ViRA-Diffusion posts 92.3 EPDMS on NAVSIM v2 navtest

ViRA-Diffusion scores 92.3 EPDMS on NAVSIM v2 navtest, outperforming "recent methods in our comparison by at least 1.9 points," according to a new preprint led by Zihao Zhang. The framework behind it, ViRA, is described as "a planner-agnostic visual representation alignment framework that keeps the planner architecture and inference cost unchanged."

The paper's underlying question is whether swapping in a stronger visual foundation model actually buys you better driving. The answer the authors land on is layered. VFM-guided representations "consistently improve driving performance across diverse end-to-end planners, with gains extending to zero-shot closed-loop evaluation," they write. But the choice of VFM target matters, and alignment to a different VFM "can further benefit planners with pre-trained VFM encoders."

One finding stands apart. Adding auxiliary perception supervision reduces sensitivity to which VFM you align to, "narrowing the EPDMS spread across five targets from 2.7 to 0.5 points" and, in the authors' phrasing, "potentially compensating for less effective VFM targets." The abstract names neither the five targets nor the baselines the +1.9-point margin was measured against.