LeRobotDataset: video-encoded robotics datasets on Hugging Face Hub

AI Impact Summary

LeRobotDataset introduces video-encoded visuals for robotics data, addressing the lack of scalable lightweight formats. Datasets average 14% of their unencoded size, with compression possibly shrinking to 0.2% in best cases, while decoding single frames remains comparable to PNG and batched decoding runs 25–50% faster than image loading. The approach relies on Hub integration and visualization tools to share and browse datasets; teams will need to adjust data loaders to streams of encoded video and ensure alignment with proprioception/state vectors. Migration considerations include codec/container choices, keyframe settings, and preserving quality metrics to avoid degrading training results.