REDWOOD CITY, CALIF.» On a recent morning, I knocked on the front door of a handsome two-story home in Redwood City, Calif.. Within seconds, the door was opened by a faceless robot dressed in a beige bodysuit that clung tight to its trim waist and long legs.

This svelte humanoid greeted me with what seemed to be a Scandinavian accent, and I offered to shake hands. As our palms met, it said: “I have a firm grip.”

When the home’s owner, a Norwegian engineer named Bernt Børnich, asked for some bottled water, the robot turned, walked into the kitchen and opened the refrigerator with one hand.

Artificial intelligence is already driving cars, writing essays and writing computer code. Now humanoids, machines built to look like humans and powered by AI, are poised to move into our homes so they can help with the daily chores. Børnich is CEO and founder of a startup called 1X. Before the end of the year, his company hopes to put his robot, Neo, into more than 100 homes in Silicon Valley and elsewhere.

His startup is among the dozens of companies planning to sell humanoids for homes and businesses. Investors have poured $7.2 billion into more than 50 startups since 2015, according to PitchBook, a research firm that tracks the tech industry. The humanoid frenzy reached a new peak last year, when investments topped $1.6 billion. That did not include the billions that Elon Musk and Tesla, his electric car company, are pumping into Optimus, a humanoid they began building in 2021.

Entrepreneurs like Børnich and Musk believe that humanoids will one day do much of the physical work that is now handled by people, including household chores like wiping counters and emptying dishwashers, warehouse work like sorting packages, and factory labor like building cars on an assembly line.

Simpler robots — small robotic arms and autonomous carts, for instance — have long shared the workload at warehouses and factories.

Now companies are betting that machines can tackle a wider range of tasks by mimicking the ways that people walk, bend, twist, reach, grip and generally get things done.

Because homes, offices and warehouses are already built for humans, these companies argue, humanoids are better equipped to navigate the world than any other robot. The push toward humanoid labor has been building for years, fueled by advances in both robotic hardware and AI technologies that allow robots to rapidly learn new skills. But these humanoids are still a bit of a mirage.

Internet videos have circulated for years showing the remarkable dexterity of these machines, but they are often remotely guided by humans. And simple tasks like loading the dishwasher are anything but simple for them.

“There are many videos out there that give a false impression of these robots,” said Ken Goldberg, a robotics professor at the University of California, Berkeley. “Though they look like humans, they aren’t always behaving like humans.”

Neo said “Hello” with a Scandinavian accent because it was operated by a Norwegian technician in the basement of Børnich’s home. (Ultimately, the company wants to build call centers where perhaps dozens of technicians would support robots.)

The robot walked through the dining room and kitchen on its own. But the technician spoke for Neo and remotely guided its hands via a virtual reality headset and two wireless joysticks. Robots are still learning to navigate the world on their own. And they need a lot of help doing it. At least for now.

‘I saw a level of hardware that I did not think was possible’

I first visited 1X’s offices in Silicon Valley nearly a year ago. When a robot named Eve entered the room, opening and closing the door, I could not shake the feeling that this wideeyed robot was really a person in costume.

Eve moved on wheels, not legs. Yet it still felt human. I thought of “Sleeper,” the 1973 Woody Allen sci-fi comedy filled with robotic butlers.

The company’s engineers had already built Neo, but it hadn’t learned to walk. An early version hung on the wall of the company’s lab.

In 2022, Børnich logged on to a Zoom call with an AI researcher named Eric Jang. They had never met.

Jang, now 30, worked in a robotics lab at Google’s Silicon Valley headquarters, and Børnich, now 42, ran a startup in Norway called Halodi Robotics.

A would-be investor had asked Jang to gather some information about Halodi to see if it was worth an investment. Børnich showed off Eve. It was something he had dreamed of building since he was a teenager, inspired — like many roboticists — by science fiction (his personal favorite: the 1982 movie “Blade Runner”). Jang was entranced by the way that Eve moved. He compared the Zoom call to a scene in the scifi television drama “Westworld” in which a man attends a cocktail party and is shocked to learn that everyone in the room is a robot.

“I saw a level of hardware that I did not think was possible,” Jang said.

The would-be investor did not invest in Halodi. But Jang soon convinced Børnich to join forces.

Jang was part of a Google team teaching robots new skills using mathematical systems called neural networks, which allow robots to learn from data that depicts real-world tasks. After seeing Eve, Jang told Børnich that they should apply the same technique to humanoids.

The result was a cross- Atlantic company they renamed 1X. The startup, which has grown to around 200 employees, is now backed by more than $125 million in funding from investors that include Tiger Global and OpenAI.

‘All of this is learned behavior’

When I returned to the company’s lab about six months after meeting Eve, I was greeted by a walking Neo. They had taught it to walk entirely in the digital world. By simulating the physics of the real world in a video-game-like environment, they could train a digital version of their robot to stand and balance and, eventually, take steps.

After months spent training this digital robot, they transferred everything it had learned to a physical humanoid.

If I stepped into Neo’s path, it would stop and move around me. If I pushed its chest, it stayed on its feet. Sometimes, it stumbled or did not quite know what to do. But it could walk around a room much like people do.

“All of this is learned behavior,” Jang said, as Neo clicked against the floor with each step. “If we put it into any environment, it should know how to do this.”

Training a robot to do household chores, however, is an entirely different prospect. Because the physics of loading a dishwasher or folding laundry are exceedingly complex, 1X cannot teach these tasks in the virtual world. It has to gather data inside real homes.

When I visited Børnich’s home a month later, Neo started to struggle with the refrigerator’s stainlesssteel door. The robot’s Wi- Fi connection had dropped. But once the hidden technician rebooted the Wi-Fi, he seamlessly guided the robot through its small task. Neo handed me a bottled water.

I also watched Neo load a washing machine, squatting gingerly to lift clothes from a laundry basket. And as Børnich and I chatted outside the kitchen, the robot started wiping the counters. All of this was done via remote control.

‘What we are selling is more of a journey than a destination’

By guiding Neo through households chores, Børnich and his team can gather data — using cameras and other sensors installed on the robot itself — that show how these tasks are done. Then 1X engineers can use this data to expand and improve Neo’s skills.

Just as ChatGPT can learn to write term papers by analyzing text culled from the internet, a robot can learn to clean windows by pinpointing patterns in hours of digital video.

Building a humanoid like Neo costs about as much as building a small car — tens of thousands of dollars.

To reach its potential, Neo must capture video of what happens inside homes. In some cases, technicians will see what happens in real time. Fundamentally, this is a robot that learns on the job.

“What we are selling is more of a journey than a destination,” Børnich said. “It is going to be a really bumpy road, but Neo will do things that are truly useful.”

When I asked Børnich how the company would handle privacy once the humanoids were inside customers’ homes, he explained that technicians, working from remote call centers, would only take control of the robot if they received approval from the owner via a smartphone app.

He also said data would not be used to train new systems until at least 24 hours after it was gathered. That would allow 1X to delete any videos that customers did not want the company to use.

“We want you to give us your data on your terms,” Børnich said.