Datasets:
Data-quality audit on DROID-100 — 66% REVIEW, 70.7% median idle
#3
by 375720783jin - opened
I ran a trajectory-only data-quality audit on this dataset (100 episodes) using RDA, an open-source CLI tool: https://github.com/liesliy/rda
Results:
- 34 PASS / 66 REVIEW / 0 EXCLUDE
- Action spikes in 1,428 frames across 99% of episodes
- Median idle: 70.7%
Integrity (NaN/Inf, timestamps, schema) is clean across all 100 episodes. The behavior layer flags 66% for review — mostly action discontinuities and idle ratios above 65%.
Full benchmark: https://github.com/liesliy/rda/blob/main/docs/benchmark.md
Question: Does this REVIEW rate match what people have seen when training on DROID-100? Would anyone be willing to spot-check 5–10 of the flagged episodes to confirm whether the flags correspond to lower-quality demonstrations? That feedback would directly validate the tool's accuracy.