Wouter Kouw develops probabilistic machine learning systems that learn from interaction with their environment. His central premise is that inference and decision-making should run on the same computational substrate: Bayesian inference as message passing on factor graphs. From this he derives online system identification algorithms for nonlinear dynamical systems — autoregressive and Gaussian process-based physical models — and, on the same graph, agents that plan by minimizing expected free energy. Much of this work concerns actively pursuing informative data points or trials. Applications so far include multi-joint dynamical systems, heat transfer processes, maritime traffic, radar tracking, and neuromorphic computing.

Through the CONTACT-AI and FEPQuad projects, the aim is to bring these agents onto legged robot platforms that use contact to resolve perceptual ambiguity and negotiate unknown terrain — ultimately to automate work that is dangerous or infeasible.

“Intelligence is more than just computation at scale - it’s about making sense of the world under uncertainty.”