Adaptive Probabilistic Inference of Human Intentions in Smart Manufacturing via Discrete Active Inference

Human-Robot Collaboration (HRC) in dynamic environments presents challenges in intention inference, adaptability, and reducing human cognitive workload. Existing methods often assume known transition models, static human intentions, or rely on explicit instructions, limiting their real-world applicability. We propose an Active Inference (AIF) framework that models human intentions as evolving latent states and integrates a trust-driven mechanism to balance autonomy and human guidance. In controlled simulations of a factory assembly scenario, the agent continuously updates its internal model under uncertain and noisy categorical observations, including human location, hand movements, and voice commands. The framework demonstrates efficient operation despite persistent uncertainty, homeostatic adaptation from incomplete models, rapid learning of novel transitions, reversal learning in response to sudden changes, and dynamic regulation of internal parameters such as learning and forgetting rates, all bounded by Variational Free Energy (VFE). Quantitative results indicate that modeling both aleatoric and epistemic uncertainties yields a task success rate of 79%, compared to 86% when only epistemic uncertainties are considered. These findings highlight active inference as a scalable paradigm for robust, adaptive, and autonomous HRC in complex real-world environments.