Artificial intelligence has made tremendous progress in areas like image and speech recognition, largely by training the machines on large sets of labeled data. But robots that have to navigate the real world face unique challenges. As a result, robots are still largely limited to highly structured environments like factories, where they perform repetitive tasks.
Credit: T. Haarnoja et al, Google/UC Berkeley
Sergey Levine, an assistant professor in electrical engineering at UC Berkeley, says that if robots are ever going to find their way into homes and our broader daily lives, they need to teach themselves the common sense that would let them navigate unknown and unstructured environments.
Speaking at MIT Technology Review’s EmTech Digital conference in San Francisco, Levine gave several examples of robots making remarkable progress in teaching themselves how to navigate the world without labeled data or human supervision. In one recent example, a quadrupedal robot used an AI technique called deep reinforcement learning to learn how to “walk” after only two hours.
How long before robots are capable enough to live in our homes? Hard to predict, says Levine. In the near term, however, he sees them being increasingly used in various delivery tasks and in places such as hospitals for tasks like making beds.
