Robot Motion Planning For Warehouse Automation

With the rapid development of electric business and logistics industry,warehousing automa-tion has become a popular research direction in the field of robotics.By using Kiva robots,Ama-zon has transformed traditional warehousing into two parts,automatic shelves moving and picking goods in shelf by human.Right now,the later one has become the bottleneck of warehousing au-tomation.And how to achieve the second part by robots instead of human has become an urgent problem to be solved.Amazon has held many picking challenges to solve this problem since 2015.To solve this problem,much work need to be done,such as real-time motion planning,rapid calibration of the robot system and grasp planning etc.With the background of warehousing automation,the main contributions of this paper are as follows.1.A robot system composed of single axis,Kinova manipulator and KG-3 gripper is designed.A multi-objective planning algorithm is proposed to accelerate trajectory planning.The algorithm accelerates motion planning by simultaneously planning different poses in the container or different poses around the pre-grasping pose.By using this algorithm in the stage of robot arm mounting configuration design,the time-consuming of planning all poses in one container of the shelf is reduced from 3 hours to about half an hour,which greatly improves the efficiency of the design of robot arm mounting configuration.2.Aiming at the problem that existing hand-eye calibration algorithms have a large error on the z-axis,a hand-eye automatic calibration algorithm based on depth information fusion is proposed.The algorithm uses the point cloud information corresponding to RGB corner points on the calibration-board to estimate the 3D pose of the calibration-board only by geometric transformations such as plane rotation and translation.The algorithm not only achieves higher precision calibration result on all of x,y and z axes,but also greatly improves the convergence speed of hand-eye automatic calibration progress.3.A new learning-based grasp planning algorithm is proposed,which solves the problems as follows.1)The optimal grasping pose produced by analytical grasp planning algorithms is not accurate.2)Existing learning-based grasping planning algorithms cost too much time and its performance varies greatly when light varies.Grasp synthesis is accomplished by using normal vector of point cloud convex hull’s triangle surfaces as the z-axis of grasping pose candidates.Depth projection is generated by projecting the point cloud of the scene to candidate grasping poses,and then depth neural network is trained to complete the grasp selection.Our algorithm does not need to model objects to be grasped,and is illumination irrelevant.At the same time,the algorithm is not only theoretically faster than existing learning-based grasping planning algorithms,but also 5 times faster than Cornell planning algorithm after CPU and GPU acceleration.