These autonomous systems don't know how to handle open-ended uncertainty yet
In a San Francisco warehouse, a faceless robot stacks blocks and quietly rehearses a future its makers are already selling to militaries. Foundation Robotics is among the first companies to build humanoid machines explicitly for warfare — from logistics to lethal engagement — joining a global race that raises questions not merely about what these machines can do, but about what humanity will permit them to become. The technology remains far from battlefield-ready, yet the contracts are signed, the field tests are underway, and the ethical frameworks to govern what comes next have yet to be written.
- A startup with $24 million in US military contracts is already testing humanoid combat robots with Ukrainian forces, compressing the timeline between prototype and battlefield in ways that outpace public debate.
- Current models cannot survive rain, cannot stand up after falling, and struggle with tasks that seasoned warehouse robots handle poorly — the gap between the vision and the machine is enormous.
- Experts warn that the humanoid form may be the wrong solution entirely, with drones and quadrupeds already proving more practical for most military roles the technology is meant to fill.
- The deeper alarm is ethical: autonomous systems that look human may erode the psychological and political resistance to deploying lethal force, while diffusing accountability when something goes wrong.
- Calls for international regulation are growing louder, but the arms race logic — China is advancing, the West cannot fall behind — is already shaping decisions faster than diplomacy can respond.
In a San Francisco warehouse, a black, faceless humanoid robot sits on the floor stacking colored blocks. This is Phantom, built by Foundation Robotics, and its founder Sankaet Pathak describes it not merely as a product but as the leading edge of a transformation in how wars might one day be fought. His company, he claims, is the only American firm building humanoid robots explicitly for the full range of military applications — from supply runs and casualty recovery to what he calls "frontline weaponisation."
The appeal of the humanoid form is practical: a world built for human hands is a world these machines could navigate without redesign. A robot that can open a door, pick up a weapon, or operate a vehicle offers a kind of universal utility that specialized drones cannot match. Foundation already holds $24 million in US military research contracts and is testing units with Ukrainian forces, with ambitions to produce 40,000 robots annually by late 2027 at under $20,000 each. Pathak frames the effort as deterrence — a technological force so formidable that conflict becomes unthinkable — and as necessity, pointing to China's parallel development.
But the machine in front of him tells a more humbling story. The current prototype has no battery, cannot handle rain, and cannot recover from a fall. The next generation promises waterproofing, six hours of runtime, and hands capable of gripping a trigger — but experts remain skeptical. Robotics advisors note that commercial humanoids still struggle with warehouse packing; the leap to navigating an unfamiliar building under fire, landing on rubble, operating in darkness, is not an incremental improvement but a different problem entirely. The AI system powering Phantom, called Cortex, performs well in controlled conditions it has been trained for. Real warfare offers no such conditions.
The ethical stakes sharpen the technical ones. Critics like Nicole van Rooijen of Stop Killer Robots argue that lethal autonomous systems lower the threshold for going to war, strip human judgment from life-and-death moments, and obscure who bears responsibility when something goes wrong. The humanoid form adds a further complication: as these machines become familiar in civilian life, people may come to trust them in ways that mask the danger they represent on a battlefield. Pathak insists humans will remain in the decision loop for lethal force — with allowances for exceptions. It is a principle that sounds reassuring until examined closely. The deeper question is not whether humanoid robots will eventually reach the battlefield, but whether humanity will have decided, before they arrive, what it wants them to do.
In a San Francisco warehouse, a black humanoid robot with no face sits on the floor, methodically stacking colored blocks. This is Phantom, and it is not preparing for war—at least not yet. It is learning how the world works, one interaction at a time, while its makers at Foundation Robotics dream of a future where machines like it could reshape the battlefield entirely.
Sankaet Pathak, who founded Foundation Robotics two years ago, walks me through the 80-kilogram steel-framed robot with the measured confidence of someone pitching not just a product but a vision. His company, he says, is the only American firm building humanoid robots explicitly for military use across the full spectrum of defense applications. That spectrum is wide: supply runs, reconnaissance, casualty recovery, hazard inspection. And then there is the part that makes people pause—what Pathak calls "frontline weaponisation." Robots that could enter buildings, search rooms, engage threats. Robots that could, in theory, do the work soldiers do now, but without the risk to human life.
The appeal is straightforward. A machine that can use human tools—a screwdriver, a weapon, a doorknob—because the world was built for human hands. A machine that could be more precise than an airstrike, more careful about collateral damage. A machine that could absorb the cost of failure in ways soldiers cannot. But the Phantom MK-1 sitting in front of me is nowhere near that future. It has no battery. It cannot survive rain. If it falls, it stays fallen. The second generation, being built in a locked section of the facility, will fix some of this: waterproofing, six hours of runtime, the ability to get back up. Better hands are coming too—hands that can grip a trigger.
Pathak's ambitions are enormous. Foundation has $24 million in research contracts with the US military and is already testing units with Ukrainian forces. The company aims to manufacture 40,000 units annually by the end of 2027, at a cost eventually below $20,000 each. Hundreds of thousands of AI-driven robots, he imagines, forming a ground force to match the drone swarms already filling the sky. He frames this not as aggression but as deterrence—a technological wall so formidable that conflict becomes unthinkable. He also frames it as necessity. China is pursuing this technology, he argues. The West cannot afford to lag.
But the technical reality is far messier than the vision. The robot's artificial intelligence, called Cortex, works by combining two types of AI models: one trained on specific tasks to interpret goals and plan actions, another trained on internet videos and the robot's own interactions to predict how the environment will respond. It sounds elegant. In practice, it means Phantom can handle tasks it has been specifically trained for in controlled conditions. Real warfare is neither controlled nor predictable. Dean Fankhauser, who advises on humanoid robotics, points out that today's commercial humanoid robots struggle with warehouse packing. The idea that they could effectively fight in a Taiwan conflict tomorrow is, he says, "fanciful." Even in five or ten years, the gap between a robot stacking blocks and a robot navigating an unfamiliar building under fire remains vast.
Robert Griffin, who works on humanoid robots at the Florida Institute for Human and Machine Cognition, identifies the core problem: unpredictable environments. A robot can be trained to jump through a window of known height onto level ground. But a window of unknown height, landing on rubble, then navigating an unfamiliar interior in darkness while under fire? That is not a training scenario. That is the real world, and autonomous systems do not yet know how to handle it. There is also the matter of power. Every humanoid robotics company faces the same wall: locomotion and moving joints consume enormous amounts of energy. Six hours of runtime would be, Griffin says, "very impressive." Whether Foundation can build hands that can actually manipulate a weapon designed for human grip remains an open question.
The ethical questions are sharper still. Nicole van Rooijen, executive director of Stop Killer Robots, a coalition of NGOs focused on autonomous weapons, sees lethal autonomous systems as inherently dangerous—they lower the barrier to warfare, remove human judgment from life-and-death decisions, and blur accountability when things go wrong. But she finds the humanoid form particularly troubling. As these machines become more human-like and more common in civilian life, people may come to trust them, may misread the danger they represent. The answer, she argues, is international rules to de-escalate the arms race before it truly begins.
Pathak counters that humans should remain "in the loop," approving lethal force before the system acts—though he allows for exceptions where autonomous firing might prevent catastrophic outcomes. It is a reasonable-sounding principle that dissolves quickly under pressure. The real question is not whether humanoid robots will eventually reach the battlefield. It is whether we will have decided, as a species, what we want them to do when they get there.
Citações Notáveis
Humans should be in the loop, approving any use of lethal force before the system can act, though exceptions exist where autonomous firing might prevent catastrophic outcomes— Sankaet Pathak, Foundation Robotics CEO
Lethal autonomous weapons lower the barrier to warfare, dehumanise conflict, and blur accountability. The humanoid form is extra worrying because human-like machines may appear familiar and trustworthy— Nicole van Rooijen, Stop Killer Robots
A Conversa do Hearth Outra perspectiva sobre a história
Why humanoid robots specifically? Why not just use drones or quadrupeds, which seem more efficient?
Because the world is built for humans. A screwdriver, a doorway, a weapon—all designed for human hands and bodies. A humanoid robot doesn't need to reinvent the tool. It just needs to learn to use what already exists.
But the robots you saw couldn't even get up if they fell. How far away is actual deployment?
Years, maybe a decade. The technical hurdles are immense. Real warfare is chaos—unpredictable terrain, darkness, things breaking. These robots are trained in controlled conditions. They don't know how to handle genuine uncertainty yet.
So why is the military interested now?
Because the potential is real, even if it's distant. A machine that could enter a building instead of a soldier, that could absorb casualties without losing lives—that changes the calculus of risk. And there's the competitive pressure. If China is building them, the West feels it has to keep pace.
What worries you most about this?
The ethical blur. Once these machines become common in civilian life—delivery, construction, care work—people will trust them. They'll seem safe, familiar. Then the military versions arrive, and the line between tool and weapon becomes invisible. That's when accountability disappears.
Is there a way to prevent that?
International rules, some argue. Agreements that slow the arms race before it accelerates beyond control. But that requires countries to agree on something they all see as existential advantage. That's a hard sell.