Multi-Source Information Fusion Navigation Based on Dynamic Attribute Scene Graphs

Scene understanding is crucial for autonomous navigation since it forms the interface between perception and planning. Effectively extracting and integrating information from a priori scene maps into the agent's observations for scene understanding is challenging. Traditional navigation methods rely on the policy network's capacity to implicitly encode such information and establish associations between them; however, the practical efficacy of these approaches is constrained. In contrast, the scene graph presents a higher-level representation of the environment, enabling explicit depiction of the semantic and geometric topological relationships present in the scene map. In this work, we propose a scene graph navigation framework using dynamic attribute scene graphs to enhance the agent's comprehension. The approach establishes connections between the scene graph and agent observations through attribute updates and utilizes graph neural networks for effective feature fusion. Experimental results demonstrate its effectiveness for UGV multi-target search tasks in diverse environments. Compared to the baselines, our method not only improves agent policy learning efficiency but also achieves more precise task completion. Additionally, we conducted experiments to verify the interpretability of the method.