LeRobot

LeRobot is an open source project for robot learning that provides models, datasets, and tools with the explicit goal of making the field more accessible. The idea is to lower the barrier to entry so that more people can contribute to and benefit from shared datasets and pretrained models, in the same spirit that open toolkits opened up other areas of machine learning. Alongside the library sits the LeRobot dataset format, a standardized way of storing robot learning data that is designed to be easy to work with and to share.

The format reflects a particular set of priorities. Camera data is stored as compressed video rather than raw frames, while low-dimensional, high-frequency signals like joint states and actions are stored in efficient tabular files, and rich metadata supports indexing, search, and visualization. This makes datasets compact and convenient to download, explore, and train on. In contrast to fully lossless research standards, which prioritize preserving every detail of an episode, LeRobot leans toward accessibility, efficient storage, and practical deployment, which is part of why it has become popular for hands-on robot learning.

In the LeRobot format, the visual and non-visual parts of an episode are stored in the ways best suited to each. Camera frames from an episode are encoded together into compressed video, which keeps large image streams manageable, while joint states, actions, and similar high-frequency signals are kept in compact tabular files. Metadata ties everything together and enables browsing and searching collections of datasets. The library builds on widely used machine learning ecosystems, so datasets can be explored and loaded conveniently, and it offers practical operations for working with episodes, such as splitting, merging, and windowing data around a given moment for training.

LeRobot matters because accessibility is itself a driver of progress, and by making robot data and models easy to share and use, it broadens who can participate in robot learning. For anyone entering physical AI, it offers a practical on-ramp, and it illustrates a real design trade-off in how robot data is stored, between the compact, deployment-friendly approach it favors and the exhaustive fidelity of lossless research formats. Understanding where LeRobot sits on that spectrum helps clarify how to choose a data format for a given goal.

It is optimized for accessibility, efficient storage, and deployment. Camera data is stored as compressed video and other signals in compact tabular files, which makes datasets easy to download, explore, and train on.

RLDS prioritizes preserving the complete temporal detail of an episode as a research-grade, lossless standard. LeRobot leans instead toward being lightweight and convenient, using video compression and a practical structure, which favors ease of use and deployment over exhaustive fidelity.

It is aimed broadly at people who want to work on robot learning without a heavy setup burden, from researchers to newcomers. Its guiding goal is to lower the barrier to entry by sharing models, datasets, and tools in an accessible way.

RLDS, Open X-Embodiment, Episode / trajectory / rollout, Data curation, Imitation learning / behavior cloning.