Optimization of obstacle avoidance using reinforcement learning

Walking through narrow space for multi-legged robot is optimized using reinforcement learning in this paper. The walking is generated by the virtual repulsive force from the estimated obstacle position and the virtual impedance field. The resulted action depends on the parameter of the virtual impedance coefficients. The reinforcement learning is employed to find an optimal motion. The temporal walking through motion consists of each parameter optimized for a situation. Optimization of integrated walking through motion is finally achieved evaluating walking in compound encountering obstacle on simulator. The resulted motion is implemented to a real multi-legged robot and results show the effectiveness of the proposed method.