Analysis of Cost Functions for Reinforcement Learning of Reaching Tasks in Humanoid Robots

In this paper, we present a study on transferring human motions to a humanoid robot for stable and precise task execution. We employ a whole-body motion imitation system that considers the stability of the robot to generate a stable reproduction of the demonstrated motion. However, the initially acquired motions are usually suboptimal. To successfully perform the desired tasks, the transferred motions require refinement through reinforcement learning to accommodate the differences between the human demonstrator and the humanoid robot as well as task constraints. Our experimental evaluation investigates the impact of different cost function terms on the overall task performance. The findings indicate that the selection of an optimal combination of weights included in the cost function is of great importance for learning precise reaching motions that preserve both the robot's postural balance and the human-like shape of the demonstrated motions. We verified our methodology in a simulated environment and through tests on a real humanoid robot, TALOS.