C-VLA: A Hierarchical Vision-Language-Action Architecture for Energy-Efficient In-Cabin Agentic AI on Automotive Edge Hardware

Driver monitoring systems have traditionally relied on reactive, single-model architectures that continuously invoke large language models regardless of driver intent complexity, resulting in excessive energy consumption and latency unsuitable for automotive edge deployment. We present C-VLA (Cabin Vision-Language-Action), a hierarchical two-brain architecture that decouples reflex gating from agentic reasoning. A lightweight CNN classifier (System 1) processes every frame at 7.9ms and 126.6 FPS on the NVIDIA Jetson Orin Nano Super, invoking a Llama-3.2-1B reasoning module (System 2) only when complex driver intent is detected. Evaluated on the DMD driver monitoring dataset (6,405 frames, 6 classes), System 1 achieves 92.8% accuracy (F1: 91.9%) while reducing LLM invocations by 65% and mean response latency by 97% compared to a single-model baseline. All code and model weights are released publicly. Index Terms—driver monitoring, edge AI, vision-language-action, hierarchical inference, automotive, Jetson Orin Code and model weights available at: https://github.com/devender-sharma-emb/C-VLA-Cabin-AgentModel weights: https://huggingface.co/devvender/cvla-cnn-gatekeeper