AEDP: An Interaction-Aware Decision-Making and Motion Planning Architecture for Complex Traffic Environments

Autonomous driving in dense traffic requires decision-making and motion planning that accounts for uncertain predictions, interactive responses, and real-time constraints. This paper presents AEDP, a hierarchical planning architecture that connects maneuver-intention inference, dynamic Stackelberg interaction modeling, multimodal trajectory proposals, risk-aware evaluation, and constrained iterative linear quadratic regulator (iLQR) refinement in a receding-horizon loop. The contribution lies in the interfaces and closed-loop integration of established methods, rather than a new reinforcement-learning, generative-model, or iLQR algorithm. In the controlled highway-merging simulation, the reported full-system success proportion is 28/30 trials (93.3%; 95% Wilson CI: 78.7–98.2%). Lateral oscillation is reported as 0.34 ± 0.06 m, alongside 200 km of urban vehicle testing. Extended ablation studies examine CVAE-based multimodal proposals, CVaR risk evaluation, adaptive safety corridors, warm-starting, trajectory reuse, and GPU parallelization. AEDP has the highest reported aggregate success-rate point estimate; its mean latency is lower than MCTS-POMDP but higher than RBP and MPC-IDM. Inferential claims are limited where paired trial-level records are unavailable for recalculation. The study does not establish formal collision-avoidance guarantees or functional-safety compliance. Because of collaborative enterprise intellectual-property restrictions, source code, complete experimental configurations, and original trial-level datasets are not publicly available; non-proprietary configuration excerpts, algorithm pseudocode, derived aggregate counts, and a separately labeled synthetic statistical test example are supplied to support methodological checking.