Our work focusses on building robustly autonomous
robots and other cyber-physical systems,
capable of intelligently working with humans and
other agents, in dynamic and continually changing
The types of applications we target include:
Human-robot collaboration in domains such
as customisable manufacturing and personal robotics.
A typical use case is where the robot assists a human
domain expert, whose own understanding of the domain
is not only incredibly rich, but also hard to formalise
in ways that could be easily programmed into the robot.
So, our challenge is to make the robot capable enough
at learning and interacting, so that it can acquire models
as it goes and learn to use them to act with increasing
levels of skill.
Active sensing and predictive modelling platforms,
applied within energy and environmental systems. A representative
use case might involve the use of a mobile robot or team
of robots acting as a sensor network, coupled with
algorithms that actively learn models of an underlying physical
process (e.g., wind flows) and other dynamics (e.g., demand patterns)
in order to enable real-time decisions and process optimisation.
This requires development and creative use of
Compositional and incremental methods for model learning and
learning to act in multi-scale, dynamic environments. Significant
contributions along these lines have included:
Use of computational topology and geometric methods for unsupervised
learning and clustering, using data as varied as joint level trajectories
to optimal policies of Markov Decision Processes.
Rapid transfer and policy reuse for decision making in changing
environments, through Bayesian choice and online learning over
a space of policy fragments.
Mechanisms of extended interaction for model selection, structure
learning and coordinated action in the face of unknown unknowns. Significant
contributions in this area include:
Type-based learning methods for multi-agent interaction with little or
no prior coordination. We have developed a new model (Harsanyi-Bellman
Ad Hoc Coordination) that brings together the notions of Bayesian Nash
equilibrium and Bellman optimality, to provide optimal actions from
beliefs over a set of "policy types".
Prediction and hypothesis testing methods for interactive decision making
The following 1-minute video, prepared for the University's Research in a Nutshell project,
broadly outlines our research agenda and perspectives. See also the associated description:
Major funding for our research work has come from the following sources (grants are listed in reverse chronological order):
DARPA Grant, Explainable AI Program: Common Ground Learning and Explanation (COGLE)