A Mobile Grasp Planning Method Based on Iterative Optimization

A mobile grasp planning method based on iterative optimization is proposed, which needn't model the target object in advance. The 3D model of the target object is measured and modeled online by the point cloud camera, and the deep convolutional neural network is used to evaluate the success probabilities of the alternative grasp locations generated by the target point cloud. The positions and orientations of the robot base and gripper are optimized iteratively until the robot reaches an optimal configuration when grasping the target object. Then A^* algorithm is used to plan a path from the robot current position to the target position. Finally, a heuristic random path approaching algorithm is used to plan the arm motion based on the robot path, so that the robot can walk and grasp at the same time. The deep learning based evaluation algorithm of success probabilities of the alternative grasp locations achieves 83.3% accuracy on Cornell data sets. The proposed motion planning algorithm can get a smoother, shorter and more favorable path for subsequent movements.