Until recently, it seemed that the world was still comfortably confident that humanoid robots, when they finally became a thing, would rise as one and murder us all in our sleep. That science fiction-stoked paranoia has dimmed of late. Now, the internet is full of videos of these bipedal automata marching across factory floors, running half marathons, dancing, trolling their way across US and Chinese cities, or else being roundly kicked or punched by overly-enthusiastic engineers – and, very disconcertingly, not toppling over.
These are but harbingers of our robot-filled future, says Silicon Valley – one galvanised by breakthroughs in the ability of automata to perceive and interact with our world and put their metal hands to more or less any task. In March, Nvidia CEO Jensen Huang proclaimed that humanoids are less than five years away from “wandering around”, citing manufacturing as a first adopter. Tesla CEO Elon Musk had previously said the firm’s Optimus robot “can basically do anything,” including “babysitting your kids.” Even Amazon seems confident that automata will play a crucial role in its business, building a ‘humanoid park’ to test out the feasibility of robots springing out of its vans to deliver packages.
What suggests this is more than periodic Silicon Valley hype-mongering, however, is the parallel enthusiasm for these bipedal robots in China. At the World Robot Conference in Beijing this August, over 50 humanoid robot companies exhibited, reflecting the importance this technology has been given within the country’s national industrial strategy. Indeed, according to Janus Henderson Investors, global shipments of humanoid robots have reached an inflexion point, with the firm predicting that 8,000 mainly Chinese-made units will be shipped by 2040. Goldman Sachs, meanwhile, has estimated that, by 2035, the humanoid robot market could reach $38bn.
“I just wonder, if I’m a company asked to buy this, what is the advantage of the humanoid?” says Jeff Burnstein, who has been President of the Association for Advancing Automation (A3) for eighteen years. “Right now, we’re not even sure what it should look like: there are many bipedal ones, but that doesn’t mean it’s the winning form factor.”
How to use and abuse your average humanoid
Robots were always meant to be humanoid – that much is clear from photos of the 1920s play that invented the term, which derives from the Czech term for corvée labour – but working examples only date from the past few decades. Not all of them were practical. When put to work as a tour guide in a Tokyo museum, for example, Honda’s Asimo unit had trouble recognising the difference between visitors raising their hand to ask a question or petitioning the famous running robot for a selfie.
Slowly, though, humanoids have been filling in for humans in more mundane occupations. Burnstein recollects how he was waited on by one such automaton in his hotel during a recent trip to China. “Running upstairs with a toothbrush or a blanket,” he says, “are tasks people don’t really love doing.”
It’s the factory floor, however, where enthusiasts predict that humanoids will begin to live up to their long-vaunted potential. Long the domain of their one-armed or no-armed contemporaries – see the many influencers videoing their slack-jawed amazement at Ocado’s London warehouse – engineers and investors argue for the deployment of their bipedal brethren to automate the remaining dull, repetitive tasks that afflict the shrinking human workforce. According to the International Federation of Robots (IFR), that moment couldn’t come too soon, with demand for robot helpers sky-high in developed economies thanks to ‘persistent labour shortages and impending demographic shifts.’
Expectations vs reality
Carmakers seem especially enthusiastic early adopters. Tesla’s Optimus, for example, is undergoing testing in its Fremont, California, gigafactory, with owner Elon Musk predicting that “thousands” of units will be working at the facility by the end of this year, while Mercedes-Benz has used Apptronik robots to transport components at select manufacturing sites. BMW, too, has dabbled with humanoid robots, collaborating with the startup Figure to deploy the latter’s automata at its South Carolina factory – though the extent to which these units have actually been doing busy work has been vigorously disputed.
Currently, these robots are largely confined to fetching and carrying. Some roboticists, like Burnstein, remain convinced that this will remain the extent of humanoid achievement for a little while yet. While all of these units can walk and, in most cases, talk, they’re still not very dextrous. “Things like picking up a wine glass or changing a light bulb,” says Burnstein, “robots just can’t do.”
That’s down to a translation problem. Unlike large language models (LLMs), which can ably crunch swathes of the internet to train themselves to perform more or less any task that requires a written prompt and answer, robots require additional coaching to be able to learn from visual data provided by, say, a YouTube video in order to perform a physical task. According to a recent paper by the Berkeley roboticist Ken Goldberg, current data curation methods along these lines mean that it would take roughly 100,000 years for an all-purpose humanoid to emerge – a little too long for most venture capital firms.
RoboForce’s founder and CEO, Leo Ma, acknowledges that this data gap is a problem, but not an insurmountable one. In May, it introduced Titan, a semi-humanoid robot designed for real-world industrial deployment in demanding outdoor environments. Ma, who previously worked as an autonomous driving architect at Baidu’s R&D centre, says the data gap exists because the domain intelligence data comes from the actual business context, which is hard to acquire, and because capable hardware to collect the data is also missing. His team plans to overcome this hurdle by working with industry partners to collect data crunched by advanced physical AI models. The standardised nature of much of this work should also make it easier to collect and apply this data, he adds.
“The objects that workers handle day-to-day on a mine site or a solar farm are very different to the tools used to collect data in a lab,” says Ma. “Therefore, no matter how much data you collect in the lab, it’ll still be missing for the actual business case.”
Myriad challenges
But access to training data is far from the only hurdle for humanoids. While it’s expected humanoids could potentially do a multitude of tasks on factory floors such as picking up boxes or filling in for sick staff, it’s an open debate as to whether manufacturers would invest in them over specialised robots already in use.
“The simple truth is that specialised automation is cheaper and easier to maintain,” says Andrew Kinder, industry strategy lead at the UK enterprise software company Info. “They’ve been optimised over years to handle specific tasks, whether that’s welding, moving goods, or picking in warehouses.”
Surely such concerns are mitigated by the obvious advantage of employing a salary-less workforce? Not as yet. While the up-front cost of individual units is becoming more accessible to large enterprises – Unitree Robotics’ Unitree RI, for example, reportedly costs under US$6,000 (£4,400) – battery power is still typically in the realm of two hours.
“Humanoids must be lightweight in order to be agile, but moving things need significant power,” explains Seth Hutchinson, a computer science professor at Northeastern University. That, he adds, “typically translates to larger motors, which then add weight, creating a kind of paradox.”
The most obvious solution? Wheels, argues Ma. RoboForce’s Titan, for example, can work for up to 8 hours because it doesn’t need to be light on its feet and can therefore house bigger batteries. Wheeled robots, though not as agile as their bipedal counterparts, could also address significant safety concerns, says Burnstein, who recalls an A3 event where a humanoid by a “well-known” developer unexpectedly fell forward onto the ground (fortunately, no one was hurt.)
“If you’re working next to a humanoid and it loses its power, it could fall over carrying something heavy and land on somebody”, says Burnstein. “There are no standards yet to say what they need to do and how they need to stop.”
Humanoids by 2030?
The robot safety standardisation body at the International Organisation for Standards has started drafting a safety standard, as has A3, but it will take time. By that point, might we see humanoid robots roaming the streets? Ma thinks not, but does suggest that what he terms “robot internships” in the workplace may increase.
Hutchinson, meanwhile, is sceptical, suspecting that large technology companies are making glitzy investments in robotics startups to create new – and largely hypothetical – future use cases for their core products. “You have to ask, what’s the benefit to that person if I believe them?” he says. “Learning and failing is expensive” for such physically and computationally complex machines, he adds, “which can be a serious impediment to putting real robots into the physical world.”
Burnstein is similarly sceptical. “I’m more on the, ‘It’s not going to be in the next five years’ camp,” he says. “I think it will be longer for several reasons, including safety, but also, what will it cost to have this humanoid, and is it faster, cheaper, more consistent than a traditional industrial robot, or even a mobile robot with an arm? Are there other solutions that might be as efficient and more cost-effective?”
Eight years, says Burnstein, might be a more realistic timeframe. Even then, you’ll be a long way from convincing people to put these bipedal robots in their homes, when the most they’ll stretch to is a hubcap-sized Roomba.
Read more: Cyber insurance is big business. But modelling risk is still tough.
