A PR2 folds fifty towels in a row
A PR2 robot at Berkeley picked a randomly tossed towel off a table, found its corners from image geometry, re-grasped, untwisted and folded it — and did so successfully in all 50 end-to-end trials on towels it had not seen before.
Why it matters
Cloth had been the standing failure case: it has no fixed shape to match against, and no end-to-end success had been published. After this a crumpled towel in an arbitrary configuration was a solved problem with a measured success rate.
Jeremy Maitin-Shepard, Marco Cusumano-Towner, Jinna Lei and Pieter Abbeel presented a corner-detection algorithm for cloth that relies only on geometric cues robust to pattern and texture. Parameters were tuned on three towels and tested on 25 others: 25 distinct colours and patterns, 16 materials, 11 sizes, lengths from 32 to 78 cm and widths from 32 to 48 cm. Each towel was tossed onto the table twice in a random configuration, for 50 trials in all. The paper states it plainly: "A perfect success rate of 100% was achieved on the 50 end-to-end trials of previously untested towels." The intermediate figures are given separately: across the 50 trials the robot made 153 corner-detection attempts, of which 147 were correct, a detection rate of 96%, and 81% counting detection together with the grasp. Individual steps did fail, but the procedure was built to recover from failures and recovered every time. The robot also handled a pile of five towels. The hardware was a prototype of the Willow Garage PR2 platform.