Autonomous driving planners must select trajectories that are safe, rule-compliant, and comfortable in dense multi-agent settings. Existing diffusion- and RL-based planners often rely on generic rewards and use relational structure mainly for representation, which limits how planning-critical rollouts guide policy improvement.
GRID-Planner addresses this by integrating graph-structured relational signals directly into the reward and update loop. It uses graph-aware importance weighting, a time-indexed relational reward surrogate, and an advantage-weighted diffusion objective with on-policy self-distillation, creating a rollout-return-advantage-update loop without explicit likelihoods.
On the nuPlan and interPlan benchmarks, GRID-Planner outperformed a diffusion-only baseline: collision rate dropped 35.8%, lane deviation 60.8%, and jerk 27.4%, while planning score improved 3.4%. The authors say the framework aligns generative planning with structured relational signals, pointing toward more robust closed-loop decision making in planning-critical scenes.