ACVRL: An Autonomous Causal Vision-Reinforcement Learning Framework for Self-Evolving Manufacturing Intelligence

Modern industrial AI systems excel at detecting and predicting manufacturing defects, yet they rarely determine why defects occur or autonomously recommend corrective actions. This paper introduces ACVRL (Autonomous Causal Vision–Reinforcement Learning), a novel manufacturing intelligence framework that integrates computer vision, causal reasoning, reinforcement learning, digital twins, distribution-shift detection, and self-reorganization into a unified closed-loop architecture. The framework formalizes autonomous manufacturing correction as a Causal-Augmented Partially Observable Markov Decision Process (CA-POMDP) and introduces four mathematical formulations governing causal-effect estimation, causal-weighted policy optimization, adaptive distribution-shift detection, and causal-guided architectural self-reorganization. To ensure reproducibility, the framework is evaluated using an open synthetic automotive paint-shop digital twin with a known causal graph and engineered confounding variables. Experimental results validate the causal reasoning component while transparently discussing current limitations of the control policy under aggressive distribution shifts. This work contributes a formal research framework for autonomous manufacturing intelligence and provides an openly reproducible foundation for future research on causal reinforcement learning in industrial automation.