Does Imitation Learning Preserve Temporal Robustness in Dexterous Manipulation? An Expert-Learner Comparison Across Task Execution Speeds
ACT matched the expert at nominal speed, then fell much harder as execution sped up.
In ParcelStow, both the scripted expert and ACT learner reached 100% task success at nominal speed. At the maximum demonstrated speed, expert success was 84%, while ACT was 53%. Most ACT failures there were insertion misalignments, and acquisitions without force closure never completed the task. The paper’s point is narrow: equal success at normal speed did not mean the learner preserved the expert’s temporal robustness. HF Daily Papers' note
In ParcelStow, both the scripted expert and ACT learner reached 100% task success at nominal speed. At the maximum demonstrated speed, expert success was 84%, while ACT was 53%. Most ACT failures there were insertion misalignments, and acquisitions without force closure never completed the task. The paper’s point is narrow: equal success at normal speed did not mean the learner preserved the expert’s temporal robustness. HF Daily Papers' note
score 4