Five humanoids, one competition: for the US Army the hard part is the body, not the brain

On 28 September 2026, the US Army published the results of xTechHumanoid, a competition launched to explore what humanoid robots could do in a military context. What follows is information and technology only: no operational content, no tactics, nothing exploitable.

The result is interesting for a reason that is not the obvious one: the winners are small prototypes, and the main difficulty they had to demonstrate was protecting the body, not autonomy.

The shape of the competition

Three dates are easy to conflate here, and the distinction matters. The demonstrations took place on 9 and 10 September, the selection of the winners was announced on 11 September, and the summary press release that this article draws on is dated 28 September. By the time the summary is published, the event itself is nearly three weeks old. What is new in the summary is the Army's own account of the results, not the demonstrations themselves.

The programme is also the continuation of a funding line rather than a one-off. The press release states more than $3 million committed since 2018 across more than five competitions, which places xTechHumanoid in a series. A competition that survives its first edition is a different kind of signal from a demonstration video.

The five winners

Sulphur Springs Entity Robotics Technology Inc., with a small humanoid built for load handling and close-range interaction.

The Institute for Human and Machine Cognition (IHMC) at the University of West Florida, with the robot Alex: the most widely covered demonstration of the five. It carries ballistic protection on the torso and is built around 29 joints, driven by quasi-direct drive actuators for fast response. Quasi-direct drive is a transmission choice: the motor acts almost directly on the joint instead of routing torque through a long chain of gears, which reduces the compliance in the chain. Our explainer on how humanoid robots work sets out what a joint count does and does not tell you about a machine.

The University of Texas Rio Grande Valley, with a lightweight humanoid.

INIT Machina, with a humanoid built for manipulation.

Dyna Robotics, with a humanoid built for locomotion.

The real subject: the body armour

This is the point the press release emphasises most, and it deserves a closer look.

Alex carries ballistic body armour at torso level. One of the other finalists presents heavy armour. Another highlights shock absorption built into the chassis design.

In other words, the difficulties that designers judge hardest to solve in a military context in 2026 are physical, how to take the impact of a projectile without transmitting the whole force into the structure, how to size a protection that does not destroy mobility, and not cognitive. Nobody in this competition presents a robot that decides, that fires, or that selects a target.

That is a useful technical signal for a general audience. On the 2026 agenda, the problem is not that military AI is ready. It is that the body of the robot is the weak link.

It also sits comfortably with the small size of the machines on show. The winner named for load handling is explicitly a small-format humanoid, and the tasks the competition describes, carrying, handling, opening, are the tasks a smaller machine is built for. Weight is a liability in exactly the environments this programme targets.

What the competition does not show

Precision matters here, because the vocabulary used around announcements of this kind tends to drift towards over-reading.

• No robot was deployed in operations. These were two-day demonstrations.
• No task success rate has been published across the winners.
• No decision autonomy has been shown in public.
• Performance under live fire has neither been measured nor disclosed.

A two-day demonstration is a legitimate thing for a programme to run, and it would be wrong to discount it. It is also not the same artefact as a verified capability, and the press release does not claim it is. Five winners out of 103 submissions is a 4.9% overall rate, and five out of ten finalists is 50%: this was a selective process, but selectivity is a statement about the applicants, not a measurement of the machines.

Why a civilian reader should care

Because this military programme funds exactly the problem that holds response robots back. A firefighter entering a burning building, a bomb disposal technician, a nuclear maintenance worker: all of them need a machine able to manipulate in a hostile environment while staying upright, under constraints of load and survivability.

The same capabilities, grasping, carrying, absorbing impacts, opening a door without hurting itself, are precisely the ones that are missing. Funding those building blocks through a military competition is an indirect route to the civilian problem, and a cheaper one than waiting for the civilian market to fund them on its own.

There is also a detail worth keeping in view. None of the tasks described here involves a weapon, a target or a decision to engage. The programme as described asks a machine to carry, to handle and to open. That is the vocabulary of a logistics and emergency response problem, and it is the reason a civilian reader can take something useful from a military announcement instead of leaving the page.

For the wider frame of robotics in hazardous environments, see our file on humanoid robots, factories and safety and our analysis of the future of military robots.