There’s a new way to have robots learn from their mistakes
By thinking of every incorrect action in one task as a way to do part of a different one, we can give AI the gift of hindsight.
Background: When humans mess up, they can learn several things: that an approach to a task didn’t work, but also that the method they just tried might be helpful for some other job. But when robots try to master tasks by themselves, they typically only learn by getting a reward for every step of a job they do correctly.
Useful mistakes:IEEE Spectrumreport that OpenAI, a nonprofit research company, released free software called Hindsight Experience Replay (HER) that lets an AI’s “failures” become successes. It does that by looking at how every attempt at one task could be applied to others. (The software also includes virtual environments where AIs can practice things like picking up objects or holding a pen.)
More realistic robo-training: HER doesn’t give robots rewards for getting a step of a task right—it only hands them out if the entire thing is done properly. That’s closer to how robots will learn in real life, but it usually slows training right down. Still, because every failed attempt can also get used for another job, that’s less of a problem in OpenAI’s system.