Teaching robots where to drive using real world data

Thesis

In this work we teach a robot where to drive from measured, real world data. In Chapter 1 we measure energy usage on a pedestrian robot and use high-level, publically available maps to predict energy con- sumption on unobserved routes with a heteroscedastic Gaussian Process model. These energy predictions can be used to make route-planning decisions in energy constrained situations, for example by learning to avoid areas with rough surfaces or that are likely to be crowded. Our method performs well, particularly in its ability to capture uncertainty in the data, however it is reliant on well annotated maps.

In Chapter 2 we consider the case where well annotated maps and good prior information about an environment are not available. Again we use measured, real world data by introducing vertical variance in a laser pointcloud as a proxy for drivability which can be labelled automatically in an online setting. We propose a method for using this online stream of data in a computationally efficient manner to learn where to drive. We then test this method against models trained in other environments and an existing method which utilises the most recent data.

Authors: Bartlett, O

• DOI: 10.5287/ora-or5n42b02
• Type of award: MSc by Research
• Level of award: Masters
• Awarding institution: University of Oxford