Learn the Task, Not the Schedule: Reducing Spurious Cross-Arm Coupling in Bimanual Imitation Learning
Learn the Task, Not the Schedule: Reducing Spurious Cross-Arm Coupling in Bimanual Imitation Learning
This paper presents Constraint-preserving Temporal Recombination (CTR), a method to reduce spurious cross-arm coupling in bimanual imitation learning. By retiming recorded segments to preserve task-required constraints while maintaining arm-local sensorimotor content, CTR improves worst-state success rates from 56% to 76% across six real-robot initializations. The approach was evaluated on four simulated and four real-robot tasks using Diffusion Policy, Action Chunking with Transformers, and π0.5, with code available at https://github.com/MINT-SJTU/CTR.
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.
Reported numbers
delay transfer
2.56
View original evidence
d evaluate them on four simulated and four real-robot tasks. Across six real-robot bimanual progress-state initializations, CTR raises worst-state success from 56% for the strongest non-CTR schedule mixture to 76%. With $\pi_{0.5}$ on an Independent task, a 4 s pause transfers 2.56 s of delay to the opposite arm under the Concurrent schedule distribution, but only 0.25 s with CTR. On real robots, CTR violates task co
Open source S4delay transfer
0.25
View original evidence
d evaluate them on four simulated and four real-robot tasks. Across six real-robot bimanual progress-state initializations, CTR raises worst-state success from 56% for the strongest non-CTR schedule mixture to 76%. With $\pi_{0.5}$ on an Independent task, a 4 s pause transfers 2.56 s of delay to the opposite arm under the Concurrent schedule distribution, but only 0.25 s with CTR. On real robots, CTR violates task co
Open source S4
- code: Reported pending Open source S5
- paper: not_reported
Source excerpts and review record
Automatically extracted; no manual editorial approval recorded.
d evaluate them on four simulated and four real-robot tasks. Across six real-robot bimanual progress-state initializations, CTR raises worst-state success from 56% for the strongest non-CTR schedule mixture to 76%. With $\pi_{0.5}$ on an Independent task, a 4 s pause transfers 2.56 s of delay to the opposite arm under the Concurrent schedule distribution, but only 0.25 s with CTR. On real robots, CTR violates task co
Open source S4
nstraints in only 0--4% of trials, versus 36--98% for narrow-schedule baselines. Code will be released at https://github.com/MINT-SJTU/CTR.
Open source S5
Source：arXiv Robotics — research abstracts · arxiv.org