Xiaomi unveils Auto World Model, blending reconstruction and generation for self-driving

Xiaomi Auto has officially launched the Xiaomi Auto World Model, a new framework that deeply integrates 3D reconstruction and video generation for autonomous driving. It breaks the industry's traditional separate approach by using 'reconstruction to anchor geometry, generation to fill imagination,' achieving SOTA results on benchmarks like Waymo and nuScenes.

Current world models have two main paths: Reconstruction (WorldRec) offers high fidelity and consistency but can't imagine new scenes; Generation (WorldGen) can predict unseen events but lacks explicit 3D structure and drifts over time. Xiaomi's model combines both: reconstruction provides geometric anchors for stable generation, while generation extends beyond reconstruction boundaries, creating a 1+1>2 synergy.

The model is already used in three Xiaomi Auto scenarios: synthetic data generation (over 100,000 clips delivered), closed-loop simulation testing (reproducing real accidents for optimization), and driver assistance training (generating dynamic instructional videos). Technical homepage and paper are now public.