MSN: Mapless Short-Range Navigation Based on Time Critical Deep Reinforcement Learning
Bohan Li, Zhelong Huang, Tony Weitong Chen, Tianlun Dai, Yalei Zang, Wen-Bin Xie, Bo Tian, Ken Cai
Abstract
Automated vehicle(AV) based on reinforcement learning is an important part of the intelligent transportation system. However, currently, the performance of AV heavily that relies on the quality of maps and mapless navigation is one potential method for navigation in a strange and dynamic changing environment. Although many efforts are made on mapless navigation, they either need prior knowledge, rely on an exceptional constructed environment or simple feature fusion mechanism in the networks. In this paper, we proposed a deep reinforcement learning method, namely TC-DDPG, which is consisted of DDPG, multi-challenge deep learning networks and time-critical reward function. By comparing to existing approaches, TC-DDPG takes the cost of time into consideration and achieves better performance and converges more easily. A new open source simulator is proposed and extensive experiments are conducted to demonstrate the performance of the TC-DDPG, which outperforms comparing methods and achieves 62.9% less in time cost, 12.0% less in distance cost and about 90% fewer in numbers of model parameters.