Incorporating High-Resolution Tactile Perception for Performative and Generalized Robotic Manipulation Through Compliance Estimation and Hardware Design

Abstract

In robotics, replicating the natural proficiency with which humans perform manipulation tasks has proven challenging. Modern control schemes are predominantly learning-based and thus depend heavily on data collected via teleoperated demonstrations. Humans rely on our tactile perception to perform contact-rich and dynamic manipulation tasks. By more seamlessly incorporating high-resolution tactile sensing and haptic feedback into teleoperation interfaces, we can work to create stronger demonstration data to support the development of more effective learned control policies. In this thesis, we present two contributions toward this goal. First, we develop an algorithm to estimate the compliance of grasped objects in real-time from tactile images to provide haptic feedback to remote users. This algorithm combines both analytical and learning-based approaches to better generalize across both object shapes and materials. Second, we create a 1-DoF robotic gripper design with integrated tactile sensing. Inspired by the principle of self-similarity, this gripper is designed to better conform to complex object geometries than traditional designs and more securely grasp objects of many shapes and sizes. Together, these contributions can be utilized to create robust, tactile-aware teleoperation platforms. These platforms would facilitate more effective data collection and thereby promote the development of more performative autonomous action in generalized robotic manipulation scenarios.