Google Arm Farm

Google's 2016 project where more than a dozen robot arms attempted grasps over 800,000 times to learn grasping from the data.

This is work from Google's Sergey Levine, Peter Pastor, Alex Krizhevsky, and Deirdre Quillen, released in March 2016 (an extended version of an ISER 2016 paper), nicknamed the “arm farm” for the sight of a whole row of robot arms working at once. Over about two months, 6 to 14 robot arms autonomously attempted more than 800,000 grasps, with success or failure judged automatically, with no manual labeling. This data was used to train a convolutional neural network: given a monocular camera image and a candidate gripper motion, it predicts whether that motion would lead to a successful grasp; at execution time, the system continuously picks the best motion and adjusts as it watches (visual servoing), with no camera calibration needed. Closed-loop grasping cut the failure rate from 34% under open-loop control down to 18%. This opened up the “large-scale real-robot data plus deep learning” approach that QT-Opt and MT-Opt later followed.

ExampleIf the gripper drifts off target mid-grasp, the robot corrects course in real time based on what it sees; faced with a pile of objects crowded together, it will even nudge one aside before grasping it.

Also called

Arm Farm, Learning Hand-Eye Coordination for Robotic Grasping with Deep Learning and Large-Scale Data Collection

Related

Google Arm Farm · Hand-Eye Coordination · Visual Servoing · Self-Supervised Learning · QT-Opt · MT-Opt

Sources

Learning Hand-Eye Coordination for Robotic Grasping with Deep Learning and Large-Scale Data Collection (arXiv 1603.02199)
Deep Learning for Robots: Learning from Large-Scale Interaction (Google Research Blog, 2016-03)

As of

2016-03