Eight years ago, not even a million dollars went into startups focused exclusively on humanoids. Today, the total amount of private capital for the industry has climbed over $1.4B. Just this month, Citi’s leading investment analysts published “The Rise of AI Robots: Physical AI is Coming for You.”
In summarizing his report, contributing author Fei Wenyan declared on a Yahoo podcast last week, “By 2050, we are looking at a $7T market for humanoids.” His contributing researcher, Rob Garlick, supported Wenyan’s bold statement by justifying that the biggest difference today is the speed of generative artificial intelligence (GenAI).
“Robots are hardly new, but there’s a number of new developments that are happening. Artificial intelligence [is] probably the most important,” stated Garlick. Wenyan and Garlick identify 50 humanoid companies around the world developing a variety of capabilities, “The Unitree H1 can move at 3.3m/s and can dance; should you want a robot that can do backflips and parkour, Atlas from Boston Dynamics can; Phoenix’s Gen 7 can learn complex tasks in under 24 hours; and Unitree’s G1 humanoid can fold itself up to fit into a cupboard.”
The report offers a worrisome comparison of human wages versus the favorable economics of robots using Tesla’s Optimus price point: “The combination of growing intelligence and growing dexterity means that humanoids could substitute for an increasing number of jobs… if Elon Musk’s prediction of a $25,000 ($20-30k) price point for Telsa’s Optimus is correct, a 36-week payback period is possible using the lowest US minimum wages of $7.25. Minimum wages in California ($16), average factory wages ($28), and average US nurse wages ($41) are added to highlight other payback scenarios. The conclusion in each is clear, that humanoids could be very compelling economically.”
This past week, I sat down with Nic Radford of Persona AI, one of the newest entrees into the humanoid arena. While Radford and his team are starting their business, he is no stranger to the science of humanoids, as for over a decade, he led NASA’s Robonaut and Valkyrie programs – one of the first humanoids deployed in the galaxy.
“I was fortunate enough to work on the research at NASA. It actually had economics that supported us looking at humanoids to partner with astronauts. So it’s about $300,000 an hour for an astronaut to be inside the Space Station, and it’s over a million dollars an hour to have an astronaut go EVA [extravehicular activity] or outside the Space Station. So when you’re dealing with those economics, it makes sense to have a multi-million dollar humanoid helper,” shared Radford.
I nudged Radford to explain what has changed today that wasn’t possible years earlier. “When you look at robots out in the world today, and you compare them with robots 10 and 15 years ago, from an electromechanical standpoint, there’s not much difference. What’s really changed is what’s between the ears of the humanoid, how it can perceive the environment very quickly, how it can make determinations about what to do,” he described.
Radford’s comment is illustrated by the uptick of new GenAI robot tools on the market, superficially to train robots to navigate our human world. One new standout is Hilbot, spun out of Henrik Christensen‘s lab at UC San Diego. Hilbot markets itself as a paradigm shift in robotic training, “By using a simulation-based approach, Hillbot is able to rapidly generate the vast amounts of training data necessary to train ‘foundation models’ for robotics with versatile robotic capabilities. Foundation models, similar to those used in language processing, provide a generalizable skill set that can be adapted for specific applications, reducing the need to train robots from scratch for each new task.” Hilbot’s claim is further elaborated in its blog: “The use of simulation-driven training not only accelerates the development process but also reduces costs associated with manual training in real-world settings. Moreover, as Hillbot continues to refine its foundation models, the company aims to build a framework where robots can quickly be adapted to new tasks by simply retraining their models in simulation.” Hilbot joins a growing list of other startups, including Bezos-backed Physical Intelligence, Texas startup Lucky Robots, and New York’s own Standard Bots.
These new foundational models enable the hardest nut to crack – dexterity. While folding laundry is a big step forward, it is far from the skills of a seven-year-old tying his shoes. Radford stressed that the goalposts have now moved beyond cool demos of machines pouring soda or folding shirts, “You’re not seeing a lot of examples yet of fine manipulation, which turns out to be where there are a lot of the higher paying jobs. Are we going to see this ubiquitous deployment of humanoids everywhere? Well, I actually believe so. It’s just a question of when.” While Persona AI’s website is deliberately vague on the company’s approach, he conveyed, “Persona has its niche. I’m excited to be able to unveil that when the time is right. If you want to draw a car analogy, we’re not building a family sedan. We’re building a truck. We’re building something that has much higher utility, and a little bit more ruggedness.”
Radford’s design perspective is in line with the feats he accomplished at NASA and its ground-breaking robotic hand that turned switches and dials on the Space Station. As he further elaborated, “A humanoid platform is nothing more than an embodiment that carries around a pair of manipulators and a pair of hands. And it turns out the human hand is fascinating. It’s fascinatingly useful, and it’s hard. It’s extremely hard to grasp. It’s extremely hard to manipulate objects. Therefore, we [the robotics community] started with some tasks where we only needed to clamp onto materials and move them around, whether it was a box or piece of sheet metal or whatever. And then, on the other side of that continuum, we’re doing a hand that was dexterous enough to put together an iPhone using tiny screwdrivers, tiny screws. I’ve always said that we’ll know when we’ve got dexterity solved when you can reach your hand in your pocket and pull out the penny over the nickel.” In my opinion, Radford’s penny challenge has thrown down the gauntlet of a new Turing test milestone for humanoids.
Radford’s optimism is joined by many in the field, including Elon Musk, who predicted, “By 2040, there will be at least 10 billion humanoid robots.” Musk further estimated that the sale of his fleets of autonomous systems could make his already hugely successful company achieve new revenue heights: “Robotic taxis makes Tesla about a $5T. The Optimus Robot, I think, makes Tesla a $25T company.” Persona’s founder agreed with his fellow Texan entrepreneur but added a few cautionary words, “I think when you’re talking about a multi-trillion dollar market, the investment we’re seeing into it is warranted because the market potential is so high. I’ve spoken to investors who call this the largest TAM of our lifetime. But it’s going to take a coalition of the willing from the financiers and the customers and the technology providers to truly make it happen. Because there’s going to be some moments where everybody goes shit, this isn’t going to work. I can only imagine in the self-driving community, there were some times where, like, this is never going to freaking work. It’s too variable. There are too many conditions and too many edge cases. And yes, I still have to monitor my Model X’s autopilot, but it did drive me home last night.”
