- LeRobot is the dominant open format for robot learning data, but decoding frames is expensive and memory-intensive.
- The original Daft LeRobot reader opened each MP4 shard per frame, causing high latency (3 seconds per frame) and linear cost growth.
- The fix uses a batch UDF that processes 16 consecutive rows at a time, grouping rows by shard, sorting and clustering targets by timestamp, and seeking once per cluster.
- The batched reader is 4–13× faster on diverse datasets and up to 15× faster on a 1080p dataset (632 frames in under 2 minutes vs. 29 minutes).
- A hand-tracking pipeline using the optimized reader improved end-to-end time from 44.8s to 9.8s with identical detections.