---

TALK: “Active Inference for Adaptive Robots: A First-Principles Approach to Perception, Planning, and Control”

Abstract
Robots operating in the real world must cope with uncertainty, incomplete information, and constantly changing conditions. Traditional policy-learning and control methods often require vast amounts of task-specific data and lack the flexibility to adapt to novel environments or unforeseen scenarios. In this talk, I will present active inference, a first-principles approach grounded in the free energy principle, as a unifying framework for adaptive robotics. By equipping robots with an appropriate generative model, we can integrate perception, planning, and control into a single process of inference: continually updating beliefs about the world and body to minimize variational free energy. I will discuss which forms of generative models are best suited for this purpose, enabling both rapid learning and efficient inference. This perspective offers a principled path toward truly open-ended, intrinsically motivated robot behavior, where autonomy emerges naturally from the physics of self-organizing systems.

---

Short Bio

Tim Verbelen received his Ph.D. in Computer Science from Ghent University in 2013, where he focused on intelligent distributed systems. He then served as a post-doctoral researcher at imec, Belgium’s leading digital research and innovation hub, where he developed distributed intelligence architectures for robotics and the Internet of Things and began pioneering research into applying the active inference framework to embodied AI. He is now Director of the Intelligent Systems Lab at VERSES AI, where his work focuses on novel and scalable applications of active inference to build embodied agents that interact with both the real world and with each other.