LeRobot Community Datasets: The “ImageNet” of Robotics — When and How?

Hugging Face · 1 year ago · 17

LeRobot, a robotics platform, is building a community-driven dataset collection system to provide diverse training data for robot learning models, similar to how ImageNet transformed computer vision. The platform has simplified data collection and uploading to Hugging Face Hub, resulting in rapidly growing contributions including datasets for So100 and Koch robotic arms, with examples like chess games and drawer manipulation tasks. LeRobot's automatic curation pipeline identifies quality issues such as incomplete task annotations, inconsistent feature naming, and low-quality episodes, publishing best practices for future contributors including standardized camera naming conventions, 30 FPS recording, and detailed task descriptions.