Feature The first telephone call in 1876 was marked by Alexander Graham Bell’s request to his assistant, Thomas, “Mr. Watson, come here. I want to see you.”
The first message over the internet in 1969, then known as ARPANET, was “LO” – which would have been “LOGIN” had the system not crashed.
And the first robotic foundation model API call in 2023 was, “Put the eggplant in the pot,” according to Sergey Levine, co-founder of Physical Intelligence.
Levine kicked off the Humanoids Summit at the Computer History Museum in California, an event intended to explore the state of humanoid robotics and artificial intelligence.
As far as milestones go, vegetable placement lacks drama, but it represents a meaningful marker on the road to making humanoid robots relevant to the general public. Manipulating aubergine without incident is not a trivial technical challenge.
Roving robots have already arrived for those in San Francisco, California; Phoenix, Arizona; Los Angeles, California; or Austin, Texas, in the form of Waymo autonomous vehicles.
Humanoid robots have made a showing in laboratory and factory settings. Fourier debuted its GR-2 line of humanoid robots in September. Figure’s O2 humanoid robot has been deployed by BMW. And Boston Dynamics’ Atlas has starred in various videos doing stunts.
Image of a Fourier GR1 robot – Click to enlarge
Soon – the exact date to be determined – human-like bots will be available to the general public.
Based on slides presented by Bernt Børnich, cofounder and CEO of 1X Technologies, the firm’s Neo humanoid robot is scheduled to debut in 2025 for a set of early adopters in the San Francisco Bay Area.
That comes with a caveat, however. Neo, in its initial rollout, will not be fully autonomous. Rather it will be teleoperated.
“It will have some tasks that will be done by AI,” explained Jorge Milburn, VP of sales for 1X. “Navigation, for sure, pick-and-place, and other stuff will come out of the box.”
But, he said, Neo will be paired with a human teleoperator capable of directing the robot from afar – probably on a one-to-one basis. “Over time, the idea is more supervision than actual tele-op,” he said.
Milburn said 1X is likely to be conservative about where it places Neo initially – meaning homes with children would be avoided due to the higher safety concerns. Neo is only 5′ 4″ and 66 lbs, with four hours of battery life, but it could still do damage if it fell on a child. The biz is looking specifically for customers who want to provide feedback and help it gather the data to improve Neo.
The main bottleneck is the diversity of data … The path to widespread adoption in society for robotics is going to be through the home, because you need that data
Beyond assisting with various mundane tasks, Milburn said he expects the companionship aspect of Neo will be important, which would entail conversing either with the resident AI model or the human teleoperator.
“We will also probably offer the ability to teleport yourself,” he said, meaning that those away from home with a VR headset could see through the eyes of their home-bound bot and presumably take over some teleoperation or communicate.
The limited autonomy of the first robots headed into people’s homes follows from the fact that, like autonomous cars, a lot of data is required for a machine to operate in a residential environment – and that data has not yet been gathered. Simulated data can help, but for humanoid robots to navigate effectively in the human world, they have to stumble and fumble before they can walk among us.
“The main bottleneck is the diversity of data,” said Børnich. “The path to widespread adoption in society for robotics is going to be through the home, because you need that data.”
Just as Waymo robotaxis initially were accompanied by a human supervisor, Neo bots and other domestic human-like machines will head out into the world, awkwardly, under human oversight, to gather the data required to allow them, eventually, to operate on their own. And as with mobile phones, expect that the privacy policies imposed by robot makers will not afford much privacy – the robot’s video record of dropping the breakfast you requested will help company engineers improve its performance.
Levine’s firm, Physical Intelligence, has been working on an AI model that should help robot makers move beyond teleoperation training wheels. It’s a general-purpose robot foundation model called π0 (pi-zero).
It’s not just a vision-language (multimodal) model along the lines of Google’s Gemini 2.0 Flash. It’s a vision-language-action model, meaning its responses can be translated into commands that control robot hardware – similar to a software agent system that has access to physical hardware rather than software applications.
“The concept for a robotic foundation model is a model that is trained on data from a huge range of different robotic systems performing a wide variety of different tasks,” Levine explained.
“It’s pre-trained on internet-scale data that gives an understanding of the vision and language. And it can then be prompted to perform tasks or fine-tuned to particular application domains with a lot less data or a lot less effort to take a spin off those applications from scratch.”
This approach has been shown to work – a robot armed with the π0 model can fold laundry, bus tables, and assemble boxes. Human workers are still better and faster at these tasks, but robots are improving.
The employment consequences of humanoid robots came up in various panel discussions. Much was made about aging populations and jobs going unfilled as a rationale for developing capable robots. Those selling mechanized labor often said their machines would take jobs no human wanted to do or would simply help people focus on more fulfilling aspects of their jobs.
A lot of the humanoids we see out there are a hammer in search of a nail
No doubt there will be some of that, but it’s equally true that companies have shown little loyalty to human employees when they have the opportunity to reduce labor costs through offshoring. Where robots can do jobs more affordably than people, with at least the same level of competence and efficiency, expect humans to lose out over time.
Some skepticism was raised about whether the humanoid form factor is the right choice, given that in some scenarios, a fixed robot arm or wheeled machine might make more sense.
“A lot of the humanoids we see out there are a hammer in search of a nail,” observed Leo Chen, director of US operations for Engineered Arts, during a panel discussion.
Chen pointed to the Boston Dynamics Atlas robot, one of the most advanced robots out there, saying that the most revenue the company earned from it came from entertainment – from putting it in a Super Bowl ad – rather than from its ability to perform any practical application.
It would be really, really, really nice to get some diversity in our robot designs
As far as the tech industry is concerned, the right form factor is whatever customers will buy. And certainly for scenarios that require operating in environments designed for people, there’s a good argument for making robots in our image.
The form factor of those making robots also came up. Ross Mead, founder and CEO of Semio, observed that most humanoid robots reflect his boyhood ideas about what robots should look like. Just as the six people on his panel were all men, he said, current robots tend to reflect the design priorities of dudes. “It would be really, really, really nice to get some diversity in our robot designs,” he said. ®
