As an example of advanced sensors and signal-processing close to the sensors to reduce the amount of data communicated with a central processing unit (or the brain), in BOS we develop a machine learning (ML) module to process tactile and tremor signals generated byneuromorphic electronic skins (ne-skin). Our starting point is our recent work on a novel ne-skin that uses triboelectric nanogenerators technology and neuromorphic circuit design .
The neuromorphic hardware represents a convolutional neural network that converts analog tactile signals into electrical pulses, and performs in-parallel computation in distributed synaptic circuits. Superior perception capability of e.g. the human hand is associated with the unique signal processing mechanisms in biological tactile nervous systems. The latest studies on biological somatosensory peripheral nerves suggest that the computation for extraction of geometric features starts at the level of the branched first-order tactile neurons close to the sensor points in the skin. In fact, this system represents an event-driven system and a typical convolutional neural network (CNN) that operates using spikes as the information carriers. The pre-processed tactile signals are sent to the sensory cortex for higher levels
of cognitive computations.