Abstract:
Recent advancements in electronic skins, e-skins, have begun to allow artificially intelligent robots
to interact with delicate objects and sustain safe human-robot interactions. However, the largescale
integration of e-skins into the wider robotic world remains an issue. The development of eskins
has taken many forms harnessing a variety of technologies including capacitive tactile
sensors. E-skins using capacitive sensors can be simple and robust, but as these e-skins gain more
spatial resolution the addition of individual connection cables and electronics can be problematic.
Complex sensor matrixes, flexible transistors, and ornate structures can become quite expensive.
Studies have approached this issue by using a multifrequency approach to help determine the
location of sensor deformation without the need for additional wiring. Scalability of this method
still remains a question. In this study, an alternative approach is presented that combines the
multifrequency approach while leveraging the computational power of supervised machine
learning. This project presents a simple soft capacitive monolithic sensor that is easy to fabricate
and leverages machine learning to extract complex information from only two sensor wires.
Machine learning models were developed that could both localize a finger press within a 5-zone
area and determine the magnitude of force and location in a 3-zone area. In a live testing
environment, all models correctly predicted >90% of button presses. Based on the high accuracy
of these models, pairing soft capacitive sensors with machine learning is a promising avenue to
address the challenge of e-skin scalability.