LeRobot
LeRobot is Hugging Face's open source library for real-world robot learning. It combines the LeRobotDataset format for storing robot demonstrations, implementations of imitation learning and vision-language-action policies, and a common interface for recording data from robots and deploying policies to them. LeRobot datasets are shared on the Hugging Face Hub, which has made the format a common way to publish robot learning data.
In this article
Building visual or physical AI?
Let's talk.
What is LeRobot?
LeRobot is an open source project from Hugging Face, written in PyTorch and released under the Apache 2.0 license, that aims to lower the barrier to entry for robot learning. It brings together three pieces that teams used to assemble on their own: a standard dataset format, ready-to-train policies such as ACT, Diffusion Policy, and SmolVLA, and a hardware-agnostic interface that supports low-cost arms like the SO-101 as well as humanoids and mobile robots.
The LeRobotDataset format reflects a set of practical priorities. Camera streams are stored as compressed MP4 video, joint states and actions are stored in Apache Parquet files, and metadata files describe the schema, tasks, and episode boundaries. The current version, v3.0, packs many episodes into each file and supports streaming directly from the Hugging Face Hub. Compared with fully lossless research formats, LeRobot favors compact storage and ease of sharing, which is part of why community contributors have published thousands of LeRobot datasets.
Key takeaways
- LeRobot is Hugging Face's open source library and dataset format for real-world robot learning.
- LeRobot datasets store camera video in MP4, state and action data in Parquet, and metadata that defines episodes, tasks, and features.
- The library covers the full loop: recording with teleoperation, sharing on the Hub, training policies, and evaluating them.
What LeRobot provides
| Component | What it does |
|---|---|
| LeRobotDataset | Stores robot demonstrations as episodes, with video, tabular state and action data, and metadata |
| Policies | Implementations of imitation learning, reinforcement learning, and VLA policies such as ACT, Diffusion Policy, π0, and SmolVLA |
| Hardware interface | Records from and controls supported robots, including the SO-101, Koch, LeKiwi, Reachy 2, and Unitree G1 |
| Hugging Face Hub integration | Publishes, streams, and visualizes datasets and trained policies |
How it works
A typical LeRobot workflow starts with teleoperation: a person moves a leader arm while a follower arm mirrors it, and LeRobot records joint positions, commanded actions, and synchronized camera frames as episodes. The episodes are written into a LeRobot dataset and pushed to the Hugging Face Hub. From there, the same dataset feeds training, where policies learn to map observations to actions, and evaluation, where trained policies run in simulation benchmarks or on the physical robot. Evaluation rollouts can be recorded in the same format, so failures become new training data.
Why it matters
LeRobot matters because it gives robot learning a shared format and a shared place to publish data. Small datasets recorded on inexpensive arms can be pooled into training mixes no single team could collect, and SmolVLA was pretrained on hundreds of these community datasets. Pooled community data also varies in quality, naming, and recording setup, which makes inspecting and curating LeRobot datasets an essential step before training.
Frequently asked questions
What is the LeRobot dataset format?
LeRobotDataset is the format LeRobot uses to store robot demonstrations. It organizes data into episodes, with state, action, and timestamp data in Parquet files, camera streams in MP4 video, and metadata describing the schema, tasks, and episode boundaries. The current version is v3.0.
How does LeRobot differ from RLDS?
Both are episode-based formats for robot learning data. RLDS stores episodes as ordered steps in TensorFlow Datasets and is the format behind Open X-Embodiment. LeRobot stores tabular data in Parquet and camera streams as compressed video, loads natively in PyTorch, and streams from the Hugging Face Hub. Converters exist in both directions.
Who maintains LeRobot?
Hugging Face's LeRobot team maintains the library, with contributions from a large open source community. The official documentation is at huggingface.co/docs/lerobot, and the source code is on GitHub.
Related terms
Director of Growth
Jesse Mostipak is the Director of Growth at Voxel51, where the work is helping humans find and trust what the brand knows, and teaching the Google knowledge graph and the LLMs answering on their behalf to do the same. That question, how knowledge gets built inside a system, is one Jesse has been chasing for years. Earlier versions of it ran through a New York City high school science classroom, data science and machine learning, and developer relations at Kaggle, Posit (formerly RStudio), and Baseten. The answer doesn't change much depending on whether the learner is a teenager, a software engineer, or a knowledge graph. Jesse holds a Master's in Education from CUNY Hunter College. LinkedIn
See all articles by Jesse Mostipak
Last updated October 8, 2026
Building visual or physical AI?
Let's talk.
© 2026 Voxel51 All Rights Reserved