Autonomous driving world models often try to predict every detail of future scenes, but most of that detail is irrelevant to choosing a safe action. The authors of DA-WAM argue that predictive representations should focus on the consequences that matter for driving decisions. Their model couples a Joint-Embedding Predictive Architecture with candidate-level trajectory scoring, so each possible trajectory gets its own future latent that is used to estimate planning factors and overall utility.
DA-WAM is trained jointly: the JEPA objective keeps future representations grounded in observed scene dynamics, while planning objectives push those representations to encode distinctions related to safety and utility. The authors also introduce a safety-critical post-training step using hard negatives—geometrically similar trajectories with different safety outcomes—to sharpen the model's ability to tell them apart.
On the NAVSIM-v1, NAVSIM-v2, and Bench2Drive benchmarks, DA-WAM achieves state-of-the-art performance among learned camera-based planners. Ablation studies confirm that each component contributes to driving decision-making, though the paper's abstract does not detail the magnitude of those contributions or compare the method against non-learned baselines.