BROWSE
Detail
Conference Paper
AROM:Control-Theoretic Learning for Resilient and Decentralized Edge Intelligence
- Authors
- Eunju Jeong, Dong-oh Kang
- Issue Date
- 2025-11
- Citation
- International Conference on Mobile Computing and Networking (MobiCom) 2025, pp.1-7
- Language
- English
- Type
- Conference Paper
- Abstract
- Federated learning (FL) in mobile robotic systems faces persistent challenges, including network instability, agent dropout, and non-i.i.d. data distributions. We propose AROM (Adaptive Robust Optimization with Momentum), a fully decentralized optimization framework that enables scalable and resilient learning without centralized coordination. AROM mitigates gradient staleness through momentum-based smoothing, enhances resilience via adaptive consensus penalties, and reduces communication overhead using scheduled peer-to-peer updates. These mechanisms are unified under a control theoretic interpretation, wherein each agent's update emulates a proportional-integral-derivative (PID) controller: gradient descent serves as proportional feedback, momentum introduces derivative damping, and adaptive penalties enforce integral correction. Importantly, each component is deliberately designed to counter a specific form of system degradation: momentum addresses temporal inconsistency caused by communication delay, adaptive penalties respond to structural volatility from topology shifts, and communication scheduling alleviates synchronization bottlenecks. Together, these elements form a cohesive mechanism that systematically ensures convergence in dynamic, uncertain, and asynchronous environments. In contrast to prior methods that assume static graphs or synchronous updates, AROM dynamically adapts to network variability and agent dropout, enabling robust deployment in realistic distributed robotic systems. Experiments on LiDAR object detection and semantic layout reconstruction (CubiCasa5K) show that AROM outperforms decentralized and federated baselines in convergence speed, consensus accuracy, and generalization. A comprehensive sensitivity analysis further confirms AROM's robustness under varying communication and topology conditions.
- KSP Keywords
- Adaptive consensus, Communication Scheduling, Communication overhead, Control-theoretic, Data Distribution, Decentralized optimization, Distributed robotic systems, Edge intelligence, Federated learning, Gradient Descent, Network variability
This work is distributed under the term of Creative Commons License (CCL)
(CC BY)
(CC BY)
