A perception-guided approach to motion and manipulation planning

Abstract

Rapidly-Exploring Random Trees (RRT) have been successfully applied to many different robotics systems for motion and manipulation planning under non-holonomic constraints. However, the conventional RRT algorithm may perform poorly in the presence of noise and uncertainty. This thesis proposes a modified form of the algorithm that seeks to reduce the robot's uncertainty in its estimate of the target by choosing solutions that maximize the opportunities for the robot's sensors to perceive the target. This new perception-guided technique will be tested in simulation and compared to the conventional RRT as well as other approaches taken from the literature. The ultimate goal is to integrate this method with a semi-autonomous robotic forklift charged with the task of approaching and picking up a loaded wooden pallet over rough terrain.