Edge-Intelligence With Inference-Aware Offloading for Latency-Sensitive Agents in IoT Networks

This letter analyzes the outage probability in edge intelligence (EI)-enabled networks supporting embodied agents, such as robotic assistants (RAs) powered by foundation models (FMs). Closed-form expressions for outage probability are derived under Rayleigh fading for both inference-aware offloading (IAO) and cloud-only (CLD) strategies. The analysis models RA and interfering user locations as binomial point processes and incorporates Gamma-distributed latencies in the wired backhaul. Unlike prior works focused on latency or energy optimization, this study explicitly accounts for FM inference accuracy and processing dynamics at both edge and cloud levels. Through analytical modeling and simulation, suitable timeout thresholds, base station (BS) inference accuracy levels, BS computation rate levels, and maximum tolerable wired backhaul latencies are identified to ensure reliable operation in both urban and indoor environments. The results show that the IAO strategy consistently outperforms CLD in terms of outage probability across all evaluated scenarios, validating its effectiveness for real-time FM-based embodied agents.