4D-WAM: New world model brings 4D consistency to autonomous driving, beats NAVSIM benchmarks
Geometric supervision fixes 2D blind spots, improving trajectory planning with zero added inference cost.
4D-WAM, developed by researchers at Nankai University, Shanghai AI Lab, Huawei Noah's Ark Lab, and Xi'an Jiaotong University, targets a core weakness in autonomous driving world models: traditional World-Action Models (WAMs) train on video, which is a 2D projection of the real 4D scene (3D space plus time). This causes models to generate visually plausible but physically inconsistent future frames, which can mislead downstream trajectory planning.
To solve this, 4D-WAM feeds the model's predicted future frames into a geometric foundation model, then uses 4D-aware responses to compute a consistency loss. This supervision teaches the model to represent and predict structurally sound 4D scenes during training, with zero additional inference cost. The team also identified an "early-decision" phenomenon where driving decisions are largely formed in the early, high-noise stages of prediction. Their decision-oriented timestep sampling strategy concentrates supervision on these critical stages, significantly improving trajectory planning. Experiments show 4D-WAM achieves state-of-the-art performance on the challenging NAVSIM-v1 and NAVSIM-v2 benchmarks.
- 4D-WAM uses geometric foundation models to compute a 4D consistency loss, making future predictions physically consistent without extra inference cost.
- The proposed decision-oriented timestep sampling targets early, high-noise stages where driving decisions form, improving trajectory planning.
- Achieves state-of-the-art results on both NAVSIM-v1 and NAVSIM-v2 benchmarks for autonomous driving.
Why It Matters
Physically consistent world models are crucial for safe autonomous driving, and this approach improves planning without adding compute overhead.