keeps its source and date; unknown means we couldn’t confirm it.