Robust Urban Navigation for Autonomous Vehicles using Vision based Segmentation Graph Attention and Hierarchical Learning

In the modern era, Autonomous Vehicles (AV) navigation in complex urban environments has become a significant challenge due to dynamic obstacles, occlusions and localization issues in Global Positioning System (GPS) denied areas. Even though, previous researchers suggested a model using traditional Reinforcement Learning (RL) which face challenges in capturing long-range dependencies in visual scenes, modeling adaptive spatial relationships and generalizing decision policies to diverse driving scenarios. To address these challenges, a hybrid framework integrating Vision Transformer (ViT), Graph Attention Network v2 (GATv2) and Hierarchical Reinforcement Learning (HRL) is proposed for scalable decision-making which is named as ViT-GATv2-HRL. Initially the multi-modal sensor data is considered which fused into an input frame. Subsequently, semantic maps are generated using ViT that converts input frame data into scene graphs and then incorporates GATv2 to extract relational embeddings. Finally, these embeddings fed into an HRL model, where high-level policies select navigation subgoals and low-level policies compute the control commands. Hence, the proposed ViT-GATv2-HRL model captures dynamic spatial relationships, enhances visual understanding and improves navigation robustness across varied urban conditions, achieving safer and more adaptive AV behavior in real-time simulations and complex urban settings.