| In recent years, reinforcement learning in various fields have got more and more attention, especially in the field of robotics. And biped robot has got great attention in robot research, how to apply reinforcement learning to control biped robot to have biped robot control had intelligent, and achieve better control effect is of great significance.In this paper, to overcome the biped robot’s continuous state and action space are not conducive to the characteristics of the traditional reinforcement learning applications, we use Sparse Online Gaussian Process Regression method which is supervised learning for the reinforcement learning’s value function fitting and the reinforcement learning’s environment model constructing, and then propose two reinforcement learning methods, one is model free reinforcement learning methods based on the Sparse Online Gaussian Process, another is model based reinforcement learning method also based on Sparse Online Gaussian process.And combining the features of Sparse Online Gaussian, we propose two action planning methods respectively for these two reinforcement learning methods to solve the exploration vs exploitation problem. Then, we use these two reinforcement learning method designed two biped robot walking balance controller, and verify these two controller by means of experimenting them on the biped robot simulation model for comparing the advantages and disadvantage between model free reinforcement learning and model based reinforcement learning when used in robot control. In addition, to further address the problem of biped robot’s learning is hard to converge and the time complexity is great for its state and action space are too large, we use the PID controller to get the initial training experience for reinforcement learning, and to specify the optimization initial point in action selection of reinforcement learning, thereby improving the performance of these two reinforcement learning controller. |
PDF Full Text Request