Faster Imitation by Re-timing Demonstrations Within the Controller Limit
Faster Imitation by Re-timing Demonstrations Within the Controller Limit
This paper addresses the collapse of imitation policies when demonstrations are sped up beyond two to three times their original rate.
The authors identify a mechanical failure mode in imitation learning where re-timing demonstrations creates commands exceeding the low-level controller's displacement bounds, proposing a greedy merging rule constrained by realizability to preserve demonstrated geometry while accelerating execution.
Source: arXiv Robotics — research abstracts · Read original article ↗
Loading article text…
What the source reports
Publisher-reported claims, with original evidence. These results have not been independently verified by RoboSignal.
- Environment
- SimulationOpen source S4
Reported numbers
success improvement
14
View original evidence
alisable. Every command is then realisable by construction, the demonstrated geometry is preserved, and the speed-up lands only where the demonstrator left slack. On four simulated robomimic tasks, 200 held-out seeds per cell, a merge cap of two is at or above the frozen policy's success on all four, by 14 points on Square, and shortens the successful episodes. At matched tempo, uniform re-timing loses 9 and 16 point
Open source S4loss on Square
9
View original evidence
alisable. Every command is then realisable by construction, the demonstrated geometry is preserved, and the speed-up lands only where the demonstrator left slack. On four simulated robomimic tasks, 200 held-out seeds per cell, a merge cap of two is at or above the frozen policy's success on all four, by 14 points on Square, and shortens the successful episodes. At matched tempo, uniform re-timing loses 9 and 16 point
Open source S4loss on Can
16
View original evidence
alisable. Every command is then realisable by construction, the demonstrated geometry is preserved, and the speed-up lands only where the demonstrator left slack. On four simulated robomimic tasks, 200 held-out seeds per cell, a merge cap of two is at or above the frozen policy's success on all four, by 14 points on Square, and shortens the successful episodes. At matched tempo, uniform re-timing loses 9 and 16 point
Open source S4merge cap
2
View original evidence
alisable. Every command is then realisable by construction, the demonstrated geometry is preserved, and the speed-up lands only where the demonstrator left slack. On four simulated robomimic tasks, 200 held-out seeds per cell, a merge cap of two is at or above the frozen policy's success on all four, by 14 points on Square, and shortens the successful episodes. At matched tempo, uniform re-timing loses 9 and 16 point
Open source S4reference motions
429
View original evidence
y tracker: re-timing 429 reference motions within its velocity envelope succeeds on 71% of clips at 2.3x, against 51% for uniform time scaling at 2x.
Open source S6seeds per cell
200
View original evidence
alisable. Every command is then realisable by construction, the demonstrated geometry is preserved, and the speed-up lands only where the demonstrator left slack. On four simulated robomimic tasks, 200 held-out seeds per cell, a merge cap of two is at or above the frozen policy's success on all four, by 14 points on Square, and shortens the successful episodes. At matched tempo, uniform re-timing loses 9 and 16 point
Open source S4
- paper: not_reported
Source excerpts and review record
Automatically extracted; no manual editorial approval recorded.
alisable. Every command is then realisable by construction, the demonstrated geometry is preserved, and the speed-up lands only where the demonstrator left slack. On four simulated robomimic tasks, 200 held-out seeds per cell, a merge cap of two is at or above the frozen policy's success on all four, by 14 points on Square, and shortens the successful episodes. At matched tempo, uniform re-timing loses 9 and 16 point
Open source S4
y tracker: re-timing 429 reference motions within its velocity envelope succeeds on 71% of clips at 2.3x, against 51% for uniform time scaling at 2x.
Open source S6
Source：arXiv Robotics — research abstracts · arxiv.org