Robotic tactile object perception based on adaptive multikernel sparse representation

Abstract

<p indent="0mm">The traditional tactile object perception using multifinger robots often ignores the force correlation among multiple fingers and among the multiple tactile sensors of each finger. To address this problem, we present a novel adaptive multikernel sparse representation (AMSR) method. First, multiple basic kernel functions are used to map all training samples into high-dimensional Hilbert space, which captures the nonlinear feature similarities of different tactile samples, and then the corresponding kernel matrix of each basic kernel is calculated. Next, an adaptive kernel weight calculation method is proposed to compute the adaptive weight of each basic kernel. A linear composite kernel, the linear combination of multiple basic kernels obtained using computed adaptive weights, is established to calculate the multidimensional kernel matrix. Finally, the same sparse pattern is applied to the coding vector of multiple finger tactile data to further explore the force correlation among multiple fingers. The proposed AMSR method is compared with the latest methods on our collected tactile dataset, and the object recognition result shows its effectiveness.