LeRobotDataset

LeRobot's standard dataset format, storing state and actions in tables and camera footage as video.

LeRobotDataset is the dataset format used by LeRobot, Hugging Face's open-source robotics library, and can be shared directly on the Hub. It splits data into three parts: low-dimensional signals such as state, actions, and timestamps are stored as Parquet tables; camera footage is encoded as MP4 video; and a meta directory records field definitions, frame rate, normalization statistics, task text, and each episode's start and end position. Version 3.0, released in September 2025, packs multiple episodes into the same file (v2.1 used one file per episode) and locates them by metadata, supporting much larger scale, and it can also be streamed directly from the Hub without downloading. Models in LeRobot such as ACT, Diffusion Policy, and SmolVLA all read this format directly.

ExampleUse lerobot-record to teleoperate an SO-101 arm while recording; the data is saved in v3.0 format and pushed to the Hub. During training, LeRobotDataset loads it, and each sample is a tensor dictionary containing observation.state, action, and camera images.

Also called

LeRobot Dataset, LeRobotDataset v3.0

Related

LeRobot · Hugging Face · Apache Parquet · Hierarchical Data Format version 5 · RLDS (Reinforcement Learning Datasets) · SO-100 / SO-101 Arm

Sources

Hugging Face LeRobot Docs: LeRobotDataset v3.0
Hugging Face Blog: LeRobotDataset v3.0

As of

2025-09