Learning Adaptive Escape Behavior for Wheel-Legged Robot by Inner Torque Information

Autonomous robots used for rescue or exploration needs to work in unknown environment. Such robots should select appropriate actions corresponding to their environments. In this research, we develop a wheel-legged robot getting better actions in unknown environment with reinforcement learning. We used values of external force measured on the robot's legs as the definition of states and rewards. For the quick convergence, the number of states and actions are reduced by using the characteristics of the robot's structure. To evaluate the performance of our learning system, we carried out some experiments with a simulator using a physics engine. The results of the experiments show the effectiveness of our system.