Robot Tactile Data Classification Method Using Spiking Neural Network

摘要

The ability of tactile perception is one of the future development directions of robotics, and the research on the classification method of tactile data is to explore the concrete manifestation of its ability. Owing to its event-driven characteristics, the spiking neural network can better process event-based tactile sensing data, but the backpropagation of the network will fail because of the nondifferentiable influence of the propagation function. In this paper, a spiking-graph neural network model for tactile perception is improved, and various differentiable functions are used to approximate the propagation function of the neural network. Finally, the function that achieves the best performance is selected as the backpropagation function through experiments. Our model improves the classification accuracy of tactile data and simultaneously accelerates the convergence speed and reduces the convergence value. The experimental results of the tactile data set collected by the robot show that the performance of the improved model in this paper is better than the original model and classify household objects more accurately.