ABSTRACT
Modern robotic platforms provide immense amounts of data that need to be quickly integrated into probabilistic models representing the environment autonomous systems operate in. In this talk I will show statistical machine learning methods for online spatial and spatial-temporal learning that are able to integrate information from heterogeneous sources, scaling gracefully to very large datasets. With these representations, I will demonstrate how Bayesian optimisation and the principle of modelling uncertainty can be used to mitigate risks in decision making for motion planning and policy search in dynamic environments.