Neuromorphic Computing Systems for Tactile Sensing Perception

Touch sensing plays an important role in humans daily life. Tasks like exploring, grasping and manipulating objects deeply rely on it. As such, Robots and hand prosthesis endowed with the sense of touch can better and more easily manipulate objects, and physically collaborate with other agents. Towards this goal, information about touched objects and surfaces has to be inferred from raw data coming from the sensors. The orientation of edges, which is employed as a pre-processing stage in both artificial vision and touch, is a key indication for object discrimination. Inspired on the encoding of edges in human first-order tactile afferents, we developed a biologically inspired, spiking models architecture that mimics human tactile perception with computational primitives that are implementable on low-power subthreshold neuromorphic hardware. The network architecture uses three layers of Leaky Integrate and Fire neurons to distinguish different edge orientations of a bar pressed on the artificial skin of the iCub robot. We demonstrated that the network architecture can learn the appropriate connectivity through unsupervised spike-based learning, and that the number and spatial distribution of sensitive areas within receptive fields are important in edge orientation discrimination. The unconstrained and random structure of the connectivity among layers can produce unbalanced activity in the output neurons, which are driven by a variable amount of synaptic inputs. We explored two different mechanisms of synaptic normalization (weights normalization and homeostasis), defining how this can be useful during the learning phase and inference phase. The network is successfully able to discriminate between 35 orientations of 36 (0 degree to 180 degree with 5 degree step increments) with homeostasis and weights normalization mechanism. Besides edge orientation discrimination, we modified the network architecture to be able to classify six different touch modalities (e.g. poke, press, grab, squeeze, push, and rolling a wheel). We demonstrated the ability of the network to learn appropriate connectivity patterns for the classification, achieving a total accuracy of 88.3 %. Furthermore, another application scenario on the tactile object shapes recognition has been considered because of its importance in robotic manipulation. We illustrated that the network architecture with 2 layers of spiking neurons was able to discriminate the tactile object shapes with accuracy 100 %, after integrating to it an array of 160 piezoresistive tactile sensors where the object shapes are applied.