Posts from this author will be added to your daily email digest and your homepage feed.
This is Sources by Alex Heath, a newsletter about AI and the tech industry, syndicated just for The Verge subscribers once a week.
Building humanoid robots is Meta’s next “AR size bet,” a top executive told me recently. That suggests the company plans to spend billions of dollars on the effort.
During a recent conversation at Meta’s headquarters, CTO Andrew Bosworth said he stood up a robotics “research effort” earlier this year at the direction of CEO Mark Zuckerberg. The team’s existence has been reported on before, but Bosworth hadn’t discussed its strategy in-depth until our interview.
“I don’t think the hardware is the hard part,” he told me ahead of Meta’s recent Connect conference. “I’m not saying the hardware isn’t also hard, but it’s not the bottleneck. The bottleneck is the software.”
To demonstrate, Bosworth picked up my glass of water from a table between us. “If you know robotics, one of the biggest problems that you have is dexterous manipulation,” he said. “These robots, they can stand, they can run, they can do a flip, because the ground is a super stable thing.” By contrast, a robot trying to pick up the glass of water would likely “immediately crush it or spill all the water.”
“I’m not saying the hardware isn’t also hard, but it’s not the bottleneck. The bottleneck is the software.”
While Meta is currently building its own humanoid, or “Metabot” as it’s called internally, Bosworth envisions the company licensing its software platform to other robot manufacturers. “I don’t care about us being the hardware manufacturers,” he explained. Instead, Meta plans to approach it similarly to Google, which licenses its software to phone makers. The idea, according to Bosworth, is to take the software blueprint Meta develops and let any company use it “as long as your robot meets these specs.”
He said that Meta’s new Superintelligence AI lab is collaborating with the robotics group to build a “world model” that can “do the software simulation required to animate a dexterous hand.” (Google DeepMind’s Demis Hassabis has also talked frequently about his goal of building a world model that brings spatial awareness to AI.) The “sensor loop doesn’t exist” for a humanoid to be able to gingerly fetch a set of keys out of a jeans pocket like a human could, Bosworth explained. “So you have to build that data set.”
“We wouldn’t be doing this if we didn’t have the lab, if we didn’t think we were going to be building the models that could do it,” he said, referring to the new team being led by ex-Scale CEO Alexandr Wang.
... continue reading