AUGUstus

Autonomous Driving VLA Model with
Unified Generation and Understanding

Yuqi Ye^1,*Caigao JIANG^2,‡Zhuonan Wang³Congyuan Pei²Chao Qu³Changhao Peng¹Junhong Lin¹Qize Yu¹Zetian Song¹Guangzhao Fu²Chongruo Wu²Shiyu Huang²Shuchang Zhou²Honggou Yang²Wei Gao^1,†

¹ Peking University

² XPeng Motors

³ Fudan University

* First author ‡ Project lead † Corresponding author

Email: yeyuqi0303@stu.pku.edu.cn and cjiangad@connect.ust.hk · WeChat: yuki-hahaha-yuki and Callmejohnjiang

Paper Code Models

Public BenchmarksNAVSIM v194.6PDMSwithout scorer
NAVSIM v290.7EPDMS1-step denoising
nuScenes0.27 mAverage L20.05% collision

In-House Training

NVIDIA H800512GPUs
Driving video318K+Hours
Training513B+Tokens

Training about 2 days

DeploymentReal vehicle12 HzOn-vehiclePreliminary test

Abstract

Vision-language-action (VLA) models have shown strong potential for end-to-end autonomous driving by leveraging world knowledge and understanding to achieve great performance in long-tail scenarios. However, current VLAs suffer from two major limitations: 1) Degeneration into vision-action (VA) models. To achieve low latency, regression-based and flow-based methods directly generate actions, yet they do not sufficiently leverage the VLM's understanding, leading to unsafe and unexplainable actions. 2) Autoregressive high latency. Autoregressive methods sequentially generate chain-of-thought (CoT) and action tokens, fully leveraging the VLM's understanding, but high latency makes deployment difficult. To enable low-latency deployment while properly using CoT to assist action generation, we propose AUGUstus¹, an Autonomous-driving VLA model with Unified Generation and Understanding. Specifically, we first unify diffusion-based action generation and autoregressive CoT understanding in a single model, where understanding is optimized only to serve generation. More concretely, joint CoT supervision preserves the VLM's spatial understanding in the shared representation and keeps it useful for safer and more accurate trajectories. Second, we propose AUGUstus-mode, which orders the action generation before the CoT so that the CoT supervises the reliable action during training while inference needs only the action output, enabling low latency. Finally, we introduce AUGUstus-GRPO, a unified GRPO algorithm that jointly optimizes the action and CoT, further improving action performance. Extensive experiments demonstrate that AUGUstus achieves 94.6 PDMS on NAVSIMv1 and 90.7 EPDMS on NAVSIMv2, establishing state-of-the-art (SOTA) planning performance without scorer module. On nuScenes, our method achieves an average L2 error of 0.27 m and an average collision rate of 0.05%, both the best among the compared methods. To validate the practicality of AUGUstus, we train it on our in-house dataset using 512 NVIDIA H800 GPUs for approximately two days, then evaluate it on an internal 1,000 km dataset and preliminarily deploy it on a real vehicle, achieving 12 Hz operation.

¹ AUGUstus refers to Augustus (Octavian), the first Roman emperor, whose consolidation of Roman authority symbolizes unification. Here, it denotes a unified generation and understanding VLA model.

Architecture

The difference between AUGUstus and other autonomous-driving VLA paradigms, panel 1 — Figure 1.The difference between AUGUstus and other autonomous-driving VLA paradigms.

Overview of AUGUstus, panel 1 — Figure 2.Overview of AUGUstus.

Supplementary AUGUstus architecture — Figure 3.Detailed AUGUstus module.

The pipeline of AUGUstus-GRPO, panel 1 — Figure 4.The pipeline of AUGUstus-GRPO.

Public Benchmark Results

Table 1: Performance comparison on NAVSIMv1 benchmark — Table 1.Performance comparison on NAVSIMv1 benchmark.

Table 2: Performance comparison on NAVSIMv2 benchmark — Table 2.Performance comparison on NAVSIMv2 benchmark.

Table 3: Performance comparison on nuScenes benchmark — Table 3.Performance comparison on nuScenes benchmark.

In-House Benchmark Results

Table 4: In-house 1,000 km evaluation — Table 4.In-house 1,000 km evaluation.

Table 5: Inference time — Table 5.Inference time.

Driving Demonstrations

Public Benchmark Videos

In-House Driving Rollouts

12 Hz Real-World Driving

Corresponding multi-camera and BEV image for real-vehicle test 01 — Figure 5.Real driving test visualization.

Corresponding multi-camera and BEV image for real-vehicle test 02 — Figure 6.Real-vehicle qualitative examples.

Corresponding multi-camera and BEV image for real-vehicle test 03 — Figure 7.Real-vehicle qualitative examples.

Qualitative Results

Public Benchmark: Comparison with Other Methods

Public method comparison · Demo 01 — Figure 8.Comparison of AUGUstus with ReCogDrive and AutoDrive-P³ across public benchmark scenes.

Public method comparison · Demo 02 — Figure 8.Comparison of AUGUstus with ReCogDrive and AutoDrive-P³ across public benchmark scenes.

Three-Stage Training: SFT, RL Stage I & RL Stage II

Three-stage training · Demo 01 — Figure 9.Trajectory comparison across SFT, RL Stage I, and RL Stage II.

Three-stage training · Demo 02 — Figure 9.Trajectory comparison across SFT, RL Stage I, and RL Stage II.

In-House: Comparison with Other Methods

In-house method comparison · Demo 01 — Figure 10.Comparison of AUGUstus with VLM+regression and VLM+flow matching on the in-house dataset.

In-house method comparison · Demo 02 — Figure 10.Comparison of AUGUstus with VLM+regression and VLM+flow matching on the in-house dataset.

In-House Driving Examples

In-house driving · Demo 01 — Figure 11.AUGUstus trajectories in representative scenes from the in-house 1,000 km evaluation.

In-house driving · Demo 02 — Figure 11.AUGUstus trajectories in representative scenes from the in-house 1,000 km evaluation.

CoT Attention

Trajectory Comparison: With and Without CoT

CoT trajectory comparison · Demo 01 — Figure 13.Trajectory comparison with and without CoT supervision in collision-risk and progress-preserving scenarios.

CoT trajectory comparison · Demo 02 — Figure 13.Trajectory comparison with and without CoT supervision in collision-risk and progress-preserving scenarios.

Citation

BIBTEX

@misc{ye2026augustusautonomousdrivingvla,
  title={AUGUstus: Autonomous Driving VLA Model with Unified Generation and Understanding},
  author={Yuqi Ye and Caigao Jiang and Zhuonan Wang and Congyuan Pei and Chao Qu and Changhao Peng and Junhong Lin and Qize Yu and Zetian Song and Guangzhao Fu and Chongruo Wu and Shiyu Huang and Shuchang Zhou and Honggou Yang and Wei Gao},
  year={2026},
  eprint={2610.16985},
  archivePrefix={arXiv},
  primaryClass={cs.CV},
  url={https://arxiv.org/abs/2610.16985},
}