You can't joystick a robot on Mars in real time. Radio signals from Earth take up to 22 minutes to arrive. This simple fact has stalled space construction for decades.
The dream of building in space has long remained just that: an aspiration, a vision of an extraterrestrial jumping-off point for exploring the cosmos.
But the gap between dream and reality is narrowing, driven by rapid advancements in AI and robotics. Startups, labs and large companies are pursuing technologies to execute physical tasks like autonomously refueling a satellite with a drone, welding accurately in space and complex docking maneuvers.
Momentum is accelerating. High-profile spacecraft launches by Elon Musk and Jeff Bezos have captured public attention, while NASA is planning the first permanent Moon site with its Artemis program.
To enable construction without real-time radio contact, the industry is turning to physical AI — the combination of models, sensors, cameras and robots that enable autonomous construction beyond Earth. This technology is the only way to bridge the communications gap, which spans 2.5 seconds to the Moon and from 3 to 22 minutes to reach Mars, according to a 2011 NASA paper.
Robots must make real-time decisions: navigate around or through a crater, distinguish between boulders and shadows, and adjust to unexpected terrain. Training for such decision making is taking place in harsh conditions on Earth, “as the minimal prerequisite to make sure their physical AI is game time,” said Cosmic Robotics’ CEO and co-founder James Emerick.
“The best AI space construction companies will be deploying a large fleet of systems here on Earth, all to ensure that the equipment is reliable over long durations,” he said. The effort is “in service of generating internet-scale field data to train physical AI models with.”
Autonomy is not optional
It’s these scenarios that physical AI, its models fed by innumerable iterations and simulations, is uniquely designed to manage.
Industry forecasts suggest space construction could start within a decade, provided that between now and 2035, these machines can ingest enough data to master autonomous operation. “Autonomy is going to be critical for building in space,” Isaac Arthur, president of the National Space Society, a space advocacy organization, said. With the hardware largely ready to go, the countdown has begun.
Two San Francisco startups are part of this effort. Spacer Robotics and Cosmic Robotics are testing their equipment on Earth in terrestrial jobs for eventual Moon and Mars missions.
Spacer Robotics’ machines automate the tying of rebar, a mostly manual, time-consuming task of creating steel-rod forms for the pouring of concrete. Cosmic Robotics’ robots automate the installation of solar panels.
Space vs. Earth Purposes
Once launched into space, the machines’ purposes will change. Building with concrete on the Moon won’t require the use of rebar, which is too difficult to transport from Earth, according to Spacer founder and CEO Lesya Hendrix.
Spacer is primarily using earthbound tasks to train its robots for work on the Moon, where gravity is one-sixth of Earth's and hydraulics are useless. The robots, while performing tasks, are also hoovering up data, building models and doing simulations in order to build the physical AI’s brain.
“For space, things are different” than on Earth, said Hendrix, whose company is named after genetically modified humanoids in Isaac Asimov books including Robots and Empire. “The thing that carries over is, in fact, the AI model.”
Spacer has collected photos from construction sites and created a data set to train its visual model for the terrestrial environment. “That same technology, as we stumbled upon talking to NASA, is also very useful for the Moon,” she said.
“Basically, the AI model is the same, but we switch the data set from construction pictures to lunar pictures, and then the robot suddenly learns to navigate a lunar surface and understand that, oh, this is a boulder, I should drive around it; this is a crater, I should also go around it.”
Spacer is in talks with NASA for a contract to provide its technology, Hendrix said.
“You need massive, internet-scale amounts of data”
Cosmic Robotics is similarly gathering data while testing its robotic solar panel installers in California deserts and creating data sets to “train models to build more of a general purpose platform,” Emerick said. “That then can be potentially fine-tuned with a much smaller data set once you land a system on the Moon or Mars,” he said.
“You need massive, internet-scale amounts of data,” Emerick said. “When you go and send a rover or other piece of equipment to a different planet, we're not going to be sending a large fleet, we're going to be sending one or two. And the amount of data that they are able to collect and train on initially is just not going to be enough to have a robust and reliable model.”
Cosmic's ultimate goal is to build a city on Mars. In the interim, the company secured a grant from NASA earlier this year to develop its Particle physical AI engine. It is also working on the agency’s Tall Lunar Tower project, a plan to build a 100-meter power and communications tower for the Artemis project on the lunar surface.
“We’ve got to start here on Earth with building the right autonomy and gaining some trust around it,” said Emerick, who projects at least another 10 years before the first space construction begins. “And then pretty quickly, we'll have autonomous systems and the data to be able to start to put them on rockets and get them out of here.”
“Nobody wants to be the failure”
Projecting the arrival of such visions is inexact at best — history is littered with missed deadlines and failed efforts. For physical AI in space construction, it will be a matter of whether the funding and momentum continue.
“It is correct to think of the problem as both very simple and very complex,” said Arthur. “We need more actual physical experience with how machines moving across the lunar surface hold up. There’s hesitation because nobody wants to be the failure who has that first operation that blows up.”
Despite the risks, the incentive is powerful. Cosmic continues its brutal desert testing, while Spacer has launched a pilot with terrestrial builder Teamwrkx Construction.
Morgan Stanley projects the space economy to reach $1 trillion by 2040, and that’s just the beginning. The companies that successfully bridge the gap between terrestrial training and extraterrestrial autonomy won't just build the next Moon base — they'll define the infrastructure of the future.
