They marched in line, interacted with guests, and poured drinks while making small talk. The vision of the future Elon Musk put forward with his Optimus humanoid robots at Tesla’s recent Cybercab event earlier this month was as astounding as it was unbelievable.
That latter word was important: It turned out the robots weren’t operating under their own steam, but were instead remotely operated by humans. While the robots were able to walk down the street in Tesla’s grand parade to show off its tech under their own steam, pretty much everything else they did that night involved human intervention.
“We are in the proof of concept phase for robotics. While we’ve seen many demos online and on stage, the capabilities of these machines have been designed for controlled environments and limited use cases,” says Nayan Jain, executive director of AI at digital product studio ustwo. “If you look deeper at live demos and videos, the generic walk cycles and manipulation tasks like pick and place or tool use are technically impressive, but not robust enough to work independently.”
Jain was optimistic about the potential future of robots—but believes that for some of the higher-profile public examples, it’ll take some time to come to fruition. “We are not at the point yet of true automation with humanoids,” he says. “The robotic future is coming, but it may take longer than we expect.”
But while Tesla’s vision of the future may be a little while off, there are indications that—outside of the limelight that Musk draws to himself and his companies—humanoid robots are on the cusp of mass adoption. In factories in the United States right now, human-like robots work on the same factory floors as humans, albeit in their own fenced-off areas. And elsewhere in the world, the robotic future is even closer to reality, and in some instances, is already upon us.
In August, China hosted the World Robot Conference in Beijing, more than 30 humanoid robots from private companies and university labs were on display. The name is a misnomer: Almost all the 30 representatives came from China, but it highlighted how developed the technology is in the country. The Chinese government announced at the event that they hoped the most developed humanoid robots would be entering mass production in 2025, and would be allocating government funding to the fight.
China is the powerhouse of the humanoid robot industry at present, with half of all robotic installations worldwide in 2022 within the country’s borders. The country—and its cities—are willing to fund the development of robots significantly.
There has always been an interest in humanoid robots throughout history, and in theory, their development shouldn’t be too taxing as the world is designed by—and inhabited by—humans. But it’s not quite so easy, says Jonathan Aitken, an expert in robotics at the University of Sheffield. “The complexities of the real world have often proved significant roadblocks to humanoid development, whether that be the quality of actuators to produce motion, or decision-making being too slow to catch-up with real world events,” he says.
Historically, that’s required programming—and a lot of it. The hardware required for humanoid robots has been around for years, but it’s the way that the robot moves and interacts that has proven trickier to tackle. Companies like Boston Dynamics have had to carefully code decision making processes within their robots, including Atlas, its humanoid robot, to make it work. But that could change with the development of foundation models, thanks to the arrival of generative AI.
“Programming a humanoid is a complex task, especially because of the number of degrees of freedom that these robots have, which means that applying precise control at this scale is extremely difficult,” says Aitken. “Foundation models offer a shortcut through this problem.” Nvidia has developed its own foundational AI model for robots, Gr00t, which stands for Generalist Robot 00 Technology, to try and help improve the rollout of humanoid robots.
Aitken points out that they would need more checks and balances to ensure any foundation models powering robots are thoroughly tested, checked and verified, so there’s no odd behavior traits that come out when such robots are out in the wild. Yet if that can be achieved, it could be a game-changer. “The ability to generate code to control a robot at this scale of degrees of freedom, interacting efficiently with a wide variety of sensors with optimal decision making is a very desirable set of qualities,” he says.
Despite those challenges, there are American success stories when it comes to robots in the U.S. Third-party logistics provider GXO Logistics is testing humanoid robots from three different providers—Agility Robotics, Apptronik and Reflex—in its facilities across the country. Agility, whose robot Digit is one of the best-known examples of a humanoid robot, and which has ties to Oregon State University through its founder, has said it has several dozen so-called “alpha units” across a number of institutions and organizations, including GXO.
“I would say humanoids are having a moment right now, and much of that is because of the introduction of AI that definitely delivers value, just like it’s doing in other industries,” says Peggy Johnson, CEO of Agility Robotics. “We use AI to teach our robot new skills very quickly, rather than having engineers program it. With these LLMs like ChatGPT now, we can teach our robot Digit skills very, very quickly.”
“As the first logistics provider in the world to deploy humanoid robots into live operations, we are paving the way for the entire logistics industry and beyond,” says Adrian Stoch, chief automation officer at GXO Logistics. The company’s Operational Incubator program is testing the robots from all three companies in a variety of roles to see where they’re best suited.
So far, GXO’s Operational Incubator program claims credit for providing feedback and data that has resulted in longer battery life for such robots, the ability to handle heavier payloads, and enhanced gripping and stability of prototype robots. “Right now we aim for par with humans, but as we train it with new skills, we’re going to see the robot doing better than humans, maybe areas that it can operate faster than humans, with the introduction of more AI,” says Agility CEO Johnson. “So that that’ll be a real accelerant for various types of tasks across different industries.”
Others are also seeing robots start to move onto their production lines. In Spartanburg, South Carolina, Figure 02, the eponymous robot from robotics startup Figure AI, was tested within a BMW plant. Figure 02 has onboard visual language models, understanding and comprehending how to react to the input it receives from six on-board cameras, learning as it goes. Figure AI raised $675 million in Series B funding in February 2024 from notable investors like Jeff Bezos, Microsoft, Nvidia, and OpenAI, bringing its valuation to $2.6 billion.
That level of investment is significant—and is just the beginning, reckon analysts at Morgan Stanley. The bank believes the total addressable market for humanoid robots will reach multi-decatrillion-dollar levels. That investment will help bring down the cost of building and running robots to levels that make them competitive with the human labor force. Morgan Stanley estimates that a humanoid robot costing $50,000 with a lifespan of 50,000 hours equates to an operational cost of approximately $1 per hour, making it cost-effective compared to human labour, which is typically around $20 per hour and is increasing due to wage inflation.
However, being on the cusp of a movement doesn’t mean that it’s ready for prime time all the time. “While we’re incredibly excited about the opportunities presented by humanoids,” Stoch says, “there is a long way to go before mass deployment is achievable.”
Related Topics: