Supporting Social Engagement in Long-Term Care Through Paired Socially Assistive Robot-Virtual Reality Activities for Older Adults with Cognitive Impairment and Apathy Authors: Miroslava Migovich, Ritam Ghosh, Nibraas Khan, Cathy A Maxwell, Paul A Newhouse, Douglas W Scharre, Alai Tan, Judith A Tate, Lorraine C Mion, Nilanjan Sarkar Published: 2026 ABSTRACT AbstractIntroductionOlder adults living in long-term care settings, particularly those with cognitive impairment and apathy, frequently experience reduced social engagement. Socially assistive robots and virtual reality systems may help support engagement, but little is known about whether they can sustain engagement during real-world deployment while also facilitating interaction between older adults. This study examined whether socially structured socially assistive robot-virtual reality (SAR-VR) systems could sustain engagement and support human-human interaction during repeated use in long-term care settings.MethodsThis analysis focused on the SAR-VR intervention arm of a larger eight-week study and included 60 older adults with cognitive impairment from 23 long-term care sites. Participants were exposed to two SAR-VR systems featuring different robot embodiments (dog-like and humanoid), roles, and activities. Across 756 repeated intervention sessions, engagement was observationally assessed using observed ratings of attention, activity attitude, robot attitude, and conversational targeting.ResultsSustained attention and activity attitudes remained consistently positive across both SAR-VR systems over repeated exposure. Peer-to-peer and research-staff directed speech occurred across both systems. While robot-directed speech was more common in the dog-like condition, partner-directed communication remained stable, indicating that increased human-robot interaction did not occur at the expense of human–human interaction. Baseline cognitive functioning was associated with attentional engagement but not with attitudinal evaluations. Baseline technology acceptance did not strongly predict post-interaction impressions. Participants who anthropomorphized the robots at baseline demonstrated more positive engagement during sessions.DiscussionThese findings suggest that socially structured SAR-VR systems can sustain meaningful engagement while preserving peer interaction in naturalistic dementia care. The results also support the feasibility of designing SAR-VR activities around shared goals and coordinated actions to promote reciprocal human engagement, while recognizing that the study design does not allow definitive conclusions about clinical efficacy or the effects of specific robot features.Keywords: collaborative technologies, multimodal intervention, human-centered design, dyadic activities, cognitive impairment, apathy Introduction Social engagement is a critical component of well-being for older adults residing in long-term care settings.1 However, many residents, particularly those living with dementia, experience limited social engagement. Dementia is accompanied by progressive declines in cognitive, functional, and social abilities that substantially affect the quality of life for older adults in long-term care settings. As cognitive impairment advances, opportunities for social interaction and engagement in activities often diminish, contributing to increased social isolation and disengagement from daily life. Among the neuropsychiatric symptoms associated with dementia, apathy is both highly prevalent and particularly detrimental. Apathy affects an estimated 82% of older adults with dementia residing in long-term care and is characterized by diminished motivation, reduced initiative, and decreased responsiveness to social and environmental stimuli.2,3 The presence of apathy is closely intertwined with reduced participation in activities and social interaction, creating a negative cycle in which disengagement further exacerbates cognitive and functional decline, reduced quality of life, and elevated burden on caregivers and staff.1,4,5 Despite the recognized importance of addressing apathy, effective treatment options remain limited. Pharmacological approaches have shown inconsistent and modest effects, and no medication has been established as the standard treatment for apathy in dementia.6,7 Consequently, effective nonpharmacological interventions remain central to addressing apathy. The most effective nonpharmacologic interventions in enhancing engagement and reducing apathy are activities that combine cognitive, physical, and social components.8–13 However, delivering effective multimodal and socially engaging activities in long-term care contexts is challenging. Lack of staff as well as inadequately trained staff often result in activities that are limited in scope, rely heavily on passive participation, and struggle to sustain social engagement over time.14,15 Residents in US long term care facilities (LTCs) receive an average of 3.90 nurse staff hours/per day/per resident, falling short of the 4.94 nurse staff hours/per day/per resident required to ensure that residents receive adequate clinical care.16 Furthermore, on average, LTC residents only receive 25.6 minutes/per day with activities personnel.17–19 In response to the scarcity of LTC staff, there has been growing interest in the use of technology-based approaches to support engagement among older LTC residents.20 Socially assistive robots (SARs) are one such technology that have demonstrated potential to promote interaction through embodied social presence, expressive behaviors, and nonverbal cues that attract and sustain attention.21,22 In parallel, non-immersive virtual reality (VR) systems offer the ability to deliver a wide range of adaptable, activity-based experiences that can be tailored to cognitive and physical abilities.23,24 However, each modality presents limitations when used in isolation, including restricted activity diversity for SARs and reduced social facilitation in VR-based interventions. Thus, integrating SARs with non-immersive VR offers a promising approach to leveraging the complementary strengths of both modalities by enabling structured, multimodal activities that create shared experiences, support coordination, and encourage interaction between participants, while the robot provides embodied guidance and reinforcement. In this way, the combined SAR–VR system functions not only as an engagement tool, but as a facilitator of human–human interaction, with technology mediating and supporting social exchange rather than replacing it. In our prior work, we employed a user-centered design approach, engaging both domain experts (eg clinicians, engineers, and care staff) and older adults themselves, to develop and deploy integrated socially assistive robot–virtual reality (SAR–VR) systems for older adults with mild to moderate cognitive impairment and apathy residing in long-term care settings.25 This prior publication described the initial design, development, user-centered evaluation, and initial feasibility of the SAR-VR systems and activities. The present study builds on that work by examining more established versions of the systems during sustained, multisite deployment in long-term care. Across both systems, the design was grounded in a shared activity framework, intentionally integrating cognitive, physical, and social components, reflecting evidence that multimodal activities are most effective for supporting engagement in this population.6,13,26,27 Activities were constructed as combinations of perceptual, motor, and collaborative subtasks, with embedded rules that required coordination and mutual support between paired participants.25,28–30 Two SAR–VR system configurations were developed and deployed. In both systems, older adults engaged in paired, collaborative activities within non-immersive virtual environments, requiring coordination, communication, and shared task completion. However, the systems differed in robot type, robot role, interaction architecture, feedback modality, and task type (see Section 3.5 for complete description). The humanoid and dog-like robots were chosen to represent two distinct approaches that are commonly used in socially assistive robotics31 and in order to support different interaction roles within the systems. Additionally, prior work showed variable interest and responses to different robot embodiments by older adults,32 supporting the inclusion of multiple robot configurations to accommodate variability in how older adults may engage with socially assistive technologies. The first system paired a humanoid robot with non-immersive virtual activities in which the robot provided direct instruction, feedback, and encouragement, guiding participants through task execution and supporting performance. The second system paired an animal-like robot (dog) with non-immersive virtual activities in which the robot functioned primarily as a motivational and reward-based agent, reinforcing task completion through expressive behaviors and contingent responses, while an on-screen avatar provided the instruction and feedback. Importantly, both SAR-VR systems were designed not as isolated technologies, but as seamlessly integrated platforms in which robot behavior, virtual activities, and collaborative task structure were tightly coupled. While the specific robot roles and activities differed, both systems required common design elements: 1) activities that were paired and interdependent;33–36 2) SAR-VR communication feedback structured to enhance collaboration between the older adults (human-human interaction) as well as between the SARs and older adults (human-robot interaction)37–40 and 3) reward mechanisms for achieving the activity goals.40,41 An advisory board consisting of nurses, activity professionals, occupational therapy, informatics technology specialist, and long-term care administrators met regularly to review and approve each activity from conception to final product. Our previous works established the proof of concept through electroencephalography (EEG) signal, physiological data (galvanic skin response), real-time gaze estimation;42 and observation of verbal, visual, behavioral, and emotional engagement measures).43 Initial feasibility of a SAR-VR system within real-world care environments40 was conducted with three activities over a 3-week period. Findings revealed that type of engagement (eg visual vs verbal) varied by activity. However, while initial feasibility was demonstrated, the nature of older adults’ engagement with the SAR-VR system activities during sustained deployment remained unexplored. There is a dearth of knowledge regarding older adults’ user-level responses to SAR-VR systems in naturalistic care settings. Observed attention, attitudes toward the robots, and attitudes toward the VR activities are critical indicators of meaningful engagement and are essential for informing future system design. This need is particularly salient in dementia care, where engagement is impacted by individual, environmental, and technological factors. Moreover, while SAR–VR systems can vary in embodiment and interaction structure, it is not yet well understood how such differences relate to engagement when the underlying activity design and collaborative structure are similar. These questions are especially important in longitudinal, real-world deployments, where engagement must be sustained over time rather than elicited in short laboratory sessions. As part of a larger multi-site study examining the effects of SAR–VR systems on apathy among older adults living with cognitive impairment,28 the present work focuses specifically on engagement-related outcomes. We examined observed measures of engagement—including attention, robot attitude, and activity attitude—among older adults with cognitive impairment and apathy participating in paired SAR–VR activities over an eight-week period. By analyzing repeated observational measures collected within long-term care environments, rather than brief laboratory sessions, this study captured engagement over time within participants’ everyday contexts. The key gap addressed by the present study is the limited understanding of whether engagement with integrated SAR-VR systems can be sustained over repeated use in real-world long-term care settings, particularly when the systems are designed to support both human-robot and human-human interaction. This work addresses this gap by analyzing repeated observational measures of engagement, HRI, and HHI across an eight-week, multi-site deployment of two distinct SAR-VR system configurations. Collectively, this work advances the field through two key contributions. First, it provides support for SAR-VR as a combined interaction paradigm that expands the space of possible engagement by leveraging the complementary strengths of socially assistive robots and virtual environments. By integrating robot-mediated interaction with adaptable virtual activities, SAR–VR systems enable a broader and more flexible range of cognitive, physical, and social interactions than either modality alone. Second, this work demonstrates that sustained engagement can be observed across distinct SAR–VR system configurations deployed in real-world settings, even when robot embodiment, role, interaction architecture, and feedback modality differ. By moving beyond feasibility to characterize user-level engagement during sustained, real-world deployment, this study provides design-relevant insight into how SAR–VR systems can support meaningful engagement in dementia care. In particular, the findings highlight the role of collaborative task structure as a stabilizing factor for attention and participation across cognitive variability, and emphasize the importance of designing for reciprocal, socially mediated interaction rather than performance accuracy alone. To further characterize engagement within these socially scaffolded SAR-VR systems, the study was guided by the following research objectives: Research Objective 1: To determine the extent of participants’ engagement with each type of SAR–VR system configuration. Research Objective 1: To determine the extent of participants’ engagement with each type of SAR–VR system configuration. Research Objective 2: To determine the extent of participants’ Human–Human Interaction (HHI) and Human–Robot Interaction (HRI) by type of SAR–VR system configuration. Research Objective 2: To determine the extent of participants’ Human–Human Interaction (HHI) and Human–Robot Interaction (HRI) by type of SAR–VR system configuration. Research Objective 3: To examine engagement, HRI, and HHI over repeated exposures to the SAR–VR systems. Research Objective 3: To examine engagement, HRI, and HHI over repeated exposures to the SAR–VR systems. Background Related Works Socially Assistive Robots with Older Adults With advances in robotics and closed-loop systems capable of monitoring and responding to human interaction in real time, there has been a growing body of research examining the feasibility and usability of robots to support various aspects of care for older adults, including social companionship, wayfinding, medication reminders, activities of daily living, promotion of healthy eating, health monitoring, and engagement in physical, rehabilitative, and cognitive activities, as highlighted in several systematic reviews.31,44–47 In an umbrella review of 35 systematic reviews spanning populations from children to older adults, ten systematic reviews identified 13 unique studies that specifically examined SARs designed to encourage social interaction.31 Social interaction in these studies encompassed engagement with the robot itself, other participants, family members, or staff. All thirteen studies reported positive outcomes, with social interaction observed regardless of whether the robot was animal-like or humanoid in form. Factors such as level of cognitive impairment and age group (eg child, young adult, older adult) were identified as important moderators of SAR effectiveness, highlighting the need to better characterize engagement across populations and contexts. Regarding paired human interaction, which is pivotal for fostering social interactions,34,35 relatively few studies have examined group-based, closed-loop robot-mediated interventions for older adults. In a recent systematic review of 25 studies involving group interactions between older adults and SARs, 19 were facilitated by a healthcare professional using the SAR as assistive technology, whereas only six were led autonomously by the SAR,36 highlighting an important gap in literature that this work aims to address. Socially Assistive Robots Combined with Non-Immersive Virtual Reality Environments Few studies have examined the use of socially assistive robots combined with non-immersive virtual reality environments.36 Most of these studies that combine robotics with non-immersive virtual reality focused on specific rehabilitation needs (eg post-stroke gait) or cognitive training. For neurorehabilitation needs, robotic devices tend to be end-effector (eg robotic arms or tools) or exoskeleton (ie wearable robotic units). Thus, most rehabilitation or therapeutic robot-VR studies consist of a single user with a robot (ie one-on-one interactions).48 In contrast, the SAR-VR systems examined in the present study were designed around paired, interdependent activities in which the virtual environment structured a shared task and the robot supported participation and social interaction. This work, therefore, extends prior SAR and VR research by examining both human-robot and human-human engagement during repeated use in long-term care. Challenges of Robot-Mediated Social Interaction for Older Adults Several challenges present themselves when introducing SARs to older adults and can be broadly discussed from the Human-Robot-Interaction framework. From the perspective of the older adult, trust is essential49–51 and is linked to acceptance and positive experiences with the SAR. In service-oriented SARs, trust emphasizes reliability and capability. For socially interactive SARs, trust emphasizes the relational basis. Characteristics of the robot, such as its appearance and communication style, contribute to the older adult’s sense of trust. Older adults’ perceptions of new technology may be a challenge to the acceptance of the SAR and introducing a SAR needs to be carefully considered. In recent systematic reviews, ongoing presence of professionals were considered comforting and providing a sense of security for older adults.36,52 Other challenges related to the older adult’s ability to interact with the robot include sensory changes (vision, hearing), cognitive function, physical function, and presence of painful conditions, such as arthritis, prohibiting prolonged interaction with the SAR. Implementation within long-term care settings also introduces additional practical challenges including, but not limited to, staff availability, competing facility priorities, physical space for running sessions and storing equipment, technology setup, and network connectivity. These implementation considerations were examined in detail in our previous user-centered design work25 and informed the development of the present systems. Accordingly, the current study builds on several conceptual frameworks detailed below. Conceptual Frameworks Guiding the Study This study is grounded in established models of engagement in older adults with cognitive impairment, as presented by Cohen-Mansfield’s seminal body of work on engagement,1,33,53–55 which conceptualizes engagement as active involvement with external stimuli, including activities, objects, and social partners. Within long-term care settings, engagement is shaped by the interaction of personal factors (eg cognitive and physical function), environmental conditions, and stimulus characteristics. Social interaction within activities is the most potent stimulus for persons with dementia, and prior work has demonstrated that small-group contexts, particularly paired interactions, are especially effective in promoting sustained engagement. Building on this foundation, the present study operationalizes engagement through the integration of socially assistive robots (SARs) and non-immersive virtual reality (VR) within collaborative, task-based activities. In this context, SARs function as socially embedded stimuli that guide and reinforce interaction, while VR environments structure shared tasks that require coordinated participation between individuals. This design reflects a shift from individual, stimulus-driven engagement toward socially scaffolded engagement, in which interaction between participants is central to the activity itself. The system architecture is informed by a body of work in robot-mediated interventions,56–58 with and without virtual reality, that has progressively advanced toward adaptive, closed-loop human-robot interaction, including the Adaptive Robotic-mediated Intervention Architecture (ARIA).59 ARIA is a closed-loop framework that enables real-time adaptation of the system behavior based on multimodal user input. This line of research has demonstrated enhancement of interaction across populations, including individuals with autism spectrum disorder,60 those undergoing stroke rehabilitation,61 and older adults living with dementia.40 Subsequent work has extended these concepts to multi-user and collaborative contexts, including robot-mediated interaction with older adults and early SAR–VR systems designed to promote activity participation and social engagement.32,40,42 The present study builds on this foundation by implementing two SAR–VR systems within a real-world long-term care setting and examining engagement within structured, collaborative interactions between participants. While ARIA provides the conceptual basis for closed-loop and adaptive interaction, the current system employs a structured interaction framework in which the robot delivers prompts, guidance, and/or reinforcement within predefined activity flows. Within this context, the SAR–VR systems function as socially scaffolded interventions designed to support engagement through coordinated task structure and facilitated interaction. In addition to system-level design, factors influencing user interaction and sustained use are considered through an adaptation of the Unified Theory of Acceptance and Use of Technology (UTAUT).62 Perceived usefulness and ease of use are key determinants of technology adoption, while prior work in socially assistive robotics further highlights the role of affective responses (eg enjoyment, trust) and anthropomorphism in shaping engagement.50,63 These factors are particularly relevant in SAR-mediated contexts, where social perception of the robot may influence both human–robot and human–human interaction. Finally, this work is informed by evidence from nonpharmacological interventions targeting apathy, which consistently demonstrate the benefits of multimodal activities that target multiple dimensions; these dimensions are cognitive activities that target executive planning, physical activity that serves as a catalyst for metacognition and empathy, and a social component of the shared activity to challenge the individuals to participate in interpersonal exchanges.6,26,27 Activities that are varied and adaptable to individual preferences further enhance sustained participation in long-term care settings.64 Together, these frameworks informed the design and evaluation of the SAR–VR systems as a socially mediated, adaptive interventions, in which engagement is not solely an individual response to a stimulus, but an emergent property of interaction between participants, technology, and task structure. The novelty of the present study lies in integrating these established frameworks within paired SAR-VR systems and examining how they operate during repeated, real-world deployment with older adults with cognitive impairment and apathy in long-term care settings. In particular, the study extends prior applications of these frameworks by examining engagement across both human-human and human-robot interactions within collaborative activities. The established frameworks provide the theoretical basis for system design, while the present work contributes empirical evidence regarding their application within a socially scaffolded, multi-user SAR-VR context. Methods Study Design and Rational As part of a large multi-site Stage I/II randomized clinical trial designed to evaluate the effects of socially assistive robot–virtual reality (SAR–VR) interventions for older adults with cognitive impairment and apathy residing in long-term care (LTC) settings28 (NCT05178992), we examined older adults’ engagement with the paired, collaborative SAR-VR systems. Pairs of participants randomized to the SAR-VR arm interacted with two SAR-VR systems (humanoid and dog*(For brevity and clarity, we use robot type labels (humanoid and dog) as shorthand to distinguish between the two SAR-VR systems.)), completing 4 weeks (8 sessions) with each system, with sessions occurring twice weekly for a total of 8 weeks (16 sessions). Pairings of individuals were not fixed and were determined based on participants’ availability to maximize participation and relieve scheduling burden. The order of the SAR-VR system configuration was randomized by site to counterbalance ordering effect. Data were collected from February 2023 to October 2025. Participants were observed at each session by trained research staff and data were collected using the Group Observational Measurement of Engagement (GOME) tool.33 Baseline interviews were conducted to gather information on demographics, attitudes toward robots, cognitive status, and apathy. We proposed collaborative activities involving two older adults based on theoretical models of engagement that emphasized meaningful engagement arises when activities align with individual’s abilities and interests and when they provide opportunities for agency and interpersonal connection.55 Social interaction plays a central role in sustaining engagement and prior work has shown that shared activities, cooperative goals, and reciprocal communication can enhance motivation and attention.33 Two types of robots were used based on our earlier work demonstrating varying interest among older adults to the robot. The exposure to the SAR-VR twice weekly sessions for 8 weeks was based on other studies involving SAR in long term care and in providing sufficient intervention for those with apathy.28 We adhered to the Strengthening the Reporting of Observational Studies in Epidemiology checklist to report the findings.65 Study Participants and Study Settings Older adults were eligible to participate if they were 60 years of age or older, resided in a participating long-term care facility for at least three months, and had evidence of mild cognitive impairment or dementia as determined through screening procedures. Participants were required to demonstrate symptoms of apathy on screening measures and to be able to engage in seated interactive activities with the SAR–VR system. Exclusion criteria included severe cognitive impairment precluding meaningful participation, uncorrected visual or hearing impairments that would interfere with interaction, acute medical instability, or physical limitations preventing safe engagement in seated virtual activities. Individuals unable to provide informed consent or assent (or without an available legally authorized representative, when applicable) were excluded. Eligibility criteria for settings included licensed by their respective state, within an hour driving distance from the investigators’ offices, physical space for set-up and secured storage, and certified by the US Centers for Medicare and Medicaid. Exclusion criterion was LTCs designated as a subacute care, rehabilitation, or hospice facility. Recruitment and Screening Site Recruitment Using IRB-approved communication script, emails were sent to sites introducing the study with up to two reminder emails, one week apart. An IRB-approved memo was attached to the Email that provided more information on the study purpose and role of the site and site personnel. For those who responded as potentially interested, an initial video meeting was arranged between the research team and key site personnel (ie activity personnel, director of nursing, administrator). The Screening Form was shared and discussed with personnel. If the site expressed interest, the administrator signed a commitment letter and returned it to the investigator team. Older Adult Recruitment Recruitment activities of older adult residents were site-specific. At all sites, site personnel approached potential participants who met the inclusion criteria using the study Screening Form as a guide. IRB-approved flyers were distributed. At several sites, investigators were invited to speak at resident meetings. Interested older adult residents and family contacted the research personnel for further discussion. Research assistants (RAs) met with older adults and family to explain the study, conduct screening eligibility, obtain written informed consent and/or assent (from participant or family member), conduct baseline data collection, and then randomize to Usual Care (UC) or SAR-VR+UC. Ethical Considerations A centralized Institutional Review Board (SMART IRB) reliance initiative was undertaken with The Ohio State University Institutional Review Board serving as the IRB of record (OSU IRB protocol #2020B0342). The study was conducted in accordance with the Declaration of Helsinki. All recruited sites accepted and agreed to rely on the primary site’s approval. Sites that participated received a small incentive ($2,000) for staff continuing education as a thank you for providing physical space to conduct the robot sessions, providing secured storage space for the equipment, and answering brief questionnaires regarding the resident’s usual activity and level of apathy. Participants received a prorated incentive (up to $125) for attending the SAR-VR sessions. All participants and their designated family member signed an informed consent document prior to the conduct of any research activities. SAR-VR System Architecture We designed the socially assistive robot-virtual reality (SAR-VR) systems to support paired social interaction and sustain engagement by explicitly combining physical, cognitive, and social components within each activity. Physical elements involved arm, hand, and/or head movements, such as moving their arm to control a custom designed input device, holding buttons, and reaching. Cognitive elements included attention, sequencing, matching, and problem-solving. Social elements were expressed through task rules that required participants to coordinate actions, divide responsibilities, or communicate to achieve a shared objective. The design emphasis was on creating shared experiences that naturally encouraged communication and coordination with physical and cognitive components, while integrating robots in ways that ranged from task guidance and feedback to motivational support and reward across activities. In total, there were seven designed activities with two robot embodiment types, a humanoid robot and an animal robot (dog). Each activity had several increasing levels of difficulty. Robot behavior during the sessions was controlled by the integrated SAR–VR system rather than by a human operator. Trained research assistants set up the robot and computer equipment and used the system interface to select the activity and level of difficulty prior to the start of each session. Once the activity began, the system controlled the robot’s behavior. The system continuously monitored participant performance and engagement during the activities (eg timing accuracy, missed actions, or inactivity) and used predefined conditions to trigger appropriate feedback. When these conditions were met, the system generated messages that were translated through an intermediate communication layer into robot-specific commands, enabling the robot or on-screen avatar to provide demonstrations, encouragement, corrective feedback, or rewards (eg celebratory movements or tricks) in response to participant actions. Trained research assistants remained present during sessions to supervise the interaction, provide clarification or additional guidance if participants did not understand the instructions delivered by the SAR–VR system, and troubleshoot any technical issues if they arose. Additional details regarding the system architecture are provided in prior works.25,30,32,42 The following subsections briefly describe the robot platforms, robot roles, and the activities. For more details regarding the system design or the full protocol, please refer to earlier publications.25,28 Robot Platforms and Roles The SAR-VR systems incorporate two socially assistive robot embodiments to support paired interaction during the collaborative virtual activities, as can be seen in Figure 1. In our earlier works,30,32,42 we found that 63 older adults varied in their responsiveness to and enjoyment of different robot embodiments, motivating the inclusion of both humanoid and animal robots in the present study. For both systems, the robots were integrated as facilitators of engagement rather than as the primary focus of engagement. Across both SAR-VR system configurations, the robot is intentionally positioned as a supportive facilitator of interaction rather than the central conversational partner. The humanoid robot emphasized task guidance and feedback while the animal robot emphasized reward and motivational support. The SAR-VR activities used in the present study were developed from activities introduced in our prior system-development work.25,29,30 The descriptions provided here summarize the versions implemented during the current eight-week deployment, which incorporated refinements made through prior user-centered design and feasibility testing. SAR-VR system with Humanoid robot [left], animal robot [right], and VR activity menu [center] shown. Humanoid Robot: NAO We used the humanoid robot NAO from SoftBank Robotics (https://us.softbankrobotics.com/) as an instructional and facilitative robot platform.44,45,66 NAO has been widely used in research in education, and more recently, in healthcare including children with autism, older adult care, and rehabilitation.44,45,66 NAO’s open architecture allows for system-level modifications towards closed-loop intelligent interaction, thus providing a cost-effective approach for conducting research that examines human-robot interactions in a variety of settings, including long-term care. Additionally, the limitations of NAO, such as restricted facial expressions, are not detrimental to the goals of supporting collaborative interaction and engagement.45 In the humanoid SAR-VR system, NAO provides verbal task instructions, real-time feedback, and encouragement throughout each activity. Robot communication is directed to each individual by name and encourages participants to help each other when relevant. The following activities were implemented in the humanoid SAR-VR system, shown in Figure 2. VR activities for Humanoid SAR-VR system.(A) Music activity. (B) Painting activity. (C) Fishing activity. (D) Booking sorting activity. Music Activity (Figure 2a) Participants collaboratively played virtual drums in synchrony with visual musical cues from a selected song. Songs were chosen from a library of familiar genres, such as country or swing, to enhance personal relevance and sustained engagement. Successful performance required attention to timing and coordinated action between partners. NAO provided pacing support, corrective feedback, and encouragement during gameplay. Painting Activity (Figure 2b) Participants completed a shared paint-by-numbers image using the handheld input device as a virtual paintbrush. Each participant was assigned specific colors, requiring coordination and sequencing to fill the image correctly. NAO provided guidance if participants attempted to use one another’s assigned colors or were not progressing, and offered encouragement as the painting was gradually completed, reinforcing shared accomplishment. Fishing Game Activity (Figure 2c) Participants assumed complementary roles to retrieve and transfer virtual fish. One participant used a virtual fishing rod to pull fish from the water, while the other used a net to move the fish into a bucket. Task completion required communication and coordination between partners. NAO provided reminders to collaborate, along with feedback and guidance as needed to support successful transfer. Book Sorting Activity (Figure 2d) Participants sorted color-coded books into matching bins within the shared virtual environment. Increasing levels of difficulty introduced collaborative constraints, such as requiring participants to exchange books or move books in synchrony. NAO provided instructional support and performance feedback throughout the activity. Animal Robot: Aibo For the animal-like robot platform, we used Aibo from Sony (https://us.aibo.com/), which looks and acts like a dog. Aibo has been used in research examining human-robot interaction, as well as in medical and therapeutic contexts and supports integration into experimental systems.67,68 In contrast to the humanoid SAR-VR system, Aibo did not provide direct task instruction or gesture-based interaction. Instead, instructional guidance, prompts, and feedback are delivered through a virtual avatar within the virtual environment. Aibo functioned primarily as a motivational and reward agent, responding to successful actions with expressive behaviors that serve as social reinforcement. This design allowed the robot to support engagement without drawing attention away from the shared task or interaction between participants. The following activities, shown in Figure 3, were implemented in the animal (dog) SAR-VR system. VR activities for animal SAR-VR system. (A) Puzzle activity. (B) Path planning activity. (C) Spelling activity. Puzzle Activity (Figure 3a) Participants collaboratively completed shape-matching puzzles displayed in the virtual environment. Each participant was assigned specific shapes, requiring coordination and communication to correctly place all pieces. As puzzle elements were completed, Aibo performed a corresponding trick. Path Planning Activity (Figure 3b) Participants worked together to guide a virtual representation of Aibo through a color-coded maze. Each participant was assigned specific movement directions and path colors, necessitating joint decision-making and communication to navigate the correct route. As participants successfully progressed through segments of the maze, Aibo performed brief tricks with a more expressive celebratory behavior upon reaching the end of the path. Spelling Activity (Figure 3c) Participants collaboratively spelled target words by selecting letters of assigned colors in the correct sequence. The activity required turn-taking, shared attention, and coordination to ensure correct ordering. The virtual avatar provided instructional guidance and feedback throughout the task, while Aibo performed a corresponding trick upon successful completion of each word, reinforcing collaborative achievement. Across both SAR-VR systems, the robot was intentionally positioned as a supportive facilitator of interaction rather than the central conversational partner. The humanoid robot emphasized task guidance and feedback, while the animal robot emphasized reward and motivational support. This variation in robot role and embodiment enabled examination of how different forms of robotic support influence social interaction and engagement between the participants and the robots, their partners, and the activities. Session Delivery Participants attended two SAR–VR sessions per week for eight weeks, for a maximum of 16 sessions. At each session, two trained research assistants were in attendance; one to monitor the technology and delivery of the SAR-VR activity and one to monitor the older adults’ performance and assist as needed. Several sessions were offered each week to allow participants a choice of time. Participation in each session was voluntary, and missed sessions were logged to reflect real-world attendance patterns in long-term care settings. Each session lasted approximately 20–30 minutes and was structured around paired participation. At the start of each session, NAO or the VR avatar for Aibo provided a greeting and a short tutorial. NAO/avatar taught the older adult how to interact with the virtual environment through incremental learning and demonstration. Complex tasks were broken down into subtasks, such as wand movement (left, right, up, down), clicking on virtual objects (pressing wand button), dragging objects across the environment, and releasing the object (pressing wand button). As NAO explained each subtask, it demonstrated the movements needed to accomplish the subtask and the systems evaluated the older adults’ performance in real time to provide appropriate guidance. If one older adult was having difficulty, NAO/avatar would prompt the older adult’s partner to show them how to perform the subtask. If there were three consecutive corrections and the older adult continued to demonstrate difficulty, then the RA would assist by demonstrating with the wand. With repeated exposure to the system, older adults were able to accomplish the tasks. All SAR–VR sessions were designed as collaborative paired experiences. Activity selection and level of difficulty for the activity were not prescribed. Rather, participants selected one or more activities from the menu (see Figure 1) and could also choose the level of difficulty. Participants engaged in the shared virtual environment alongside a partner, with task completion requiring coordinated interaction. Pairings were not fixed across sessions and were determined based on participant availability to maximize participation and reduce scheduling burden. When a resident partner was unavailable, a trained research assistant, care staff member, or family member served as the interaction partner to maintain the collaborative structure of the session. Participants were exposed to both SAR-VR systems across the intervention period, engaging in four weeks of sessions with the humanoid-based system and four weeks with the animal (dog)-based system. SAR-VR Intervention Fidelity We incorporated recommendations from the NIH Behavior Change Consortium to maximize treatment fidelity across the multiple sites and RAs over time.69 In brief, we used theoretical models to address engagement of individuals with cognitive impairment and apathy; conducted standardized training for RAs with ongoing quarterly reviews; utilized standardized protocols; and required RAs to alternate travel quarterly between Nashville and Columbus to conduct simultaneous sessions to minimize drift. These training and review procedures were applied to all observational measures used in the study. All sessions were video recorded and RAs reviewed coded sessions with one another at least once a month to assess consistency in observational ratings. Discrepancies were discussed at the weekly co-PI-RA meetings as well as at the quarterly full investigator meetings to reinforce consistent application of the observational coding criteria across raters and sites. RAs completed field notes during each session documenting the session date, time began and ended, and notes on issues encountered during the session and evaluated the amount of effort the older adults exerted to accomplish tasks. Measures The present analyses focused on baseline cognitive status and robot acceptance with the observed engagement during SAR–VR sessions. Observed Engagement and Interaction Engagement and interaction were assessed during SAR–VR sessions using a modified version of the Group Observational Measurement of Engagement (GOME) tool,33 previously used in studies involving socially assistive robots and older adults.43,54,66,70 The GOME captures engagement across five dimensions using ordinal scales: attendance duration (0 = none to 6 = entire activity), attentiveness and participation (0 = none to 7 = very attentive and participating), attitude (1 = very negative to 7 = very positive), cognitive difficulty (0 = very difficult to 5 = not at all difficult), and dexterous difficulty (1 = not at all difficult to 5 = very difficult). Attentiveness/participation and attitude were each scored separately for engagement directed toward the robot and toward the activity. Ratings were recorded for both the level observed during the majority of the session and the highest level reached. Participants’ verbalizations were noted as absent or present (0.1) during each session and further coded by the target of each participant’s verbalizations (eg directed toward the robot, partner, researcher, others in the room, or activity) to characterize interaction patterns during collaborative sessions and are referred to as TalkTarget within the Results. There were seven possible TalkTargets: Robot, Staff, Resident/Older Adult Partner, Research Staff, Family, cannot Tell, or Did Not Talk. Trained research assistants conducted direct observations during active participation and recorded engagement and interaction ratings accordingly. Cognitive Status Cognitive functioning was assessed at baseline using the Self-Administered Gerocognitive Examination (SAGE),71 a brief screening instrument that evaluates multiple cognitive domains and provides a global score reflecting cognitive status. Total scores range from 0 to 22, with lower scores indicating greater cognitive impairment. SAGE scores were treated as a continuous variable. The baseline SAGE assessment was administered after informed consent was obtained but prior to participation in any robot-mediated intervention. Baseline SAGE scores were used to characterize variability in cognitive functioning and to examine associations between cognitive status and observed engagement during SAR–VR sessions. Robot Acceptance Survey (RAS) Baseline attitudes toward robotic technology were assessed using a standardized Robot Acceptance Survey72 informed by the Unified Theory of Acceptance and Use of Technology framework (63). The survey consists of 25 items, rated on a 7-point scale ranging from 1 (most positive) to 7 (most negative). The measure can be summed to yield a total score ranging from 25 to 175, with lower scores reflecting more positive acceptance of robotic technology. Five of the items examined the extent to which the participant attributed human/dog characteristics to the SARs (anthromorphising); scores ranged from 5 to 35. Baseline robot acceptance variables were examined in relation to observed engagement and interaction during collaborative SAR–VR sessions. To examine the relationship between baseline anthropomorphism and engagement during SAR–VR interactions, a composite HumanCharacteristics score was computed from selected items of the Robot Acceptance Survey (RAS) reflecting the extent to which participants attributed human-like qualities to the robot prior to interaction. Higher scores indicated lower levels of anthropomorphism. This participant-level measure was merged with engagement outcomes derived from observational ratings collected during sessions with both robot systems, humanoid and dog, including Activity Attitude, Robot Attitude, and Attention. Data Analysis Data were structured and prepared for analysis in MATLAB § (MATLAB version 9.14 (R2023a) https://www.mathworks.com) by synchronizing all measures using participant identifiers and organizing variables by SAR-VR system (dog vs humanoid). Although interaction sessions were conducted in pairs, observational coding was performed at the level of the individual participant given the varying nature of attendance. Accordingly, session-based variables were structured at the participant–session level rather than at the paired-session level. Descriptive Analyses Attendance rates were calculated separately for each SAR-VR system. For each participant, the rate was calculated as the number of sessions attended dived by the total number of sessions offered. These participant-level rates were then averaged as the final attendance rates. For engagement variables derived from the Robot GOMES (Activity Attitude, Robot Attitude, and Attention), participant-level means were computed within each SAR-VR system by averaging across that participant’s attended sessions. Group-level descriptive statistics (means and standard deviations) were then calculated across participants. Inferential analyses were selected according to the structure and distribution of the outcomes. Linear mixed-effects models were used for session-level longitudinal analyses to account for repeated observations within participants. Wilcoxon signed-rank tests were used for paired comparisons of participant-level outcomes between SAR-VR systems because the observational measures were ordinal and non-normally distributed. Spearman rank-order correlations were used for association analyses involving ordinal or non-normally distributed variables. Multiple-comparison corrections were applied within defined families of related comparisons, as described below. To examine changes in engagement over time while accounting for repeated observations within participants, linear mixed-effects models were conducted for each outcome (attention, activity attitude, and robot attitude). Session number (RepeatInst) and system condition (Dog vs Humanoid) were included as fixed effects, along with their interaction. Participant was included as a random intercept to account for within-subject correlation across sessions. Thus, the participant-specific random effect accounted for the within-person clustering of repeated session observations. Models were fit using maximum likelihood estimation. Additional sensitivity analyses were conducted to examine whether primary findings were affected by the interaction partner type or study site. Partner type was included as a fixed-effect covariate. Because the 60 participants were distributed across 23 study sties, site was evaluated as a random intercept rather than a categorical fixed effect. These sensitivity analyses were used to determine whether accounting for partner type and site-level clustering materially altered the primary findings. Activity type was not included as an additional covariate because activity type was structurally linked to the SAR-VR system configuration and therefore could not be independently estimated from system condition. Accordingly, condition effects were interpreted as differences between the overall SAR-VR configurations rather than as effects attributable specifically to robot embodiment or activity type. To complement the longitudinal analysis, within-participant comparisons between SAR-VR systems were conducted using Wilcoxon signed-rank tests on participant-level mean scores aggregated across sessions. Effect sizes for these comparisons were calculated using standardized rank statistics. The sample size for the parent randomized control trial was determined based on the study’s primary outcome and was not specifically powered for the secondary engagement outcomes examined in the present analysis. Accordingly, no separate a prior sample size calculation was performed for these engagement analyses. For TalkTarget variables, proportions were computed at the participant level separately for each SAR-VR system using only sessions that were a) marked as attended and b) in which the interaction partner was an older adult participant; this restriction was applied because, at the time of analysis, the dataset did not allow us to reliably distinguish between instances in which speech directed toward a researcher reflected assistance-seeking behavior vs instances in which the researcher served as the interaction partner. Limiting the analysis to sessions attended by two older adult participants reduced ambiguity and allowed more accurate reflection of the target of the participant communication within the intended collaborative context. Specifically, for each participant and SAR-VR system the number of attended sessions in which a given TalkTarget category was marked was divided by the total number of attended sessions for that participant in that condition. This yielded a within-participant proportion reflecting the relative frequency of each communication target while accounting for missed sessions. Participants were included in the TalkTarget analysis only if they had attended valid sessions in both SAR-VR systems under these criteria. These participant-level proportions were then averaged across participants to obtain overall mean proportions for each TalkTarget category in the Dog and Humanoid conditions. Standard deviations were calculated across participants. Baseline SAGE and RAS scores were summarized using descriptive statistics (mean, standard deviation, range). RAS scores were additionally categorized according to predefined acceptance groupings (Accepting, Moderate, or Hesitant/Not Accepting of Robots) for descriptive reporting. Comparisons Across SAR-VR Systems Nonparametric statistical tests (Wilcoxon signed-rank tests)73 were used for within-participant comparisons between Dog and Humanoid conditions due to the ordinal nature of the observation ratings and non-normal distribution of the variables. For Activity Attitude, Robot Attitude, and Attention from the Robot GOMES, Bonferroni correction74 was applied following the Wilcoxon signed-rank tests to control the family-wise error rate across this defined family of three related engagement outcomes (). This adjustment reduced the likelihood of Type I (false positive) error inflation due to multiple comparisons within the same outcome domain. Effect sizes were calculated using: where z represents the standardized test statistic and N represents the number of paired observations. Participant-level paired comparisons between SAR-VR system configurations were conducted using Wilcoxon signed-rank tests, and effect sizes were computed. Because the TalkTarget variables represented multiple related comparisons within the same communication domain, Bonferroni correction was applied across these comparisons to control the family-wise Type I error rate. To further examine the relationships between conversational targets, additional sub-association analyses were performed. Pairwise Spearman correlations were conducted across talk target proportions to assess if conversational engagement was distributed across targets within sessions. Robot preference counts were derived by comparing participant-level mean Robot Attitude scores across conditions. Association Analyses To examine the relationship between baseline characteristics and engagement outcomes, Spearman rank-order correlations were used due to the ordinal nature of several variables and violations of normality assumptions. Baseline Robot Acceptance (RAS) Spearman correlations were computed separately for each SAR-VR system to examine associations between baseline RAS scores and participants’ averaged Robot Attitude ratings. Associations between HumanCharacteristics and engagement outcomes were assessed separately for each SAR-VR system using Spearman’s rank-order correlations (ρ), given the ordinal nature of the engagement measures and non-normality of the data. Analyses were conducted using within-participant data, with each participant contributing one set of engagement scores per condition. Missing values were excluded using pairwise deletion. To account for multiple comparisons across six correlations (three outcomes per SAR-VR system), a Bonferroni correction was applied, with statistical significance evaluated at . To evaluate whether baseline acceptance differentially influenced attitudes toward distinct SAR-VR systems, analyses were stratified by SAR-VR system, humanoid and dog. To control for Type I error across these parallel tests, Bonferroni correction was applied Baseline Cognitive Status (SAGE) Spearman correlations were also computed separately for each SAR-VR system to examine associations between baseline cognitive status, as measured by the SAGE, and engagement-related outcomes (Activity Attitude, Robot Attitude, and Attention). Stratifying analyses by SAR-VR system allowed exploration of whether cognitive status was differentially associated with engagement depending on system. Because these analyses were exploratory rather than part of a prespecified family of confirmatory comparisons, Bonferroni correction was not applied. Applying a conservative correction across these exploratory associations could substantially increase Type II (false negative) error. Accordingly, effect sizes were interpreted alongside unadjusted p-values when evaluating these associations. Observed Task Difficulty Spearman correlations were computed separately for each SAR-VR system to examine associations between researcher-observed Cognitive Difficulty and Dexterity Difficulty ratings and participant-reported Activity Attitude and Robot Attitude. Analyses were stratified by SAR-VR system to explore whether the relationship between task demands and participant attitudes differed depending on system. Given the exploratory nature of these analyses and the number of related comparisons conducted within each SAR-VR system, Bonferroni correction was not applied to reduce the risk of inflating Type II error. Accordingly, Spearman’s ρ and unadjusted p-values were considered when interpreting these associations. Results Participants A total of 71 older adults were initially enrolled in the SAR–VR arm of the larger trial. Six participants withdrew from the study due to acute illness, relocation from the facility, or loss of interest. An additional five participants had incomplete observational data or an incomplete robot acceptance measure and were excluded from the present analyses. The final analytic sample consisted of 60 participants across 23 long-term care sites. Participant enrollment in the SAR-VR arm within individual sites was relatively sparse, with 13 sites contributing two or fewer participants and 17 sites contributing three or fewer participants. The sample included 42 females and 18 males. Participants had a mean age of 84.8 years (SD = 8.32), with ages ranging from 63 to 101 years. Descriptive Analysis A total of 756 sessions were completed across both SAR-VR system configurations, with similar distribution between the Dog (Aibo) and Humanoid (NAO), as seen in Table 1. Participation was high across the study, with participant-averaged attendance rates above 97% for both SAR-VR system configurations, indicating consistent participation and enabling direct comparisons between the systems without concern for differential participation or adherence effects. Table 1Session Attendance by SAR-VR System Configuration (N = 60 Participants; 756 Sessions)MeasureOverallDog (Aibo)Humanoid (NAO)Total sessions attended (n)756380376Mean sessions per participant12.67 (± 3.01)6.336.27Mean attendance rate per participant (%)–97%98% Session Attendance by SAR-VR System Configuration (N = 60 Participants; 756 Sessions) Baseline characteristics (Table 2) indicate that participants represented a range of cognitive functioning (SAGE: M = 7.46, SD = 4.72) and attitudes toward robots (RAS: M = 103.52, SD = 34.41). One-third of the participants were strongly favorable and accepting of robots (RAS scores 25–85), 38% were moderately accepting (RAS scores 86–115), and 28% were hesitant or not accepting (RAS scores 116–175). Table 2Baseline Cognition and Acceptance of Robots (N = 60)MeasureMean (SD)RangeSAGE (Cognitive Status)a7.46 (± 4.72)0–20RAS (Robot Acceptance)103.52 (± 34.41)33–175Notes: aSAGE scores 17–22 indicate normal thinking, scores 15–16 indicate mild impairment, and scores < 15 indicate potential dementia. Baseline Cognition and Acceptance of Robots (N = 60) Notes: aSAGE scores 17–22 indicate normal thinking, scores 15–16 indicate mild impairment, and scores < 15 indicate potential dementia. Engagement and Interaction Patterns Across both SAR-VR systems, participants demonstrated generally positive engagement with the SAR–VR system. The observed mean Activity Attitude scores were similar across systems (Dog: M = 4.80; Humanoid: M = 4.63), indicating moderately positive evaluations of the activity in both systems. Observed Attention to the activity was high for both systems (Dog: M = 5.87; Humanoid: M = 5.74), suggesting sustained engagement regardless of system. Observed Robot Attitude scores showed greater descriptive separation, with higher average ratings in the Dog system (M = 5.17) compared to the Humanoid system (M = 4.62). These descriptive engagement outcomes are illustrated in Figure 4. Observed attitudes and attention averages by SAR-VR system. To evaluate how observed attention, activity attitude, and robot attitude varied across the four-week per SAR-VR system deployment period, session-by-session trajectories were examined for attention, activity attitude, robot attitude, and attendance rate across interaction instances (1–8) for both systems (Figure 5). Across the Dog condition, engagement metrics remained relatively stable over time. Attention scores showed minor fluctuations, ranging approximately from 5.7 to 6.1, with no consistent upward or downward trend. Activity attitude remained similarly stable (approximately 4.5–5.2), with a slight increase observed at interaction instance 6, followed by a return to baseline levels. Robot attitude demonstrated modest variability, with a small peak at instance 6 but otherwise consistent scores across sessions. Attendance rates remained high throughout the deployment (approximately 0.9–1.0). A similar pattern was observed in the SAR-VR Humanoid system. Attention scores remained stable across interaction instances (approximately 5.6–6.0), with minimal variability. Activity attitude and robot attitude followed parallel trajectories, both showing small increases at instance 6 and slight declines thereafter, but without evidence of systematic change over time. Attendance rates remained consistently high (approximately 0.9–1.0). Across both systems, no monotonic decreases in attendance, attention, attitudes, or attendance were observed over the eight interaction instances. Mean attendance rates, attention, activity Attitude, and robot Attitude across sessions. To further evaluate these patterns, linear mixed-effects models were conducted (Table 3). Session number was not a significant predictor of attention (β = 0.027, SE = 0.020, p = 0.160) or activity attitude (β = 0.009, SE = 0.019, p = 0.651), and showed only a marginal effect for robot attitude (β = 0.037, SE = 0.020, p = 0.068). There was no significant interaction between session number and SAR-VR system condition across outcomes, indicating that engagement trajectories did not differ between the Dog and Humanoid systems. Within the linear mixed-effects model, a significant main effect of condition was observed for both activity attitude (β = 0.381, SE = 0.154, p = 0.014) and robot attitude (β = 0.332, SE = 0.160, p = 0.038), with lower ratings in the Humanoid system condition. Table 3Linear Mixed-Effects Models Predicting Engagement OutcomesOutcomePredictorβSEpAttentionSession (RepeatInst)0.0270.0200.160Condition (Humanoid)−0.0680.1560.622Session × Condition−0.0230.0260.320Activity AttitudeSession (RepeatInst)0.0090.0190.651Condition (Humanoid)−0.3810.1540.014*Session × Condition0.0150.0230.505Robot AttitudeSession (RepeatInst)0.0370.0200.068Condition (Humanoid)−0.3320.1600.038*Session × Condition−0.0460.0240.056Notes: β = fixed effect estimate; *p <0.05.Abbreviation: SE, standard error. Linear Mixed-Effects Models Predicting Engagement Outcomes Notes: β = fixed effect estimate; *p <0.05. Abbreviation: SE, standard error. Sensitivity analyses accounting for partner type and study site did not materially alter these findings. For attention, session number, SAR-VR system condition, and their interaction remained nonsignificant after adjustment. The significant condition effects for activity attitude and robot attitude were also retained after adjustment for partner type and after additionally accounting for site-level clustering. Partner type was not significantly associated with any of the three engagement outcomes. Participant-level intraclass correlation coefficients were 0.388 for attention, 0.346 for activity attitude, and 0.320 for robot attitude, indicating meaningful within-participant correlation across repeated observations. Comparisons Across SAR-VR System Type Activity Attitude showed a small difference favoring the Dog system (p = 0.0279, r = 0.28), but this effect did not remain statistically significant after Bonferroni correction. In contrast, Robot Attitude differed strongly by SAR–VR system, with significantly higher ratings for the Dog (p = 7.72 × 10−7, r = 0.64), indicating a consistent preference across participants. Detailed within-participant comparison statistics are presented in Table 4. This pattern was further reflected in preference counts derived from Robot Attitude ratings: 46 participants (76.7%) preferred the Dog, 11 (18.3%) preferred the Humanoid, and 3 (5.0%) showed no preference. Table 4Within-Participant Comparisons Across Type of SAR-VR System for GOME MeasuresOutcomeDog MeanHumanoid MeanZprBonferroni SigActivity Attitude4.804.632.200.02790.28NoRobot Attitude5.174.624.947.72 × 10−70.64YesAttention5.875.741.390.1630.18No Within-Participant Comparisons Across Type of SAR-VR System for GOME Measures The target of conversation for each participant further characterized engagement during sessions. Analysis was limited to only those sessions that consisted of two older adults, as described in the Methods, resulting in a sample of 55 older adults. Five participants were excluded due to uneven distribution of participants within a site (ie odd number of participants), which resulted in no available sessions in at least one of the SAR-VR systems where their partner was an older adult participant. These 55 older adults completed 285 sessions with the Dog SAR-VR configuration system and 291 sessions with the Humanoid SAR-VR configuration system for the target of conversation analysis. Participants directed speech one or more times toward the older adult partner in both conditions (Dog: M = 0.680; Humanoid: M = 0.602). There was no significant difference between SAR-VR system conditions for partner-directed speech (Z = 1.77, p = 0.076), indicating that peer engagement remained stable across SAR-VR systems. Speech directed toward the robot itself occurred in more sessions with the Dog condition (M = 0.415) compared to the Humanoid condition (M = 0.282). This difference was significant at the uncorrected level (Z = 2.66, p = 0.0077, r = 0.36), but did not remain significant following Bonferroni correction (α = 0.0071). Speech directed toward research personnel was also frequently observed across both conditions (Dog: M = 0.818; Humanoid: M = 0.768); however, this difference did not reach statistical significance (p = 0.115). Speech directed toward staff, family members, or instances categorized as “cannot tell” or “did not talk” occurred at low proportions in both conditions and did not significantly differ between SAR-VR system types (all p > 0.05; none survived Bonferroni correction). To examine relationships among conversational targets, pairwise Spearman correlations were computed across talk target proportions. There was no evidence of an inverse relationship between research-directed and resident-directed speech (Dog: ρ = 0.21, p = 0.13; Humanoid: ρ = 0.27, p = 0.049, not significant after correction). Robot-directed and resident-directed speech were positively associated (Dog: ρ = 0.33, p = 0.013; Humanoid: ρ = 0.23, p = 0.096). Association Analyses Baseline Robot Acceptance and Robot Attitudes Spearman rank correlations were computed to examine whether baseline robot acceptance (RAS) was associated with post-interaction Robot Attitude ratings for each SAR-VR system. Lower RAS scores reflect greater initial acceptance. Accordingly, negative correlations indicate that participants with greater baseline acceptance tended to demonstrate more positive Robot Attitude Ratings. Baseline robot acceptance was modestly associated with post-interaction Robot Attitude for both systems. For the Dog system, RAS scores were negatively correlated with Robot Attitude (ρ = − 0.276, p = 0.0325), indicating that participants with more accepting baseline views tended to rate the Dog robot more positively. A similar pattern was observed for the Humanoid system (ρ = −0.265, p = 0.0406). However, after Bonferroni correction for the two correlation tests (α = 0.025), neither association remained statistically significant. These findings suggest that baseline acceptance was not significantly associated with participants’ attitudes towards the robot during engagement with either robot. Spearman correlations were conducted to assess the relationship between baseline anthropomorphism (HumanCharacteristics) and engagement outcomes within each SAR-VR system. Higher HumanCharacteristics scores reflect lower anthropomorphism. Across both SAR-VR systems, all correlations were negative, indicating that participants who anthropomorphized the robot to a greater extent prior to interaction demonstrated more positive engagement during the sessions. In the Dog condition, anthropomorphism was significantly associated with Activity Attitude (ρ = − 0.373, p = 0.0033), representing a moderate relationship that remained significant following Bonferroni correction (α = 0.0083). Associations with Robot Attitude (ρ = − 0.278, p = 0.0317) and Attention (ρ = − 0.240, p = 0.0646) followed the same directional trend but did not remain significant after correction. In the Humanoid condition, anthropomorphism was significantly associated with Activity Attitude (ρ = − 0.325, p = 0.0113), Robot Attitude (ρ = − 0.329, p = 0.0103), and Attention (ρ = − 0.262, p = 0.0435). Although these relationships were consistent in direction, none remained significant after correction for multiple comparisons. Baseline Cognitive Status (SAGE) and Engagement Outcomes Spearman correlations were conducted to examine associations between baseline cognitive functioning (SAGE) and engagement-related outcomes. For the Dog condition, baseline cognitive status was positively associated with observed Attention (ρ = 0.302, p = 0.0190), indicating that higher cognitive functioning was related to greater sustained attention during sessions. No significant associations were observed between SAGE scores and Activity Attitude (ρ = 0.166, p = 0.204) or Robot Attitude (ρ ≈ 0.000, p = 0.9996). For the Humanoid condition, SAGE scores showed a similar pattern. Baseline cognition was positively associated with Attention (ρ = 0.275, p = 0.0332), while associations with Activity Attitude (ρ = 0.248, p = 0.0565) and Robot Attitude (ρ = 0.181, p = 0.167) were not statistically significant. Observed Task Difficulty and Attitudes Associations between researcher-observed task difficulty and attitudinal outcomes were examined separately by SAR-VR system using Spearman correlations. For the Dog condition, greater observed cognitive difficulty was positively associated with Activity Attitude (ρ = 0.258, p = 0.0433), suggesting that participants who were rated as experiencing greater cognitive challenge nonetheless reported more positive attitudes toward the activity. Cognitive difficulty was not associated with Robot Attitude. Greater observed dexterity difficulty in the Dog condition was associated with less positive attitudes toward both the activity and the robot (correlation coefficients and p-values consistent with prior analyses), indicating that motor-related challenges may have negatively influenced evaluative impressions specifically in the Dog-based system. In contrast, for the Humanoid-based system, no significant associations were observed between cognitive or dexterity difficulty and attitudinal outcomes. Discussion We designed two SAR-VR system configurations based on our earlier studies and interdisciplinary theories of engagement and interaction as a nonpharmacological treatment to address apathy in older adults with cognitive impairment. Both systems were designed as a socially embedded dynamic process involving more than one person and of similar duration and exposure (4 weeks/2 sessions per week for each system). Both systems required three elements: paired and interdependent activities, communication to enhance collaboration and interaction, and reward mechanism. Each activity required cognitive, physical and social dimensions. The systems differed, however, in SAR type (humanoid or dog), robot role (coach or companion), interaction architecture, feedback modality, and task type. Accordingly, differences observed between conditions should be interpreted as differences between the overall SAR-VR configurations rather than isolated effects of robot embodiment or any single system feature. Results demonstrated both systems worked as designed in real-world long-term care settings involving older adult residents who had symptoms of apathy and mild to moderate cognitive impairment. Moreover, the results remained stable over the 8 weeks of exposure, with engagement scores remaining consistently high for attention (Dog: M = 5.87; Humanoid: M = 5.74), moderately positive for activity attitude (Dog: M = 4.80; Humanoid: M = 4.63), and positive for robot attitude (Dog: M = 5.17; Humanoid: M = 4.62), and no significant effects of session number observed in the linear mixed-effects models. Importantly, differences between conditions should be interpreted as differences between the overall SAR-VR configurations rather than as isolated effects of robot embodiment. The systems differed across multiple dimensions, including robot embodiment, robot role, interaction architecture, feedback modality, and task type. These features could not be independently disentangled in the present study. Sensitivity analyses accounting for partner type and site-level clustering did not materially alter these primary findings. Several explanations come to mind for the positive findings. First, the SARs themselves may have been perceived as pleasant companions for the older adults. We selected NAO (humanoid) and Aibo (dog), based on prior work of older adults’ acceptance of these two SARs as well as their open architecture for software programming.44,45,66–68 Older adults in general show positive reactions to NAO, often perceiving it as entertaining and non-threatening and providing meaningful activities.25,32,44,45,49 The Aibo puppy by Sony is highly interactive and simulates life-like dog behavior.75 A potential mechanism for its positive acceptance among older adults is the affect-generating aspect of the human-animal bond often found in pet therapy as well as providing engaging activities.76 Indeed, Krueger and colleagues75 posit that studying human-animal behaviors, such as attributing emotions or personality onto a pet, will guide the ongoing development of SARs to serve as companions of humans in the future. Second, the duration and exposure to each SAR-VR system bears comment. Studies of SARs with older adults have varied in their duration and exposure and ranged from 3 to 12 weeks, one to three sessions per week, and 10 to 30 minutes per session, all dependent upon the focus of the study outcomes, such as fall prevention exercises, care support and assistance, or social engagement.77 Our choice of 4 weeks by SAR-VR system balanced pragmatics of conducting the study with sufficient exposure to determine sustained interest in the SAR-VR activities — always a challenge when dealing with individuals who have apathy. Future studies are needed with longer and more frequent exposure to SAR-VR systems. Third, the activities themselves may have contributed to the sustained interest. Following the key principles for enhancing engagement among older adults with apathy and cognitive impairment, all activities were multi-modal in nature that incorporated both cognitive and physical aspects. Second, a variety of activities were offered that could tap into a gaming element (eg, the fishing game), problem-solving challenge (eg, spelling or path finding), or artistic elements (eg, music, painting). Offering a variety of activities capitalized on individuals’ preferences and functional abilities.64 Last, the activities were designed to enhance social connectivity since only by collaboration could one successfully complete the task; participation in social informal activities has been shown to be particularly effective in promoting psychosocial health and well-being (79)). Our results demonstrated that human–human interaction (HHI), whether directed toward another older adult partner or toward research personnel, occurred in the majority of sessions and remained stable across SAR-VR system conditions. This suggests that different SAR-VR configurations did not substantially alter patterns of conversational engagement or reduce opportunities for peer-directed interaction. A notable observation across both conditions was the high proportion of sessions in which speech was directed toward research personnel. Because the present analysis captured only the occurrence of speech and not its content or intent, it is not possible to determine whether this reflected task-related assistance, clarification, or broader social interaction. This distinction warrants further investigation. Additionally, we assessed relationships between conversational targets at the participant level. These analyses did not indicate an inverse relationship between research-directed and resident-directed speech, and associations were small and positive in direction. Similarly, robot-directed speech was positively associated with resident-directed speech, suggesting that engagement with research personnel or the SAR-VR system did not occur at the expense of peer interaction. Human–robot interaction (HRI) was also frequent and occurred in a greater proportion of sessions in the Dog SAR-VR system. This may reflect differences in robot role or user response to the system. Across both systems, however, HRI did not replace HHI, supporting the interpretation that SAR-VR systems can provide opportunities for social interaction without displacing peer-directed communication. The high frequency of research-directed speech also highlights the likely role of facilitation in supporting SAR-VR sessions. From an implementation perspective, this underscores the importance of considering the level and type of support needed for implementation in long-term care. Future work should incorporate more detailed coding of conversational content and interaction dynamics to better distinguish peer-driven from staff-mediated engagement. The difficulty level of the activities was designed as an increase in cognitive load and greater physical dexterity. In the humanoid SAR-VR system, no significant associations were observed with the level of cognitive or dexterity difficulty and attitudinal outcomes. However, for the dog SAR-VR system, greater observed dexterity difficulty was associated with less positive attitudes towards the activity and the dog robot. This pattern may relate to differences in how instruction and support were delivered across the two system configurations, although contribution of these features cannot be separated from other differences between the systems. In the humanoid condition, the NAO robot provided direct instruction, feedback, and physical demonstration of task-relevant movements, offering more embodied guidance, whereas in the dog condition, instruction and feedback were delivered by an on-screen avatar, with the robot functioning primarily as a motivational and reward-based agent without physically demonstrating the actions. For the dog SAR-VR, cognitive difficulty resulted in more positive attitudes toward the activity. It may be that cognitive difficulty could be successfully overcome given the social interactive collaborative nature of the activity and the entertainment function of the dog. Future studies will need to examine the role of the robot in relationship to the activity characteristics, such as cognitive difficulty and dexterity issues. We are unable to comment on which particular activities were most successful in engagement given we designed the study to enhance participant agency in choosing both the specific activity as well as level of difficulty. Indeed, researchers have noted that tailoring activity content and delivery is more important than the activity type in enhancing well-being among persons living with dementia.64,78,79 Baseline robot acceptance and anthropomorphism demonstrated modest associations with observed robot attitudes and may play an important role in shaping engagement with socially assistive robots. Participants who entered the interaction with a greater tendency to attribute human-like and dog-like qualities to the robots demonstrated more positive attitudes toward both the activity and the robot itself. This pattern was consistent across both systems, indicating that anthropomorphism may act as a general facilitator of engagement rather than being specific to a particular robot form. Baseline cognition was also positively associated with observed attention for both systems but was not meaningfully associated with attitudinal evaluations of the activity or the robot. Others have reported on baseline technology (robot) acceptance, anthropomorphism and cognition influencing the interaction with SARs; additional older adult characteristics include higher education, being a woman, prior technology experience and convenience.50,63,75,77 Limitations Several limitations should be considered when interpreting these findings. First, the precise ways in which the two systems were successful as a nonpharmacological approach to engage individuals with apathy requires further examination. The process of engagement is influenced by a complex interplay of the individual’s characteristics, the activity’s characteristics, the environment’s characteristics, and the social characteristics.1,6,8–10,80,81 Greater insight in understanding these factors will enhance architectural planning and design. Second, our study was designed as a short-term intervention. Studies examining the frequency and duration of SAR-VR interaction is needed to determine the exposure necessary to achieve success. Future work employing longitudinal modeling approaches may further clarify how engagement evolves with sustained exposure. Third, engagement measures relied on researcher-observed ratings and coded conversational targets. While these measures provide ecological validity within real-world care environments, they remain subject to observer interpretation. Furthermore, several members of the research team were involved in the design, development, implementation, and evaluation of the SAR-VR system. Although standardized training, observational protocols, and cross-site review procedures were used to promote consistency, this investigator involvement may have introduced potential observer or interpretive bias and should be considered when interpreting the positive findings. In addition, conversational coding identified who participants directed their speech toward, but did not capture the content, quality, reciprocity, or intent of those interactions, limiting conclusions about the nature or meaningfulness of the social exchanges. The integration of complementary objective measures—such as gaze tracking, physiological indicators, or automated speech analysis—would strengthen future investigations of engagement processes. Associations involving baseline robot acceptance, anthropomorphism, cognition, and task difficulty should also be interpreted as exploratory relations. Finally, although we found similar degree of engagement between the two systems, future work is needed to better understand and clarify those most likely to benefit from types of robots and VR activities. Comparisons were limited to two SAR-VR configurations that differed across multiple bundled characteristics, including robot embodiment, robot role, activity type, feedback modality, and interaction architecture. Accordingly, observed differences between conditions cannot be attributed to any single system component. Broader comparisons that systematically vary individual system features are needed to isolate their independent contributions to engagement. Although sensitivity analyses accounting for partner type and site-level clustering did not materially alter the primary findings, larger samples within individual sites would allow these contextual factors to be examined more directly. These findings should therefore be interpreted as evidence of sustained engagement during repeated SAR-VR use rather than evidence of clinical efficacy or isolated robot-embodiment effects. Beyond these limitations, future research should examine the SAR–VR approach over substantially longer periods and with larger cohorts within individual long-term care sites. Although engagement remained stable across the eight-week intervention, longer deployments are needed to determine whether engagement and acceptance are sustained after prolonged exposure and to further distinguish persistent effects from potential novelty effects. Larger intervention and control cohorts would also enable more robust examination of site-level variability and participant characteristics that may influence engagement. Future studies should therefore account for individual factors such as age, cognitive and physical function, sensory ability, prior technology experience, and medication use or changes; intervention parameters including the type of robot, role of the robot, and activity type; and environmental factors such as interaction partner, time of day, and staff involvement. An important next step will also be to evaluate the system with progressively reduced researcher involvement to determine the level of staff facilitation required for routine use. Such longer-term implementation studies should consider not only engagement and acceptance, but also adherence, technical reliability, staff burden, resource requirements, and scalability to establish the feasibility of sustained SAR–VR deployment in long-term care settings. Conclusion Within long-term care settings, where dementia, apathy, and social withdrawal are prevalent, there is a critical need to examine ways for technology to complement personnel in the delivery of activities to engage older adults and contribute to their quality of life. SAR-VR systems intentionally structured around integrated cognitive, physical, and social components may provide pathways for sustaining reciprocal engagement among older adults with cognitive impairments within long-term care environments. Through an eight-week deployment of paired SAR–VR systems in long-term care settings, engagement was examined using observed measures of attention, activity attitudes, robot attitudes, and conversational targeting across two SAR-VR system configurations. By focusing on user-level engagement during sustained real-world implementation, these findings extend beyond feasibility to explore how socially scaffolded robotic–virtual systems function in practice. Importantly, human–human interaction remained central across SAR-VR systems, reinforcing the conceptualization of engagement in this context as inherently social. When designed to distribute responsibility across pairs and to prioritize participation over performance accuracy, SAR–VR systems can create opportunities for reciprocal and peer-directed interaction without displacing peer exchange. This human-human component is particularly important within long-term care, where cognitive impairment and apathy may reduce spontaneous initiation, conversational engagement, and participation in shared activities. In this context, the value of SAR-VR may extend beyond engagement with the technology itself by providing a structured setting in which residents are prompted to attend to, respond to, and coordinate with one another. Rather than replacing interpersonal interaction, the technology may therefore function as a scaffold for peer engagement by creating shared goals, opportunities for reciprocal exchange, and repeated occasion for social participation. Together, these findings demonstrate that SAR–VR systems can be designed to engage older adults with apathy in dementia care through the integration of socially assistive robots, collaborative virtual activities, and structured social interaction. Because the evaluated systems differed across multiple design features, these findings support the overall SAR-VR configurations rather than the independent effect of any single system component. In long-term care environments, where sustaining initiative and reciprocal interaction can be challenging, thoughtfully designed SAR–VR systems may serve as an important tool for staff in providing the crucial meaningful social participation for their older residents. Acknowledgments The authors wish to thank Abby Kilpatrick, Sadie Hunter, Courtney Grisez, Susan McCabe, and Kelley Colopietro for assisting with research activities and the long-term care settings that graciously allowed us access to their environments. Funding Statement The study was supported by the National Institute on Aging of the National Institutes of Health under award number R01AG062685. All authors are supported by National Institutes of Health grant funding. MM and NK received financial support from the National Science Foundation. REDCap is supported in part by the National Institutes of Health (NIH/NCATS) UL1 TR000445 at Vanderbilt University. Abbreviations ARIA, Adaptive Robotic-Mediated Intervention Architecture; EEG, electroencephalography; GOME, Group Observational Measurement of Engagement; LTC, long-term care facilities; RAs, research assistants; RAS, Robot Acceptance Survey; SAR-VR, socially assistive robot–virtual reality; SARs, socially assistive robots; UC, usual care; UTAUT, Unified Theory of Acceptance and Use of Technology; VR, virtual reality. Author Contributions Judith A Tate, Lorraine C Mion and Nilanjan Sarkar share last authorship. All authors made a significant contribution to the work reported, whether that is in the conception, study design, execution, acquisition of data, analysis and interpretation, or in all these areas; took part in drafting, revising or critically reviewing the article; gave final approval of the version to be published; have agreed on the journal to which the article has been submitted; and agree to be accountable for all aspects of the work. Disclosure Dr Miroslava Migovich reports meeting/travel support from National Institute on Aging of the National Institutes of Health, outside the submitted work. The author(s) report no other conflicts of interest in this work. References