Efficient planning in non-gaussian belief spaces and its application to robot grasping

Abstract

The limited nature of robot sensors make many important robotics problems partially observable. These problems may require the system to perform complex information-gathering operations. One approach to solving these problems is to create plans in belief-space, the space of probability distributions over the under-lying state of the system. The belief-space plan encodes a strategy for performing a task while gaining information as necessary. Most approaches to belief-space planning rely upon representing belief state in a particular way (typically as a Gaussian). Unfortunately, this can lead to large errors between the assumed density representation of belief state and the true belief state. This paper proposes a new sample-based approach to belief-space planning that has fixed computational complexity while allowing arbitrary implementations of Bayes filtering to be used to track belief state. The approach is illustrated in the context of a simple example and compared to a prior approach. Then, we propose an application of the technique to an instance of the grasp synthesis problem where a robot must simultaneously localize and grasp an object given initially uncertain object parameters by planning information-gathering behavior. Experimental results are presented that demonstrate the approach to be capable of actively localizing and grasping boxes that are presented to the robot in uncertain and hard-to-localize configurations.