Preprint proposes adaptive device-edge inference for VLA robotics

An arXiv preprint presents EcoVLA, a device-edge co-inference framework for vision-language-action models used in robotics. The authors say it predicts latency and energy across device, edge and network conditions, then selects a scheme that meets real-time constraints. In experiments under a 20 Hz action-output constraint, they report energy-efficiency gains of up to 236% over existing co-inference approaches while maintaining service-level objectives during changing network and edge workloads.