AIoT-driven federated meta-reinforcement learning for energy-aware task orchestration in large-scale edge–cloud cyber-physical systems

Abstract

The rapid evolution of Artificial Intelligence of Things (AIoT) applications is transforming large-scale edge–cloud cyber-physical systems into highly dynamic and distributed environments. In such settings, heterogeneous IoT tasks must be orchestrated under strict latency, energy, cost, and service-level agreement (SLA) constraints while preserving data privacy. However, conventional centralized or static reinforcement learning approaches struggle to adapt to non-stationary workloads, mobility-driven channel variations, and large-scale distributed coordination requirements. This paper proposes an AIoT-driven federated meta-reinforcement learning framework with adaptive sub-goal decomposition for energy-aware task orchestration across the cloud-to-thing continuum. The framework integrates meta-initialized policy learning for rapid adaptation, privacy-aware federated aggregation of policy updates, and attention-based sub-goal allocation to align local decisions with global multi-objective constraints. Extensive simulations under realistic 6G IoT conditions demonstrate that the proposed approach reduces energy consumption, end-to-end latency, and operational cost while improving task success ratio and energy–delay efficiency compared to strong reinforcement learning and federated baselines. The proposed framework provides a scalable and privacy-aware AIoT orchestration solution for next-generation intelligent edge–cloud cyber-physical systems.