A humanoid-robot revolution is coming. Don't worry - here's why it will take a while.

By William Gavin and Christine Ji

Robots that mimic humans are set to create a $5 trillion market. But it will take years and a lot of improvements to get there.

Humanoid robots are expected to be deployed in factories and households - carrying out both basic and, potentially, advanced tasks.

Bringing robots that look and move like humans to households and factories is closer than ever. But despite promises from a growing number of startups, fully autonomous humanoid robots are still years away.

There's a few good reasons for that - chief among them being that it's really difficult to make a good robot, much less a humanoid one that can operate on its own.

"Humanoid robots are so hard," Brett Adcock, who leads the robotics startup Figure AI, said recently. "Both the dynamics and high dimensionality make it such a highly complex system-level problem."

While some companies are aiming to ramp up their efforts over the next several years, fewer than 1 million humanoid robots are expected to be sold in the U.S. by the end of the decade, according to Morgan Stanley's predictions, while it will take until after 2036 for more than 1 million to be sold in a single year. By 2050, though, Morgan Stanley predicts the market could be worth more than $5 trillion.

Alberto Rodriguez, the director of robot behavior for Boston Dynamics' Atlas humanoid, says most humanoid robots are still in the early stages of development - with companies focused on figuring out how to refine and apply their technology.

The biggest issue holding robotics companies back, Rodriguez said, is ensuring reliability. After all, what's the point in putting a robot on the factory floor or in a home if it can't keep up with humans?

"It's the most important mission" of modern robotics, said Rodriguez, who predicted that it will take about four years for thousands of Atlas units to be deployed in the field. Boston Dynamics' owner, Hyundai Motor Group (KR:005380), plans to buy and deploy "tens of thousands" of its robots in the coming years for industrial purposes.

Why humanoids?

From C-3PO to the Terminator, humanoid robots have dominated our science-fiction dreams for decades. However, some experts say the fixation on the human form is misplaced, to an extent.

Optimizing a robot to look like a human often results in the robot performing poorly at specialized tasks, according to Tom Chi, a founding member of Alphabet's (GOOGL) (GOOG) R&D factory Google X, where he worked on self-driving cars and Google Glass. Now a founding partner at At One Ventures, Chi points to the Roomba as an example: It's a simple disc that can clean floors much more efficiently than a humanoid robot, for a fraction of the price.

Andrew Anagnost, CEO of the engineering-software firm Autodesk (ADSK), expressed similar sentiments. Certain types of industrial robots don't need to mimic the human form to be effective, Anagnost told MarketWatch. It's often more efficient to design a factory around the capabilities of specialized machines, rather than retrofitting humanoids into human spaces, he said.

Others argue that it makes sense to design robots that can, eventually, integrate into new environments designed for humans, by humans. The promise of replicating a future previously only really witnessed in science fiction has also helped fuel investment in more than a dozen startups working on humanoid robots.

But there are serious issues that the industry needs to grapple with - including safety, which roboticist Christian Hubicki said is one of the hardest problems "in all of robotics." Researchers have a lot to consider on that front, including how to prevent robots from falling in a way that can harm someone and detecting people or even pets.

There's also the issue of a robot knowing its strength, especially considering companies plan to deploy humanoids on factory floors alongside workers. Figure, a startup with a $39 billion valuation, was sued last month by a former engineer who claims a malfunctioning robot had "carved a one-quarter-inch gash into a steel refrigerator door." The company has disputed the engineer's claims.

Companies are also having a tough time working out kinks in their hardware. Tesla (TSLA), for example, has had difficulty replicating the human hand - which some have labeled the "last-mile problem" facing the industry - for its Optimus humanoid expected to go into production next year.

Read: What's ahead for Tesla as it doubles down on AI

That's because of the complexity of both the human hand itself as well as the human sense of touch, according to University of Delaware professor Charles Dhong, whose research involves replicating touch.

"It's an incredibly difficult thing, especially to create a hand that is as dexterous and capable as the human hand," said Tesla CEO Elon Musk during an October earnings call. "The human hand is an incredible thing."

Ayanna Howard, a roboticist and dean of the Ohio State University's engineering school, said that the human hand, five fingers and all, isn't "necessarily the optimal configuration" for a robot. What's important is understanding how human hands can grasp an object, and then extrapolating that to whatever synthetic appendage has been chosen, she added.

Some, like Tesla, Figure and 1X Technologies, have committed to making human-like hands with five fingers. Others have decided to go with more human-inspired approaches.

Boston Dynamics gave its Atlas robots "grippers" with just three digits, one of which can rotate to also serve as a thumb to help it hold objects. Agility Robotics, which is focused on making humanoids for factories, eschewed hands altogether, opting to directly attach tools to its Digit robot's arms instead.

Further advancements in artificial intelligence are also likely needed. Today's leading large language models don't have the capability to power humanoid robots, Anagnost argues, due to the fact that they lack spatial reasoning. While LLMs can predict the next word in a sequence, they don't possess an understanding of physics, and they can't visualize and manipulate objects and spaces in three dimensions.

Read on: As ChatGPT turns 3, here's what's crashing the party

That's "something that language is never going to solve," Anagnost said. Instead, a growing number of AI researchers are exploring "world models," or AI that can absorb and process large quantities of sensory data. Anagnost points to the work of Stanford professor Fei-Fei Li and her startup, World Labs, as the "next frontier" for the industry.

For Atlas, Boston Dynamics is developing large behavior models, a type of AI system that focuses on replicating specific actions. Researchers rely heavily on teleoperation - remotely controlling robots - to collect data for processing, said Rodriguez.

That method, which many players use, including Tesla, is time-intensive and requires mass repetition of tasks to collect data. Other options companies have explored include training on synthetic data and through simulations, which provide an easier way to scale but can offer more imprecise results.

Sunday Robotics, a startup that only recently exited stealth mode, says it has a more efficient method than traditional teleoperation. By using custom gloves that mimic the hand of its Memo robot and employing workers in various homes to complete tasks, CEO Tony Zhao says costs can be dramatically reduced.

According to Zhao, his company's approach also provides better feedback on how much force is needed for tasks like folding socks. That can be a game changer when training robots for more delicate processes, such as picking up wine glasses or putting away eggs.

Ready for testing

Despite glaring issues, it's never been cheaper to get a humanoid robot. Just don't expect them to be very useful.

Most cheap full-sized humanoids come from companies based in China - such as Unitree Robotics, which sells some models for just thousands of dollars. There's also a market for smaller robots, such as Beijing-based Noetix Robotics' child-sized robot that costs less than $1,500. Although consumers can purchase such products, they're most often being bought by researchers or hobbyists.

"This is a very new phenomenon and it is really help[ing] push the field forward, because now anyone basically could have a humanoid robot to play around with," said Hubicki, a professor at the Florida A&M-Florida State University College of Engineering. "That's only been good for the field."

While most U.S. companies have yet to slap a price tag on their humanoids, a handful have a rough estimate of what they would charge for a stay-at-home robot. Tesla wants to sell Optimus for between $20,000 and $30,000, Musk has said.

1X Technologies has started taking $200 deposits for its home robot, Neo, through an early-adopter program. The California company is offering the humanoid for either a one-time $20,000 payment that comes with a three-year warranty, or for a $499-per-month subscription.

But there's a catch: The reportedly clumsy, but cozy-looking, robot will be controlled at times by a 1X teleoperator. According to the Wall Street Journal, which tested Neo in October, Neo is not smart enough yet to operate totally independently.

Bernt Børnich, the company's CEO, told the Journal that Neo should be capable of doing "most things in your home autonomously" next year, but that the quality may not match what humans are capable of. 1X will also have safety guardrails to ensure customer privacy, he said.

Sunday Robotics is another company planning to put limited quantities of humanoids in U.S. homes next year. It's started a free beta program that consumers can apply to join, with trials set for late 2026. Sunday says it costs about $25,000 to make its Memo robot prototypes, but that should be cut down to less than $10,000 as it scales production.

(MORE TO FOLLOW) Dow Jones Newswires

12-27-25 0900ET