Event
Proposing ActiveAlign: A Data Adaptation Framework for Active Perception and VLA Policies
1 reports · 1 sources
Jiawei Geng and other researchers propose the ActiveAlign data adaptation framework, which trains vision-language-action (VLA) policies using the observation state at the termination of active perception, addressing the mismatch between initial observations after active camera movement and the policy training distribution. The paper reports that on 40 LIBERO tasks, ActiveAlign achieves a 70.7% success rate with pi_0 and 72.7% with pi_0.5, which is 40.0% and 26.4% higher than their respective standard training baselines; in real-world physical occlusion scenarios, the success rate improves from 35/90 to 83/90. The paper points out that if the camera repositioning module is directly attached in front of pi_0.5 without adaptation, the success rate under perturbed initial conditions drops from 46.3% to 11.5%. At deployment, the same viewpoint module is placed before the adapted VLA without modifying its architecture.
Report timeline
- ActiveAlign bridges active perception and VLA policies via observation-aligned data adaptation
The paper proposes ActiveAlign, a data adaptation framework that addresses the mismatch between the initial observation state produced by active camera motion and the manipulation policy's training distribution (first-frame distribution sensitivity): on LIBERO, directly prepending camera repositioning to pi_0.5 reduces the success rate from 46.3% to 11.5%.
arXiv cs.RO ·