Sensing and Actuation Technologies for Dexterous Manipulation in Constrained Environments

Abstract

A primary reason why dexterous robot manipulation remains challenging is because we havenot equipped our robots with sufficiently robust perception and control systems. In all but the most structured and unconstrained environments, roboticists must make compromises in designing these systems due to an artificially limited set of available sensing and actuation modalities. Since most environments cannot be outfitted with large arrays of cameras, we accept the use of one or a few cameras that are sensitive to lighting conditions and whose view can easily become occluded by environmental clutter or even the robot itself. The range of motion demanded by dexterous manipulation require robots to be highly articulated, forcing us to choose between heavy, power consumptive robots that use many motors or significantly less controllable robots that use fewer. Here, we develop alternative forms of sensing and actuation, and demonstrate their ability to facilitate dexterous manipulation in real robotic systems. We design a fingertip embedded proximity sensor that is robust to occlusion, and propose a framework that allows the robot to estimate the pose of general objects using these types of sensors. With respect to actuation, we design electrostatic brake equipped joints that are four times lighter and a thousand times more power efficient than their motor driven counterparts. We demonstrate that underactuated robots augmented with such brakes achieve independent positional control of each of their joints while maintaining low weight and power consumption. Finally, we integrate both fingertip embedded proximity sensors and electrostatic brakes into a single robot hand. By combining both of these technologies with adapted Bayesian state estimation and model predictive control algorithms, we demonstrate how alternative forms of sensing and actuation increase the speed and precision of in-hand manipulation.