**1. Introduction**

As the second decade of the 21st century is ending, artificial intelligence and machine learning solutions are finding their place more and more in daily life. Besides the wide application of machine learning in data processing, planning, and decision-making, intelligent machines and robots are expected to perform all the physical tasks that humans do. These robots are required to reproduce human capabilities to be reliable substitutes for them. To achieve this, a huge research effort is devoted to both discover human brain functions and develop computational models to align technology on biological patterns. As such, the creation of artificial sense of touch is an important topic of interest. Many different tactile sensors are nowadays designed, produced, and available on the market [1]. However, data acquisition and processing techniques that make their use viable in practical applications are still required to evolve toward higher levels of efficiency.

In the case of human tactile perception, neuroscientists have revealed six different exploratory procedures that humans are make use of in order to perceive a stimulus by touch, among which "contour following" and "enclosure" are used for object recognition as they help exploring the shape of objects [2]. Klatsky et al. [3] mentioned the contribution of two sensory modalities in tactile perception, namely, cutaneous and kinesthetic cues. Cutaneous cues sensed by mechanoreceptors in the skin can obtain information about the texture, roughness, vibration, and temperature of a surface, while kinesthetic cues are provided by joints, bones, and muscles and supply information about the weight or the object's shape [3].

On the other hand, haptic exploration of objects is believed to be more reliable when visual data is available [4]. Moreover, Amedi et al. [5] refer to multimodal cells in human brain responding to both visual and tactile data, concluding the close contribution of human visual and tactile systems.

With inspiration from the visuo-haptic contribution in human sensorial loop and in accordance with the contour following strategy that humans use for object recognition, in this work we simulate the process of tactile data acquisition from a dataset of 3D models where the tactile data includes both of the two tactile sensory modalities employed by humans (cutaneous and kinesthetic). To align the process of tactile data acquisition with reality, an adaptive procedure is introduced to reduce the size of sensor while increasing the spatial resolution to probe the locations where a real sensor of a determined size will not be able to acquire data due to geometrical features. A similar probing strategy is also used in humans when using fingertips to touch finer details of objects and the palm for larger surfaces. The acquired sequential data from object contours are then used for the purpose of object recognition. A computational model of visual attention (i.e., a biologically inspired model selecting relevant areas in a visual scene for further exploration and analysis, as in human visual system) is advantageously employed to guide the process of contour following and results are compared to the case where visual information is not available (blind contour following in Section 5). The originality of the work with respect to the literature is found, more specifically, in the technique we propose to employ visual data for object recognition.

The main contributions of the work are as follows. (1) Simulation of tactile images (cutaneous cues) using a virtual tactile sensor relying on working principle of Force Sensing Resistor (FSR) arrays. (2) Adaptive simulation of tactile images according to geometry of objects. (3) Guidance of the process of contour following by engaging visual data from an enhanced model of visual attention. (4) Classification of sequential tactile data using different machine learning approaches including Convolutional Neural Networks (CNN), support vector machines (SVM), and k-Nearest Neighbors (kNN). (5) Fusion of cutaneous and kinesthetic data for making decision on object classes based on the probability values obtained using the CNNs.

The paper is structured as follows. Section 2 briefly discusses the current literature on the topic. A concise explanation of the framework we are proposing is provided in Section 3. The process of tactile data acquisition is detailed in Section 4. Section 5 provides more information about the implementation of contour following. The classification of the acquired tactile data is discussed in Section 6. Results are reported and discussed in Section 7, and Section 8 concludes the work.
