Robotic Learning

Academic Contact: Matteo Leonetti
Academic Staff: Dr Yanlong Huang, Dr Chengxu Zhou, Dr Mehmet Dogar, Dr Andy Bulpitt, Dr Syed Ali Raza Zaidi, Professor Anthony Cohn, Professor David Hogg, Professor Mark Mon-Williams, Professor Richard Romano, Professor Shane Xie

Our research in robot learning encompasses a large variety of topics from perceptions to behaviours.

Reinforcement learning: We devise algorithms for life-long learning of skills on autonomous mobile service robots. The robots explore different behaviours by trying actions in the environment, while maximising a reward signal which gives them a measure of the quality of their actions. The service robots of the future will be required to be versatile, performing a continuously increasing number of tasks. We study life-long learning algorithms, which allow to transfer knowledge from different tasks, taking advantage of past experience when a new problem arises.

Learning the physics of object manipulation: When we humans push an object, hit a ball, or cut a slice of bread, we have predictions about the forces we need to apply and the physical effects these forces have on the objects.

We are interested in robots learning to predict such physical effects, so that they can use these predictions and plan new intelligent physical actions.

Unsupervised learning of activities: Robots must be able to operate in environments inhabited by humans, such as home and offices. For this reason, they need to understand and react correctly to human behaviour. Our service robots can observe the environment for long periods of time, and build a model of activities that are commonly performed by people. This allows the robot to anticipate what the person is about to do, and provide help if needed.