How We Made Our LeRobot Video Reader Up to 15× Faster

We tackled the bottleneck of slow video decoding in robot learning pipelines by redesigning our LeRobot reader in Daft. Instead of opening remote files for every single frame, we now batch decode requests by shard, intelligently clustering targets to minimize seeks. This optimization reduced decoding time from linear to flat scaling, delivering up to 15× speed improvements across diverse datasets while keeping our GPU resources fully utilized.