Abstract

In robotics, the ability to learn predictive models of systems and environments in an unsupervised, data-driven way has emerged as a promising research direction. World models in particular harbour the potential to serve as an interactive experience store for autonomous agents, enabling direct trajectory optimisation as well as model-based learning in imagination. In this talk I will describe our recent work in creating robust and versatile world models. In particular, I will demonstrate that causally inspired inductive biases provide sufficient structure to achieve efficient adaptation to intervened environments. I will then demonstrate that the structure encoded in a learnt latent space already provides a powerful and intuitive way to disentangle and manipulate task-relevant factors of variation. I will describe how this structure can be exploited to capture a purely data-driven model of complex robot platforms and demonstrate that this not only casts a novel light on affordance learning, but also reveals a framework capable of learning a versatile unified representation for quadruped locomotion deployable in the real-world.

The seminar will be held in the JDB Seminar Room, Department of Engineering, and online (zoom): https://newnham.zoom.us/j/92544958528?pwd=YS9PcGRnbXBOcStBdStNb3E0SHN1UT09