GITSR: graph interaction transformer-based scene representation for multi-vehicle collaborative decision-making

This study proposes GITSR, a Graph Interaction Transformer-based Scene Representation framework for multi-vehicle collaborative decision-making. In mixed traffic environments where connected automated vehicles (CAVs) and human-driven vehicles (HDVs) coexist, effective scene understanding and interaction modeling are essential for safe and efficient decision-making. GITSR first extracts driving environment features under an intelligent connected vehicle setting. Then, an agent-centric dynamic occupancy grid is encoded by a Transformer module to obtain local scene representations. Meanwhile, vehicle motion information is modeled as graph structures, and spatial interaction features among traffic participants are extracted using a Graph Neural Network (GNN). Finally, the collaborative decision-making process is formulated as a Markov Decision Process, and driving actions are generated through reinforcement learning algorithms. The proposed framework is validated in a challenging highway off-ramp scenario. Simulation results show that GITSR effectively captures scene representations and spatial interaction information, outperforming baseline methods across multiple evaluation metrics.