Mutual capacitance tactile sensor design and sim-to-real bridging for robotic grasping

Abstract

The humans’ tactile sensing ability is remarkable and it is crucial for a multitude of real-life tasks. Transferring this exceptional capability to robots presents a lot of potential across various applications, especially in dynamic and unknown environments where computer vision is not practical. As a result, the aim of this project is create a cost effective tactile sensor that can entirely cover an anthropomorphic robotic manipulator and test its functionality by assessing its ability to represent primitive shapes as point clouds in the 3D space and to discern these shapes. This project represents a first step towards unlocking the future potential of enabling a robotic manipulator to rummage in cluttered environments, even in instances of limited or absent computer vision capabilities. Furthermore, we recreated the sensor’s functionality in a simulation environment to reproduce synthetic tactile data that are close to the ones generated in real-life for an efficient future AI models training. Contemporary and forthcoming AI models require large datasets for a proper training and simulations can leverage the power parallel programming among many other tools to expedite and streamline the datasets generation process , thereby enhancing its efficiency. After designing the tactile sensor, we mounted it on a real-life robotic workstation to perform objects grasping and shape recognition experiments.Furthermore, a simulation replica of the robotic system was to generate synthetic tactile data. Minimizing the sim-to-real gap is a paramount objective, ensuring the generated synthetic tactile data closely mirrors real-world interactions. The developed sensor was successfully able to capture the shapes of some primitives objects and to discern them in both simulation and reality and the initial sim-to-real gap was significantly reduced. In sum, we have developed a cost-effective tactile sensor that can be integrated with robotic systems for shape recognition without relying on vision and we have developed a simulation framework to efficiently generate realistic synthetic tactile data that can be used to train AI (Artificial intelligence) algorithms in the future. However, some modifications should be implemented to improve the reliability and performance of the sensor in both real-life and simulation.

Date	21 Dec 2023
Original language	American English
Awarding Institution	École de technologie supérieure
Supervisor	Jean-Philippe Roberge (Supervisor)