Adaptive resource scheduling for hierarchical edge intelligence via self-attention and self-supervised deep reinforcement learning

Abstract

Mobile edge computing (MEC) provides an effective platform for supporting delay-sensitive intelligent services in dynamic environments. In UAV-assisted robotic systems, sensing tasks and computing requests vary significantly over time, while communication and computation resources remain heterogeneous and constrained. These characteristics may lead to inefficient task allocation, resource imbalance, and unstable system performance. To address these issues, this paper presents an SS-DDPG-based task scheduling framework for a collaborative robot-UAV-MEC architecture. A hierarchical computing model is established in which robots generate tasks and collect environmental information, UAV provide communication assistance and task forwarding, and MEC servers offer powerful computing resources. To improve decision-making quality, a self-attention mechanism is integrated into the Actor-Critic structure to capture interactions among distributed nodes and exploit global state information. In addition, a self-supervised learning strategy is introduced to provide auxiliary guidance during policy optimization, enabling adaptive scheduling under dynamic workloads. A reward design considering both task latency and system stability is further developed to improve long-term performance. Experimental results demonstrate that the proposed method achieves lower service delay, better workload distribution, and stronger robustness than conventional reinforcement-learning-based scheduling approaches.