Human Teleoperation of a Social Robot: Key Design Challenges

Abstract

The use of robots outside of factories and industrial automation has been steadily growing in recent years. While the range of functional tasks that these systems can carry out has been steadily increasing, their ability to collaborate and work alongside people has been extremely limited, and has created a barrier prohibiting their broader adoption. The development of effective social interfaces for robots remains an inherently challenging interdisciplinary problem, requiring detailed consideration of both engineering and socio-cultural factors. Rather than focusing on artificial intelligence directly, something that could accelerate adoption of these robots is social robot teleoperation, allowing remote operators to pilot robots that interact socially. This could overcome workforce limitations in customer-facing jobs in distant or hazardous environments, while also enabling those with mobility constraints or otherwise to engage socially as if in-person from the comfort of their home. Importantly, teleoperation data collected during use of the system could also inform the design of social artificial agents, accelerating progress towards automated social robots. To enable adequate social interaction, a substantial amount of research and development on teleoperation design is yet required. Task-driven, non-social robot teleoperation has received focus in human-robot interaction studies, but outside of minimally-embodied telepresence, social robot teleoperation is undeveloped in research, particularly in regards to experimental rigour and repeatability. This thesis aims to address key research issues which currently impede the development of social robot teleoperation interfaces. Through analysis of the modalities of social interaction, the task is approached along three key areas of research, with a focus on repeatable, universally-applicable methodologies. First, a framework for evaluating and analysing perceived robot affect expression is developed and trialled, enabling statistically reliable, data-driven selection of robot affect expression via teleoperation. It is shown that the proposed statistical methods yielded easily understandable graphs on a robot's expressed emotional range, providing valuable insights to developers. Second, robot speech comprehension is evaluated to determine whether operator prosody should be conveyed in a speech teleoperation interface. By employing methods from the field of audiology, it was shown that speech comprehension did not change by conveying operator prosody for participants local to it, however comprehension was negatively affected by altered prosody for those unfamiliar with the prosody of the operator. Discussion of results provides recommendations and precautions to robot designers who wish to mitigate this issue. Finally, by adapting standard situation awareness measurement techniques, a tool for objectively measuring social situation awareness was developed, and used to evaluate the benefits of spatial audio feedback during robot teleoperation. Analysis of results alongside secondary measurements show that further research may be required to refine and validate the technique, but benefits to spatial audio feedback were nonetheless observed in subjective and qualitative results. These three investigations yielded valuable, sometimes counter-intuitive insights for robot designers, while also providing robust, re-usable statistics and methodologies that can evaluate teleoperation capabilities for any social robot platform. While the path towards practicable social robot teleoperation is yet long, this thesis hopes to enable through its contributions a more reliable and expedient traversal of this path in future research.