The mind that makes the machine useful

I. The care gap is a market failure

About 55 million people live with dementia worldwide, and the number is projected to reach 139 million by 2050. The global cost of dementia care already exceeds $1.3 trillion a year, and most of it falls on unpaid family carers, who burn out, fall ill and die earlier than their peers.

One older adult in four experiences social isolation. Depression in later life is underdiagnosed and undertreated. Children with autism wait months or years for therapy that costs their families everything. Mental-health services are rationed by geography and income. And the World Health Organization has projected a shortfall of about 13 million health workers by 2035, a gap no plausible hiring plan closes.

These are care problems before they are technology problems. But the care system is buckling under a demographic pressure that hiring alone cannot meet. So here is the question I think robotics keeps asking wrongly: why are we building better robot bodies when the bottleneck is the robot's mind?

II. The robot brain gap

The socially assistive robot is a real product category. Furhat builds robots with projected faces and conversational AI. PAL Robotics ships platforms with on-board language models. ElliQ offers proactive companionship for older adults. PARO, the therapeutic seal, has been reducing anxiety in people with dementia for more than a decade.

The hardware is ready. The sensors, the actuators, the cameras, the microphone arrays: ready.

What is missing is the mind.

Today's assistive robots ship with fixed cognitive capabilities chosen by their manufacturer. Want a robot that can run cognitive-behavioural exercises? You need a different product from the one that provides dementia companionship. Want to move from elderly care to autism support? Buy another robot.

It is like selling smartphones with every app soldered to the motherboard. No store, no updates, no ecosystem: just whatever the factory decided you needed. The opening is in the layer between the hardware and the person: cognitive models, the minds and personalities that make a robot genuinely useful for social assistance, distributed so that anyone qualified can create, validate and deploy them.

III. What a cognitive model is

A cognitive model is more than a chatbot running on a robot. It is a whole architecture.

Understanding, of context, emotional subtext and social cues, across verbal and non-verbal channels.

Long-term memory: who you are, what you talked about last Tuesday, that you prefer to be called "nonna" instead of your legal name, that you get anxious in the afternoon.

Adaptive behaviour: shifting communication as dementia progresses from early to late stages, adjusting the difficulty of an exercise to performance, modulating its responses to the person's state.

Explainable reasoning: when the robot changes the subject, redirects attention or escalates to a human carer, it can say why. A partner you can question, instead of a black box.

Personalisation: a model of the individual that deepens through interaction and feedback.

Call the integrated architecture the cognitive core. On top of it sit domain packs: dementia care that understands the stages of decline and adapts from verbal to non-verbal communication as abilities change; mental wellbeing that delivers evidence-based exercises with emotion recognition and crisis detection; autism support that models social scenarios through predictable, non-threatening interaction; and companionship for older adults that starts conversations, manages reminders and keeps families connected.

The research for each of these exists. Frameworks for adaptive dialogue in human-robot interaction are published; reinforcement learning for social behaviour works in labs; vision-language models can explain a decision in terms of what the robot actually perceived. What nobody had done, when I wrote the first version of this in February 2026, was assemble them into something downloadable, deployable and open.

IV. Why these four domains

I chose them on three criteria: urgent unmet need, evidence that assistive robots help, and a population large enough to justify the work.

Cognitive impairment. Carers burn out at very high rates, and robots like PARO already reduce anxiety and lift mood with a companionship that does not tire. They are limited by fixed capabilities. A model that adapts to the person's stage of decline, remembers their history and knows when to call a human is a step change.

Mental health. AI coaching already delivers positive-psychology and CBT exercises with measurable results, but without a body. A robot that reads your face, modulates its tone and sits with you in silence when that is what you need is a different thing from a chatbot on a phone. The non-negotiable requirement is crisis detection and escalation: these models must know their limits.

Children's development. Children with autism respond remarkably well to robot-mediated social-skills training: a robot's patience and predictability make a safe space to practise eye contact, gesture and response. The critical constraint is transfer: skills learned with a robot have to carry over to people.

Companionship in old age. The fastest-growing application: robots that start the conversation instead of waiting for it, manage medication, connect families, run cognitive exercises. If the minds are distributed openly, a care home no longer has to wait for one manufacturer to build exactly the right profile for its residents.

V. A store for minds

The structural move is to separate the mind from the body, the way the smartphone separated the app from the phone.

Model developers publish cognitive models, the core plus domain capabilities, described by what they do, whom they are for, which platforms they run on and how they were validated. Validators, domain experts, ethicists, safety engineers, review and certify them, and stake their own reputation and value on the certification: certify accurately and earn, approve something unsafe and lose. Robot owners, care facilities, schools, families, license the models and install them; a care home could run one pack in its memory unit and another in its common rooms. And people who choose to contribute anonymised interaction data can do so in a way where the models train on the data without the data ever leaving its owner.

Every model carries a cryptographic fingerprint checked at download, every license is enforced by code, every update arrives through a secure channel. The models live off-chain; the ledger does what it is good at, which is verification, rights and value exchange.

VI. Safety is the product

This is AI that works with the most vulnerable people in society, so the safety architecture is the product itself, and no feature list can stand in for it.

A human in the loop, by design. Every model carries configurable escalation: when it reaches its limits, a crisis, an ambiguous emotional state, a medical concern, it hands over to a carer, a therapist or an emergency contact. A model is judged partly on how well it escalates.

Explanation as a requirement. Families and carers need to know why the robot did what it did. "I noticed Maria seemed agitated after lunch, so I suggested we listen to her favourite music" is the standard: behaviour you can read.

Fairness, audited. Models are checked for bias across age, gender, ethnicity and income, and diverse training data is a requirement.

Regulation, built in. GDPR for European deployment, health-data rules wherever the robot works, the EU AI Act's high-risk obligations for safety-critical systems, medical-device pathways where they apply. An immutable audit trail is exactly what regulators increasingly ask for.

In February 2026 I attached a token design and a marketplace roadmap to this argument. I have left them out: they were one venture's plan, and the argument does not depend on them. A bad model here fails a person with dementia, and every mechanism has to be judged by that.

VII. Better minds

The venture of the Fucina Nexus Program working on this is Amarchord, whose function is cognitive models for socially assistive agents and robots. It is in the pipeline, and its first functional proof is a versioned cognitive pack running one bounded assistive task, with entitlement, intervention and rollback, on software or a simulator; a physical robot is optional.

The care crisis is already here. Tens of millions of people living with dementia. Children waiting for therapy. A population ageing faster than the workforce trained to look after it.

The robots exist. The AI exists. What is missing is the layer that connects them: validated, continuously improving cognitive models that turn a general-purpose robot into a patient, adaptive partner in care. Better hardware will not close the gap. Better minds might.