Data-quality audit on 3 ALOHA sim datasets — 90% REVIEW, 100% spike rate

#4
by 375720783jin - opened

I ran a trajectory-only data-quality audit on all three ALOHA sim datasets in LeRobot format using RDA, an open-source CLI tool I built: https://github.com/liesliy/rda

Results (50 episodes each):

Dataset PASS / REVIEW / EXCLUDE Action spikes Median idle
aloha_sim_insertion_human 5 / 45 / 0 1,338 (100% eps) 70.7%
aloha_sim_transfer_cube_scripted 5 / 45 / 0 2,489 (100% eps) 64.0%
aloha_sim_insertion_scripted 1 / 49 / 0 1,633 (100% eps) 63.7%

Integrity (NaN/Inf, timestamps, schema) is clean across all 150 episodes. The behavior layer flags 90%+ for review — driven by action discontinuities in literally every episode and 63–70% idle ratios.

Full benchmark (11 datasets, 4,909 episodes): https://github.com/liesliy/rda/blob/main/docs/benchmark.md

Question for the community: Has anyone noticed whether the 100% spike rate affects policy training on these datasets? Are the action discontinuities inherent to the ALOHA sim controller, or a conversion artifact? I'd be happy to share the per-episode CSV if anyone wants to spot-check a few REVIEW episodes.

Sign up or log in to comment