Social Robots for Sleep Health: A Scoping Review

We defined our search string after an initial, limited review of the titles and abstracts of relevant literature to identify keywords and phrases. We initially built our search string with the following terms: “social robots” and “sleep”. We further refined our search string and selected relevant databases in consultation with two academic librarians—one with an expertise in computer science and engineering literature and the other with an expertise in medical literature; the selected databases included IEEE Xplore, PubMed, the ACM Digital Library, Embase, Scopus, Ei Compendex, and APA PsycINFO.

We utilized the following search string to retrieve relevant articles from the aforementioned databases: ((companion OR assistive OR social OR service OR caregiver) AND (robot OR robot*)) AND (((sleep OR sleep*) OR ((“sleep initiation and maintenance disorder” OR (sleep AND initiation AND maintenance AND disorders) OR insomnia)) OR (apnea OR apnoea))).

The search of the selected databases on March 9th, 2023 using the above search string yielded 641 results, which were then uploaded to Covidence, a systematic review software that helps organize and screen articles. The details of our study screening process are described in Fig 1.

Two research team members screened the titles and abstracts of our search results based on the following inclusion and exclusion criteria:
Inclusion criteria:

1. 1.

   The study is published in a peer-reviewed journal or conference proceeding.

2. 2.

   The study is written in English.

3. 3.

   A full-text version of the study is available.

4. 4.

   The publication assesses aspects of sleep health outcomes using social robots.

Exclusion criteria:

1. 1.

   The study does not report on sleep health interventions or diagnoses.

2. 2.

   The publication solely describes robot development.

3. 3.

   The publication focuses on a surgical intervention.

4. 4.

   The publication is a review of existing literature.

5. 5.

   The study solely describes sleep intervention protocol or recruitment strategies.

The articles remaining after this screening were retrieved for a full-text review; the same two team members read and discussed the appropriateness of each article to determine their inclusion in this scoping review. Any disagreements were resolved via further discussion of the criteria until a consensus was reached. The research team additionally conducted a reference check among the publications selected for full-text review to identify further articles of relevance. We note that no publication time limits were included in the above criteria, as the concept of “social robots” has only emerged in recent years.

For each of the publications that passed the full-text screening, we extracted their 1) study characteristics, 2) research methodologies, and 3) research findings. Study characteristics include basic information such as the authors’ names, the countries in which the studies were conducted, year of publication, and purpose of the study, while research methodologies focus on study design, setting, sample, intervention, duration, and outcome measurements. Finally, we summarized the main findings of the studies that aligned with the purpose of our review.

Full-text data extractions were conducted by three researchers—one with a background in engineering and robotics, the other two with backgrounds in health science. Each article underwent data extraction performed by one of the three researchers; the extraction was then double-checked by another member of the research team to ensure that the extracted contents were accurate. The team met weekly to discuss their progress and resolve any disagreements regarding data extraction via mutual consensus.

Each publication was independently appraised by two of the three researchers using the Joanna Briggs Institute critical appraisal tools [26]; specifically, the Checklists for Randomized Controlled Trials and Quasi-Experimental Studies were used to determine study quality. The three research team members then met and discussed to reach an agreement on the final quality appraisal. Please refer to Tables 5 and 6 in the Supporting Materials section for our quality appraisal outcomes.

Additionally, we used the ROBIS tool to assess the risk of bias in our scoping review [27]. Three of this review’s authors collectively answered the signalling questions and resolved disagreements through discussion and mutual consensus. The tool ultimately yielded a low risk of bias in our assessment. Table 4 in the Supporting Materials section presents the outcome of our ROBIS assessment and the rationales for any potential concerns.

Comparing the social robots used in the reviewed studies helps to identify attributes that are critical in effectively driving sleep health interventions. A total of six different social robots were used in the ten reviewed studies (Figure 2). Four studies used PARO, a robotic baby harp seal [38]; one study [29] used the PaPeRo (“Partner-type-Personal-Robot”) robot [39]; two studies [31, 32] used a humanoid robot called Sota [40]; and one study [36] employed the Kabochan Nodding Communication Robot [41]. All of the aforementioned social robots were developed by Japanese firms and/or academic institutions. Nao, developed by Aldebaran Robotics, was another humanoid robot used in a study [37]. Finally, one study [35] included in this review did not provide a detailed enough description of the robot it used.

Despite primarily executing tasks through social rather than physical interactions, a social robot’s physical embodiment greatly impacts its performance and how others perceive it [15]. Social robots are designed to faciliate social and empathic interactions, yet their physical attributes may differ greatly in terms of degree of human likeness—yet the shared physical characteristics (e.g., adorable eyes, pleasing shape, etc.) between the animal-like PARO, the childlike Kabochan, and the more robotic Sota, Nao and PaPeRo serve as evidence of the significance of a robot’s embodiment in driving positive interactions. PARO resembles a baby harp seal and is designed to be cute and calming with a soft, white, synthetic fur exterior [38]. Kabochan sports a similarly soft exterior while resembling a 3-year-old boy in form, voice, and movement [41]. PaPeRo, Nao and Sota, on the other hand, have hard-shell bodies but are equally adorable in shape. All of these robots also have charming eyes to further their friendly demeanor; the eyes of PaPeRo, Nao and Sota additionally feature embedded LED lights to augment their social interactions. These robots’ pleasant, non-threatening embodiments help to drive positive interactions, while their animal-like or non-realistic humanoid designs temper user expectations; this is critical given the sensitive application domain and current limited capabilities of even state-of-the-art artificial intelligence [42].

Sensors help robots perceive the world and are therefore crucial in both powering social interactions and effectively driving interventions. The robots used in the evaluated studies were all equipped with cameras or light sensors and microphones, allowing them to “see,” “hear,” and respond appropriately. PARO, Nao, and Kabochan also have an array of tactile sensors installed across their bodies to detect and react to users’ touches. PARO has additional temperature sensors, similar to PaPeRo’s temperature and humidity sensors; although it is unclear whether these capabilities were utilized in the reviewed experiments, these environmental sensors hold the potential to modify a social robot’s intervention based on the state of the physical environment. Sota may also be linked with smart sensors (e.g., smart watches), thus allowing researchers to potentially personalize interactions and interventions based on factors such as heart rate and body temperature. Overall, the sensors available to a social robot are fundamental in facilitating dynamic interactions and intervention execution.

The user experience provided by a social robot is a critical factor in the effectiveness of its interventions and directly affects its ability to drive engaging social interactions. A common social interaction feature across the robots in the reviewed studies was face tracking, which in some cases also allowed for face recognition. Face tracking allows for the simulation of eye contact, which significantly improves interaction quality by communicating attentiveness and indicating interpersonal interest [43]. Furthermore, all of the robots additionally expressed themselves to users via the sound modality, with PARO being limited to animal-like sounds and the more humanoid robots like PaPeRo using natural-language-based speech interactions with varying degrees of autonomy.

Social robots also tailor their physical motions relative to their sensor inputs to produce lively, friendly interactions. The goal of personalizing a robot’s interactions to its user’s actions is the development of a positive relationship and consequent enhancement of the intervention experience. PARO leverages its tactile sensors to power interaction through movement in its tail, flippers, and eyes during petting; it also responds to sounds, including its name and any words its user frequently repeats. Similarly, Kabochan can verbally respond—as well as sing and nod—in response to its user’s touch and spoken words. PaPeRo interacts by blinking the LEDs on its cheeks, mouth, and ears to express rich emotions; it also executes face tracking in response to human words. Similarly, Sota interacts using a cloud-based speech dialogue system and is able to track and remember specific faces. Nao’s interactions are similarly dynamic, powered by its speech dialogue and two 2D camera systems.

Social robots are equipped with sensors designed primarily to enable social interactions; therefore in order to track and/or support users’ sleep health, additional external sensors were used in the reviewed studies. See Table 3 for a summary of the sleep variables evaluated and the sensors used to measure them.

Wrist- and arm-based actigraphy sensors were the most widely utilized, with four of the nine included studies relying on them to collect sleep statistics including sleep time, sleep efficiency, and wake after sleep onset (WASO) by utilizing accelerometer data. Despite being designed as less intrusive relative to electroencephalography (EEG) sensors, these on-body sensors were reported to not have been well-tolerated by certain users [34, 33]. Furthermore, actigraphy methods have been criticized for overestimating sleep, thus undermining their reliability—especially for individuals with chronic conditions [44].

Given the sensitive user groups these investigations have targeted so far, some studies also elected to use completely off-body sensors. For instance, a series of infrared sensors was deployed in participants’ living environments to collect their wake-up times, bedtimes, and sleep duration [29]. Similarly, sheet-shaped body vibrometers (SBV) allow for constant, noninvasive monitoring of care recipients’ sleep patterns while in bed [31]; in one reviewed study, an SBV informed a social robot about the nighttime awakenings of study participants in a nursing home, allowing the robot to intervene when necessary [31]. Although SBVs and infrared sensor networks are plausible off-body alternatives for collecting sleep data, they require considerable effort to deploy and are not yet commercially available like on-body actigraphy sensors are.

Three studies did not use any sleep sensors at all, instead relying on observations from care providers to estimate subjects’ sleep habits [35] or utilizing questionnaires to collect data on sleep outcomes, such as nocturnal sleeping hours and difficulty in initiating sleep [36, 37]. Although these survey-based methods do not provide the same degree of fidelity and granularity in their measurements relative to on- and off-body sensors, they do provide insights into the more user-centered and subjective aspects of sleep health.

In this section, we describe the design details of the reviewed robot-driven interventions to provide key insights toward understanding their effectiveness and interpreting their results. All of the studies included in this review measured the change in certain sleep variables as a result of their designed interventions either as a primary or secondary focus. Tables 1, 2, and 3 summarize the key intervention details of the reviewed studies. Below, we first present the results of the interventions and then discuss their limitations.

The reviewed studies did not clearly establish how or to what degree robot-driven interventions impact sleep health; however, they did provide several promising and statistically significant indicators for certain key sleep parameters, highlighting the potential for social robots’ future in this domain. Table 3 details the sleep parameters investigated by the reviewed studies and indicates whether a statistically significant result was reported for each parameter.

Most of the reviewed studies acknowledged the limitations of their experiments to a certain extent; limitations reported across the studies included a limited number of participants [36, 29, 28], restricted participant demographics [32, 29], inadequate intervention duration [36, 34], lack of sustained impact post-intervention [30], uncertainty in collected data [30, 33, 28, 37], insufficient analysis of sleep data [29], inadequacies in robot interactions [35, 32, 31, 37], the absence of a randomized control trial design [35], and high risk of the novelty effect [28]. It should be noted that the shortcomings reported above are common across many studies in this review and not only the ones wherein they were explicitly acknowledged.

The most commonly unacknowledged limitations were inadequate descriptions of the finer details of the explored robot-driven interventions and minimal discussion on the nature of the interactions between the subjects and the robots, rendering it very difficult to fully contextualize the interventions and the robots’ roles within them and thus complicating the assessment of the robot’s true impact on its user’s sleep health. On a similar note, several studies failed to provide detailed descriptions of the sleep scales they used or only used investigator-developed sleep measures; the absence of such key details, along with a lack of standardized sleep measures, hinders a proper comparison of the results of the different studies. It should also be noted that long-term post-intervention effects were not studied by the reviewed publications in any meaningful manner. There was also a general lack of acknowledgement or discussion on the potential privacy issues that arise when deploying digital agents in personal living spaces. Moreover, the studies were limited to only Japanese, Northern European, and Oceanic contexts, which may pose an issue as cultural implications may affect human-robot interactions and thus the efficacy of any of these robot-driven sleep health interventions.

As previously mentioned, many of the analyzed studies were limited to small sample sizes and short intervention durations, adding uncertainty to their reported results and additionally impeding the development of a robust understanding of the longer-term impacts of social-robot-driven interventions on improving sleep health. Long-term randomized controlled trials should be attempted to neutralize the effects of factors such as the novelty effect and to help establish clinical results.

A major factor contributing to the small sample sizes in studies involving social robots is the cost of these robots in general—on average, the social robots used in the reviewed studies cost around $6,000 USD each. Furthermore, commercially available social robots are not very customizable for research purposes; on the other hand, developing an in-house social robot may reduce costs and allow it to be tailored to the requirements of a given experiment—but also requires a significant amount of financial investment, engineering work, and human expertise. Cost concerns drastically reduce the number of research groups that are able to conduct large-scale studies involving social robots; therefore there is a critical need for a flexible yet affordable social robot platform to drive further research in this area.

Our analysis also indicated that the social robots used in the reviewed studies were not specifically designed for sleep health interventions; currently, researchers must adapt their interventions to the capabilities of social robots designed for general social interactions, a fact reflected in the general designs of the reviewed interventions. The social robots presently available for use are viable for exploratory studies, but as this field matures and targets more specific sleep mechanisms, the functional requirements for these social robots will shift. Either the robots must be designed more specifically and with requirements stemming from a deeper understanding of the sleep health field or they must be modular or customizable enough to be easily converted to meet such requirements. The additional research requirements expected as this field matures further underscore the need for a flexible yet affordable social robot platform to foster more robust research.

The reviewed studies largely provided limited descriptions of the roles the robots played within their interventions and failed to sufficiently characterize the nature of the robots’ interactions with their human users, severely inhibiting accurate analysis and assessment of the robots’ actual impact on experimental outcomes. This deters the establishment of any clinical relevance of social-robot-driven sleep health interventions and hinders further work in this field as other researchers do not have a precise or definitive foundation to build upon. Future studies should elaborate on the details of their robots’ intervention roles and the nature of their interactions to allow for a more accurate appraisal of their impact on any reported outcomes.

On a similar note, the reviewed studies failed to report the individual sleep health backgrounds (i.e., existing sleep disturbances, self-reported poor sleep quality, insomnia, etc.) of their participants. They also did not report any information on participants’ relevant personal histories, such as their typical diet, exercise regimen, physical activity level, medication use, or alternative interventions previously or concurrently tried. Although most studies elaborated on the sleep health challenges their targeted sub-populations are generally known to face, the absence of individual participants’ sleep and personal histories deprives the reported outcomes of valuable context; future research should collect and present these crucial data-points to underscore their results.

Finally, most of the reviewed studies only provided shallow justification for their intervention design choices; specifically, there was a lack of robust grounding of intervention plans and details in any previous work—especially within any sleep science literature. This is somewhat acceptable given the exploratory nature of the studies conducted so far; however, future work should attempt to rationalize intervention designs with relevant references to sleep science research in order to validate their experimental hypotheses and improve study outcomes.

A key justification for deploying social robots in the healthcare domain is their potential to boost user motivation to maximize compliance with interventions through effective interactions. Good communication is critical for maintaining motivation [57] and is a fundamental aspect of effective human-robot interaction (HRI). Good communication becomes even more potent in the context of long-term deployments where the novelty effect eventually wanes. Negative effects of sub-optimal human-robot interactions were reported in the reviewed studies as a result of a robot’s interactivity being incompatible with a user’s physical capabilities [35]; a lack of smooth, rich, context-appropriate conversations [32, 31]; and discomforting physical appearance [31]. It is interesting to note that with the aid of an expert observer, interactions with social robots tended to be more positive [35]; thus, to maximize the positive impact of a social robot’s deployment—especially in a sensitive health care domain like sleep health—researchers should consider user experience of past experiments and attenuate prior shortcomings for their studies. Furthermore, there is a need for better sensing and communication models to foster more engaging, situation-suitable interactions with users in the long term. Additionally, the ability to understand a user’s preferences regarding the physical embodiment of their robotic companion and the personalization of a robot’s appearance in accordance with said preferences may further enhance user experience.

Another fundamental factor that may impact the efficacy of an intervention is how and where a social robot engages with its users, which in turn depends on the robot’s mobility; for instance, one of the reviewed studies detailed how mobile robots in nursing homes may increase opportunities for user interaction [35], while another study reported on the development of a smart, mobile alarm clock to counter oversleeping and sleep fragmentation [58]. Empowering social robots with more robust mobility may open new avenues for their application and efficiency; however, socially aware, safe navigation is a nuanced and challenging problem [59] which requires further research before deployment in the real world. Two recent studies reported mixed results of insomnia interventions using Somnox, a non-social sleep robot with haptic interactions [60, 61]; exploring haptic interactions for social robots for sleep may be an interesting avenue for future research.

Only five studies from this review sample cited sleep health outcomes as the primary focus of their interventions [32, 28, 33, 31, 37]. The remaining studies reported positive sleep health outcomes, but had originally designed general interventions without a focus on sleep; these interventions generally did not leverage prior work in sleep medicine to inform their design. Despite the lack of a scientific focus on sleep health, the resulting positive outcomes are encouraging; it may be that leveraging past research in sleep science to design and drive future interventions will help yield even better outcomes.

There is a large focus on older adults in this research area, with only one of the reviewed studies focusing on a younger age group. There is valid justification for such a strong focus, as large portions of older adults, especially ones with chronic conditions, tend to experience overall poorer sleep quality; for instance, 71% of people with dementia have sleep disorders such as insomnia, daytime sleepiness, restless leg syndrome, REM sleep behavior disorder, etc. [62]. Still, different age groups and those with certain medical conditions have different sleep requirements and challenges, demanding further work toward understanding how social robots may facilitate the sleep needs of younger adults and children. There is also a lack of research on treating specific sleep disorders (e.g., snoring, bruxism, restless leg syndrome) which impact more than fifty million people in the US alone [63]. Using social robots to drive health interventions for a wider population may help improve sleep interventions generally—for instance, by driving the discovery of better models for motivating intervention adherence and collecting larger data sets to help classify specific sleep disorders.

Lastly, none of the reviewed studies attempted to diagnose any sleep disorders through the use of social robots. Having captured a wide range of sleep data by deploying these robots in people’s living spaces, there is a lost opportunity to leverage such data to detect and characterize sleep disorders and customize interventions around them.

Despite being the gold standard for non-invasive on-body sleep health measurements, arm- and wrist-based actigraphy sensors are cited to be unreliable in their recordings and have been criticized for requiring long wear time to ensure the validity of their data [64, 65]; furthermore, it was reported that these sleep trackers may not be suitable for use with elderly individuals [33]. On the other hand, deploying a completely non-invasive network of infrared sensors for the collection of sleep statistics requires a significant installation effort and may not be suitable for many common living situations [29]. A sleep measure that has been increasingly used in sleep research is the Sleep Profiler, an in-home, three-channel (EEG, electromyography [EMG], and electrooculogram [EOG]) sleep monitor that allows for detailed sleep assessments beyond the basic sleep measures collected via actigraphy; this system facilitates the collection of deeper information regarding sleep spindles, REM latency, and micro-arousals—all data which help contextualize an individual’s sleep health—yet still suffers from the same usability shortcomings as other actigraphy methods.

Therefore, there is an urgent need for a cheap, comprehensive, easily deployable, non-invasive, off-body sleep tracking system. Recent work has explored using radio frequency (RF) signals for non-intrusive sleep state detection and other health measurements [66, 67]; further research into improving RF-signal-based health statistic measurements may lead to the development of the accessible sensors required for social robots to drive more dynamic and personalized interventions—especially for the most vulnerable and sensitive populations.

There are several ethical dilemmas that must be addressed when deploying social robots to drive health interventions. For vulnerable populations such as older adults, researchers must account for the emotional bond that may form between a deployed system and its human user and quantify the degree of trauma that may be caused by their separation at the end of the study. Additionally, there is a growing danger of inducing Turing deceptions in participants as social robots become increasingly more competent at social interactions [68]. There are also several privacy issues with deploying robots in the home health care context that should be addressed promptly [68], as users’ privacy concerns may hinder their trust in and reliance on intervention systems and thus dampen the efficacy of those interventions. There must also be further discussion on the role of social robots in sleep health interventions; for example, whether social robots should independently drive interventions or if they should facilitate interventions in collaboration with human experts is a critical question yet to be addressed.

Our review has its own limitations. First, our search was limited to articles published in English; given the relatively large proportion of work in this domain originating from Japan, we may have missed certain articles of relevance published in different languages. Moreover, our search only included papers that were published during or before March 2023; thus we may have missed articles of relevance that have been published since. It should also be noted that this review was not registered as registration was not required for a scoping review following the methodology by Peters et al. [24].