Instead of programming robots to handle each step of a new job, Pieter Abbeel, an assistant professor at UC Berkeley, has created robots that can observe humans demonstrating a task and then mimic them, or learn from pictures how to handle a piece of flexible material they’ve never seen before. His robots have learned to perform flying acrobatics, tie surgical sutures, and neatly sort socks.
Abbeel’s key innovation was to program the robots so that they can reliably infer the underlying intent of their instructors, filtering out the “noise”—irrelevant variations, or even slight mistakes, in the instructors’ demonstrations. Each robot is usually shown around 10 demonstrations before it can extract general rules of behavior. Even without an instructor, it can sometimes work out what to do. For example, Abbeel taught one robot how to fold laundry by giving it some general rules about how fabric behaves, and then showed it around 100 images of clothing so it could analyze how that particular clothing was likely to move as it was handled. After that, the robot could fold towels and sweaters without further instruction. —Kristina Grifantini