AUGUstus unveils autonomous driving VLA model with unified generation and understanding
autonomous
| Source: HF Papers | Original article
AUGUstus presents a unified vision‑language‑action model that merges generation and understanding, targeting VLA drift to vision‑action and latency limits in autonomous driving.
XPENG‑AI, together with researchers from Peking University and Fudan University, has unveiled **AUGUstus**, a vision‑language‑action (VLA) model that merges diffusion‑based action generation with autoregressive chain‑of‑thought (CoT) understanding. The new architecture is designed to keep the rich reasoning capabilities of large VLMs while delivering the low‑latency performance required for on‑vehicle deployment. According to the authors, training on 512 H800 GPUs over roughly two days consumed more than 318 000 video hours and 513 billion tokens, and the resulting system runs at 12 Hz on a car‑mounted processor.
The announcement tackles two persistent shortcomings of existing VLA approaches. First, many models collapse into pure vision‑action pipelines, losing the benefits of language‑driven reasoning. Second, autoregressive generation of CoT and action tokens typically incurs prohibitive latency, limiting real‑time use. AUGUstus resolves this by ordering actions before CoT during training, so inference can rely on a single action output. A unified reinforcement objective, dubbed **AUGUstus‑GRPO**, jointly optimizes both components, preserving understanding while streamlining execution.
Performance metrics underscore the impact. On the NAVSIMv1 benchmark, AUGUstus achieved a 94.6 PDMS score without employing a scorer, setting a new state‑of‑the‑art result for VLA‑driven autonomous driving. The model’s ability to handle long‑tail driving scenarios—situations that deviate from typical training data—suggests a step toward more robust, end‑to‑end self‑driving systems.
What to watch next includes field trials that move beyond simulation, integration into production‑grade vehicle stacks, and further benchmarking against competing VLA frameworks. Industry observers will also monitor whether the low‑latency, reasoning‑rich approach can influence regulatory discussions around safety verification for autonomous driving software.
Sources
Back to AIPULSEN