Work Day and Night: A Learning Based Illumination Irrelevant Grasp Planning Method

This paper presents a new grasp planning method to solve grasp synthesis and grasp selection problems. A depthprojection algorithm is proposed to solve synthesis problem. Using only depth information of working scene, the algorithm establishes the relationship, which is expressed as depthprojections, between point cloud and grasp poses. And a learning network is constructed and trained to accomplish the grasp selection, i.e. selecting the best one from depth-projections in this paper. Besides parallel acceleration methods on both CPU and GPU are applied to further improve the efficiency of grasp planning. Experiments suggest that 1) There is no need to model gripper or target object, which is inaccurate or even impossible for some deformable objects; 2) Our method is illumination irrelevant with 83.3% test accuracy, therefore, robots can work no matter day and night, and no matter in household or outside environments; 3) Grasp planning for one object costs only a few seconds with the CPU and GPU acceleration, meaning that our method is applicable to piratical environment online.