ST-GICM: A Spatiotemporal Graph Learning Framework with Intrinsic Curiosity for Robust Autonomous Exploration

With recent advances in deep reinforcement learning (DRL) and graph neural networks (GNNs), graph-based autonomous exploration methods have significantly improved decision-making performance in complex environments. However, under partial observability and sparse-reward conditions, existing methods still struggle with long-horizon decision-making and sustained exploration. To address these challenges, we propose a spatiotemporal graph learning framework, termed ST-GICM, that improves the robustness and efficiency of autonomous exploration by integrating graph-structured encoding, temporal memory, and an intrinsic curiosity mechanism. Specifically, a Graph Attention Network (GAT) and a Spatiotemporal Reasoning Core (STRC) are employed to dynamically encode the viewpoint graph and fuse temporal memory, thereby alleviating perceptual aliasing in graph-based exploration. In addition, an Intrinsic Prediction Module (IPM) is designed to generate intrinsic rewards based on the prediction error of graph-level latent representations, thereby encouraging sustained exploration. Experiments conducted in procedurally generated complex topological environments show that the proposed method outperforms existing baselines in terms of coverage rate, success rate, repeated revisit rate, and oscillation count, while maintaining trajectory costs comparable to those of the baselines. These results demonstrate the effectiveness and superiority of ST-GICM in partially observable environments under sparse-reward conditions.