In plain words
When a robot takes time to plan its next set of movements, it may keep moving while the scene changes. The new movements can then be based on an old view.
PACT-VLA prepares several possible movement chunks and uses a newer view, just before switching, to choose one. In simulated dynamic-control tasks, its average success at a four-step delay was higher than RTC’s, where RTC is a method designed to keep action chunks robust to delay. This result is an average for that test setting.
The research
When a robot is still working out its next movements, it may continue acting and the scene can change. Can it use what it sees closer to the moment of switching to choose a chunk that better fits the current situation?
PACT-VLA selects one complete robot action chunk from a fixed candidate pool by scoring it against a newer observation just before execution.
On Kinetix, a simulation of dynamic control tasks, researchers compared action-chunk selection at a four-step inference delay. PACT-VLA averaged 77.2% task success across 12 tasks, versus 62.2% for RTC, a delay-robust action-chunk method, with the same released checkpoint and execution protocol.
The higher mean supports selecting a chunk with information received closer to execution. It applies to the tested procedures and tasks; the comparison does not isolate candidate generation from candidate selection.
The Kinetix comparison with RTC changes the action-chunk selection procedure and does not isolate candidate generation from candidate selection. The reported success differences have no confidence intervals, so they do not quantify statistical uncertainty.
Analysis
Deep
- Original paper authors
- Yixiang Zhu, Ruilong Wu, Yonghao Chen, Jiaxiang Zou, Yusong Hu
- Submitted to arXiv
- License
- arXiv non-exclusive