Safe and Robust Robot Learning in Unknown Environments
Learning can be used to improve the performance of a robotic system in a complex environment. However, providing safety guarantees during the learning process is one of the key challenges that prevents these algorithms from being applied to real world applications. We aim to address this by combining robust and predictive control theory with probabilistic modelling techniques such as Gaussian Process Regression. In addition, some of our recent work has focused on enabling these systems adapt safely to new conditions that arise as a natural part of deploying robots in realistic outdoor environments for long periods of time. Our algorithms have been evaluated on ground and aerial vehicles as well as on mobile manipulators for human-robot collaboration.