Explainable and transferable deep reinforcement learning for adaptive patrol of rail-guided robot system

Intelligent facility management systems can reduce the workload of human operators by enabling autonomous operation. However, the lack of transparency in existing machine learning-based systems often hinders user trust, especially in safety-critical environments such as industrial and public facilities. To ensure reliability and accountability, autonomous systems must not only perform effectively but also provide human-understandable explanations for their actions. This article presents an explainable deep reinforcement learning framework for a rail-guided patrol robot that adaptively controls its speed based on the visual complexity of its surroundings. The proposed system employs the Deep Deterministic Policy Gradient (DDPG) algorithm to learn a continuous speed-control policy directly from image-based observations. To enhance transparency, Gradient-weighted Class Activation Mapping (Grad-CAM) is integrated into the actor network to visualize which spatial regions of the input most strongly influence speed decisions, providing post hoc explanations of the model’s decisions. To support real-world deployment, we incorporate a Cycle-Consistent Generative Adversarial Network (CycleGAN)-based domain adaptation module that transforms real camera images into a simulation-compatible visual style, enabling the trained policy to operate without additional retraining. Grad-CAM is also used to assess the semantic consistency of translated images and verify that domain adaptation preserves task-relevant visual cues. Because the proposed framework is designed around lightweight visual inputs and compact neural networks, its computational demand remains modest and suitable for embedded execution. Grad-CAM analysis is used for explainability rather than for action generation, and its computation does not affect the timing of the control loop. The framework is evaluated through extensive experiments in both simulation and a physical testbed environment. Results demonstrate that the robot successfully adjusts its patrol speed in response to scene complexity and that the learned policy provides coherent and meaningful visual explanations. These findings highlight the potential of combining deep reinforcement learning, visual domain adaptation, and explainable AI to realize trustworthy and adaptable autonomous patrol systems.