RAPiD: Real-Time Diffusion Planners via Reward-Guided Consistency Distillation

A new framework called RAPiD has been developed by researchers for real-time autonomous driving trajectory planning. This system transforms a pretrained DiffusionPlanner into a quicker consistency student model, significantly cutting down the iterative denoising delay that often obstructs real-time applications. RAPiD supports multi-modal trajectory generation and facilitates two-step inference. Its training incorporates deterministic teacher denoising steps from the static diffusion planner and a low-noise data anchor to ensure trajectory alignment with expert demonstrations. To improve safety, an Implicit Q-Learning critic is trained using a balanced set of ground-truth log-replay and DiffusionPlanner rollout trajectories, each evaluated with a modified PDM-style reward, offering trajectory-level guidance beyond standard imitation learning. This framework is elaborated in a recently updated arXiv paper (2602.07339) and addresses the significant latency issue in diffusion-based planners, potentially allowing their application in closed-loop autonomous driving systems.