04 / Research

Publication

CoRL 2026Conference on Robot Learning

An Empirical Study of Data-Loading Bottlenecks (and Non-Bottlenecks) in Robotic Imitation Learning

Mrinal Jain1, Arpit Gupta2, Animesh Garg1 · 1Georgia Tech  2General Motors

TL;DR Faster data loaders don't shorten robot-policy training. Unindexed video seeking does lengthen it. We profiled end-to-end LeRobot training (ACT, SmolVLA) against a Rust data engine we built to be as fast as possible. The loader turned out not to be the bottleneck.

1–3.5%of each training step is spent waiting for data. Model compute takes the rest.
3.94×isolated loader speedup over torchcodec, but 324 s vs 326 s end to end, within noise.
5.6×slower training from PyAV's unindexed AV1 seeking: the one real bottleneck.
Chart: share of each training step spent on compute versus data wait. Data wait is 1.1 to 3.5 percent in standard configurations and about 488 percent with PyAV seeking in AV1 video