OccStress: Stress-Testing the 4D Occupancy Forecasting Chain
OccSTeP-WM is a method for autonomous driving that builds a persistent understanding of 3D scenes by handling temporal disturbances and future actions. It uses a tokenizer-free representation, linear complexity attention (Mamba), and incremental fusion to predict safe ego-motion sequences. The model shows state-of-the-art performance on challenging real-world driving scenarios.