Driving · 2026-10-20

Scene-Adaptive VLA: Efficient Autonomous Driving via Dynamic Layer Routing

Yujia Yang, Quan Yuan, Jiawei Fan, Yang Li, Tiange Fu, Xiaoyuan Fu, Guiyang Luo, Jinglin Li

AI-assisted summary based on the paper. Consult the linked source for methods and limitations. Benchmark numbers appear after editor verification.

TL;DR

Scene-Adaptive VLA is an efficient VLA-based autonomous driving framework that uses dynamic layer routing to adjust active parameters based on real-time scenes. It characterizes model-perceived scene complexity using temporal scene variations and decision deviations encoded into scene-aware tokens, then uses a budget-conditioned routing mechanism to allocate an inference budget and select layer combinations. Adapted to SimLingo on Bench2Drive, it reduces inference-time computational overhead by 53.5% with negligible degradation.

Why it matters

The paper targets reducing computation for VLA autonomous driving under real-time constraints. Instead of only sequential adaptation, it introduces structural adaptation that dynamically activates parameter subsets per frame based on scene complexity, aiming to lower onboard compute/energy while preserving driving and language capabilities.

Method

Characterize model-perceived scene complexity via temporal scene deviations and inter-frame decision deviations, encoded into scene-aware tokens.
Use a budget-conditioned layer routing mechanism with budget and routing queries processed in shallow LLM layers to determine an active parameter ratio and select layer combinations (binary routing mask).
Train end-to-end with two-stage fine-tuning: SFT (randomized budgets) and GRPO reinforcement fine-tuning balancing driving proficiency and computational overhead.

Abstract (from arXiv)

Vision-Language-Action (VLA) models offer a promising paradigm for autonomous driving. However, the massive parameter scale of VLAs strains onboard computational resources during real-time inference. This directly increases energy consumption and limits vehicle battery range, making it critical to minimize computational overhead without compromising driving capability and explainability. Current research on efficient VLA-based driving focuses on sequential adaptation (\textit{e.g.}, Chain-of-Thought and token adjustment), leaving structural adaptation underexplored despite parameter redundancy across dynamic driving scenes. To address this gap, we propose Scene-Adaptive VLA, an efficient framework leveraging dynamic layer routing to adjust active parameters based on real-time scenes. Inspired by the temporal continuity of scenes and model predictive uncertainty, our approach characterizes model-perceived scene complexity via frame-level temporal scene variations and decision deviations. To uncover latent scene-to-parameter correspondences, we encode this characterization into scene-aware tokens, alongside two specialized queries. A budget-conditioned layer routing mechanism evaluates these refined queries to allocate a budget that determines the active parameter ratio and subsequently selects appropriate layer combinations. The entire framework is optimized end-to-end. Extensive evaluations on the Bench2Drive benchmark show that when adapted to a state-of-the-art VLA model, our approach significantly reduces inference-time computational overhead by 53.5\% with negligible degradation in driving and language capabilities. Code is available at https:github.com/yang-luvu/Scene-Adaptive-VLA.

Publication record

Bibliographic evidence describes where and when a record appeared. It does not validate scientific claims.

First public release: 2026-10-17 · Indexed by Pub-AI: 2026-10-20 UTC

preprint