Vision-language-action (VLA) models are seen as a promising paradigm for autonomous driving, but their large parameter scale creates a serious bottleneck: real-time inference strains onboard computational resources and directly increases energy consumption. This makes it hard to deploy them in vehicles with limited computing power.
The paper proposes Scene-Adaptive VLA, which uses dynamic layer routing to address this inefficiency. Instead of running the full model on every input, the routing mechanism selectively activates layers based on the complexity of the current driving scene. This allows the model to allocate computation where it is needed, potentially reducing the average workload and energy draw.
The abstract does not include experimental results or quantitative comparisons, so the actual efficiency gains and any impact on driving performance are not yet reported. The work is presented as a method for making VLA-based autonomous driving more practical on resource-constrained hardware.