Training robot policies to act after they move the camera

When a robot moves its wrist camera to see around clutter, the view changes along with the robot’s pose. Can a manipulation policy trained on ordinary demonstrations still continue reliably from that new combination of visual and state cues?

ActiveAlign trains existing robot action policies on successful continuations from camera-movement endpoints, preparing them to manipulate from the views they will encounter.

In a simulation test of 40 LIBERO manipulation tasks, each policy faced perturbed starting poses and was measured by average task-completion success across tasks.

Simply adding camera movement to standard-trained policies lowered that average; ActiveAlign-trained policies reached 70.7% for π₀ and 72.7% for π₀.₅, versus 30.7% and 46.3% for standard training.

Because the adapted policies train on successful continuations beginning at the camera module’s endpoints, they performed better in the same shifted-start setting.

This supports the claim that preparing a manipulation policy for the new view and robot state can make active camera movement useful; the comparison changes both training data and use of the viewpoint module.

The LIBERO results apply to the tested policies and perturbed-start simulation protocol. A separate physical evaluation tested π₀.₅ across three apple-manipulation tasks and three occlusion levels, with 90 trials total.

Offline adaptation requires successful alignment and replay, and the deployed system selects a viewpoint once before manipulation.