Open Access

ARTICLE

Graph-Based Constrained PPO for Low-Latency and Energy-Aware AI Agent Migration in Internet of Vehicular Agents

Kanyang Jiang¹, Yingkai Kang², Ming Li^2,*

1 School of Automation, Guangdong University of Technology and Key Laboratory of Intelligent Detection and the Internet of Things in Manufacturing, Ministry of Education, Guangzhou, China
2 School of Automation, Guangdong University of Technology, Guangzhou, China

* Corresponding Author: Ming Li. Email:

(This article belongs to the Special Issue: AI-Driven Optimization for Secure and Sustainable Edge IoT Services)

Computers, Materials & Continua 2026, 88(2), 61 https://doi.org/10.32604/cmc.2026.083294

Received 01 April 2026; Accepted 03 May 2026; Issue published 15 June 2026

Abstract

The Internet of Vehicular Agents (IoVA) interconnects distributed AI agents across vehicular networks to deliver real-time intelligent services for vehicular users. Due to the limited computing capacity of vehicles, AI agents are deployed on nearby RoadSide Units (RSUs) to perform computation-intensive inference. As vehicles traverse RSU coverage boundaries, AI agents must migrate to target RSUs to maintain service continuity. However, the communication and computing resources at each RSU are shared among multiple co-served vehicles, creating coupled allocation decisions that jointly determine system latency and energy consumption. To address this challenge, we propose a low-latency and energy-aware AI agent migration framework that models the end-to-end system latency and vehicle energy consumption in the IoVA. Since the cumulative nature of energy consumption introduces long-term constraints that cannot be handled by instantaneous optimization, we formulate the resource allocation problem as a constrained Markov decision process and develop a Graph-based Constrained Proximal Policy Optimization (GCPPO) algorithm to solve it. GCPPO employs a bidirectional graph attention network to extract the relational features between heterogeneous vehicles and RSUs, thereby enabling topology-aware resource allocation, and adopts a Lagrangian dual mechanism to adaptively enforce the long-term energy constraints. Simulation results demonstrate the effectiveness and scalability of the proposed algorithm, which achieves a 31.3% reduction in average system latency over baselines while attaining a 96.4% constraint satisfaction rate.

---

Keywords

---