Robotic Data Capture

点击翻译按钮获取中文摘要

This proposal addresses the development of robotic systems that can move about in our world and record events and phenomena of interest to us: the robotic photo-documentarian.  The objective is to develop systems that combine the attributes of photo-journalist, wedding photographer, scientific explorer, and court stenographer.  The key to all these problems is an awareness of which phenomena are of interest to human obervers, and then getting to the right places in space and time to make the requisite observations. One key way to both learning how to do that, and making the right observations is to track and follow human guides who have the same interests.  This can be done be decomposing this task into a few key algorithmic components, as described below. The proposal focusses on basic algorithmic issues associated with allowing robotic systems to document our activities while working in close proximity to humans and responding in a socially-adaptive manner.  This includes known what to observe, when to make those observations, how to do so in a socially acceptable manner, and how to cope with obstacles and moving targets by predicting their behaviour, and how to adapt to changes in lighting and physical impediments.  The span of the work begins with theoretical development, then simulation, and finally deployment on a range of real robotic vehicles including underwater vehicles already under development by the applicant. The techniques to be used span sensor-based optimization, algorithmic game theory, deep-reinforcement learning and inverse reinforcement learning.  The physical deployment targets include a Husky terrestrial vehicle, a family of robotic catamarans (boats), and the Aqua underwater/amphibious vehicle all of which are already in use by the applicant.