Adaptive Regulation through Model-Observation Mismatch: From Process Control to Cognitive Neurorobotics

Adaptive behavior requires biological and artificial systems to respond when their internal models no longer match what they observe. How can such discrepancies be used not only to correct predictions, but also to regulate how a system interacts with its environment?

In this talk, I will explore this question across two research directions spanning process control and cognitive neurorobotics. First, I will introduce a model predictive control framework in which a controller actively guides a nonlinear process toward informative conditions to distinguish among competing models while maintaining safe operation. I will then move to cognitive neurorobotics, where brain-inspired computational principles are used to investigate perception, prediction, and action in embodied artificial agents. In physical human-robot interaction, I will show how a predictive-coding-inspired recurrent neural network enables a humanoid robot to regulate the balance between its learned expectations and sensory information arising from human-guided movements. In particular, recent mismatches between sensory observations and the robot's internal model are used to regulate how strongly it relies on its learned dynamics or accommodates ongoing interaction. Together, these studies illustrate how model-observation mismatch can provide a general signal for adaptive regulation, connecting ideas from control engineering, artificial intelligence, neuroscience, and embodied cognition.