ABSTRACT When resource‐intensive deep neural networks (DNNs) are deployed on mobile devices with short battery life (such as drones and augmented reality AR headsets), a trade‐off between high performance requirements and limited onboard energy needs must be achieved. Current offloading techniques mainly rely on static linear optimization, but these methods fail to fully consider the nonlinear dynamic changes of battery depletion and the actual performance degradation caused by hardware instability. To address these limitations, we propose a robust and battery‐aware edge‐cloud collaborative inference framework based on a Double‐Dueling DQN (3D‐QN) agent. Additionally, a cubic battery penalty mechanism is introduced to simulate the nonlinear urgency of battery depletion, significantly increasing the energy multiplication when approaching the critical threshold, thereby encouraging active energy saving. This framework also explicitly considers the limitations at the device end, including thermal throttling and random operating system interference, to ensure the continuity of services. We implemented a full‐stack cross‐layer prototype connecting the Android client and the edge server to validate the system. The proposed framework effectively approximates the Pareto frontier of energy‐delay trade‐offs in a large number of tracking‐based simulations. Moreover, compared with the most advanced baseline methods such as PPO, A2C, and SARSA, this framework demonstrates superior sampling efficiency and decision stability. Even in the presence of severe network jitter, the system can maintain low latency. Therefore, this work provides a deployable solution for reliable edge‐cloud inference systems.
Che et al. (Mon,) studied this question.