Caine

Development of a real-time digital twin for an autonomous AGV fleet operating in a warehouse environment. Robots self-organize without a central controller to detect and resolve floor incidents while safely navigating around racks and obstacles. Coordination emerges through Active Inference and free-energy minimization over a continuously learned shared world model. A supervisor LLM, deployed on Crusoe Managed Inference, converts the live operational state into clear, actionable recommendations for non-technical operators and continuously improves through operator feedback and Bayesian evidence updates. Built as a scalable proof of concept for industrial and operational environments using pure Python (NumPy and Pygame) with no GPU requirements.