yangluvu/scene-adaptive-VLA

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

This repository contains the model presented in Scene-Adaptive VLA: Efficient Autonomous Driving via Dynamic Layer Routing.

Scene-Adaptive VLA is an efficient Vision-Language-Action (VLA) framework for autonomous driving that dynamically routes layers based on scene complexity, reducing inference-time computational overhead by 53.5% with negligible degradation in driving and language capabilities.

Code and evaluation instructions are available at github.com/yang-luvu/Scene-Adaptive-VLA.