#Industry News
DEEP LEARNING ROBOT: 10 DAYS TO LEARN HOW TO GRASP
Cobot Baxter Learns How to Grasp Over 150 Objects Without Human Intervention at Carnegie Mellon University
When compared to many animals, human babies are surprisingly vulnerable in their first months, however they do have a high capacity for self-learning and are able to master a variety of skills within their first year. Something as basic as a child learning how to pick up toys that are a range of shapes and sizes is in fact an impressive feat. Now researchers have observed similar behavior in a two-armed industrial robot called Baxter. The researchers, Lerrel Pinto and Abhinav Gupta at Carnegie Mellon University, provided Baxter with deep learning capabilities, arranged a table full of objects in front of it, and walked away, leaving it to learn on its own.
Baxter is a human-safe industrial cobot that is designed to perform repeatable tasks, such as line loading, packaging and material handling. Each arm is equipped with a two-finger parallel gripper and the robot also possesses a high resolution camera to see what it is doing. The robot was programed to grasp an object by separating it from its neighbors, then to select a random point, rotate its grippers in line and attempt a grasp. For every point, Baxter repeats the grasping process some 188 times, each time adjusting the gripping angle by 10 degrees.
Pinto and Gupta then left the robot in front of the table of objects to figure out how it should grip each one. During a period of 700 hours, Baxter completed 50,000 grasps on 150 different objects, and discovered which approach worked best with which object. The objects included a variety of household items such as plastic toys and a tv remote. Baxter was able to predict whether a grasp would be successful almost 80 percent of the time, which was considerably better than other approaches Pinto and Gupta tried. The experiment was conducted in a cluttered, unstructured environment and provides an important insight into how robots can learn and adapt without human intervention.
