DREAM: Defensive risk-aware enhanced maneuver planning for autonomous vehicles in heterogeneous traffic

Autonomous highway overtaking and merging remain hazardous when a large vehicle masks a fast-approaching agent until its path intersects the ego maneuver. Reacting only after line-of-sight recovery leaves little braking margin, whereas uniform caution produces disruptive deceleration toward shadows that are in fact empty. Existing risk-field planners largely score a static or externally supplied map and couple it to one planning layer, which precludes coordinated pre-reveal intervention. We introduce DREAM, which treats occlusion-conditioned latent hazard as a transported physical state governed by an inertial advection–diffusion–reaction partial differential equation, and injects the resulting Cartesian field into Frenet-domain model predictive control with control barrier functions through maneuver-level decision veto, risk-adaptive barrier modulation, and horizon-wise cost shaping. Evaluation spans controlled Python episodes, paired trailer-occlusion ablations, sensor-mediated CARLA experiments, naturalistic Bird-eye-view dataset trajectories replay, and a ROS 2 platform demonstration. The propagated field withdrew the requested merge about one second before the hidden actor became planner-visible, raised minimum hidden-actor clearance by 0.428 m, and lowered collision-or-near incidence from 29/30 to 19/30 in dense traffic, at a bounded empty-shadow speed cost. Under inD replay it tracked logged human speed profiles more closely than non-field and reduced-field variants, indicating calibrated anticipation rather than uniform conservatism. This evidence supports improved anticipatory hidden-interaction margins. The code implementation and selected scenario demonstrations for this study could be found at https://github.com/SAS-HKU/DREAM.git .