Datasets:
Failure-onset timing analysis on this dataset: two failure shapes (fully reproducible)
#8
by makambo1 - opened
We ran an execution-monitoring analysis on the 52-episode PushT set and published the results with figures and scripts: https://robot-eval-platform.vercel.app/blog/evaluation-1-failure-timing
Method: the success envelope is the 10th percentile of coverage reward across the 25 successful trials at each 0.1 s step; a failure's anomaly is the first 0.5 s sustained drop below it (in the spirit of Park et al., ICRA 2016), with a coverage-peak fallback for trials that never leave the envelope.
Findings:
- 25/52 trials reach max coverage >= 0.9 (48.1%)
- Only 8 of the 27 failures diverge from the envelope mid-episode (median 8.0 s, about 52% of episode time)
- The other 19 track the success envelope and simply run out of time, flagged at their coverage peak (median 97.8% of episode time)
- The same envelope rule fires on 10/25 successes (40%) at this threshold, which the post discusses
Limitations (teleop demonstrations, single signal, small n) are stated in the post. Thanks to the LeRobot team for the dataset.