Researchers led by Yoobin Shin have built TapTacTouch, which reconstructs the 3D shape of an object from a few touches of a robot hand, with no camera, no training and no shape models for particular kinds of object, according to a paper on arXiv. Vision-based reconstruction fails when objects are hidden or cannot be seen.
A robot arm with six degrees of freedom and a hand with 20 collects touches with 3x3 pressure sensor arrays on its finger pads. The robot's kinematics turn each touch into a 3D contact point, and a multi-scale Gaussian process surface fills in the shape, using confidence weights, weak guesses at the surface and a geometric prior to keep local detail and guide the parts it has not touched.
In real tests on six objects it reached a Chamfer distance of 1.610, an Earth Mover's distance of 7.738 and an F-score of 95, taking about 50 seconds per object.



