Enhancing the Generalization Performance and Speed Up Training for DRL-based Mapless Navigation
share
Training an agent to navigate with DRL is data-hungry, which requires millions of training steps. Besides, the DRL agents performing well in training scenarios are found to perform poorly in some unseen real-world scenarios. In this paper, we discuss why the DRL agent fails in such unseen scenarios and find the representation of LiDAR readings is the key factor behind the agent's performance degradation. Moreover, we propose an easy, but efficient input pre-processing (IP) approach to accelerate training and enhance the performance of the DRL agent in such scenarios. The proposed IP functions can highlight the important short-distance values of laser scans and compress the range of less-important long-distance values. Extensive comparative experiments are carried out, and the experimental results demonstrate the high performance of the proposed IP approaches.
READ FULL TEXT