| Layer | Metric | Episodes measured | Pass rate |
|---|---|---|---|
| Layer 1 · Integrity | missing/dropout | 10 | 100% |
| Layer 1 · Integrity | NaN/Inf | 10 | 100% |
| Layer 1 · Integrity | schema | 10 | 100% |
| Layer 1 · Integrity | timestamps | 10 | 100% |
| Layer 1 · Integrity | joint limits | 0 | n/a |
| Layer 2 · Temporal & Motion | timestamps | 0 | n/a |
| Layer 2 · Temporal & Motion | sensor sync | 0 | n/a |
| Layer 2 · Temporal & Motion | velocity/acc | 0 | n/a |
| Layer 2 · Temporal & Motion | action spikes | 0 | n/a |
| Layer 2 · Temporal & Motion | sufficiency | 0 | n/a |
| Layer 3 · Dataset Utility | idle ratio | 0 | n/a |
| Layer 3 · Dataset Utility | coverage | 0 | n/a |
| Layer 3 · Dataset Utility | distribution | 0 | n/a |
| Metric | Finding | Evidence |
|---|---|---|
action_discontinuity | Observed action discontinuity: 1712 spike(s) across 10 episode(s) (100%); statistical risk signal, not confirmed corruption | RISK_SIGNAL |
idle_ratio | Observed low effective motion: median 30.0% effective motion (70.0% idle); statistical risk signal | RISK_SIGNAL |
velocity_acceleration | Observed extreme acceleration spikes: 164 across 10 episode(s); statistical risk signal | RISK_SIGNAL |
sensor_synchronization | Sensor sync: N/A (no per-stream timestamps in any episode) | UNVERIFIABLE |
REVIEW (10): 0, 1, 2, 3, 4, 5, 6, 7, 8, 9
EXCLUDE (0): —
REVIEW is a statistical signal worth human inspection before training — not a verdict of "bad data". EXCLUDE means an integrity-layer hard failure or unreadable episode.
D:/workbuddy-data/reports/agibot_2026/rda_audit_rl_hgdagger_realrobot.json · RDA audit runs fully offline on LeRobot v2.1/v3.0 datasets · pip install robot-data-audit to reproduce · GitHub: github.com/liesliy/rda