EcoVLA: Energy-Efficient Device-Edge Co-Inference for Vision-Language-Action Models

A new research paper on arXiv (ID: 2608.15502) introduces EcoVLA, an adaptive device-edge co-inference framework designed for Vision-Language-Action (VLA) models. The framework aims to maximize system energy efficiency while meeting real-time constraints in robotic systems. VLA models, which are foundational for Embodied AI, face high inference costs that challenge deployment. On-device inference is limited by compute capacity and energy budgets, struggling to balance real-time control and energy efficiency. Offloading inference to edge servers introduces unpredictable latency due to fluctuating system conditions. EcoVLA addresses these issues by proposing a unified co-inference framework that jointly considers real-time constraints and system-level energy efficiency. The paper highlights the scarcity of systematic research on co-inference tailored to VLA models. The framework is presented as a solution to optimize energy efficiency without compromising real-time performance. The research is published on arXiv, a preprint server, and is available at the provided URL.