Deep Reinforcement Learning Based Autonomous Exploration under Uncertainty with Hybrid Network on Graph

Abstract

This paper mainly focuses on the autonomous exploration of unknown environments for mobile robots with deep reinforcement learning (DRL). To accurately model the environment, an exploration graph is constructed. Then, we propose a novel S-GRU network combing graph convolutional network (GCN) and gated recurrent units (GRU) based on the exploration graph to extract hybrid features. Both the spatial information and the historical information can be extracted by using S-GRU, which could help the optimal action selection by employing DRL. Specifically, In S-GRU, one GRU is performed to extract the inner information related to the historical trajectory, and another is used to combine the current and historical inner information as the current state feature. Simulation experimental results show that our approach is better than GCN-based and information entropy-based approaches on effectiveness, accuracy, and generalization.