Research On Control Algorithm Of Self-reconfigurable Robot Based On Reinforcement Learing

Self-reconfigurable robots can change their configuration and motion according to the task and environment.Due to their modularity,versatility and self repair ability,they have great potential to realize cooperative tasks in complex environment.The ability of self-reconfiguration is the key to the variant of robot.Therefore,the control algorithm and structure design of self-reconfigurable robot are worth exploring.This paper mainly studies the multi-body path planning involved in the self-reconfigurable robot composed ofn individuals on the plane.Several individuals can be explored separately.When necessary,it forms snake,ring,rectangle and other configurations to cross specific terrain.For example,the robot can explore the environment for search and rescue.Once the trapped person is found,it forms a snake to pass through the crackto reach the trapped person’s position,observe the trapped person and provide supplies.However,in the process of their dispersion to aggregation,it is necessary to consider the possibility of robot collision caused by path conflict and coordinate their actions.In addition,the premise of global path planning for multiple robots is to master the pose and other information of all individuals and obstacles.However,robots sometimes can not obtain global information,so they need to be able to work independently by relying on their own sensors to obtain local information.Therefore,this paper designs a hierarchical framework to solve the global path planning of multi-agent based on environmental prior knowledge and the local path planning of single agent based on sensor information in uncertain environment.According to the complexity of the task,it is divided into overall planning layer and sub task execution layer from top to bottom.The top layer uses FA-WoLF-PHC to plan the global path for each robot.According to the top layer’s decision and the environmental information detected by the sensor,the bottom layer divides the robot working area with grid.The reward function of Q-learning algorithm for local path planning is designed,so that each robot can work independently and have environmental adaptability.In order to provide an experimental platform and verify the effectiveness of the algorithm,the docking mechanism and control system of the robot are designed in this paper.Based on the physical robot,the mobile navigation docking experiment and self-reconfiguration experiment are carried out.The simulation and experimental results verify the effectiveness of the algorithm and the reliability of the robot,provide the corresponding algorithm basis and hardware platform for the future control research work.