Research On Motion Planning Of Robot Arm Based On Deep Reinfocement Learning

Multi-degree-of-freedom manipulator has the characteristics of flexible motion,and its motion planning is a research hotspot in the field of robotics.When the manipulator moves in the complex environment with obstacles,such as grasping,handling and man-machine cooperation,it is necessary to plan the movement path and the attitude of the manipulator arm during grasping.This paper focuses on the path planning and posture planning of the manipulator based on deep reinforcement learning.Aiming at the problem of long training time and large training samples,a motion planning algorithm of the manipulator with transfer learning is proposed.Firstly,aiming at obstacle avoidance of manipulator,a path planning algorithm for the right arm of NAO robot based on Deep Deterministic Policy Gradient(DDPG)is proposed.An obstacle-free and an obstacle simulation environment are built based on MuJoCo simulation platform.In the simulation environment,the path planning control strategy of the manipulator is automatically learned by using DDPG algorithm through setting the reward function.The end-to-end control of the manipulator from input to output is realized,and the path planning for obstacle avoidance of the manipulator is achieved.Secondly,in the obstacle-free environment,aiming at the pose and posture planning problem of multi-degree-of-freedom manipulator when grasping objects,a pose and posture planning algorithm based on DDPG is proposed.According to the different grasping poses and postures of NAO right arm,the reward function of learning task is designed.Furthermore,because the DDPG algorithm needs a lot of time and data samples in the training process,a DDPG based on transfer learning algorithm is proposed.By comparing the training results of DDPG algorithm and DDPG based on transfer learning algorithm,it is shown that the DDPG based on transfer learning algorithm has a faster training speed.The LINEMOD algorithm is used to obtain the pose and posture of the target object,and the experiment of DDPG based on transfer learning algorithm is completed on the NAO manipulator.Finally,in the obstacle environment,aiming at the problem of the pose and posture planning for multi-degree-of-freedom manipulator when grasping objects,the pose and posture planning algorithm based on DDPG of transfer learning is proposed in the obstacle environment.The results of simulation show that the training speed of DDPG based on transfer learning algorithm is faster than that of traditional DDPG algorithm in the obstacle environment.In order to validate the effectiveness of the proposed algorithm,the experiment of the pose and posture planning for NAO robot's right arm in the obstacle environment is carried out.The experimental results show that the DDPG based on transfer learning algorithm can effectively achieve obstacle avoidance planning during the process of the pose and posture planning under the obstacle environment.