Learning robot stiffness for contact tasks using the natural actor-critic

This paper introduces a novel motor learning strategy for robotic contact task based on a human motor control theory and machine learning schemes. Humans modulate their arm joint impedance parameters during contact tasks, and such aspect suggests a key feature how human successfully executes various contact tasks in variable environments. Our strategy for successful contact tasks is to find appropriate impedance parameters for optimal task execution by reinforcement learning (RL). In this study recursive least-square (RLS) filter based episodic natural actor-critic is employed to determine the optimal impedance parameters. Through dynamic simulations of contact tasks, this paper demonstrates the effectiveness of the proposed strategy. The simulation results show that the proposed method successfully optimizes the performance of the contact task and adapts to uncertain conditions of the environment.