Abstract

For humans, the sense of touch is of immeasurable importance. It starts with babies that hold on to fingers and explore their surroundings in a tactile fashion, using their hands and mouth. Touch supports seamless and safe physical interaction with the external world to accomplish tasks by grasping and confidently handling objects, detecting when they are about to slip and react accordingly. Touch information conveys the texture, shape and material of objects we interact with, complementing vision, especially in the dark, or when there are occlusions. Finally, social and affective touch, such as hugs, strokes, or handshakes, are crucial in development and influence well-being and social interactions.

With the increasing development of autonomous robots for laborious or tedious tasks, as well as for tasks that require high precision in industry, agriculture, surgery, healthcare or household, the main tasks the robots would have to fulfil require them to safely and robustly interact with objects and people. Interestingly, many of those (object manipulation, slip detection, grip strength modulation) often rely on vision-based algorithms. This is heavily related to the fact that many AI and machine learning approaches are based on visual inputs, as vision can count on large available standardised datasets and a long history of computer vision development.

However, this comes with high computational cost and low performance when the environment is dark or parts of the visual field are obstructed. Further, when purely visual information is processed, important cues such as the time and quality of contact are missed. As such, there is a need to equip robots with tactile sensors extensively. Not only robots but also prosthetic limbs benefit from the application of tactile sensors and the possibility of providing tactile feedback, to improve the sense of ownership and decrease the attentional burden of using vision to manipulate objects.

Unlike artificial tactile sensors that periodically sample the pressure applied to them, the human sense of touch is based on a family of sensors embedded in the skin (mechanoreceptors), that only provide information to the brain when a stimulus is applied. Mechanoreceptors are grouped into different categories depending on their sensitivities and responses to stimulus types. Different mechanoreceptors families rely on physical transducers with different size, placement, adaptation and responsiveness, to capture the diverse information about sensory stimuli. This encoding substrate supports the rich, robust, accurate and efficient encoding of the information about contact needed for physical interaction.

The goal of this PhD thesis is to design and investigate circuits for the implementation of artificial tactile sensors that operate similarly to their biological counterparts. As such, it focusses on contact-driven encoding of pressure information from a variety of pressure transducers, such as capacitive and piezoelectric, and on capturing information about static and dynamic pressure at different bandwidths. The approach used in the circuit development is that of Neuromorphic Engineering, with a focus on mixed-mode subthreshold circuits, to achieve low power, low latency, compactness and adaptivity.

Publications

Panzeri, S., Janotte, E., Pequeño-Zurro, A., Bonato, J., & Bartolozzi, C. (2023). Constraints on the design of neuromorphic circuits set by the properties of neural population codes. Neuromorphic computing and engineering, 3(1), 012001.

D'Angelo, G., Janotte, E., Schoepe, T., O'Keeffe, J., Milde, M. B., Chicca, E., & Bartolozzi, C. (2020). Event-based eccentric motion detection exploiting time difference encoding. Frontiers in neuroscience, 14, 451.

Janotte, E., Bamford, S., Richter, O., Valle, M., & Bartolozzi, C. (2022, June). Neuromorphic capacitive tactile sensors inspired by slowly adaptive mechanoreceptors. In 2022 20th IEEE Interregional NEWCAS Conference (NEWCAS) (pp. 119-123). IEEE.

Janotte, E., Mastella, M., Chicca, E., & Bartolozzi, C. (2021). Touch in robots: A neuromorphic approach. ERCIM News, (125), 34-51.