An Integrated XAI-RL Architecture for Ethical Decision-Making in Autonomous Robots

Abstract

Can a robot's ethical reasoning be transparent without slowing it down? That is the question we set out to answer. Autonomous robots now make decisions that affect people's lives, and if we cannot understand how they decide, trust becomes a serious problem. We built a unified architecture that couples Explainable AI (XAI) with Reinforcement Learning (RL) through three connected components: a Perception Layer for situational awareness, a Decision Reasoning Layer that breaks agent utility down into measurable ethical sub-rewards (safety, fairness, and well-being), and a contrastive Explainability Engine grounded in a real-time Ethical Ledger. We tested it against Black-Box RL and two post-hoc XRL baselines (LIME and SHAP) across two simulation environments and a physical robot pilot on a TurtleBot3 platform. Our system reached 92±3% ethical alignment in simulation and 89±4% on the physical platform, beating every baseline (p < 0.01 throughout). In a controlled, IRB-approved user study (N=128), participants rated it higher on trust (4.5±0.4), understandability (4.6±0.3), and perceived safety (4.4±0.5) than the alternatives. Explanations were also fast: 4.2±0.6 ms on average, versus 18.3±2.1 ms for SHAP and 12.7±1.8 ms for LIME on the same hardware—well within real-time limits. An ablation study and sensitivity analysis back up these architectural choices.