Fast and Accurate Tactile Object Recognition using a Random Convolutional Kernel Transform

The task of tactile object recognition is an ever-evolving research area comprising of the gathering and processing of features related to the physical interaction between a robotic system and an object or material. For a robotic system to be capable of interacting with the real-world, the ability to identify the object it is interacting with in real-time is required. Information about the object is often strongly enhanced using tactile sensing. Recent advancements in time series classifiers have allowed for the accuracy of real-time tactile object recognition to be improved, therefore generating opportunities for enhanced solutions within this field of robotics. In this paper, improvements are proposed to the state-of-the-art time series classifier ROCKET for analysis of tactile data for the purposes of object recognition. A variety of classifier heads are implemented within the ROCKET pipeline; these models are then trained and tested on the PHAC-2 tactile dataset, achieving state-of-the-art performance of 96.3% for single-modality tactile object recognition while only requiring 11 minutes to train.