Research On Jump Motion Control Of Bipedal Robot Based On Deep Reinforcement Learning

Bipedal robots are anthropomorphic and have a multi-joint structure,which has been put into use in disaster relief,commercial performances,home entertainment and other fields,and has been widely valued by researchers.In recent years,many researchers have devoted themselves to the application of deep learning combined with reinforcement learning in bipedal robots.However,most of the research still stays on the relatively simple walking motion control,and cannot fully tap the multi-joint potential of bipedal robots and play the training advantages of deep reinforcement learning algorithms.Reinforcement learning,combined with the representation ability of deep learning,can be continuously optimized and learned according to the data generated by the robot’s interaction with the environment,so as to obtain a controller without manual design.In this paper,aiming at the complex long jump motion problem of bipedal robots,this paper carries out research on the long jump motion control of bipedal robots based on deep reinforcement learning,and the main research work is:First,the bipedal long jump movement was analyzed.The bipedal long jump movement needs to have enough force in a certain direction to jump up to complete the long jump movement,and analyze the deep reinforcement learning algorithm model;The relevant simulation training environment for deep reinforcement learning is constructed,and finally the Pybullet simulation environment is selected for further research.Secondly,a human action redirection algorithm was designed.Under the gravity-free conditions of the Pybullet simulation environment,based on the three-dimensional position information of the key points in the captured human movements,a redirection algorithm is designed to redirect the human movements to the NAO bipedal robot.The results show that based on the key point information of human long jump action,the designed human action redirection algorithm can realize the long jump movement of NAO bipedal robot under nongravity conditions,and obtain the joint angle data of long jump movement.Finally,the input space,reward function,and output space are designed.Analyze the characteristics of bipedal long jump motion and the improvement of deep reinforcement learning algorithms;Based on the joint angle data obtained by redirection,the joint angle data such as shoulder and foot is used as part of the input state space,and the feedback of the sole sensor is taken as part of the reward function.Experimental results show that based on the design of the input space,reward function and output space,improve the training efficiency of the long jump movement,the convergence efficiency of the reward value and the robustness of the system,which can realize the upward jumping motion of the NAO bipedal robot.The method designed in this paper,through the learning ability of reinforcement learning and the characterization ability of deep learning,can enable the NAO bipedal robot to achieve upward jumping movement by imitating human movements,tap the potential of bipedal robots with multiple joints,and open up a new path for the complex movement of bipedal robots,which can be applied to some complex terrain scenes.