Abstract
Robots that can perform new skills by learning from human provided demonstrations, rather than being explicitly programmed with lines of code, are often seen as a pathway to the widespread use of robotics in industry and beyond. People interacting with robot learners to provide such demonstrations face a challenge, namely the communication barrier between humans and robots, leading to misinterpretation of intent and abilities on both sides of the inter-action.This thesis explores the problem of how to improve the quality of systems which learn from demonstration. This problem is approached from the view-point that gains can be made by holistically improving both the teacher, as well as the learner. In taking this dual approach, a series of contributions are presented that focus on methods, models, and metrics for improving learning from demonstration.
First, a novel learning algorithm is presented that better extrapolates tasks from limited demonstration data and thus can significantly reduce teaching effort. A model for the overall learning from demonstration process is then presented, which better accounts for the human influence on a learning system in which the person is an integral part of the process. Finally, the issue of human-robot misunderstanding is addressed through a task-agnostic evaluation method that can be adapted to provide feedback and guidance to teachers, based on the learner’s current abilities.
Each contribution is supported with extensive experimental results on both simulated and real-world data from human subject trials. It is shown that the proposed solutions offer a significant improvement on system performance, even when robots are deployed by people with no understanding of robotics or machine learning.
| Date of Award | 1 Feb 2021 |
|---|---|
| Original language | English |
| Awarding Institution |
|
| Supervisor | Matthew Howard (Supervisor) |