Look Only When the Action Could Change

Abstract

Occlusion makes a robot uncertain about many scene details, although only a subset can change which action it should execute. We propose a wrist-camera policy that queries views according to their expected reduction in action regret. The policy compares affordance hypotheses against a small library of safe actions and stops looking when the expected benefit falls below the cost of another view. We will test whether this criterion saves sensing motions at matched task success on paired, partially observable manipulation scenes.

Keywords