The nervous system is remarkably good at monitoring sensory inputs, yet it does so with signals that would bewilder many computer scientists: an irregular cascade of activity spikes that can be sent to hundreds of downstream neurons, where they are combined with other spike trains arriving from still more neurons.
Now, researchers have used spiking circuitry to create an artificial robotic skin, borrowing elements of how sensory-neuron signals are transmitted and integrated. While the design incorporates some clearly non-biological elements, it benefits from the fact that we have chips capable of running neural networks with spiking signals, allowing this approach to pair naturally with energy-efficient hardware that can run AI-based control software.
Locating inputs with spikes
The sensory network in our skin is highly intricate. It contains specialized receptors for different modalities—heat, cold, pressure, pain, and more. In most regions these signals feed into the spinal cord, where initial processing can trigger reflexes without involving the brain. Still, signals travel along specific neurons to the brain for deeper processing and potential conscious awareness.
The team behind this recent work, based in China, set out to replicate some of those functions for an artificial skin intended to cover a robotic hand. They focused sensing on pressure, but reproduced other nervous-system capabilities, such as determining the location of stimuli and damage, and implementing multiple layers of processing.
The project began by fabricating a flexible polymer skin embedded with pressure sensors, connected to the rest of the system via conductive polymers. A subsequent layer translated the pressure-sensor outputs into sequences of activity spikes—brief pulses of electrical current.
Spike trains can encode information in four ways: the waveform of each pulse, its amplitude, the pulse duration, and the spike frequency. Biological systems most often use firing rate to carry information, and the researchers employed that to represent the pressure detected by a sensor. The other encoding dimensions are used to form a barcode-like signature that identifies which sensor produced the reading.
