Interactive Perception in Clutter

In robotic manipulation tasks, perception traditionally precedes the act of manipulation, after which the two are typically interleaved. Perception thus enables efficient and effective manipulation. This paper explores the idea that manipulation enables efficient and effective perception, which in turn improves performance on the overall robotic manipulation task. We study this idea in cluttered environments in two contexts sorting small objects on a crowded tabletop and exploration of small occluded environments (e.g., the shelf of a home refrigerator). For object sorting, we present a robust pipeline that uses two simple motion primitives to manipulate a cluttered scene of Duplo bricks in ways that improve the robot’s perceptual, (and hence) grasping and sorting abilities. Using the PR2, we experimentally compare brick sorting with and without manipulation aids explicitly introduced to improve perception. We show substantial improvement on the overall task (sorting) by inserting manipulation aids into the task operation. For exploring a cluttered environment, we describe a strategy that uses adaptive look-ahead planning interleaved with object manipulation to improve perception. We present preliminary simulation results using this planner.