Simultaneous Planning of Grasp and Motion using Sample Regions and Gradient-Based Optimization

Abstract

Motion planning is an essential component of robotic systems. Gradient-based planning has been proposed to produce smooth paths under constraints for manipulation tasks such as picking and placing objects. However, it does not deal well with discontinuities, which occur in many manipulation problems, e.g. when deciding whether to pick an object from the side or from the top. Sampling-based planning is robust against such discontinuities, but often produces paths with unnecessary motions that require heavy post-processing. In this paper, we propose a novel method to solve a complete manipulation task using gradient-based optimization while preserving the advantages of sampling-based planning. By dividing the surface of the object into regions where the grasp pose can be extracted quasi-continuously, we define multiple optimization problems in parallel, which are evaluated independently. As our method generates the motion path and grasp plan for the entire task, constraints that arise from each moment of the task are propagated automatically to the optimization of the entire task, facilitating the setup. We show the effectiveness of the proposed method in simulation and with a real robot.