The open-source robotics community just got a major boost with the release of LeRobotDataset v3.0, a significant expansion of the popular Lerobot dataset. This new version dramatically increases the scale and diversity of real-world robotic manipulation data available for training and research.
Version 3.0 adds hundreds of thousands of new trajectories collected across multiple platforms and environments, including data from ALOHA, DROID, and other collaborative research projects. The dataset now covers a wide range of tasks such as assembling, opening/closing, and precise manipulation with dual-arm setups.
Key highlights of the release include:
- Scale: Over 100,000 new episodes, bringing the total to more than 500,000 trajectories.
- Diversity: Data from 15+ robot embodiments including Franka, UR5, ViperX, and custom research platforms.
- Quality: All data comes with human annotations and calibrated camera parameters.
- Accessibility: Released under permissive licenses with standardized format and tooling.
Researchers can now download the full dataset or subsets organized by task and environment directly from the Hugging Face Hub. The team behind LeRobotDataset has also updated their preprocessing and training pipelines to make it easier to use the new data with popular imitation learning algorithms.
This release is expected to accelerate progress in robot learning by providing a common benchmark for comparing methods and enabling training of more generalizable policies. The community is already integrating v3.0 into projects focused on zero-shot transfer and multi-task learning.
For detailed specifications and download instructions, visit the official LeRobotDataset documentation.