Evidence review · Updated July 2026
What clinical trials actually tell us about companion robots
Companion robots have produced encouraging results in some studies of dementia care, loneliness and social engagement. But the evidence is not a simple success story: benefits vary by outcome, device, setting and comparison group, and many trials are small or difficult to blind.
What counts as companion-robot research?
This review focuses on embodied social, companion and therapeutic robots designed to interact with people. That includes pet-like devices such as PARO and social robots such as Pepper when they are studied for engagement, wellbeing or supportive care.
It does not combine those studies with surgical robots, powered exoskeletons, industrial robots or ordinary telehealth apps. Those technologies have different mechanisms, risks and clinical outcomes. Even within companion robotics, a pet-like robot used in a nursing home is not equivalent to a conversational device used by an older adult living alone.
The evidence at a glance
Short-term engagement and agitation
Several controlled studies in dementia care report improvements in observed engagement, mood or agitation. Results are not identical across scales or comparison groups.
Loneliness and emotional wellbeing
Some recent trials report benefits, but interventions differ substantially and may include communication with family members or human operators.
Cognition, sleep and durable clinical change
Evidence is inconsistent or limited. Positive findings over a few weeks do not establish long-term disease modification or independent clinical effectiveness.
The safest overall conclusion is that some companion-robot interventions can be useful as structured, non-pharmacological activities. Current evidence does not support treating the entire product category as a proven medical replacement for human care.
Four trials worth understanding
These studies illustrate both the potential of companion robots and the reasons cautious interpretation is necessary.
PARO in long-term dementia care
Design: 415 residents across 28 long-term-care facilities were assigned by facility to PARO, a similar plush toy or usual care. Sessions lasted 15 minutes, three times weekly, for 10 weeks.
Finding: Video observations favored PARO over usual care for several engagement, pleasure and agitation measures. However, the standard agitation questionnaire did not show a difference between groups, and PARO was better than the plush toy primarily for engagement.
Takeaway: the active plush-toy comparator shows how touch, novelty and attention can explain part of the effect.
Group-based PARO for mild dementia
Design: Adults aged 65 or older with mild dementia were recruited from four dementia day-care centers and assigned to a six-week group intervention or control condition.
Finding: The researchers reported improvements across several cognitive, psychological and physiological measures. Psychological changes were still observed one month later, while reported physiological effects had diminished.
Takeaway: follow-up timing matters; different outcomes may not persist equally.
CARESSES and culturally responsive interaction
Design: A culturally adaptive system running on Pepper was tested in care homes in England and Japan. Thirty-three residents completed a two-week intervention comparing culturally adaptive, limited-adaptation and usual-care conditions.
Finding: The study reported a signal for emotional-wellbeing improvement, but it was explicitly exploratory and small.
Takeaway: personalization and cultural context may influence acceptance, yet larger and longer studies are needed.
Social robots and loneliness at home
Design: Seventy-three community-dwelling Japanese adults aged 65 or older who lived alone and experienced loneliness were randomized in a four-week study using BOCCO emo.
Context: The robot supported conversation with human operators and/or family members and provided reminders. That makes the intervention socially meaningful, but it also means the robot’s embodiment cannot be separated completely from the human contact it facilitated.
Takeaway: ask whether the benefit came from autonomous AI, human connection, reminders, the physical robot—or a combination.
What systematic reviews and meta-analyses found
A single positive trial is not enough to settle a clinical question. Reviews combine results across studies, but their conclusions are only as reliable as the studies and comparisons included.
2023 PARO systematic review and meta-analysis
The review included 12 articles and 1,461 participants aged 60 or older with dementia. It found small pooled effects for anxiety, agitation and depression, a moderate effect for medication use and a negligible effect for total sleep time. Narrative findings suggested possible improvements in apathy and sociability.
The crucial qualifier: the authors graded the evidence for all outcomes as low because of methodological limitations, small samples and wide confidence intervals. Intervention formats also differed, including group versus individual and facilitated versus non-facilitated use.
2025 meta-analysis of randomized dementia trials
This analysis reported pooled improvements in agitation and anxiety. It did not find statistically significant overall effects for cognition, neuropsychiatric symptoms, depression, quality of life, daytime step count or nighttime lying duration.
Interpretation: the results support continued investigation of targeted outcomes, not a claim that intelligent robots improve every aspect of dementia.
Evidence by outcome
The table below is an editorial synthesis of the cited research, not a formal clinical guideline. “Some support” means that controlled studies report a signal, while important uncertainty remains.
| Outcome | Current signal | Important limitation |
|---|---|---|
| Engagement | Observed interaction may increase during structured sessions. | Novelty, staff attention and comparison with a toy can influence the result. |
| Agitation | Several dementia studies and pooled analyses report improvement. | Results can differ between video observation and standardized questionnaires. |
| Anxiety | Small pooled effects have been reported for PARO. | Overall evidence quality has been rated low. |
| Loneliness | Recent community studies provide encouraging signals. | Some interventions facilitate human contact, making the active ingredient difficult to isolate. |
| Depression or mood | Some individual and pooled results are positive. | More recent pooled dementia analyses have not found a consistent overall depression effect. |
| Cognition | Individual trials sometimes report changes. | No consistent evidence that companion robots slow disease or improve cognition broadly. |
| Sleep | The 2023 PARO review found a negligible pooled effect. | Few studies and differing measurement methods. |
| Quality of life | Small exploratory studies show possible benefit. | Short duration and small samples limit confidence. |
Why companion-robot trials are difficult to interpret
Blinding is difficult
Participants and staff can usually see whether a robot is present. Expectations may change behavior and reporting.
The comparator changes the question
Usual care tests the whole experience. A plush toy, tablet or non-robotic version helps isolate embodiment or interactivity.
Novelty can fade
A two-week improvement may reflect excitement. Longer follow-up is needed to test continued use and durable benefit.
Facilitation varies
A staff-led session is not the same intervention as leaving a robot available without facilitation.
Products are heterogeneous
Pet-like robots, humanoids and communication devices use different interaction models and should not be treated as one intervention.
Outcome scales differ
Observed behavior, self-report, caregiver ratings and clinical scales can produce different answers to similar questions.
How to read a companion-robot clinical trial
- Identify the population.Age, diagnosis, living arrangement and baseline severity determine who the results may apply to.
- Define the intervention.Check the exact robot, session length, frequency, facilitator and whether human contact was included.
- Inspect the comparator.Usual care, a tablet, a plush toy and another activity answer different questions.
- Find the primary outcome.A trial may measure many outcomes; the pre-specified primary outcome carries more weight than an isolated positive secondary result.
- Look beyond statistical significance.Ask whether the size of change was meaningful to participants and caregivers.
- Check duration and follow-up.Short exposure can test feasibility but cannot establish long-term effectiveness or safety.
- Review registration and funding.Prospective registration, disclosed analysis plans and transparent conflicts make selective reporting easier to detect.
Questions before using a robot in care
Evidence from a research setting does not automatically establish that a consumer product is suitable for a particular person. Families, clinicians and care organizations should evaluate both the intervention and the operational risks.
- Goal: Is the robot intended for engagement, reminders, communication or a defined therapeutic activity?
- Fit: Was the same device studied in a similar population and setting?
- Consent: Can the person understand and accept the interaction, recording and data practices?
- Supervision: Who monitors distress, misunderstanding, falls, malfunction or inappropriate responses?
- Privacy: Where are audio, video and interaction logs processed and stored?
- Clinical boundaries: Is it clear that the robot cannot diagnose, handle emergencies or replace necessary human care?
- Evaluation: What outcome will be monitored, over what period, and what would cause use to stop?
- Total cost: Are subscriptions, maintenance, training and staff time included in the decision?
The World Health Organization’s guidance for AI in health emphasizes autonomy, wellbeing, transparency, accountability, equity and sustainability. Those principles are relevant even when a device is described as a companion rather than a medical system.
Frequently asked questions
Are companion robots clinically proven?
Some specific robots and structured interventions have supportive trial evidence for selected outcomes. The product category as a whole is not proven for every claim, population or setting.
Which companion robot has the most research?
PARO is among the most extensively studied therapeutic companion robots, particularly in dementia and older-adult care. That evidence should not be transferred automatically to unrelated consumer robots.
Can a robot reduce loneliness?
Some studies report improvement, including recent research with community-dwelling older adults. Results depend on how the robot is used and whether it facilitates contact with other people.
Can a companion robot replace a caregiver or therapist?
No. Current research generally evaluates robots as supplementary activities or support tools, not replacements for necessary human care, diagnosis or treatment.
Does a positive trial mean I should buy the product?
Not by itself. Confirm that the tested version, population, goal and setting resemble your situation, then consider privacy, support, total cost and professional advice where appropriate.
How this evidence review was prepared
We prioritized randomized trials, systematic reviews, PubMed records, open-access research reports, ClinicalTrials.gov and World Health Organization guidance. We removed unsupported institutional claims and avoided converting preliminary findings into guaranteed clinical benefits. This is a narrative evidence review, not a formal systematic review.
Research and guidance cited
- Moyle et al.: cluster-randomized PARO dementia trial
- Chen et al.: group-based PARO randomized trial
- CARESSES randomized trial report
- BOCCO emo loneliness study
- 2023 PARO systematic review and meta-analysis
- 2025 meta-analysis of intelligent-robot dementia trials
- ClinicalTrials.gov: Pepper robot versus tablet study
- WHO: Ethics and governance of artificial intelligence for health
