Inverse Attention-Weighted Model with Heterogeneous Spatio-Temporal Interaction Graph for Autonomous Navigation Systems

Abstract

We propose a Res-Mlp-based attention mechanism for robotic navigation where dynamic and static obstacles coexist, enhancing collision avoidance and robot navigation. Traditional approaches struggle with predictive foresight and real-world complexity, limiting their applicability. To address these challenges, we introduce an inverse attention-weighted module based on Res-Mlp module to refine Robot-Human and Robot-Obstacle attention, improving model robustness and sensitivity to hazards. Our approach builds upon HH attn and integrates the Gumbel Social Transformer (GST), enabling more accurate and safer robot navigation. Additionally, we construct a heterogeneous spatio-temporal interaction graph and incorporate diverse obstacle shapes to create realistic scenarios. Using curriculum learning, we improve model efficiency and generalization. Experimental results show a 96% success rate in high-density crowds and 92% in complex environments with diverse obstacles, demonstrating enhanced feature extraction and safer navigation. Our method achieves a balance between conservatism and assertiveness, making it a reliable solution for real-world deployment.