When generating driving videos for long periods, AI models commonly make roads shift, cars change identity, and lose visual quality. This is because they lack a persistent memory of the scene.
WAM-Sim uses a structured memory that permanently stores road layout, background appearance, and each car's identity. Then it applies , after initial training, to improve the model's outputs by rewarding realistic and physically correct simulations.
Over minute-long simulations, WAM-Sim preserved scene similarity at 0.910 and identity coherence at 0.681, far better than previous methods. This enables realistic long-horizon driving simulation, essential for trustworthy autonomous vehicle testing.