Somewhere in a lab at the University of Tennessee, a waist-high humanoid robot with a tablet strapped to its chest has been giving talks about wandering, bathing and sleep. It takes questions afterwards. It sets a short quiz. Then it walks the person in front of it through a breathing exercise. The robot is Pepper, a commercially available social robot, and the system running on it is called RISE, short for Robot-based Information and Support to Enhance Alzheimer's Caregiver Health. Its designers published the first account of it in Frontiers in Robotics and AI on 27 January 2026, and the ambition behind it is unusually plain: take a caregiver support programme that works but costs too much to deliver, and see whether a machine can carry some of it.
The programme in question is REACH, the most heavily evidenced family caregiver intervention in dementia care. The authors, who include Linda Nichols and Jennifer Martindale-Adams of the VA Caregiver Center in Memphis, both long associated with REACH, are candid in the paper about why it has not reached most families: cost, labour requirements, and the absence of insurance reimbursement. REACH is fundamentally person-delivered. A trained coach sits with a caregiver, personalises the material, and teaches skills the caregiver keeps using after the sessions end. That is exactly the part that does not scale. The paper frames the scale of the gap using Alzheimer's Association figures it cites from 2023: more than 6.7 million adults in the United States living with Alzheimer's and related dementias, and roughly 11 million family caregivers providing about 18.5 billion hours of care a year.
What makes RISE worth reporting is less the robot than the plumbing behind it. Generative AI let loose on a question about a father who will not get in the bath is a liability; it can invent, confidently and fluently. So the team built the system as retrieval-augmented generation, which means the model is required to fetch its answers from a fixed, verified document, the REACH Caregiver Notebook, rather than from whatever it absorbed in training. To test whether that discipline holds, they curated 52 caregiver question-and-answer pairs spanning modules such as wandering, bathing and sleeping, with answers written by domain experts as the ground truth, and scored the system against them using the open-source RAGAS library. The comparison was against GPT-4o answering the same questions with no grounding at all. RISE's grounded backend reached 87 percent correctness and 92 percent relevancy against the expert answers.
Then came the part that families should read carefully. The user evaluation ran in two phases, and neither phase involved informal dementia caregivers. Phase one was usability testing with technically literate users, students recruited through their professors, who consented under a University of Tennessee Knoxville ethics protocol and received no incentive. The authors are explicit that this was deliberate, a way to find prototype faults in a controlled setting without risking distress to real caregivers. Phase two was expert review by dementia caregiving professionals, asked to judge fidelity to the source material, clinical relevance and real-world applicability. Across all components, Likert usability scores landed between 3.6 and 4.6 out of 5.
The low end of that range is the interesting end. The slide layouts the AI generated scored 3.6 out of 5, with participants reporting that the text was too dense and the robot talked too quickly. One wrote simply, "The robot spoke too fast." The team responded by adjusting the prompt to produce shorter presentations in plainer language; Pepper's interface, they note, could not support AI-generated images or video, because their content validation so far covers text only. That is a small detail with a large shadow. A caregiver who has not slept properly in a year, standing in a clinic waiting room with a robot reciting at them, is not an audience that forgives pacing errors.
So what does a family take from this? Not a product. RISE is a research prototype, evaluated for accuracy and acceptability, and the authors themselves say the next step is a larger study with real informal dementia caregivers. Nothing here shows that talking to Pepper lowers a caregiver's stress, improves their sleep, or delays anyone's move into residential care, because none of that was measured. What has been shown is narrower and still meaningful: it is possible to build a machine that will only say what a verified caregiver manual says, and that people who know the field found the result credible enough to be worth testing properly.