Tactile Intelligence for Robot
Team members
Robot tactile intelligence research aim to build hierarchical artificial neural networks by mimicking the structure and signal patterns of the spinal nervous system of the human body. In particular, we mimic two stages of tactile processing in the nervous system at mechanoreceptors and cuneatus nuclei in medulla. In the mechanoreceptor level, we model the spiking activity of Merkel cells with slowly adapting type I (SA 1) afferent while excluding its adaptation characteristics in this study. In the cuneatus nuclei level, we developed an SNN learning method optimized for the network architecture and the application.
The Biological neuron model is a mathematical design of the cortex neuron's spiking dynamics. In simple terms, it converts a continuous sensory input(in tactile case the sensory input means mechanical pressure!) into a spike pattern of 0 and 1 by calculating membrane potential of neuron. This study extends the biological neuron model currently limited to SA-1 and FA-1 to SA-2 and FA-2. Through this research, it is expected to send more sophisticated tactile information to the robots.
(a) An experiment on the grasping objects using a robotic gripper and tactile sensors. (b) The scheme of a spiking neuron network (SNN) with a WTA-STDP(winner-take-all STDP) algorithm developed in this study. The SNN learns to cluster tactile sensor data collected while robotic hand grasps objects with different shapes. (c) Input and output spiking patterns of SNN. The example of 42 spike trains of the input layer neurons during grasping and releasing of different objects.