一种采用迭代优化的移动抓取规划方法

A Mobile Grasp Planning Method Based on Iterative Optimization

提出了一种无需对目标物体进行预建模的迭代优化移动抓取规划方法.该方法通过点云相机在线对目标物体进行立体模型测量和建模,通过深度卷积神经网络对目标点云生成的多个候选抓取位置的抓取成功率进行评价.然后,对机器人底盘和手爪的位置和姿态进行迭代优化,直到抓取目标物体时机器人达到一个最优的位形.再用A*算法规划一条从机器人当前位置到目标位置的运动路径.最后,在路径的基础上,用一种启发式随机路径逼近算法规划手臂的运动,实现边走边抓的效果.本文的深度学习抓取成功率评估算法在康奈尔数据集上取得了83.3%的精确度.所提运动规划算法能得到更平滑、更短且更有利于后续运动的路径.

A mobile grasp planning method based on iterative optimization is proposed, which needn’t model the target object in advance. The 3 D 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.