The future of [socially assistive] robots need not be autonomous
Why socially assistive robots ought to be teleoperated, and how that changes the way we think about them
This post is a summary of “Practical, Ethical, and Overlooked: Teleoperated Socially Assistive Robots in the Quest for Autonomy” which was co-authored by myself, Terran Mott, and Tom Williams and published at alt.HRI 2022.
When most people think of robots, they often think of a future where fully autonomous robots are embedded in society; serving humans without posing a threat. These robots would support people through physical assistance, by moving objects in space, or through social assistance, by talking to people. Robots that support people through social interaction are called Socially Assistive Robots [1].
Socially Assistive Robots are not science fiction. They exist today in many forms and are used in a variety of use cases. For example, the Misty robot, which has a small body and can move via wheels, has been used in classrooms to help teachers engage with children. The Paro robot, which is designed to look like a robot seal, has been used to support elderly adults with mild cognitive impairment and dementia. The Furhat robot, which has a human-like face, has been shown in customer support and education applications. These are just a few examples of how robots can already be used in practice.
While these robots clearly differ in how they look, they can also differ in how they function; are they controlled or fully autonomous. While we may think of robots as either being controlled by a person (teleoperated) or autonomous, researchers have suggested that there are many levels of autonomy (LoA) for a robot [2]. They also provide guidelines for developers and other researchers to determine the appropriate level of autonomy for a robot based on the task it is expected to perform. The first step of these guidelines is to consider three dimensions of the robot’s task:
- Task criticality: how important is it that the task is performed correctly? Or how dire are the consequences of something going wrong?
- Task accountability: how important is it that responsibility for the task’s output is clearly assigned?
- Environment complexity: how complicated is the environment in which the task is performed?
The guidelines then suggest that
- the more critical the task → the more human control should be provided.
- the higher the need for accountability of the task → the more human control should be provided.
- the more complicated the task → the more human supervision is needed and more autonomous capabilities are needed if the robot is chosen to be autonomous.
These suggestions are just the first step, and developers and researchers should continue along these guidelines to identify the specific level of autonomy that is appropriate for their robot’s task.
We used this first step as a heuristic for the range of level of autonomy that we would expect to see for Socially Assistive Robots. We wanted to see if researchers of Socially Assistive Robots were following these guidelines or if they were working towards something else. This meant that we had to answer the following questions:
- When Socially Assistive Robots are researched, in what kinds of applications are they used?
- In those applications, are the robots fully autonomous? Why and why not?
- In those applications, what do guidelines for robots’ level of autonomy recommend?
We looked at the last 5 years of publications in the two leading research venues on human-robot interaction: the International Conference on Human-Robot Interaction (HRI) and Transactions on Human-Robot Interaction (T-HRI). Here’s what we found:
Patterns in the research
The three most common applications of Socially Assistive Robots in the research are to support elderly people, in education, or in therapy.
Across these domains, we found a plurality of the papers in the research discussed fully autonomous robots. Many others were often not explicit about their choice in robot autonomy; choosing to focus on other aspects of the robot.
When authors take about the future of their robots, most of them were motivated by a future with autonomous robots. Most authors in general were not explicit about the future Level of Autonomy motivating their work or were ambivalent about it. This can make it difficult to understand the goal behind the research and future directions for the work.
Importantly, the vast majority of papers did not include the rationale for their choice of the robots’ level of autonomy. This lack of explicit description of their rationale, coupled with no explicit discussion of the future motivating the work, makes it difficult to understand the future that Socially Assistive Robot researchers as a whole are working towards.
These researchers are largely the ones defining the future of Socially Assistive Robots. We encourage them to be explicit about their choice in robot autonomy, their rationale for that choice, and the future they envision that is motivating that work. By sharing these details, they can encourage other researchers to focus on work that might benefit the community as a whole.
Importantly, researchers may choose different levels of autonomy for their robots for vastly different reasons. Researchers often choose to design experiments that rely on an experimenter using a teleoperation interface, called a Wizard of Oz setup, to minimize the burden of developing an autonomous system. However, researchers may choose to use autonomous systems for experiments to alleviate the burden of developing teleoperation interfaces.
What is the recommended Level of Autonomy for Socially Assistive Robots?
To determine the recommended Level of Autonomy for Socially Assistive Robots, we examine the three most common socially assistive domains’ task accountability, task criticality, and environment complexity.
When assisting the Elderly, robot tasks can vary from highly critical for cognitive impairment-related tasks, to low criticality when focused on companionship. Accountability also highly varies depending on the task choice. Since robots working with the elderly are often used in a static environment, environment complexity is fairly low. Level of Autonomy selection guidelines would suggest a wide range of autonomy for robots socially assisting the elderly; from fully teleoperated to fully autonomous depending on tasks.
When assisting in education, task criticality can vary based on the criticality of the content delivered and its impact on students. Students are often young children who are a vulnerable population and require high task accountability. Educational settings can vary from a highly complex and noisy classroom to a fairly static and simple one-on-one tutoring session. Level of Autonomy selection guidelines would suggest a wide range of autonomy for robots socially assisting in education; from fully teleoperated to fully autonomous depending on tasks.
When assisting in therapy, task criticality, and accountability are high since therapy is a sensitive task that often involves vulnerable individuals or individuals in a vulnerable state. The environment complexity of therapy can vary from low complexity in an office environment to high complexity in a fitness center. Level of Autonomy selection guidelines would suggest a narrow range of autonomy for robots socially assisting in therapy; mostly fully teleoperated.
Our analysis suggests that current research on Socially Assistive Robots is not in alignment with Level of Autonomy selection guidelines’ recommendations for these domains.
How are these robots developed and deployed?
We found the chosen Level of Autonomy also influences how these robots are developed and deployed.
When the robot is fully autonomous, developers partner with experts like therapists or teachers to code the robot autonomy. They usually have to limit the scope of the robot’s capabilities to allow it to be fully autonomous. When the robot is deployed, it engages with the person it is assisting directly and we can say that the individual receiving assistance is the robot user.
When the robot is teleoperated, it can either be teleoperated by a caregiver, like a therapist or teacher, or teleoperated by the person actually receiving assistance. When the robot is teleoperated by a caregiver, developers must create a teleoperation interface for the caregiver to use. The robot user is still the person receiving assistance but there is now a teleoperation interface and its user is the caregiver. When the robot is teleoperated by the assisted individual, developers must again create a teleoperation interface, now for the person receiving assistance to use. The robot’s user is now other people that the robot interacts with on behalf of the individual receiving assistance. And that individual is now the teleoperation interface user. In both these cases, the robot’s capabilities are limited by the teleoperation interface and the ability of the person operating the robot to do so effectively.
Understanding how these three scenarios change the users involved and introduce limitations is crucial to effectively designing and evaluating robots.
In summary, we found that recent research on Socially Assistive Robots focuses on autonomous robots whereas Level of Autonomy selection guidelines would suggest these robots include more teleoperation. Researchers often do not include a rationale for their choice in robot Level of Autonomy or describe a future motivating their work. The choice in Level of Autonomy changes the users, and design objectives of these solutions and should be carefully accounted for. Importantly, this choice warrants careful consideration due to its ethical ramifications; choosing autonomous robots takes power away from caregivers who are the experts equipped to handle these complex assistive tasks.
[1] Feil-Seifer, David, and Maja J. Mataric. “Defining socially assistive robotics.” 9th International Conference on Rehabilitation Robotics, 2005. ICORR 2005.. IEEE, 2005.
[2] Beer, Jenay M., Arthur D. Fisk, and Wendy A. Rogers. “Toward a framework for levels of robot autonomy in human-robot interaction.” Journal of human-robot interaction 3.2 (2014): 74.
