Thesis abstract:

«p»In recent years, through advances in machine learning techniques, robots have been able to perform increasingly complex tasks. This is due to the ability of learning techniques to capture and model underlying patterns in data, where models can be intractable or expensive to develop. These advances have begun to unlock the potential of robots for performing cumbersome, tedious, or dangerous jobs repeatedly without fatigue. Soon, it is hoped that automation will become more commonplace in manufacturing, agricultural, and medical industries where mistakes can be costly and the only way to increase productivity is to work longer hours. «/p» «p»Unfortunately, the tasks from these industries often require modelling, programming, and tuning from a robotics expert, which can be an expensive and time-consuming process. However, non-experts can teach a robot how to perform simple tasks through a technique called learning from demonstration. This process would allow laypeople to effectively control the outcome of a robot’s actions, circumventing the need for roboticists. However, it is likely that robots will have to perform a task with only a few demonstrated samples, which is a far cry from the hundreds or thousands of samples typically needed to train robotics models. Additionally, if a robot wants to explore a state-action space in search of a more optimal path to follow, this could cause harm to surrounding humans, the environment, or the robot itself, depending on the circumstances.«/p» «p»My research is focused on resolving some of these issues to improve automation of such industries.«/p»