Teaching bots how the world works
Photo: Adam Berry/Getty
Robots are getting pretty good at the repetitive, precise tasks that make up a good deal of factory and warehouse work. But place one in a home it's never seen before, or on a busy sidewalk, and it's likely to struggle to get around or do anything useful.
Driving the news: These chaotic scenarios — called "edge cases," because no two are the same — are the singular focus of a new robotics startup that was announced today. The high-powered venture wants to teach robots to think more like people in order to navigate the world.
The big picture: A wild debate has been raging in AI, and it's all about rules. One side says that machines should learn nearly everything from scratch; the other says that computers — like humans — must lean on some basic concepts about the world.
The team behind the new startup, Robust.AI, is firmly in the second camp.
- One co-founder is Gary Marcus, an NYU psychologist and AI expert who carries the banner for scientists who don't believe AI can learn how to navigate through the world without some level of prior knowledge about how it works.
- Another is Rodney Brooks, a legendary MIT roboticist who previously built Rethink Robotics, which sold factory robots meant to work alongside humans. Rethink folded last year.
No robot today can deliver a package all the way to any doorstep, or take care of an elderly person in their home. "For those kinds of situations, you need robots that can actually think for themselves — robots that can deal with an ever-changing world," Marcus says.
- He argues that deep learning — a reigning AI technique that teaches machines patterns without any hard rules — can't do the job on its own.
- "In order for these machines to reason and operate with more humanlike priors and a deeper understanding of the world, just brute-forcing deep learning is not going to get you there," says Peter Barrett, co-founder of VC firm Playground Global, which led the seed-round investment in Robust.AI.
Bringing back ideas from the era of symbolic AI — a focus on ground rules that died out in the 1980s — is a potential way forward, Barrett says. "I see it as absolutely necessary if we really want to close the gap between the tour de force mechanical capabilities of today's robots and their rather limited intellectual capacities."