DA-RL: decoupled attention-enhanced visual feature reinforcement learning for optimised decision-making in cyber-physical systems

Abstract

In dynamic and complex environments, visual reinforcement learning (Visual RL) agents struggle to extract task-relevant information, achieve stable policy optimisation, and generalise across changing conditions. This paper proposes DA-RL, an attention-enhanced and decoupled reinforcement learning framework for cyber-physical decision-making. An attention-based encoder highlights critical environmental cues, while decoupled feature and policy optimisation improves training stability and convergence efficiency. An adaptive weighting mechanism further enhances robustness under environmental changes. Experiments show that DA-RL outperforms conventional reinforcement learning methods in convergence speed, decision accuracy, and environmental adaptability, demonstrating its potential for autonomous driving, intelligent transportation, robotics, and industrial control.