Efficient Exploration for Quadruped Robots Using Graph Neural Networks With Global Attention

Abstract

Autonomous exploration in an unknown environment is a critical task for quadruped robot applications. Traditional approaches based on stacked decision rules often suffer from inefficiency due to the greedy behavior. In this work, we propose a hybrid architecture that jointly employs a graph neural network (GNN) and a multi-head self-attention mechanism within a reinforcement learning (RL) framework to address myopic decision-making across spatial and temporal dimensions. The environment is represented as a topological graph, where the GNN enables the agent to learn the spatial structure of the map by leveraging topological relations. Meanwhile, the multi-head self-attention mechanism (MHSA) captures global dependencies among subregions at different positions, enabling the robot to reason over long-range spatial relationships. The RL framework further mitigates temporal greediness by encouraging policy learning from accumulated experience, which equips the robot with the ability to reason about the spatial structure of unseen areas and strike a balance between exploiting known regions and exploring unknown ones. The proposed approach was validated in publicly available simulation environments. Experimental results demonstrate that the proposed approach significantly reduces redundant exploration and shortens the overall planned trajectory length.