Embodied artificial intelligence through free energy minimization

Abstract

Artificial Intelligence (AI) should automate dangerous jobs instead of creative ones. The International Labour Organisation reports over 300 million work-related accidents and diseases per year, with nearly 3 million being fatal. Embodied AI can reduce this drastically, for example by taking over construction site or power plant inspection (top incidences). But designing and developing AI systems that continuously learn in-situ on resource-constrained devices is a challenge, one that can’t be solved with more data. I will present a framework for designing intelligent autonomous systems based on the Free Energy Principle, a leading physics-based theory of information processing in brains. We implement this framework by performing variational Bayesian inference through message passing on factor graphs. Our software toolbox, RxInfer.jl, is used for identification, planning and control of complex mechatronic systems and mobile robots.