Reinforcement Learning Based Walking Control Of Biped Robot

With the development of technology and science,robots have been widely used in people's life.As one of the hotspots,biped robot,which has the characteristics of high flexibility and strong adaptability,is highly intergrated into human life.In the application of biped robots,stable walking ability is the basis for its completion of work tasks.However,the structure of the biped robot is complex and the walking control is difficult.Traditional gait planning and control methods have the disadvantages of high model complexity,limited application scenarios,and large differences between the generated gait and the human gait.Therefore,it is of great significance to carry out innovative research on gait planning and control methods.Aiming at the problems of the above-mentioned traditional gait planning and control methods,this project designs a walking controller based on the reinforcement learning framework to train the walking strategy of the agent in a simulation environment,and transplant the walking strategy to the real robot for experiment.The main work includes:(1)According to different experimental tasks,two biped robots have been developed.The10-DOF robot has a mass distribution and structural design that mimics the human body,and is designed to generate walking actions similar to the human gait.The 6-DOF robot with simple structure undertakes the task of verifying the walking strategy in real world.The structure design and hardware selection of the two robots were carried out,and the control system based on RTOS and ROS was built.(2)The second part illustrates the mainstream reinforcement learning algorithms and relevant characteristics analysis.And a walking controller based on PPO was designed,which includes the reward function,noise conditions and many other parts.(3)A simulation experiment environment and simulation walking training are conducted in this part.On the basis of the Gym reinforcement learning standardized environment,combined with the Mu Jo Co simulation platform,the agent is driven by the walking controller to interact with the environment to learn walking strategies.(4)In the real experimental environment,the walking strategy is transplanted to the real robot for walking experiments.Finally,the robustness of the robot gait showed the effectiveness of reinforcement learning in the gait planning process of the biped robot.This thesis proposed the application of reinforcement learning walking controller,simulation and real experiments to prove the feasibility of solving the traditional robot gait planning and control methods problem.