XPeng AUGUstus VLA Claims NAVSIM State-of-the-Art Without a Scorer Module

Researchers from Peking University, XPeng Motors, and Fudan University published AUGUstus on October 14, a vision-language-action model for autonomous urban driving that achieves a Predictive Driver Model Score of 94.6 on the NAVSIM v1 benchmark — a claimed state-of-the-art result accomplished without the scorer module that many competing approaches rely on. The paper addresses one of the central engineering tensions in autonomous driving AI: how to preserve the reasoning capacity of large language models while keeping the action-generation loop fast enough to run on a production vehicle.

The Latency-Reasoning Dilemma That AUGUstus Sets Out to Solve

Vision-language-action models promise to replace hand-coded driving rules with end-to-end neural systems that can reason through unusual scenarios in natural language before issuing steering and throttle decisions. The practical problem is that generating chain-of-thought reasoning — the step-by-step verbal logic that makes VLA decisions more interpretable and generalizable — adds latency. In a vehicle navigating an urban intersection, even modest compute delays affect the available safety margin.

Earlier driving VLA approaches split into two camps. Regression-based and flow-based models generate vehicle actions quickly but cannot draw on language-based understanding, leaving them brittle in novel scenarios. Autoregressive chain-of-thought systems improve reasoning quality by generating verbal interpretations before the action token, but because reasoning tokens must complete first, the latency budget strains real-time deployment.

AUGUstus introduces a training scheme the paper calls AUGUstus-mode: during training, the model generates the action output first, and only afterward produces the chain-of-thought reasoning tokens. Because the model has learned to produce coherent reasoning after its own actions, the inference pipeline can skip the reasoning step entirely — outputting only the action — without having been trained differently from a reasoning-capable model. The model retains reasoning-shaped internal representations from training while running at deployment without paying the token-generation cost.

How AUGUstus-GRPO Jointly Optimizes Action and Language

The second architectural contribution is AUGUstus-GRPO, built on Group Relative Policy Optimization. Standard RL fine-tuning for language models optimizes only language outputs. AUGUstus-GRPO applies a joint reward signal that simultaneously evaluates the quality of the driving action and the quality of the chain-of-thought produced after it. The joint optimization is designed to prevent the reasoning tokens from drifting toward plausible-sounding language that does not actually improve driving decisions — a failure mode that emerges when action and language objectives are trained separately.

For action generation, AUGUstus uses a diffusion-based head producing continuous trajectory outputs. The NAVSIM v2 results use a "1-step denoising" variant that approximates the full diffusion process in a single forward pass, trading some distributional breadth for lower latency. Training used 512 NVIDIA H800 GPUs for approximately two days, consuming over 318,000 hours of driving video and more than 513 billion tokens.

NAVSIM Results and What the Benchmark Actually Measures

The NAVSIM v1 benchmark Predictive Driver Model Score combines five sub-metrics: no-fault collision rate, drivable-area compliance, time-to-collision margin, ego progress toward the route goal, and a comfort score penalizing abrupt inputs. On NAVSIM v2, which extends the metric set with additional dimensions including lane-keeping and heading consistency, AUGUstus achieves an Extended PDMS of 90.7. On nuScenes, the paper reports 0.27 meters average L2 displacement error and a 0.05 percent average collision rate. A preliminary on-vehicle test demonstrates the model running at 12 Hz — an early result rather than a production-certified target.

The benchmark uses non-reactive simulation: other vehicles do not respond to the ego agent's actions, which makes evaluation reproducible but limits its ability to capture interactive driving dynamics such as merging negotiations or yielding behavior at uncontrolled intersections. High PDMS scores are a meaningful signal about trajectory quality under defined conditions, not a safety certification for open-road use.

The explicit "without scorer module" framing matters because a scorer module evaluates multiple candidate trajectories after generation and selects the highest-rated one, inflating leaderboard results relative to what the base model achieves on its own. AUGUstus's 94.6 reflects intrinsic trajectory quality rather than post-hoc candidate selection. For context, the NAVSIM paper reports a human expert reference score of 94.8 on the same benchmark. All benchmark results are company-reported in the arXiv preprint and have not been independently reproduced as of this writing.

Where AUGUstus Fits Within XPeng's Commercial VLA Roadmap

The AUGUstus paper is an academic publication from XPeng's AI research team alongside PKU and Fudan, distinct from XPeng's production autonomous driving program. XPeng's commercial VLA 2.0, which entered mass production in March 2026, took a different architectural direction: it removed language as an intermediate representation entirely, moving to direct vision-to-action mapping. XPeng reported that over 50 percent of assisted driving mileage in the first deployment month used VLA-generated actions. A second-generation production upgrade in August 2026 expanded on-device parameters by 3.5 times, according to XPeng.

XPeng is also running a Robotaxi program, with trial passenger operations beginning in China in 2026, and Volkswagen has been announced as the first external customer for XPeng's AI technology, with global VLA 2.0 deliveries planned for 2027. AUGUstus and VLA 2.0 represent different engineering philosophies within the same company — the production system removing language intermediaries entirely, the research paper arguing that language intermediaries can be retained through a training-time ordering trick that eliminates their inference cost.

Claims Requiring Independent Verification

The news brief that prompted this article contained multiple figures that do not appear in the actual paper. It attributed a "13B parameter" count to the model (not stated in the paper), described "6.2 million km of XPeng fleet data" (the paper reports 318K+ hours of driving video, not a distance figure), and cited a "94.3% intersection-negotiation success rate in Guangzhou real-world testing" (the paper's primary result is a 94.6 PDMS on NAVSIM v1 simulation, not a Guangzhou road test). The brief also mentioned "43 adversarial scenario types from the China National Intelligent Vehicle testing standard" (the paper uses NAVSIM and nuScenes). The AUGUstus project page confirms the paper was released on October 14 at arXiv identifier 2610.16985; the brief additionally cited the wrong arXiv identifier (2410.10189). None of the brief's specific statistics appear in the paper.

The model parameter count remains publicly unstated. Whether model weights will be released has not been announced, limiting independent reproducibility.

What the NAVSIM Lead Means for the VLA Research Race

The 94.6 PDMS without a scorer module matters primarily because it separates architectural quality from post-hoc optimization. If the result holds under independent reproduction, AUGUstus-mode offers a practical approach to retaining VLM reasoning benefits in driving systems without paying an inference penalty — a tradeoff that has constrained reasoning-capable VLAs in production. The significance potentially extends beyond driving: any embodied AI system where language-model reasoning is valuable during training but impractical at deployment speed could benefit from the same action-first training ordering.

Whether AUGUstus-GRPO's joint optimization translates to measurably better driving across reactive, open-road conditions — and whether NAVSIM scores correlate with real-world safety improvement in complex urban traffic — will determine how much influence this architectural approach carries on the next generation of deployed systems.

Read more: World Action Agent previews robot moves before physical execution