No Robot Is Built To Comfort

A care robot sat across the table from me and did everything the datasheet promised. It found my pulse through a fingertip and held the reading to within one beat per minute. It felt the tremor in my hand. It mapped the pressure of my palm across its own surface, a grid of contact I could read on a laptop. Then I asked it for the thing a nurse does without deciding to do it. A hand laid on a forearm. Held there. Not gripping, not checking, not measuring. The machine had no mode for that. Not a bad mode. No mode. The instrumentation is finished. The behaviour was never designed. Somewhere between the sensor and the actuator, the point of touch fell out of the build.

Two variables decide whether a robot's touch reads as care, and neither of them is resolution. The first is compliance. Users rate a robot's touch as twice as pleasant when the surface is skin-like instead of rigid, at identical force and identical motion. Skin-like means a modulus between 10 and 100 kilopascals. A rigid surface reads as machine, and the body files machine under not safe. The second is temperature. A hand at 32 degrees is rated significantly more comforting than the same hand at 22 degrees or at 15. Cold is a danger signal. That is not poetry, that is wiring. Neither variable is a sensor. Double the taxel count on the best array in the building and you lose both.

There is a nerve fibre in your hairy skin tuned to one narrow band of stimulus, and it is why a stroke works. Move across the skin at 3 centimetres a second, at skin temperature, and the C-tactile afferents fire. Oxytocin rises. Stress markers fall. That works when a robot does it, which is the finding nobody has built on. The window is not wide. 1 to 10 centimetres a second, with half a newton to 2 newtons of normal force. Then timing. Constant velocity on a fixed period feels mechanical because it is. Human stroking accelerates at the start, decelerates at the end, and never repeats. A servo tracing a perfect line is still a perfect line.

Walk the commercial shelf and the absence is total. PARO is a seal with touch sensors under the fur, and it responds to being petted. It is therapy for one narrow setting. It is not general purpose. SoftBank's Pepper carries capacitive sensors in the head and hands and can register a stroke. The surface is poor compliance with no temperature control, so what comes back is a detection, never a comfort. Sophia has no touch at all. Jibo simulated affection with animations on a screen. MIT's soft pneumatic gripper is compliant by design and has never been programmed for social touch. Every machine on that shelf was built to detect contact. Detection is a solved problem. Meaning was never on the road map.

The instrumentation is finished. The behaviour was never designed.

Set the two design briefs side by side and the conflict stops being an engineering detail. A manipulation robot wants a hard, high-friction surface, a force range from 0.1 to 100 newtons, high bandwidth for texture and vibration, and stiff position servo control. Its safety objective is simple. Do not drop the object. An affective robot wants the opposite column. Soft, warm, skin-like. A force range from 0.1 to 5 newtons. Low bandwidth, because stroking lives at low frequency. Sensing for contact presence, temperature and compliance. Soft, backdrivable actuators and compliant force servo control. Its safety objective is a different kind of problem. Do not hurt the person. No single machine optimizes both columns.

The therapeutic case is already paid for by humans doing it by hand. Tactile comfort reduces agitation in dementia patients by 30 to 50 percent, comparable to medication without the side effects. Gentle touch also lowers perceived pain, through the C-tactile pathway that quiets the pain signal behind it. Cortisol falls. Oxytocin rises. The market on the other side of that is worth billions and contains no technology. It has human caregivers in short supply and weighted blankets. I have built companies across 12 countries and restructured a industrial group. The pattern in hardware never changes. The sensing half gets funded because it demos well. The softness half gets deferred because it looks like manufacturing. The probability that a robot good enough to hold a hand ships is not zero. Most road maps price it at zero.