Efficient Planning for Object Search Task Based on Hierarchical POMDP

Searching for a target object in clutter by a robot needs to solve perception uncertainty due to occlusion among objects, the robot has to move objects to gather information about where the target may locate. Meanwhile, we hope that the robot could execute as few actions as possible to get the target. Partially observable Markov decision process (POMDP) is proper to model this task, but it is hard to solve for large scale problems. And action feasibility of robotic manipulation must be identified by motion planning, which is time-consuming. We adopt hierarchical POMDP to reduce the complexity of POMDP planning. Besides, we propose methods to reduce the times of motion planning, the key idea is assigning reward to actions related to feasibility probability, instead of identifying action feasibility at the beginning. Simulations under various scenes show that our methods are effective and efficient.