Artificial Intelligence for Tactile Object Recognition

Abstract

Tactile perception is a crucial sensory property for robots, in the context of understanding the surrounding environment and the physical attributes of encountered objects. Electrical Impedance Tomography (EIT) is a diffusive imaging technique that infers internal conductivity distribution from boundary measurements. Changes in pressure affect the conductivity properties of artificial skin materials, which can be inferred with EIT, thus giving rise to artificial touch. EIT requires the solution of the forward and inverse problems. The inverse problem, i.e., image reconstruction, is severely ill-posed and non-linear, thus affecting reconstruction quality. In this thesis, a variety of image representation approaches are proposed to further the state-of-the-art in machine learning (ML)-based inverse solutions. Specifically, to reduce the rank-deficiency of the voltage-impedance mapping, three image representation techniques are proposed. These are then coupled with the superior adaptation properties of ML to improve both image quality and robustness to noise. Processing commences by decomposing the impedance image into a number of blocks. The first representation paradigm utilizes surface representation via bivariate polynomials, while the remaining two explore the properties of sparsity and energy compaction, offered by two well-known image compression methods, i.e., Discrete Cosine Transform (DCT), and Discrete Wavelet Transform (DWT). This helps cast the target data into a smaller dimensional-space, while allowing the retrieval of fine, crisp images upon decompression. Furthermore, we introduce a task decomposition scheme to efficiently distribute the learning problem between a set of ML modules. The overarching contribution, therefore, is a complete EIT reconstruction module, that performs a ML-based voltage-impedance mapping built on the three proposed representation models.

Date of Award	Dec 2022
Original language	American English
Supervisor	PANAGIOTIS Liatsis (Supervisor)