Human Robot Collaboration in dynamic real world settings requires agents to perform high level intent inference without explicit rewards, rules, or corrective feedback. We present an Active Inference (AIF) framework formulated as a partially observable Markov decision process with discrete states. The method achieves self supervised goal recognition through minimization of variational free energy, enabling intrinsic inference of changing human intentions and autonomous action in non stationary environments.Intentions are treated as latent states. The agent integrates uncertain and noisy categorical observations that reflect human location, hand motions, and the fading of voice commands which serves as a proxy for changing trust. The model handles rapid learning of new transitions and effective reversal learning when context shifts abruptly. Simulation studies show that the approach maintains strong robustness. With both aleatoric and epistemic uncertainties the success rate is 79%. With only epistemic uncertainty the success rate reaches 86%.The method is further assessed on a physical Franka Emika Panda platform performing a dynamic handover task. Despite unmodeled noise and temporal ambiguity the robot reached approximately 68% accuracy. It shows that the AIF architecture can infer human intent from ambiguous and uncertain gestures in a manner similar to how people interpret each other, which in turn indicates that the framework can support adaptive and self supervised robotic partners.