When could humanoid robots go to war? Sooner than you think, says firm testing them in Ukraine

“I think we have this psychological reaction, which is like the Terminator, but the reality is if you really think about it from a practical sense, it's not really like that,” Foundation CEO Sankaet Pathak told Euronews Next. “It's essentially like why would you send a bunch of humanoids if your goal is to be able to wreak havoc?” Instead, the CEO specified that humanoids could take on more combat roles over time, offering a level of precision that other weapons systems, including aerial bombing, cannot match. According to him, it makes no sense to wreak havoc with a cluster of humanoids, since a bomb can do so more cheaply. Humanoids, he argued, become useful specifically when a military objective requires precision, avoiding damage to infrastructure and harm to civilians while carrying out a complex mission.

Robots on the battlefield

Pathak does not expect humanoids to replace drones, but sees them filling a gap as ground combat becomes increasingly perilous for human soldiers. “Humanoids only make sense when your mission objective is more precision, essentially trying to make sure you don't destroy infrastructure, hurt civilians, do the best of your ability as you can accomplish a very complex mission,” he said. “I do think it has become increasingly dangerous for soldiers to be on land and on ground. Secondly, I do think it is the next phase of precision that I think would be a great thing generally.” There is no specific treaty governing the use of humanoid or autonomous robots on the battlefield, and they fall under existing international humanitarian law (IHL), which requires weapons to respect the distinction between combatants and civilians. Last week, United Nations Secretary-General Antonio Guiterrez said his “main concern is with “lethal autonomous weapon systems.” “Let’s call them what they are: Killer robots,” he wrote in a post on LinkedIn. “Machines selecting and engaging their target and taking a life — without human control and judgment,” he added. Since 2023, the UN has been negotiating a dedicated treaty on Lethal Autonomous Weapons Systems (LAWS) through the Convention on Certain Conventional Weapons, with the UN Secretary-General pushing for a binding ban on weapons that operate without human control by 2026. Asked about this, Pathak said he did not see a reason humanoids would be treated differently from other precision weapons systems already in use, such as armed drones and unmanned ground vehicles.

AI on the battlefield

The humanoid uses artificial intelligence, which is built by the companies’ own world models. The company does not disclose its suppliers. World models work by learning to watch videos or using simulation data and other spatial inputs, and then building their own representations of scenes or objects. They do, of course, require a lot of data to train on, but they have different use cases to chatbots. To put it more simply, instead of predicting the next word, which large language models (LLMs) do, they build representations of physical environments and predict what would happen in the next scenario or world, modelling how things move over time. “We're heavily focused on world models that can predict the future. We think that's going to be at the core of building intuitive AI that's robust,” said Pathak. Asked whether AI seizes control of armed robots, he believes such a threat would not take the form of humanoid robots but of drones or worse. “If an AI's mission objective is to destroy humanity, I guarantee you they're not sending 100,000 humanoids. I think they're just using drones or nuclear arsenals,” he said. For him, the more immediate danger is "AI terrorism," meaning misuse of widely available AI models for cyberattacks, disinformation, or arming consumer drones, which he blamed on open source models. AI companies that say they are open source argue they are making technology more accessible to all, but proponents of closed source companies warn it is a threat to public safety, as anyone can update the models and change the safeguards. A prime example of this was when Meta in 2023 released its open-source large language model (LLM) Llama 2. Just days later, people released their own uncensored Llama 2 versions, asking questions about how to build a nuclear bomb, which the LLM could then answer. But Pathak said a scenario in which AI systems could rewrite their own directives, self-improve, and self-replicate independent of large, detectable compute clusters remains several major technical steps away. “As these models become more efficient, as they can self-replicate with less compute, they can improve themselves and also demonstrate behaviours where they can bypass and update their own directives. That's when you're getting to that territory,” he said. But he said, “we're probably like three, four, maybe five hops away from that.”

The path ahead

Nevertheless, even if not handling weapons, the CEO said humanoids could serve a purpose on the battlefield in roles such as material handling, meaning moving supplies between indoor and outdoor environments, and reconnaissance, including mapping interiors and clearing buildings. These tasks have already been tested in Ukraine. The military testing in Ukraine saw Foundation redesign its hardware for outdoor, high-stress conditions. The company is building its next-generation robot, Phantom 2, which is waterproof and dustproof, with a payload capacity that has risen from roughly 25–30 kilograms in the first version to about 80 kilograms. Its tolerance for falls, measured in G-force, has increased from 12–15 Gs to nearly 100 Gs, and it carries a 3-kilowatt-hour battery. Foundation's investor base includes US President Donald Trump’s son Eric Trump, Stripe and the venture firm Define, among other backers. Foundation currently leases Phantom robots commercially for about $100,000 (€90,000) per robot per year. Military customers purchase the robots at similar pricing.