Prioritized Reward of Deep Reinforcement Learning Applied Mobile Manipulation Reaching Tasks

In this paper, we apply deep reinforcement learning (DRL) for reaching target positions with a mobile manipulator while coordinating the mobile base and the manipulator and study the performance of different reward functions to get a higher success rate and more efficient movement. The reward is basically defined by the function of distance between the robot and the goal. We propose principles to build reward functions based on geometric series theory and discuss possible reward forms combined with different elements. We also present a prioritized reward function for mobile manipulation to weight movements of different parts and further provide a method to define the weights. Experiments are carried out in both two-dimensional and three-dimensional collision-free environments, and a further investigation into a relative task of going through an opening doorway is evaluated in the end.