Value of Potential Field in Reward Specification for Robotic Control via Deep Reinforcement Learning

摘要

View Video Presentation: https://doi.org/10.2514/6.2023-0505.vid Specifying proper state rewards plays an important role in obtaining reinforcement learning-based control policies. The objective of this paper is to design and evaluate a potential field-induced reward specification method in both virtual and real-world environments. In particular, we first describe the procedure to obtain state rewards by analyzing the relationship between potential fields and the principle of reward assignment. Then, we compare the performance of two deep reinforcement learning algorithms based on the potential field-induced state rewards with that of the classic potential field controller in both AirSim and real-world indoor drone testing environments. The comparison demonstrates the value of potential fields in reward specification while also suggesting the limitations, which will be further investigated as {one} part of our future work.