Fairness and Trust in Human-Robot Collaboration
Abstract:
This MSc project forms the first part of a PhD project proposed jointly by Swansea University Computational Foundry and QinetiQ. It takes an in-depth look at the aims and motivations outlined in the proposal and seeks to identify how best to address them, in a way that is in line with the motivations of QinetiQ and the Computational Foundry; incorporates a human-centred approach to responsible innovation; and also makes a valuable contribution to the field of Human-Robot Collaboration.
The primary focus of the thesis is a review of the literature regarding trust and fairness in the field of Human-Robot Collaboration, as well as methods of task allocation. Trust was found to be essential for effective collaboration between human and robot agents, with the level of trust needing to be matched to the capability of the robot agent, since both over-trust and under-trust can have an adverse effect on effectiveness and safety. It was also found that improvements in fairness will also improve trust. In addition, studies in the field of Human-Robot Trust are extensive, and many different methods of task allocation have already been studied. However, relatively little work has been done with regards to fairness of task allocation, so this is intended to be the main area of focus for the project.
A discussion of the review findings concluded that the amount of effort required to perform a task can be used as a metric for assessing fairness when allocating tasks, and this can be incorporated into a Q-learning algorithm which uses equality of effort between the agents as it’s reward function, and so puts fairness at heart of the learning process.
A pair matching game is investigated as an initial implementation of the algorithm in a practical application, and also a stand-alone algorithm intended to determine whether the amount of effort required to perform a task is an effective metric for assessing fairness of task allocation.
Although the proof of concept was not fully implemented, the preliminary results suggest that the strategy is worth pursuing. Future work will include completion of the algorithm, and its implementation in a practical, real-world application.