Path Planning And Implementation Of Quadruped Robot Based On Reinforcement Learning

With the complexity of robot working space,quadruped robots can replace traditional robots to work on unstructured maps with good stability and environmental adaptability.The ability of autonomous path planning is an index to evaluate the intelligence of quadruped robots,so path planning algorithms have received extensive attention from scholars.Common path planning algorithms are only applicable when the environmental information is known,but the environmental information in practical application is often unknown,so path planning algorithms need to have certain learning ability to adapt to the environment.In order to improve the intelligence of quadruped robot and make it reach the designated position accurately and safely,this paper proposes a global path planning algorithm and a local path planning algorithm for the quadruped robot,and builds a robot experimental environment in ROS to test and verify the feasibility and effectiveness of the algorithm.The main research contents are as follows:(1)The software and hardware development platform of the quadruped robot is built to provide conditions for the specific realization of the proposed path planning algorithm.First introduced the sensor selection and environment modeling methods,then designed the quadruped robot odometer and robot model,and finally introduced the overall navigation framework.(2)In terms of global path planning,in order to solve the problem of path planning in workspaces where environmental information is unknown or difficult to obtain,the Qlearning reinforcement learning algorithm based on value functions is used to explore the environment,which improves the intelligence of the robot.In view of the problem that the Q-value table of the traditional Q-learning algorithm occupies a large storage space,lock variables and four derivative properties are introduced,through which the Q-value table is updated at one time,which reduces the time complexity and storage space of the algorithm;for the shortest path problem,an improved Q-learning algorithm capable of oblique movement is proposed,and the search direction of the original algorithm is improved.In addition to the original four movement directions,four oblique movement directions are added,improving the intelligence of the algorithm and adaptability to the environment.The simulation results show that the improved algorithm can achieve better path planning,and effectively reduce the calculation amount of the algorithm and shorten the length of the path.(3)In terms of local path planning,considering the search efficiency and environmental adaptability of the algorithm,the A* algorithm based on graph search and the PRM,RRT and dual-tree RRT algorithms based on sampling are compared and analyzed from the two indicators of planning time and path length.The simulation experiment verifies the superiority of the dual-tree RRT algorithm in planning efficiency,path length and environmental adaptability.