LeRobotDataset v3.0 adds multi-episode storage and streaming in lerobot
AI Impact Summary
LeRobotDataset v3.0 introduces multi-episode files with relational metadata and a streaming access path via StreamingLeRobotDataset, enabling on-the-fly processing of large robotics datasets and reducing local filesystem pressure. The data layout separates tabular data (Parquet), video (MP4), and metadata, with episode-level pointers to support concatenation across episodes; a one-liner converter provides a migration path from the previous format. To consume the new format, tooling must support the StreamingLeRobotDataset interface and newer lerobot releases (v0.4.0+), and datasets on the Hugging Face Hub can be indexed and searched using the updated structure.
Affected Systems
- Date
- Date not specified
- Change type
- capability
- Severity
- info