Graph-Trust-Enhanced Embodied Multi-Agent Reinforcement Learning for Secure Collaborative Control of Low-Carbon Energy Behaviors in Edge IoT Systems

Abstract

Low-carbon energy control in edge IoT systems requires heterogeneous buildings, storage units, electric vehicles, renewable interfaces, and flexible loads to coordinate under partial observability, unreliable communication, and strict physical constraints. Existing multiagent reinforcement learning methods generally treat agents as homogeneous controllers and cannot jointly handle embodiment differences, corrupted messages, trust uncertainty, and action feasibility. This paper proposes GT-EMARL, a graph-trust-enhanced embodied multi-agent reinforcement learning framework that combines capability-aware agent encoding, Bayesian trust estimation, trust-weighted graph reconstruction, centralized training with decentralized execution, and differentiable safety projection. Experiments on four public energy datasets show that GT-EMARL consistently improves low-carbon operating performance, robustness to false data injection and communication disturbances, physical safety, and edge-deployment efficiency compared with representative multi-agent baselines. Ablation results further confirm the complementary contribution of the trust, uncertainty, embodiment, and safety modules. The results demonstrate that trustworthy message aggregation and certified action execution are both essential for secure collaborative low-carbon control.