GRID-Planner: Graph-Structured Relational Reward Shaping for Diffusion Planning
GRID-Planner is a diffusion planning framework that integrates graph structures and relational signals into trajectory generation for autonomous driving. It combines graph-aware importance weighting, a time-indexed relational reward surrogate, and an advantage-weighted diffusion objective with self-distillation. This approach shapes policy improvement by biasing the generative process towards safer, rule-compliant, and more comfortable trajectories derived from closed-loop interactions.