This paper presents a new approach to autonomous navigation using deep reinforcement learning (DRL). It addresses the issues of limited explainability and suboptimal performance in DRL by introducing three auxiliary tasks that improve decision-making and provide intermediate indicators. The tasks involve inferring the internal states of surrounding agents, predicting their future trajectories, and estimating the degree of influence of the ego agent on others. A spatio-temporal graph neural network is employed to encode relations between dynamic entities, enhancing state inference and decision-making. The approach is tested on an intersection driving simulator, achieving robust and state-of-the-art performance.
Publication date: 29 Nov 2023
Project Page: Not provided
Paper: https://arxiv.org/pdf/2311.16091
