DAP: A Discrete-token Autoregressive Planner for Autonomous Driving

Decoder-only Transformer Tokens L = Cross Entropy routing tokens a):Action Planning Models bev tokens traj Routing BEV Traj tokens b):Decoder-only Autoregressive Model Figure 1. Comparison of planning architectures. (a) Non-autoregressive direct mapping: encode historical observations to continuous embeddings and decode future trajectories in one forward pass. (b) DAP (ours): a discrete-token autoregressive planner that jointly forecasts future environment representations and ego trajectories, producing aligned supervision and sampling-based rollouts that improve planning quality and robustness.