Robotic Versus Human Coaches for Active Aging: An Automated Social Presence Perspective

Applications of SARs in elderly care constitute a trending research topic. Several studies with older adults test their acceptability and possible range of applications. In Iwamura et al.’s work [41], seniors used two robots as shopping partners, and the findings suggest that the robots’ social skills (e.g., conversation) improve people’s intentions to use them. McColl and Nejat [50] present an exploratory study at an elderly care facility to investigate user engagement and compliance during mealtime interactions with a robot. Participants were engaged and compliant with the robot’s instructions and also perceived the robot as enjoyable.

Other studies have investigated the evolution of seniors’ perceptions of SARs over a longer period of time. In a study of which factors determine long-term user acceptance of social robots [34], the evidence reveals that hedonic factors gain the most attention, but utilitarian factors are a fundamental prerequisite of long-term interactions (i.e., the robot must have a clear purpose). A more recent contribution investigates older adults with dementia at residential aged care facilities over four years [46]. After each trial, the robot’s services improved, reflecting staff and resident feedback, leading to statistically significant increases in engagement and robot acceptability.

Robotic coaches are platforms that engage, monitor, and support physical exercise activities, through verbal and non-verbal communication, demonstration of exercises, and real-time corrective or motivational feedback. An early study [25] explored interaction enjoyment and perceived utility, with a comparison of two different approaches. The first relied on a relational robot that praised, addressed users by name, and expressed humor and empathy. The second used a non-relational robot that just provided scores and help as needed. Users regarded the relational robot as a better companion and exercise coach than the non-relational one. Robotic coaches thus need to provide motivational support to users during the exercise. In turn, users perceive relational robots as more intelligent and more helpful. Feasibility tests of a robot coach architecture with older adults also indicate promising acceptability results [24]. This architecture includes a rule-based decision process but also can learn users’ preferences, estimate their affective state, and recall past interactions. Users’ perceptions of their interaction with the robot thus improved, relative to their initial expectations. Schneider and Kümmert [66] also investigate the differences in motivation created by a robot that solely instructs users (instructor role) and one that exercises alongside the user (companion role); people appear more motivated by the robot companion role. Finally, in a robot coaching scenario [33], an elderly user study cites engagement and enjoyment while exercising but also reveals some difficulties, such as hearing instructions, focusing on the physical aspects of the robot while ignoring verbal instructions, and confusion associated with performing sequences of gestures.

Extant studies also compare the performances of human and robotic actors in a variety of contexts. To summarize these findings, we start with studies in which participants have no direct interaction with the human/robotic actor, then continue on to studies in laboratory settings, and finally end our analysis in this section with findings from real-world field experiments.

The first group of studies relies on online surveys that present participants with written [75] or video recorded [39] scenarios, followed by a set of questions designed to evaluate their perceptions of human versus robotic actors. The foci of the studies varies, from moral dilemmas to the communication ability of the human or robot agent. For example, one online study [75] reports that participants assign higher moral obligations to human than robotic actors, such that they believe it is morally wrong for a hypothetical human actor to choose to act and save four, while sacrificing one, coal miner, but they expect the robot to make this choice when faced with the same moral dilemma. Another study [39] compares human and robotic doctors’ ability to inform patients about their health conditions, following a standardized script. Participants evaluate the robot better at communicating bad news, though these researchers predict that human social presence might invert the results in a real-life scenario. Although the design of these experiments allow for replication and comparability between humans and robots, they have limited external validity, because they refer to specific, artificial scenarios, and participants have no contact with actors or no immersion in the scenario. Such a limitation is especially important in critical scenarios [39], such as when participants do not experience a real health diagnosis.

Another group of studies relies on laboratory-based methodologies that enable participant–agent interactions. These studies generally report no significant differences in the behaviors of participants in the presence of robotic versus human agents. For example, in a shared attention memorization task with a human or robot agent [47], the participant and agent point at images and look at them together. Later, participants reviewed another set of pictures and had to identify which they had seen before. Recognition performance was better when the participant initiated the pointing behavior than when the collaborator did so, but it did not matter if that collaborator was a human or a robot. In a turn-based variant of the Tower of Hanoi experiment, conducted with a human or robot collaborator [44], participants performed significantly better with a collaborator than when playing individually, but again, there were no significant differences between the human and robot collaborator conditions. Because the collaborator always made the best move, the lack of performance differences could be because the game was too simple. Although these studies suggest that participants benefit equally from collaborating with other humans or robots, one contribution [44] assigns an advantage to humans, in terms of perceived social presence and perceived credibility. These designs are replicable and controlled, and they rely on participant–collaborator interactions. However, the laboratory setting might limit internal and external validity.

Finally, some studies use realistic field experiments to compare human and robotic teachers [28, 45]. For example, when teachers gave a tutorial on prime numbers, no significant differences emerged in the student evaluations of human or robotic teachers [45]. The robot in this study was fully autonomous but could not adapt its gestures to events in the classroom or engage in mutual gazes with students. In a similar study of programming principles [28], students between 14 and 17 years of age performed better with the robotic teacher, whereas those 10–13 years of age performed better with a human teacher. Younger students (6–9 years) showed no significant performance differences. Although they do not suffer the limitations of the previously mentioned studies, strong novelty effects are present in these studies, especially [28]. Beyond the novelty of interacting with a robot, the participants were visiting students, unfamiliar with the university environment.

This literature review reveals that prior research has focused on different domains of social interaction, including moral expectations, diagnosis communication, shared attention benefits, gaming performance effects with a collaborator, and teaching. They also provide valuable insights into the issues and methodologies that are required to achieve successful experiments involving human and robot actors. First, in particular, they reveal that people evaluate humans and robots differently, even if the robot is perceived to be intelligent [75]. Second, lab experiments allow for more control over the stimuli, but they might not be valid in real-life scenarios [39]. Third, some elements of these methodologies are standard and necessary for interaction studies, to increase their internal validity. The human actor must be trained to follow the same script and move similarly to the robot; the task should match the robots’ skills to reduce possible biases. However, literature is still sparse in this domain, and to the best of our knowledge, no study compares older adults’ perceptions of exergames when they are coached by either a human or a robot. The following section substantiates the importance and relevance of such a study.

According to Fiske et al. [31], when interacting with members of the same species, humans need to determine whether the other is friend or foe (i.e., warmth dimension), and whether it is capable of acting on good or ill intentions (i.e., competence dimension). Although robots are not members of the same species, their anthropomorphism [53] and increasing social dexterity make them resemble humans in many aspects. This study in particular focuses on SARs, which offer assistance through social interactions in a humanlike manner [26]. In a healthcare context, SARs are autonomous, understand social cues through facial and voice recognition, and can provide both child (e.g., autism therapy, [65]) and elderly (e.g., medication reminders, [13]) care. In an elderly care context, SARs offer services such as health monitoring and safety, encouragement to engage in rehabilitation or general health-promoting exercises, social mediation, interactions, and companionship [27]. By performing more socially engaging tasks, these robots take roles as companions, collaborators, partners, pets, or friends [21, 32]. We build on the idea that humans judge a social interactor (even a non-human actor) on warmth (i.e., friendliness, kindness, caring) and competence (i.e., efficacy, skill, confidence) dimensions, which then determines their emotional, cognitive, and behavioral reactions. Thus we address recent calls for research on the effects of automated social presence on the perceived enjoyment and usefulness of service interactions and the acceptance of new technologies [36, 72].

The robotic platform in these experiments is the Vizzy robot [51], developed by the Institute for Systems and Robotics (ISR-Lisboa/ISR). Vizzy is a 1.3-m-tall, wheeled robot with a humanoid upper torso and a friendly, marsupial-like design (Fig. 3). It has a total of 30 mechanical degrees of freedom (DOF). The two DOFs of the differential drive base allow it to navigate planar surfaces easily and plan its trajectories using well-established algorithms [59, 64]. The head can perform pan and tilt movements, and its eyes can do tilt, vergence, and version movements, for a total of five DOFs. Vizzy’s arms and torso also have 23 DOFs, such that it can perform human-like motions. Two laser scanners at the bottom of the mobile base capture planar point-clouds of the environment, providing valuable information for the robot’s safe navigation. Each of Vizzy’s eyes contains a camera, used for object/people detection and tracking. A depth camera is mounted on Vizzy’s chest, further enhancing the robot’s sensing capabilities for human movement analysis. Other components include a loudspeaker and a microphone, useful for human–robot interactions, and 12 tactile sensors on each hand [55], mounted on the finger phalanges.

Vizzy’s actuators and controllers allow it to navigate, grasp objects, perform gestures, and gaze in a biologically inspired way [1, 63]. The robot also has software that enables it to detect and follow people in the environment, using an implementation of the Aggregate Channel Features Detector [22] with appearance-based tracking using color features [29]. A set of RVIZ plug-ins allow the remote control of the robot (head and base) in Wizard-of-Oz (WoZ) experiments, in a gaming-like way, while visualizing the obstacles gathered by the laser and the people detected by the cameras. A web interface (see Fig. 4) allows a “wizard” to select utterances for the robot to speak while listening through the robot’s microphone. This approach supports faster replies and interactions, compared with explicitly typing utterances, though at the cost of speech flexibility.

To perform the exergames with seniors, we used the Portable Exergames Platform for Elderly (PEPE) [70], an augmented reality gaming platform that projects exergames on the floor, while a Kinect sensor captures the person’s movements to control the game elements. Users can play through PEPE without requiring any wearable sensor or controller, which minimizes the burden and complexity for older adults. Two touchscreens on the top of the platform provide additional information to staff members monitoring the exercise. The games currently need to be initiated and terminated by a person, but in the future, the social robot will be able to control the gameflow. The chosen game for our experiments is ExerPong [16], which requires players to control a green paddle (Fig. 5a) with their body movements. The objective is to hit the yellow ball with the green paddle, making sure it does not leave the game area. Colored boxes populate the gaming area. When touched by the yellow ball, the colored boxes are destroyed and yield points to the player. After destroying all the boxes, the player completes the level. Every time the player fails to hit the yellow ball, and it leaves the playing area, a box reappears. During the game, audiovisual stimuli give performance information to the player, in the form of a red visual feedback and success and failure sounds. The players can either control the paddle by walking sideways or with horizontal arm movements. Thus ExerPong can be played by people with different degrees of mobility.

In total, 58 elderly persons (42 female, 16 male) participated in the study; they ranged in age between 48 and 94 years (µ = 78.79, σ = 9.85). We excluded people suffering from severe physical (e.g., full physical immobility) or mental (e.g., dementia) health problems, as well those incapable of giving their consent. The target population comprised elderly persons living autonomously (e.g., in their own home) or in a nursing home who accept services from a care institution. Thus, the main inclusion criterion was that they were clients of the elderly care centers with which we partnered, which offer a wide range of tailored activities (e.g., card games, fitness, handcrafts). We randomly divided the overall sample into two experimental groups: 22 participants in the human coach group (16 female, 6 male), aged between 65 and 94 years (µ = 75.46, σ = 12.79), and 36 participants in the robotic coach group (26 female, 10 male), aged between 48 and 91 years (µ = 80.83, σ = 6.98). The human coach experiments were carried out at two locations (Location A: 10, Location B: 12 elderly participants), and the robotic coach experiments were carried out at three locations (Location C: 11, Location D: 10, Location E: 15 elderly participants).

The data collection, in collaboration with five elderly care institutions, took place during July–September 2017. Each participating elderly care institution provided a spacious activity area, an adjoining room for surveys/interviews, and an in-house psychologist. The activity area was essential, because we needed to fit the gaming platform and secure enough space for the robot to navigate throughout the area. We requested an additional room to ensure each informant’s privacy during the data collection. Finally, the in-house psychologist helped screen elderly clients, to evaluate their fit for the experiment (exclusion criteria: severe physical or mental health problems). The visits were scheduled during the regular activities time held by each care center.

As a part of the study, the exergame platform was installed in the main activities room of each care institution. During the regular activities time, when people usually participate in various activities offered by the institution (e.g., drawing, knitting classes, yoga, card or board games), the seniors were approached either by a human or the robot, depending on their assigned experimental group, and invited to try a new activity that reportedly was being considered as an addition to the existing activity offers. Because all the activities took place in a main activities room, peers could observe the exergame activity, though some remained engaged in their own activities. To keep the experimental scenarios as constant as possible, we developed a standardized set of steps for each actor (i.e., human/robot) to follow: (1) Approach an elderly person who satisfies the inclusion criteria; (2) introduce the new activity (i.e., exergames) and invite the elderly person to join the game; (3) if the elderly person accepts, escort her or him to the gaming area; (4) provide instructions for how to play the game; (5) motivate the elderly person by words of encouragement and feedback on game progression; and (6) ask the elderly person whether to continue or terminate the game. The locations of the coach during Steps 4–6 were standardized, as depicted in Fig. 5. Although the “wizards” directing the robots were present in the same room, they were described as technicians, available resolve any problems with the exergame system. Furthermore, they were standing in a non-visible area, out of the field of view of the participant during the game. The communication was standardized; both human and robot actors used the same sentence to introduce the exergames, give instructions, and motivate players. Past studies have shown that gendered voices can influence perceived trust, likeability [48], and competence [18] if the agent is a computer with no gender-specific characteristics other than the voice. People tend to give higher ratings to agents whose voice gender matches their own, and they also apply gender stereotypes. Therefore, Vizzy had a female voice to match the gender of the caregivers, who were all women. On average, the time participants spent playing the exergame was 08:07 min, with a standard deviation of 02:32 min. The minimum playing time was 1:58 min, and the maximum was 12:25 min. After the game, a member of the research team escorted the elderly participants to the separate room to conduct the survey and interview. Prior to being surveyed, all elderly informants signed an informed consent, outlining the main objectives of the research project and guaranteeing the anonymity of their responses. We consciously decided not to ask for informed consent before their interaction with Vizzy or the exergame platform, because we wanted to mimic a regular activity at the center, rather than create a research environment.

To test whether a robot (i.e., automated social presence) is comparable to humans in terms of motivating elderly people to engage in physical activity, we chose a mixed-method approach, combining quantitative and qualitative data. The main instrument was a questionnaire, augmented with probing questions to elicit more in-depth, qualitative insights.