Extending the Horizon: Uncertainty-Aware 4D Scene Generation for Robust Autonomous Driving

4D LiDAR scene generation for autonomous vehicles
is limited by a uniformity fallacy: existing models allocate equal
computational effort across all spatial regions, starving rare and
dynamically critical areas of representational budget. This produces three compounding failures—generation quality degrades
beyond roughly ten frames (approximately 0.5 seconds), outof-distribution detection is performed per-frame and therefore
misses temporal anomalies, and segment-first pipelines lose novel
or occluded objects at the initial detection step. We propose a
hard-first, uncertainty-guided generation framework that schedules capacity to high-entropy regions—dynamic outliers, occlusions, and unseen object classes—before resolving background
geometry. Shannon entropy serves as the routing signal, directing the denoising process to front-load difficult regions while
the model retains maximum representational freedom. Openvocabulary semantic priors are coupled with the uncertainty
routing to label and render novel object classes that segment-first
pipelines discard. We evaluate on nuScenes and report Expected
Calibration Error under out-of-distribution scenarios alongside
temporal-coherence and horizon-length metrics. By prioritizing
the regions where the model is least certain, the framework
extends reliable generation horizons, reduces calibration error
on rare-event scenarios, and enables safer downstream planning.