The Map Is Drawn By Density

The hand is under 5 per cent of the body surface and takes roughly 25 per cent of the primary somatosensory cortex. The lips take about 15 per cent, the face 10, the palm 8, the trunk 5, the leg 4. None of that allocation follows the size of the part. It follows innervation density. Fingers carry the highest mechanoreceptor density of any patch of human skin, and the map is drawn where the receptors are packed, not where the area is. A trunk covering a thousand square centimetres gets less cortical territory than a fingertip covering two. That is the first correction this field needs. A sensor array is usually specified in taxels per square centimetre and tested for spatial resolution on flat, even contact. Biology priced the same problem by density and handed the largest share of the budget to the smallest part of the surface. Penfield and Rasmussen drew that map in 1950. Nobody has overturned it.

One contact on a fingertip fires four separate channels. Merkel cells report sustained pressure over seconds. Meissner corpuscles report slip onset in tens of milliseconds. Pacinian corpuscles report vibration from 1 to 10 milliseconds. C-tactile fibres carry affective touch over 100 to 500 milliseconds. The fast discriminative path runs at 35 to 75 metres per second through the dorsal column to the thalamus and up into areas 3b, 1 and 2. The slower spinothalamic path carries pain, temperature and crude touch with almost no spatial detail, and it exists to carry salience rather than geometry. The brain does not experience four data streams. It experiences one touch. A single contact activates all four channels and arrives as a single percept. That binding is the part artificial skin has not solved. A sensor stack publishing pressure, vibration and temperature as three separate streams has reproduced the input and skipped the output. The cortical hierarchy runs from what surface am I touching in area 3b, to what object am I holding in area 2, to how do I use it in the posterior parietal cortex. Three questions, three levels of abstraction, and most arrays answer the first one only.

After amputation the cortical territory for the missing hand is invaded by the face and the trunk. Touch on the cheek is then felt as coming from the hand that is no longer there. That is the phantom limb, and it is a measurement of how much of the map was holding the hand in place. Restore meaningful touch signals through a prosthesis and the hand area re-expands. Simple pressure or touch detection measurably improves prosthesis embodiment and control. Bandwidth was never the binding constraint. Two consequences follow. Signals that follow natural adaptation rates engage plasticity better than signals delivered at maximum rate. And selective attention is not a nicety, because the brain suppresses irrelevant tactile information as a matter of course. A skin that reports everything is competing with its own useful data. The probability that a system reporting every taxel every frame beats a system with an attention layer is low. The field keeps building the firehose version.

When vision and touch disagree, vision usually dominates. The rubber hand illusion is the standard demonstration, and it works on almost everyone. Touch holds three domains anyway: material properties such as compliance, texture and weight; incipient slip and grasp stability; and anything hidden from view. For robot manipulation the division is direct. Vision is the spatial modality. Touch is the contact modality. A manipulation stack running on vision alone is missing the channel that fires when contact actually happens, which in the Pacinian band is 1 to 10 milliseconds before a grasp fails. Fusing pressure, vibration and temperature into a single percept is the algorithmic equivalent of cortical binding. It is an open problem, not a solved one. Papers that report a fused pipeline usually report stage-one fusion at the sensor board, and stage-one fusion is where the four channels stop being four channels on the page only.

A trunk that covers a thousand square centimetres gets less of the map than a fingertip that covers two.

I signed a EUR 20 million personal guarantee on an industrial roll-up and watched the company fold on Christmas Day 2008. What I took from that was not that risk is bad. It was that a technically superior system loses to the one whose signals the buyer can act on. Prosthetics research says the same thing from the other end. The right signal structure may matter more than the right resolution, and that sentence is doing more work than any datasheet in this field. Run the numbers on a dense array. A 12-bit channel per taxel across 1,000 taxels at 500 hertz is 7.5 million bytes a second before any processing. That is a data problem handed to the host under the label of performance. Bandwidth is cheap in the datasheet and expensive in the loop. A 4,096-taxel array streaming raw frames is a firehose. The same array reporting pressure gradients and slip events is a decision. Local extraction at the skin, object-level recognition at the system, and the host receives two numbers instead of a stream.

The mapping work gives an order of operations. Local feature extraction at the skin level, edges and pressure gradients pulled out at the contact surface. Object recognition at the system level, where the power and the model sit. Fusion before resolution, because four channels bound into one percept is worth more than one channel at four times the density. Two of those three are engineering work with a schedule. The third has been open since Mountcastle described cortical columns in 1957, and it decides whether a tactile system feels like touch or like telemetry. So the first product is not the densest skin. It is the one that reports four channels, binds them into one signal, and answers the two questions a host can act on: what surface is this, and is the grip slipping. Density is the easiest number to sell and the last one that matters. The hand is 5 per cent of the body and 25 per cent of the map because the map was drawn by density of use. Build for the same thing.