Research On Motion Planning Method Of Snake Search And Rescue Robot Based On CPG And Reinforcement Learning

Snake robot has the advantages of strong environmental adaptability and high flexibility,which can play an important role in the field of disaster relief.However,for the study of its motion control,it is necessary not only to coordinate multi-joint rotation to ensure multi-mode motion,but also to flexibly avoid obstacles to adapt to rescue environment.In view of the good stability and adaptability of CPG and the excellent autonomous learning ability of reinforcement learning,it is of great significance to study the motion control method of snake search and rescue robot based on CPG and reinforcement learning.In order to realize multi-mode motion control of snake robot,a motion control method based on two-layer CPG network was proposed.Aiming at the problems of the traditional path planning method,such as slow search speed and easy to fall into the local optimal state,a path planning method based on the improved DDPG algorithm was proposed,and the two methods were verified by experiments on the built snake robot.Specific research work is as follows:(1)For the output of CPG network waveform by the traditional oscillator connection,not only requires a lot of parameter analysis,but also has difficulties in motion mode switching and other problems.In this paper,a two-layer CPG network control method based on Matsuoka oscillator is proposed.Combined with the constructed joint D-H model and motion force analysis,the joint motion curve of the snake robot is obtained.Through Webots simulation and actual experiments,snaking,wriggling and turning movements were successfully realized,and the effectiveness of the proposed method for multi-mode motion control was verified.(2)Aiming at the problems such as low efficiency and poor accuracy of super parameter setting in the two-layer CPG network model.A hyperparameter optimization method based on the improved sparrow search algorithm is proposed,and the influence of related parameters on the output is clarified.The kinematic performance of the snake robot is evaluated from the sinuous velocity and displacement curves through simulation experiments and kinematic models.The results show that compared with the traditional parameter setting method,the proposed method not only improves the search ability and solving accuracy of the optimal superparameter,but also makes the center of gravity distribution of the snake robot more uniform when it snakes,which proves the effectiveness of the proposed method and provides a new idea for the optimization control of the snake robot.(3)In the path planning task of the snake robot,the traditional reinforcement learning algorithm has problems such as slow training speed,easy to fall into the dead zone and slow convergence speed in the face of multi-obstacle environment.An improved DDPG algorithm is proposed.Firstly,a multi-layer LSTM neural network model is introduced to control the degree of information memory and forgetting in the experience pool.Secondly,CPG network was integrated into reinforcement learning model by optimizing characteristic parameters,and new network state space and reward function were designed.Finally,the improved algorithm and the traditional algorithm were deployed in the Webots environment respectively for simulation experiments.The results show that,compared with the traditional algorithm,the overall training time of the proposed algorithm is reduced by 15% on average,the number of iterations to reach the target point is reduced by 22% on average,the number of times of falling into the dead zone during driving is reduced,and the convergence speed is also significantly improved.The algorithm can effectively guide the snake robot to avoid obstacles,which provides a new way for it to execute the path planning task in complex environment.