Intelligent Earth Seminar: Ingmar Posner (Oxford)

25 February 14:00
Seminar Room 1a, Doctoral Training Centre, 1-4 Keble Rd
pastedgraphic 1

Prof. Ingmar Posner leads the Applied Artificial Intelligence Lab at Oxford University. He was a Founding Director of the Oxford Robotics Institute and an entrepreneur (Oxa) until he took up a position as an Amazon Scholar in 2021, which he still holds. Ingmar’s research develops AI models that uncover structure in the world, helping robots - and increasingly scientists - make sense of the complex systems and environments. In particular, his recent work is focussed on structured world models enabling machines to discover re-usable and interpretable insights for improved generalisation through observation, action and interaction.

 

Beyond Prediction: Structured World Models for a Changing Planet

Across science, we are drowning in data but starved of insight. Recent breakthroughs in domains such as protein folding and weather forecasting demonstrate deep learning’s predictive power, yet these systems remain predictors, not explainers: they map inputs to outputs but do not reveal the mechanisms or variables that underpin scientific understanding.

In this talk, I present our work on Mechanistic World Models: a framework designed to bridge the gap between raw observation and symbolic reasoning. Drawing on lessons from robotics, I will detail our progress in developing models that move beyond monolithic prediction to extract re-usable, interpretable abstractions from observations. By decomposing complex observations into discrete latent entities and causal dynamics, we enable AI to move from mere pattern matching to true system identification. I will conclude by discussing how these models provide a blueprint for a new generation of Intelligent Earth models. In an era of climate instability, we can no longer rely on models that simply interpolate the past; we require systems that don't just mimic the planet’s data, but help us decode its governing laws to provide actionable, mechanistic foresight.