Robot grasping and handling plays a vital role in industrial production.The robot grasping and handling based on the combination of machine learning and visual information can automatically configure the given task without any manual intervention,which significantly reduces the programming work.In this paper,a robot high-degree-of-freedom grasping strategy based on depth deterministic strategy gradient algorithm is proposed.Actor and Critic networks are constructed based on pointnet++/pointnet,and the point cloud data obtained from multiple perspectives are used as state data input to alleviate the influence of occlusion and the lack of two-dimensional image information on the success rate of high-degree-of-freedom grasping.Considering long convergence time in algorithm training,optimization methods such as model pre-training based on expert strategy and priority experience replay were adopted.The simulation results show that the improved depth deterministic strategy gradient algorithm can significantly improve the training speed and final results.
关键词
深度强化学习/机器人抓取搬运/高自由度/多视角点云
Key words
deep reinforcement learning/robot grabbing and handling/high degree of freedom/multi-view point cloud
维普
万方数据