Fifty Milliseconds To Let Go

The pan is hot and your hand is already gone. The withdrawal takes about 50 milliseconds. The conscious pain arrives somewhere around 150 milliseconds later, and that is the part nobody expects: the movement is complete before the sensation exists. The wiring explains it. A nociceptor in the skin fires. The signal reaches the dorsal horn of the spinal cord and synapses there. An interneuron hands it straight to a motor neuron, and the muscle contracts. No brain sits in that loop at any point. Only after the arm has moved does the same signal travel up to be felt. The reflex layer is faster than consciousness, and it is not a degraded copy of it. It is a second channel with a different job. Stop the damage first, then report.

An artificial skin copies that split. Each taxel, or a small group of them, carries a high-threshold detector that fires when force crosses into damaging territory. Published designs put that line at 10 to 50 newtons. When it trips, the signal skips the normal readout queue and drives a motor inhibit line directly. Above that sits regional processing. A cluster of taxels shares a local unit that reads a shape rather than a spike: a sharp edge, a pinch, a steep pressure gradient. It answers in 0.5 to 2 milliseconds and it can grade the response - tighten the grip, or let go now. Then there is the centre. The full tactile stream still goes there, at 10 to 50 milliseconds, for everything that needs context. The fast channel does not replace it. It runs alongside.

Four priority bands is the pattern in the literature. Critical means force past the damaging line, or a sharp object: stop all motion, output inside 1 millisecond, the hot-surface case. High means incipient slip: tighten and re-seat the grasp inside 50 milliseconds. Medium means sustained pressure or a fast approach: slow down, go compliant, 10 to 50 milliseconds. Normal means texture and exploration, with no deadline at all. The hardware is unglamorous. A comparator against a reference voltage. A priority encoder on a logic tree. A dedicated line, or a prioritised event bus, for the signals that outrank everything else. A safety-rated path straight into the motor controller. A small buffer holding the last 1 to 100 contacts, so the reflex can be read back afterwards and learned from.

A robot on the bench has no negative signal of its own. It will keep closing a joint with a finger inside it. It will press harder when the contact resists. It will hold a hot part until the sensor or the gripper is destroyed. That is an economic problem before it is a philosophical one. The contact surface is the most expensive part of the assembly to replace, and it is the part that meets the world first. The 2025 work in PNAS on a neuromorphic electronic skin with active pain perception measured the alternative: three pain levels at taxel level, a regional rule where three adjacent taxels reporting moderate pain trigger a withdrawal, and a recoil from a sharp probe in under 5 milliseconds from stimulus onset. The pain signal was logged as well, which turns a reflex into training data.

The withdrawal finishes about 150 milliseconds before the pain exists. The decision was made in the spine, not the head, and none of that is an accident.

Spurious withdrawals are dangerous. A gripper that lets go at the wrong moment drops what it is carrying, so a threshold that trips on noise is worse than no threshold at all. The designs answer with hysteresis and a temporal filter. The threshold is not a constant either. The force that damages a surface depends on the geometry of the contact, the material of the object and the speed of the robot, so the honest version is an adaptive threshold read from context. Then the control loop. A reflex that interrupts the main controller has to hand the task back without losing it, which is a state machine - reflex, then return - rather than an emergency stop. The safety line has to stay independent of the main processor, or the fast path inherits the slow path's failure modes. The standards for a machine that flinches are still being written.

Sixteen years of field work across Nigeria, Serbia, Kazan and Fukushima taught me one thing about hardware that has to survive: a safety system waiting on a central decision arrives with the latency of the whole chain, and the chain is always longer than the diagram. The biology does not centralise the flinch, and it had millions of years to try. The probability that a machine which must not damage what it touches ends up making this decision at the point of contact is not a question of fashion. Centralising it costs the same information twice, once to notice the damage and once to replace the part.