Grounded world models: Biology's blueprint for next-gen embodied AI

A new paper from Pezzulo et al. (arXiv:2607.13560) challenges the dominant paradigm in generative and embodied AI, which relies on large-scale predictive learning over multimodal data — with language providing the primary scaffold onto which other modalities attach. The authors, drawing on neuroscience and cognitive science, argue that biological intelligence works in the opposite direction: grounded world models acquired through real-world interaction form the semantic foundation, and language is later attached to that scaffold. Current AI systems, they contend, remain fundamentally passive, learning from externally provided datasets rather than through autonomous experience.

To illustrate their thesis, the authors present five examples of neural circuits that support grounded world modelling in biology: navigation in physical and conceptual spaces, affordance-based perception and object interaction, active perception and exploratory learning, allostatic control and emotion, and the distinction between self- and world-generated outcomes. These examples highlight features missing from today's embodied AI — including intrinsic dynamics as a basis for learning, the centrality of action in aligning those dynamics with the external world, and the role of autonomous, open-ended experience. The paper also discusses how early predictive and control mechanisms in biology scaffold higher cognitive abilities like reasoning, planning, imagination, theory of mind, and communication. Finally, they propose that future embodied AI could be trained through social interaction to build world models that are not only grounded but also socially shared and aligned with human norms.