Graph Attention Memory for Visual Navigation.

Dong Li,Qichao Zhang,Dongbin Zhao,Yuzheng Zhuang,Bin Wang,Wulong Liu,Rasul Tutunov,Jun Wang

DOI: https://doi.org/10.1109/docs55193.2022.9967733

2019-01-01

Abstract:Visual navigation in complex environments is inefficient with traditional reactive policy or general-purposed recurrent policy due to the long-term memory problem. To address this issue, this paper proposes a graph attention memory (GAM) architecture consisting of memory construction module, graph attention module, and control module. The memory construction module builds the topological graph based on supervised learning by taking the exploration prior. Then, guided attention features to reach the goal are extracted with the graph attention module. Finally, the deep reinforcement learning based control module makes decisions based on visual observations and guided attention features. In addition, the convergence of the proposed GAM module for recurrent attention operation is analyzed in this paper. We evaluate GAM-based navigation system in two complex 3D ViZDoom environments. Experimental results show that the GAM-based navigation system outperforms all baselines in both success rate and navigation efficiency, and significantly improves the generalization.