2026-09-18
UNH deployed the MARSS socially assistive robot in five dementia-care households for 1-6 months and co-designed navigation, privacy, and care protocols in place.
Socially assistive robots have spent two decades mostly as companions in supervised facilities. Evidence for a robot that lives in a dementia household and runs care routines on its own, for months, is thin. A lab demo and a survey cannot surface the failures that only show up at home: the robot spinning on a rug, a missed dock at night, a camera that is unacceptable in a bedroom.
A University of New Hampshire team spanning occupational therapy, nursing, kinesiology, and computer science put MARSS (Mobile Assistive Robot with Smart Sensing) into real homes. The hardware is Hello Robot's Stretch 3. The software is theirs: distributed sensors watch people, the house, and the robot; a high-level AI planner decides whether a care service should fire; behavior trees expand that decision into an executable plan. Services include medication, exercise, and meal reminders, weather alerts, and computer-vision detection of getting out of bed. The authors say this appears to be the first study to co-design a SAR in the wild with people living with Alzheimer's disease and related dementias (ADRD) and their caregivers.
Three instruments sit on top of the NIH Stage Model for Behavioral Intervention Development.
The NIH model runs from basic research (Stage 0) through dissemination (Stage V). This paper is Stage IB. Stage IA was a focus-group prototype study. Stage IB is a home feasibility pilot whose job is to harden the technology, the care protocols, and the training before anyone runs an efficacy trial.
UTAUT (Unified Theory of Acceptance and Use of Technology) supplies the interview skeleton: performance expectancy, effort expectancy, social influence, facilitating conditions, anxiety, and trust. The same trained interviewer visited monthly for 45-90 minutes. Four researchers coded transcripts independently, then reconciled to saturation.
GAS (Goal Attainment Scaling) measures whether goals that matter to that household actually moved. Zero is the expected outcome; +1/+2 beat it; -1/-2 miss it. Scores are weighted by importance, difficulty, and baseline. A composite near 50 is treated as a sanity check that goals were set at the right difficulty.
Five spouse dyads, ten people. Four dyads had mild-to-severe ADRD; one volunteer dyad without dementia helped debug hardware early. Three lived in a retirement community, two in their own homes. Trial length ranged from 1 to 6 months. Each robot was programmed to that home's layout, daily rhythm, and care needs. Dyads were excluded if the care recipient was non-compliant or had a history of aggression that threatened people or equipment.
Qualitative analysis produced four co-design themes, each tied to a concrete change.
Technical design: the robot got stuck on rugs, failed to dock, moved unpredictably, and dropped volume until prompts were inaudible. The team improved low-profile obstacle handling, added infrared sensing next to April-tag docking, constrained paths, attached external microphones and speakers, and added a remote data logger. Appearance was deprioritized after a participant said looks do not matter if the job gets done. Images, video, and audio are processed and discarded locally; only transcripts of commands and robot activity stay on site for debugging. A large language model was later added for two-way talk, fenced by a system prompt that keeps replies short, useful, and away from harmful advice.
Care protocol design: clock-based reminders failed against dementia's irregular days, so triggers moved to context (awake, present, task due). High-priority tasks such as medication gained an activity-recognition check and a daily checklist. Cameras in bedrooms and bathrooms were refused and replaced with motion and door sensors.
Training management: families first met the robot in the lab. Caregivers were consistent that care is done by people. The robot is a complement, not a replacement. One caregiver said they would not leave the person with the robot for three days.
Complementary care: value clustered on repetitive chores, medication, dressing for a walk, taking out trash. In Dyad 2 the care recipient learned to use a coffee maker from a video on the robot's screen, took medication without prompts most of the time, and reached +2 on wearing proper attire before walks. Dyad 3's bed-exit detector, lab-tested first, produced two true positives that stopped a stand-up in a person at high fall risk from postural hypotension.
GAS is reported for the three ADRD dyads who lasted beyond two months. Dyad 1 (moderate AD, 4 months) started low because of navigation failures, rigid schedules, and a caregiver who watched reminders without acting; scores rose after docking and navigation fixes, context-triggered protocols, and a reduced exercise target. Dyad 2 (mild AD, 6 months) beat expectation on three of four goals. Dyad 3 used the robot for about one month before relocating for health reasons; both scored goals, medication checklist and bed-exit warning, reached +2.
For people building health AI and embodied care, the contribution is a process, not a leaderboard. Vision, planners, and chat are commodity parts. What actually blocked adoption was reliability, a privacy boundary the family would accept, and protocols that flex with the disease and the day. Stacking the NIH stage model, UTAUT, and GAS gives a reusable co-design loop for later Stage II/III trials.
The product lesson is blunt. If a home robot cannot dock itself or be heard over a television, caregiver load rebounds into babysitting the machine. Reliability is the admission ticket, not a bonus feature.
The authors are clear this is an ongoing pilot: five dyads, exposure from one to four months in the limitations text, while Table 1 lists Dyad 2 at six months. Participants self-enrolled with an existing interest in robotics and do not represent typical older consumers. The earliest dyad had no ADRD; their data were kept because they improved hardware and navigation.
There is no control group and no standard clinical outcome scale. GAS covers three dyads. The bed-exit detector reports two true positives and no false-positive or miss rate. Several authors have an ownership stake in Raha Tech, an LLC aimed at commercializing MARSS, disclosed as a conflict. Read this as a co-design field report, not an efficacy claim.