Robots manipulating objects in clutter must decide not only how to act, but also where to look. Moving the camera can expose an occluded target, but it also shifts the initial observation state away from what the manipulation policy saw during training. The authors of this paper call this first-frame distribution sensitivity and show that naively prepending camera repositioning to a VLA policy in LIBERO reduces success from 46.3% to 11.5%.
To bridge this gap, they introduce ActiveAlign. A learned viewpoint module first generates a camera trajectory. Offline, ActiveAlign uses closed-loop alignment to connect the trajectory endpoint to an existing demonstration, replays the remaining manipulation trajectory, and retains successful continuations for fine-tuning. At deployment, the same viewpoint module precedes the adapted VLA policy without modifying its architecture.
Across 40 LIBERO tasks, ActiveAlign achieves 70.7% success with pi_0 and 72.7% with pi_0.5, outperforming standard-training baselines by 40.0% and 26.4% respectively. Matched-state and data-matched controls indicate the gains go beyond prefix shortening and generic viewpoint augmentation. In real-world physical occlusion tests, success rises from 35/90 to 83/90, and the same trend holds for OpenVLA and GR00T N1. The authors conclude that active perception benefits manipulation when policy training is aligned with the observation and robot-state distribution induced by viewpoint selection.