Datasets:
Corrupt outlier action values in a few episodes
#1
by shadunts - opened
Some episodes contain physically impossible outliers in the action vectors. Episode 81 has action dim 15 reaching 33071.62 and episode 34 has dim 22 reaching 1895.82 (visible in meta/episodes_stats.jsonl max fields). Looks like isolated sensor/teleop glitches that should probably be filtered or fixed at the source.