But language models can only do so much, and what robots need most urgently is a new kind of model that understands the physical world the way an LLM understands the written word. And to build that model, engineers need way more data. While walking around BAAI, I see scores of workers behind desks. They’re teleoperating various robot arms and grippers to teach algorithms simple manipulation tasks such as sweeping up beans on a table, pouring liquids from a jug into different cups, and picking items from shelves. A young man wearing a virtual reality headset appears to be making tea as a camera records his every move. The idea is that with enough training data, robots will intuit how to do all sorts of things without specific training.
The trouble is, nobody quite knows what data is most useful to the robots, let alone how much they need or how best to collect it. And for humanoids to become ubiquitous, people need to invent hardware that better mimics a human hand. For a robot, doing a backflip is a lot easier than flipping a coin.
Still, Tony Zhao, cofounder and CEO of Sunday Robotics, a California-based startup, tells me he worries that companies like his don’t stand much of a chance against Chinese firms, which can bring on more workers, like BAAI’s teleoperators, to train robot models and rapidly roll out new hardware. “The iteration speed, the US is losing there,” he says. “And honestly I don’t know how we can win.”
To try to keep up, Zhao recently recruited an executive from a Chinese robotics company with deep connections and experience tapping into China’s vast and complex supply chain. “The only way we can beat Chinese companies is to build a China team,” he says.
Some US CEOs, including Demis Hassabis of Google DeepMind and Lachy Groom of Physical Intelligence, both of whom are chasing the robo-ChatGPT moment, have told me they imagine robotics development roughly mirroring that of smartphones, where China makes the hardware and the US makes the brains. (Except that Huawei is now making both.)
The answer might be that the US government needs to get involved, suggests Jonathan Hurst, cofounder and chief robot officer of Agility, which makes humanoids. He’s imagining, among other things, heavy investment in advanced domestic manufacturing, such as tax incentives for firms that use robots in their warehouses and factories, as a way to prop up domestic robotics firms. Such a strategy might start to mimic the Chinese government’s patient capital investment in its industries. “We have to be very smart about automation,” he says. “It is the only way.”
My hotel in Beijing, in the high-tech hub of Zhongguancun, didn’t have any of the wheeled robots that routinely deliver items to guests’ rooms at some big-city hotels. Instead, mine had an unfailingly polite human called Stephen. When I needed to get a shirt cleaned, Stephen got the job done in just a few hours. As I flew home at the end of my trip, I reflected on how many hands had washed, pressed, packaged, and transported the garment back at such speed. Even in China, the robots have not yet won.
What Say You?
Let us know what you think about this article in the comments below. Alternatively, you can submit a letter to the editor at mail@wired.com.
