Proactive Risk Assessment and Explainable Evasive Trajectory Generation for Safety-Critical Autonomous Driving

The widespread deployment of autonomous driving systems critically depends on guaranteeing safety, especially in complex, dynamic environments. Traditional reactive approaches often fall short in highly uncertain or rapidly evolving situations, leaving minimal margins for error and lacking transparency. To address these limitations, we propose \textit{Proactive-Scene}, a novel framework that transitions autonomous driving from passive reaction to active prevention, providing both safety and transparency. Our framework comprises two core components: the Multi-Agent Risk Prediction Network (MARP-Net), which leverages Graph Neural Networks and Transformer-based Encoders for comprehensive multi-agent trajectory prediction and proactive risk assessment, identifying critical adversarial vehicles and generating a detailed risk heatmap; and the Explainable Evasive Trajectory Generator (EETG-Module). The latter employs a Constrained Optimization Solver to generate safe, kinematically feasible evasive trajectories, coupled with an LLM-based Explanation Generator to provide natural language justifications for these maneuvers. Evaluated on diverse datasets and simulated safety-critical scenarios, \textit{Proactive-Scene} significantly outperforms existing baselines across all metrics. It achieves a notably lower collision rate, a higher recall of critical events, superior evasion timeliness, and strong explanation coherence. Our framework demonstrates robust performance across challenging scenarios and operates within real-time computational constraints, fostering increased trust and understanding in autonomous driving decisions.