After a hard shift working around the house and interacting with people and other machines, 1X Technologies’ soft, fabric-covered NEO humanoids, designed to be as endearing as a plushie, will sift out the details they need to retain from the mass of data they can discard.
“I want the robot I have at home to remember me and my kids and my family,” said Bernt Børnich, the Norwegian CEO of Palo Alto-based 1X Technologies, explaining to Newsweek that research on memory is a priority in the company’s push to develop embodied intelligence—robot brains.
“Essentially, robots will dream in the end, because they will need to take all the experiences from today and kind of back propagate it,” said Børnich.
One of the most significant technological competitions in the history of humanity is heating up: the race to build brains that will let robots think and perform multiple physical tasks alongside humans and, in so doing, revolutionize homes, factories, hospitals and other workplaces. That race is much harder than building the AI language models we’re now used to. And while China is becoming dominant when it comes to building robots that can now sprint quicker than the fastest human, the United States is by no means behind when it comes to developing their minds.
The CEO of robot brain company Skild AI likes to share a clip of robot hands carefully lighting a cigarette to make his point. The film is black and white. The year is 1957.
“The hardware was there even at that time,” Deepak Pathak told Newsweek on a midnight Zoom call from the West Coast. “It’s the brain that has been missing all these years, and this is exactly the opposite of what every robotics company has done…. What we are doing here is building a general purpose brain for robots.”
The challenge is clear at robot maker Unitree’s home in Hangzhou, China, a 2,000-year-old city that has reinvented itself as a tech hub 100 miles west of Shanghai. Unitree’s robots wowed viewers with perfectly choreographed dancing and martial arts at China’s annual Lunar New Year television gala. They competed alongside others in August’s World Humanoid Robot Games in Beijing, where a robot from another Chinese company ran 100 meters faster than any human in history. Yushu Technology Co., Ltd., which trades as Unitree, saw its shares soar more than 600 percent on its first day of trading on August 19, a sign of investor interest in robotics.
But what the robots can’t really do yet is think. At the showroom in Hangzhou, their movements are decided by a young man in a black parka standing in the corner with a game controller.
“The biggest challenge for the whole industry is the brain. All of the industry are trying to develop it,” Unitree’s Yolanda Xie told Newsweek in Hangzhou. “We want to make the foundation for the movement first. Once we have a solid foundation for the movement then we can add the brain.”
With China being the biggest testing ground for deploying humanoid robots, it has also become a source of viral videos of robots going wrong: kicking someone in a crowd or striking out at fellow workers. But the robots don’t yet have the intelligence to behave maliciously. “I don’t think we’ve made a robot as smart as a squirrel,” said Sam Kriegman of Northwestern University in Illinois, whose lab has been building and “breeding” strange robotic life forms as one route to the goal. “There’s probably 50 Nobel Prizes that need to be won and figured out to get a machine that’s that smart,” he told Newsweek.
At the heart of the challenge for developers is what’s known as Moravec’s paradox: that the level of computing needed for “hard” intellectual tasks, such as those being done better and better by large language models such as ChatGPT, is much lower than for “easy” tasks such as walking around or making a cup of coffee. That’s not to say everyone is pessimistic. Elon Musk’s Tesla, which did not respond to a request for comment, is targeting the end of 2027 for releasing its Optimus humanoid robot. 1X Technologies is aiming even earlier, with plans to deliver the first NEO later this year.
For now, a factory or warehouse is the best place for seeing the simple, repetitive and preprogrammed tasks that they can do. Such industrial sites are also becoming schools for developing robots with more advanced intelligence.
“We can leverage the fact that we don’t have to solve everybody’s problem. We just have to solve our customer’s problem,” said Pras Velagapudi, chief technology officer at Agility Robotics, which is initially focusing on industrial tasks with its robot Digit.
Agility this week unveiled its Digit 5 humanoid robot, saying it was the first that had been engineered to be able to work safely in close proximity to human workers at scale and without the physical safety barriers that had traditionally been needed.
The reward for developing a robot brain with general capabilities is immense, not only for whoever succeeds, but for the economy at large. The importance is even greater for industrialized countries with shrinking workforces, where robots could supplement human labor. That makes robotic intelligence a top priority for NVIDIA, whose processors, which power much of the AI industry, have helped make it the world’s most valuable company. It is now working with many of the companies in the race for the robot brain and is a key force behind them. “Eighty or ninety percent of the world’s GDP is in the physical world,” said Deepu Talla, head of robotics at NVIDIA, which is keeping parts of its robotics technology as open as possible to help accelerate innovation across the industry.
“Our business model is we sell the three computers: the infrastructure for training the brain, for testing the brain in simulation, and for the robot brain,” Talla told Newsweek.
“We desperately need robots for labor shortages or dangerous jobs, or a combination of robots and humans. We need a productivity increase in all of these to continue the deflation curve and improve the quality of life.”
The human brain is the product of more than 500 million years of evolution since the first neural tissue formed in worm-like animals. As individuals, we spend years learning to use them. Those are the timelines the robot brain engineers seek to compress. Problem number one is data—robots need much more than the text and images that are fed into large language models. There’s plenty of video on the internet of people performing almost any conceivable act, but the race has also spawned a whole subindustry of people with cameras strapped to their heads and wrists performing ordinary tasks to feed the demand for specialist video data just for training robots.
“We look at how if a human opens a bottle, we can see, OK, their hand is going towards it. How do you pull the bottle? Or even the bottle has to be held in this way,” said Skild AI’s Pathak.
But the visual data only explains what the world looks like, not what it’s like to interact with it as a robot—a challenge one step beyond that faced by self-driving vehicles.
“You could watch a video of how to hit a baseball, but then you have to actually go and practice the thing and try it in the real world yourself,” said Brian Ichter, a cofounder of Physical Intelligence, or Pi, which describes itself as bringing general purpose artificial intelligence into the physical world.
“The way we collect most of the data is we have people behind the robots, what we call puppeteering them. So there are two sets of arms and I’m basically moving one and marionetting the other robot. In this way, we can have basic data of the robots doing really complicated tasks, like folding a shirt or cooking a meal.”
Sensors in the fingers of 1X robots teach how much pressure to apply to an object, just as we learned as children that a baseball bat should be gripped more solidly than a banana.
Computer simulation plays an important part in testing too, but ultimately the robots need to encounter the real world—and even the robots’ greatest champions acknowledge that the current capabilities fall short of multiple impressive demo videos on YouTube recorded under favorable conditions.
The home is the hardest challenge to crack.
“You have kids, you have pets, you have different untrained folks who don’t necessarily follow any sort of rules around the robot, and so I think starting over there is very difficult,” said Agility’s Velagapudi. “We can start with some place where we know that we’ve solved enough to be valuable, and then just incrementally add improvements.”
For Pi, it meant renting dozens of Airbnbs to help its robot brains navigate everything from strange kitchen sinks to different kinds of coffee maker or wiping up spills. “Now, there’s still some challenges when we do that. It’s like mid-90 percent, so we need to get this closer to 100 percent. I think this really high performance is still a bottleneck, and particularly in general scenes,” Ichter said. “It definitely does laundry and it definitely makes the bed and I’d say coffee; it maybe depends on the type of coffee maker. We have taken the robot on the road and went out to San Diego for a conference and had it make espresso. So it transfers reasonably well, but I don’t know that it would go to your house and be able to make it just yet.”
One advantage robots have over humans is that it’s easier for them to share exactly what they have learned across their entire network. In theory, every skill a robot has learned can be taught to every other connected robot. But that creates new privacy challenges as well. How much information should a robot share with other robots or other people? How will it know how to behave?
“It’s not just that you don’t want your neighbor’s robot to know,” Børnich said. “What your robot talks to you about is not necessarily what it should talk to your wife about.”
Alongside concerns about privacy are questions of physical safety, liability for manufacturers and government regulation, as well as science-fiction fears of robots running amok and becoming masters rather than servants.
It is about much more than just what happens if you tell the robot to go to the kitchen to pick up an orange and instead it picks up a knife, said Agility’s CTO. “When we’re talking about safety, we’re talking about if this robot is just walking along and then its encoder ring, the part that tells it where its leg is, falls off—which is something that we’ve seen,” he said. “What’s it going to do? Is it going to fall on somebody? Is it going to disable itself properly? Is it going to degrade in a graceful way?
“Those are the types of safety concerns that are really kind of more fundamental,” he added.
An indication of the struggle to get it right is a four-legged and horned rescue robot equipped with a chainsaw and built by California-based Satyress that shocked social media viewers. It comes with detailed instructions on how to “kill” it if it malfunctions.
“What we care about more is a bad person misusing the robots to do a bad thing, not caring so much about the robots doing a bad thing,” said Unitree’s Xie. “In future, regulation will need to be more comprehensive.”
From safety at the individual level, there are also wider security implications for robots that will undoubtedly be put to use on the battlefield. Robot technology on the ground, as well as with drones in the air and water, is already reshaping wars from Ukraine to the Middle East. That brings a harder edge to a global competition that so far at least has involved significant collaboration between researchers and companies in adversaries China and the United States. U.S. labs use Chinese robots and Chinese companies also draw on American chips and other technologies as well as expertise.
“The argument becomes China is ahead in many ways because they are doing robotics very extensively, but this narrative is driven by hardware,” said Skild AI’s Pathak. “But it’s not like China is ahead in the brain.”
Some in the industry fear the atmosphere will change as the overall technology war comes under greater scrutiny from both Washington and Beijing, and as the robot brain technology advances towards deployment. Talk is of when it will reach the “ChatGPT moment”—when it takes off as the large language model did in late 2022. Some think that could still take decades, but the most bullish put it in a small number of years.
“We hope that in the future, we can see a robot be introduced into an unfamiliar household and it can achieve approximately 80 percent of tasks successfully through voice or text commands,” Unitree CEO Wang Xingxing told a Beijing robot conference in August.
Ichter told Newsweek that the pace of progress since founding the company in 2024 had exceeded his expectations, leading him to shorten his timeline estimates. “If I had written this out before we started Pi, I would have been at maybe five years is when robots are in the world doing something useful, and then some scaling beyond that. Instead, it’ll be like two and a half years to get there,” he said. “I feel like we’ve moved from 50 percent to mid-90 percent and really need to get these kind of things to 100 percent before they can really go out in the world.”
Updated on 09/16/2026 at 11.16 a.m. ET with Agility announcement of next generation robot