GAPL framework cuts collision rates 0.76% with LLM-guided trajectory planning

GAPL (Grounded Action-effect Policy Learning) is a new framework from Zhihong Cui and seven collaborators, designed to overcome key limitations of LLMs in autonomous driving trajectory planning. While LLMs excel at semantic reasoning, they often hallucinate, struggle with numerical precision, and lack grounding in real-world dynamics. GAPL integrates three modules: an LLM-based Effect Evaluator that produces structured action-effect estimates, a Simulation-based Effect Grounder that derives dynamics-consistent predictions from simulator rollouts, and an Effect-Aware Decision Maker that grounds LLM estimates against simulation using a distiller. This closed-loop system guides Proximal Policy Optimization (PPO) to learn robust driving policies.

Experiments across four Highway-env scenarios show GAPL consistently beats baselines, achieving average reductions of 0.76 in collision rate, 0.86 in average displacement error (ADE), and 2.00 in final displacement error (FDE), alongside an average reward gain of 1.44. These results demonstrate that grounding LLM knowledge in simulation significantly enhances both safety and accuracy. GAPL offers a practical path to integrating LLM reasoning into real-time control, potentially accelerating development of safer autonomous driving systems that can handle complex, dynamic environments.