Mobility-aware Lyapunov-guided deep reinforcement offloading for wearable edge computing

Abstract Wearable devices (WDs) are becoming an important platform for mobile intelligent services, but their limited battery capacity and computing power make continuous local inference difficult. Existing MEC offloading methods often assume fixed edge association or jointly optimize all offloading variables in a single control space, which makes it difficult to adapt to user mobility while preserving queue-aware resource control. To address the mobility-conditioned offloading problem in wearable edge computing, this paper proposes Mobility-aware Lyapunov-guided Deep Reinforcement Offloading (Mobi-LyDRO). Mobi-LyDRO uses a feasible-action-masked reinforcement-learning policy to select the discrete association between each WD and its reachable edge servers; conditioned on that association, a Lyapunov drift-plus-penalty allocator controls WD CPU frequency, offloading power and edge CPU allocation. For a stationary ergodic admissible association process that admits a queue-feasible comparator with uniformly bounded conditional second moments and service slack outside a compact backlog set, the lower Lyapunov layer retains a conditional cost-backlog tradeoff and bounded-average-queue result; this is not an unconditional guarantee for the learned association policy. Ten-seed simulations with matched mobility, fading and workload streams show that Mobi-LyDRO produces no mobility-infeasible actions, reduces handovers by 57.1% relative to Edge-only, and attains a mean ES-load Jain index of 0.962. The strong Edge-only heuristic and the infeasible-action No-mask ablation have lower modeled cost, so the measured advantage of action masking is feasibility rather than uniform cost dominance. The evaluation uses synthetic wearable-inspired arrivals with periodic and bursty variation rather than replaying measured wearable-to-edge task traces.