AUGUstus VLA Hits 94.6 PDMS on NAVSIMv1 Without a Scorer

AUGUstus, a driving vision-language-action model described in a new arxiv preprint, reports 94.6 PDMS on NAVSIMv1 and 90.7 EPDMS on NAVSIMv2, which the authors call "SOTA planning without scorer module."

The paper frames this as a resolution of a tradeoff the authors say current driving VLAs face. Regression and flow-based methods generate actions quickly but "underuse VLM understanding, causing unsafe, unexplainable actions," the abstract says, while autoregressive methods that walk through a chain of thought before emitting a trajectory suffer high latency that "hinders deployment." AUGUstus unifies diffusion-based action generation and autoregressive CoT understanding in a single model, then orders actions before CoT during training so that "inference needs only action output, enabling low latency." A unified reinforcement objective the authors call AUGUstus-GRPO jointly optimizes action and CoT.

On nuScenes the authors report an average L2 error of 0.27 m and an average collision rate of 0.05%, which they describe as "best among compared methods."

To back the practicality claim the authors say they trained on in-house data using 512 NVIDIA H800 GPUs for roughly two days, evaluated on an internal 1,000 km dataset, and "preliminarily deploy on a real vehicle, achieving 12 Hz operation." The abstract names no vehicle platform, operator or jurisdiction for that on-road run.