MotionWorld: Rethinking 4D Occupancy Forecasting through Explicit Scene Evolution

We introduce MotionWorld, a simple yet effective training-free analytical occupancy world model that decomposes 4D forecasting into static-scene propagation and instance-level motion extrapolation. For the static background, we leverage ego-motion priors to warp historical background and present Boundary-aware Conservative Completion (BCC) to conservatively fill warping-induced holes. For dynamic objects, we propose Trajectory Instance Association (TIA) to recover a sparse trajectory per instance, and Self-Validated Motion Extrapolation (SVME) to select among stationary, constant-velocity, constant-acceleration, exponentially smoothed motion and turning hypotheses via historical replay. The extrapolated instances are then fused with the propagated background to produce future semantic occupancy. We further present MotionWorld-Refine, a compact variant that learns targeted corrections using the analytical forecast as a structured prior, improving accuracy while mitigating the structural distortions of unconstrained full-scene generation. On Occ3D-nuScenes, MotionWorld attains highly competitive accuracy at 265.19 FPS on a dual-core CPU, achieving a 3.5$\times$ speedup over the previous fastest approach running on an RTX 4090 GPU. Under the GT occupancy setting, MotionWorld-Refine establishes a new state of the art, surpassing the previous best method by 1.26% semantic mIoU and 5.14% geometric IoU with only 13% of its parameters. Over an 8-second horizon, it maintains the average gains of 13.29% and 11.81%in semantic mIoU and geometric IoU, respectively.