- The paper demonstrates the feasibility of autonomous, in-home individualized Cognitive Stimulation Therapy (iCST), with 9 participants completing 31 of 63 scheduled sessions during a one-week deployment.
- The paper finds that 6 of 9 households met the benchmark of at least two sessions per week, while caregiver availability, technical failures, interface barriers, and competing priorities shaped adherence.
- The paper shows that effective robot-delivered dementia care requires personalization, responsive dialogue, family-aware interaction, reliable infrastructure, and transparent communication about privacy, but longer trials are needed to establish clinical benefits.
Motivation and positioning
Cognitive Stimulation Therapy (CST) is the leading non-pharmacological intervention recommended in UK and Canadian dementia care guidelines, with meta-analytic evidence of cognitive gains roughly equivalent to a six-month delay in decline for mild-to-moderate dementia. Individualised CST (iCST) extends this to one-to-one delivery by professionals or family caregivers, but a large multicentre RCT reported that only 40% of iCST dyads achieved the prescribed two or more sessions per week, due to caregiver time constraints, stress, and lack of training. This adherence gap between controlled efficacy and in-home practice motivates the study: can an autonomous socially assistive robot (SAR) deliver structured iCST reliably in the home?
The paper claims two contributions: (1) the first in-home deployment of an autonomous social robot delivering evidence-based CST (as opposed to generic games, reminders, or chat), and (2) a detailed behavioural and contextual analysis of what drives multi-session engagement, family involvement, and adherence. Prior robot studies largely targeted older adults without diagnosis or with MCI, took place in facilities rather than homes, or delivered unstructured content; only CARMEN offered autonomous in-home interaction, and that was compensatory cognitive training for MCI, not CST for diagnosed PwDs.
System design
Co-STAR integrates a Misty II robot, a tablet interface, a tabletop microphone, a mini-CPU, and a local router, all housed in a custom enclosure; no video is collected and all data remain on-device (privacy-by-design). Speech output uses OpenAI's TTS, while speech recognition uses WhisperD, a Whisper variant fine-tuned for disfluencies and filler words characteristic of dementia speech. Robot dialogue was pre-generated before deployment — an important constraint given later findings about perceived scriptedness.
Five iCST activities were implemented: Popular Places, Famous Faces, Common Sayings, Word Association, and Object Categorisation, targeting memory recall, language activation, and semantic processing. Sessions were designed at 30 minutes, initiated on demand via the tablet. Personal details collected before deployment (travels, jobs, hobbies) were embedded into activity content. The design follows seven principles from prior stakeholder consultations with 16 dementia stakeholders, including "connection over correction," privileging opinions over facts, autonomy, and encouraging family participation.
Study and usage results
Nine clinically diagnosed PwDs (mean age 79.7; five female) hosted the robot for one week with a target of one daily session. Participants completed 31 sessions out of 63 scheduled (mean 3.44 per person), with average session length of 32m42s closely matching the 30-minute design goal. Using the common iCST benchmark of ≥2 sessions/week, 6/9 households met it — 75% of those living with family and 60% of those living alone. The authors argue this exceeds adherence typically observed in caregiver-led iCST, though the comparison is informal rather than statistical.
Missed sessions (37 total) had identifiable causes: caregiver unavailability (10), technical failures such as power loss and Wi-Fi instability (8), loss of interest (7), interface difficulties including swipe-to-unlock barriers (6), and competing life priorities (6). The authors note that adherence was often limited by infrastructure rather than motivation, implying a higher ceiling with improved reliability — but this remains an inference, since motivation and technical failure are not cleanly separable in individual cases.
Quantitative measures were positive but based on incomplete samples. Six participants returned SUS responses, averaging 80.8 (SD ≈ 25.1, range 40–100), above the conventional benchmark of 68, though the large variance signals uneven experience. GAToRS subscales from eight participants showed low negative personal attitudes (M = 1.00, SD = 0.00), moderately positive personal trust/attitudes (P+ M = 4.53), high perceived societal benefit (S+ M = 6.25), and moderate concern about negative societal impact (S− M = 3.45). Four participants asked to keep the robot beyond the week.
Family members and the triadic reality
A central finding is that in-home therapy is effectively triadic rather than dyadic. Participants living with caregivers completed more sessions (M = 4.3, SD = 3.0) than those living alone (M = 2.8, SD = 2.3). Caregivers initiated sessions, operated the tablet, and brokered engagement: Priscilla, living alone, missed exactly the two days her daughter was unavailable; Henry, whose children visited infrequently, completed zero sessions. Yet cohabitation could also suppress intrinsic motivation when existing social networks already provided stimulation, as Ray articulated.
Family presence also produced therapeutic side effects beyond the protocol: sessions triggered previously unknown memories and intergenerational storytelling (e.g., Ray revealing he had ridden a motorbike before his daughter was born). Several participants sought social validation, asking whether others were using Co-STAR, suggesting collective adoption norms matter for acceptance. The authors derive concrete HRI recommendations: explicit autonomous versus caregiver-assisted modes, and default support for multi-user turn-taking.
Personalisation, trust, and perceived competence
Personalisation emerged as critical for sustained engagement, along several dimensions:
- Adaptive difficulty: educational background shaped preferred challenge; Ray (PhD) found tasks too easy, while Esther sought harder items.
- Content relevance: personalised prompts about places participants had lived produced the longest, most engaged responses.
- Voice accent: the American-accented voice reduced comfort and trust for some British participants, cited explicitly as reducing Julia's willingness.
- Sensory settings: volume and brightness adjustments were repeatedly requested.
Trust concerns mapped to perceived system competence rather than generic robot attitudes. Scripted, non-contingent responses ("It can't think… it's a machine") and slow startup degraded perceived value; one participant's disengagement was attributed directly to latency. Privacy questions ("Is my name going to be everywhere?") also surfaced despite the local-processing design, indicating that privacy-by-design must be made legible to users, not merely implemented.
Case studies
Four contrasting cases illustrate adoption patterns. Gideon (eight sessions, exceeding the weekly target) shows how tech familiarity and intrinsic motivation sustain use even with minor technical faults. Georgia demonstrates that high motivation does not guarantee adherence when competing priorities dominate. Benjamin represents willing users blocked by infrastructure — repeated Wi-Fi disconnections and caregiver difficulty operating the tablet capped him at one session. Priscilla shows a trajectory from apprehension to reliable five-session use with minimal but essential family support, with her daughter reporting mood and engagement gains. Collectively these cases reinforce that adherence is jointly determined by user, caregiver, and system reliability.
Limitations and open questions
The paper concedes its principal limitations plainly: the sample is small (n = 9), the deployment lasted only one week, and no cognitive outcome measures were collected — all claims about benefit rest on self-report, caregiver report, and interaction logs. Quantitative instruments had missing data (SUS n = 6, GAToRS n = 8), and the adherence comparison against human-led iCST is contextual rather than controlled. A researcher present on day one may have inflated early engagement. Whether robot-delivered iCST produces measurable cognitive or quality-of-life gains over months, and whether adherence holds once novelty effects dissipate, remain open empirical questions the authors defer to future long-term deployments.
Conclusion
Co-STAR demonstrates the feasibility of autonomous, unsupervised, in-home delivery of structured, evidence-based iCST to people with dementia, achieving adherence comparable to or better than human-facilitated baselines within a single week. Its most durable contributions are contextual: in-home therapy is triadic, infrastructure failures rival motivation as adherence barriers, personalisation and responsiveness drive trust, and privacy must be communicated as well as engineered. The central unresolved question is whether these engagement patterns translate into sustained clinical benefit over longer deployments with larger cohorts.
