The Skin That Only Reports Change

The hand carries about 17,000 mechanoreceptors and routes them through roughly 38,000 nerve fibres, and it never sends a frame of anything. It has no clock. Each fibre is its own circuit, firing when its own stimulus crosses a threshold, in its own time, silent the rest of the day. The information is in the timing of the spike, not its size. All or nothing. Every camera, every force grid and every electronic skin built on a matrix does the opposite: it reads every cell, every cycle, and ships the whole grid whether anything changed or not. That is the interconnect problem, and it was designed in by accident.

Address-event representation is the artificial version of that nerve. The rules are four lines long. Each taxel keeps a baseline reading. A comparator watches the live signal against it. When the difference crosses a threshold, the taxel emits a packet: its address, a timestamp, and one bit for rising or falling. The data line is shared, and a taxel asks for it only when it has something to say. No event means no change. No change means no bandwidth. Ten thousand taxels, one bus, no central poll. The taxel that is being touched is the only one talking.

An event-based array in 2018 ran 32 by 32 taxels, with 2 microseconds of latency and 100 microwatts per taxel. By 2022 it was 64 by 64, half a microsecond, 10 microwatts. The 2025 neuromorphic parts run 128 by 128, 0.2 microseconds, and 1 microwatt per taxel, with a threshold you can tune per taxel instead of per row. Multiply 1 microwatt by 10,000 taxels and you get 10 milliwatts for the whole skin. A frame-based array at the same density sits 10 to 100 times above that before you count a single wire. Then look at the traffic. A grasp of 100 taxels sends about 100 events at the start and then roughly 10 per second. Sliding sends 200 to 500. Scanning a texture sends 1,000 to 5,000. Idle sends zero. Under ordinary manipulation the saving against a full frame is 10 to 1,000 times. I deployed 210 energy systems across Africa and Asia, and most of that fleet spent its life doing nothing at all. The cost of a machine is set by what it does when nothing is happening. That is the whole case for a spiking skin in one line.

The 2025 work does something better than save bandwidth: it moves the decision out to the taxel. A threshold crossing can fire a local motor command directly, tightening the grip on incipient slip, in under 1 millisecond. Route the same signal to a central processor and back and it costs more than 10 milliseconds. A 2025 PNAS neuromorphic skin added a privileged channel for damaging stimuli that interrupts normal traffic and withdraws the limb. That is the spinal reflex arc in silicon, and it only works because the taxel is already deciding what matters. Push the computation to where the data is born and the centre stops being the bottleneck. The peripheral taxel is not a sensor. It is the first processor.

A sensor that says nothing costs nothing.

None of this is an open research problem. The parts exist and the numbers are published. The gaps are dull, and they are why a robot hand still ships with a frame-based pad. Every group writes its own event format, so there is no ecosystem and no second supplier. A comparator, a baseline memory and an event generator add real circuit area on every taxel, which caps how dense the array can get. The baseline drifts with temperature, humidity and age, so it has to be reset, and each reset fires a burst of events that were never touch. The learning algorithms still trail the frame-based ones, which is the gap closing fastest, because event-based vision is dragging the tooling along behind it. Ten thousand taxels on ten milliwatts is not an improvement on the old architecture. It is a different machine, and the datasheets have not caught up with it.

The purchase order will not be decided on physics. It will be decided on an interface nobody has agreed on yet. The probability that the first genuinely useful whole-body skin is event-based is not zero. Most of the industry is pricing it at zero, because every specification sheet it reads is a frame rate. So the question to ask a supplier is not how many taxels. It is what the skin sends when it is not being touched. An answer of nothing is the only one that holds past a fingertip. A skin that speaks only when something changes is not a cheaper sensor. It is the first one whose bill scales with events instead of with time. Put the comparator where the touch happens, and the rest of the machine stops paying for silence.