Foundation Models for Industrial Control Applications - Facilitating Transfer and adaptation to new contexts

Description of the Project

Recent research has demonstrated that a learnt model of system dynamics enables control policies to be optimised using synthetic data generated by the model itself. Specifically, such a "world" model is trained on observations of a system's state over time and learns to predict the state evolution of a system when conditioned on different control actions.

The aim of this project is to create versatile, action-conditioned world models of systems and environments, which are able to efficiently and effectively adapt to novel contexts such as changes in the environment or the system itself.

To achieve this aim we will train world models given data from robotics, industrial and/or environmental domains. The world models will be of sufficient quality to enable the learning of control policies via reinforcement learning from data generated by the world model itself. Once a suitable base model has been trained, we will explore representations, architectures and mechanisms for efficient transfer to novel but related domains such as intervened environments or adjacent product categories.

We expect a core challenge of this project to lie in the learning of representations suitable to capturing the various state and action spaces as well as observation modalities of different systems in a unified way. To tackle this challenge, a core aspect of this project will be investigating latent representations that embed the dynamics of differing systems to a shared latent space.

The generalising representations should enable both transfer learning to new systems as well as efficient offline to online adaptation.

Novelty learning efficient and versatile World Models for system identification in the context of Industrial Process Control is novel, data-driven direction when it comes to predictive maintenance are transfer of complex systems that are difficult to model otherwise.

EPSRC Alignment This project falls within the EPSRC Research Areas Engineering, Information and Communication Technologies and Manufacturing the Future. Collaborators The project is funded via an iCASE award to Siemens, who will be collaborating on this project.