26 December 2024 in Robotics and Autonomous Systems
29 September 2026 in Image and Vision Computing
11 May 2026 in SSRN Electronic Journal
Vision-centric world models have shown great promise in autonomous driving, but existing autoregressive paradigms based on Transformers suffer from quadratic computational complexity and severe spatial degradation in long-horizon 4D occupancy forecasting. To address these challenges, we propose DeltaMamba, a novel kinematics-aware spatiotemporal world model. By forecasting the differential changes of the environment using State Space Models rather than reconstructing dense grids from scratch, our method achieves linear O(L) complexity and effectively mitigates long-term compounding errors.Furthermore, to bridge the critical gap between world representation and motion planning, we introduce a Kinematics-Aware Dual Attention module coupled with Codebook Reconstruction Planning. Modulated by Adaptive Layer Normalization, this architecture explicitly aligns the visual features with ego-intent, forcing the network to proactively focus on navigable corridors and potential hazards. Extensive experiments on the nuScenes dataset demonstrate that our framework establishes a new state-of-the-art. It significantly mitigates error accumulation, boosting the 4D forecasting mIoU by 49.6% relatively at the challenging 3-second horizon, while achieving highly competitive planning safety. Notably, DeltaMamba strikes an optimal balance between extreme 4D perception and safe trajectory generation, achieving an inference speed of 14.1 FPS during validation, a 43.9% relative speedup over existing baselines.