Sergey Levine, an investigator at CITRIS and the Banatao Institute and an associate professor in the Department of Electrical Engineering and Computer Sciences at UC Berkeley, and his research team have developed a method to teach robots to perform complicated tasks with extreme accuracy.
Transitioning machines from repetitive, single actions to performing unpredictable tasks is one of the trickiest challenges currently facing roboticists. Levine’s lab is adopting reinforcement learning, an AI-powered training method that allows a robot to attempt a task in the real world and, using feedback from cameras, learn from its mistakes to eventually master the skill. The team’s most recent system incorporates human intervention, which allows a human to correct the robot’s course, and these corrections can also be incorporated into the robot’s training data.
The team presented the robot with a variety of tasks representing different types of uncertainty, such as flipping an egg in a pan, assembling a motherboard, and “Jenga whipping,” a dexterity exercise in which a whip is used to strike a single block out of a Jenga tower while the rest of the structure stands sound. They further pushed the robot’s adaptability by staging different mishaps. By the end of the training, the robot could execute these tasks with complete accuracy, a crucial metric for potential industry and manufacturing applications of robots that require reliability.