Data-quality audit on bridge_orig_lerobot (sampled) — 93.3% median idle
#2
by 375720783jin - opened
I ran a trajectory-only data-quality audit on a 25-episode sample of this dataset using RDA, an open-source CLI tool: https://github.com/liesliy/rda
Results:
- 4 PASS / 21 REVIEW / 0 EXCLUDE
- Action spikes in 91 frames across 84% of episodes
- Median idle: 93.3% — the highest across all 11 public datasets I tested
Integrity (NaN/Inf, timestamps, schema) is clean across all 25 episodes. The behavior layer flags 84% for review, driven almost entirely by the 93.3% idle ratio.
Full benchmark (11 datasets, 4,909 episodes): https://github.com/liesliy/rda/blob/main/docs/benchmark.md
Context: I sampled 25 episodes from the full 53k set, so this is a snapshot — but the 93.3% idle is consistent across the sample. If anyone has trained on Bridge data and noticed the policy converging toward "do nothing," this might explain why. I'd be happy to share the per-episode CSV if anyone wants to dig deeper.