A body is not a targeting algorithm
The manual’s autonomous weapons and military escalation page covers AI-directed targeting broadly — drones, loitering munitions, and the software that decides what a weapon shoots at, regardless of what carries it. This page is narrower and more literal: what physical robot bodies — legged, wheeled, bipedal, quadrupedal — militaries are actually building, buying, and in a few documented cases, arming and deploying. Embodiment changes the argument, because a machine that walks a minefield or carries a rifle raises procurement, workforce, and escalation questions a targeting model does not.
This page synthesizes public reporting — trade press, defense-beat journalism, and one investigative review — current as of September 2026. Fielded military robotics is a fast-moving, secrecy-prone beat; a program described here as active, paused, or cancelled can change within months. Nothing here is an order of battle, and nothing here should be read as an endorsement of any platform or policy.
The Army’s flagship program just collapsed
The clearest fact in this space is a cancellation, not a breakthrough. In May 2025, the U.S. Army halted development of its Robotic Combat Vehicle (RCV), the program meant to field autonomy-assisted armored ground vehicles alongside crewed formations, as part of an internal budget realignment. Maj. Gen. Glenn Dean, the Army’s Program Executive Officer for Ground Combat Systems, said flatly that RCV would stop development and that the fate of its robotic software component was unknown. Army Secretary Daniel Driscoll went further, framing the underlying vehicle concept itself as a design failure: the service wanted a small, agile machine and ended up with something closer to a heavy tank. RCV was never a humanoid or legged program — it was wheeled and tracked ground-vehicle autonomy — but its collapse is the single most concrete data point about how far U.S. Army ambitions for autonomous ground combat vehicles have actually gotten by 2025.
DARPA’s real 2025–2026 bet is wheels, not legs
A common misreading treats DARPA’s 2012–2015 Robotics Challenge — the event that produced viral footage of humanoid robots falling over — as an ongoing pipeline toward fielded military humanoids. It is not. DARPA’s own program page lists the Robotics Challenge as archived, with no legged or humanoid successor. The agency’s active ground-robotics investment is RACER (Robotic Autonomy in Complex Environments with Resiliency), which is autonomy software for conventional wheeled platforms, not walking robots. RACER completed its final testing in January 2026 and is transitioning its stack to military and commercial users; an October 2025 demonstration at Fort Hood paired the RACER Heavy Platform with an M58 MICLIC mine-clearing line charge for autonomous minefield breaching, with the Army’s 36th Engineer Brigade. The distinction matters: the government’s most active 2025–2026 ground-robotics dollars are going to autonomy for vehicles that already exist, not to the bipedal robot bodies the DRC popularized.
A startup, a warzone first, and a conflict of interest
The most consequential humanoid story of the period involves no government program at all. Foundation Future Industries, co-founded in 2024 by Sankaet Pathak — previously CEO of the fintech Synapse, which filed for bankruptcy in 2024 with customer funds still disputed — holds a combined $24 million in Army, Navy, and Air Force research contracts. Its bipedal Phantom MK-1, built for logistics and inspection in hazardous settings such as minefields, CBRN environments, and urban combat rather than direct engagement, became the subject of what TIME reported as an apparent first: two units deployed to Ukraine in February 2026 for frontline reconnaissance, resupply, and weapons-inspection support — not combat. TIME separately disclosed that Eric Trump, son of the sitting U.S. president, is an investor in Foundation and its newly appointed chief strategy adviser, a fact worth stating plainly as a real conflict-of-interest consideration around a Pentagon contractor, without further editorializing. Separately, a widely repeated claim that “50,000 humanoid robots could serve as the U.S. military’s frontline by 2027” reads as industry aspiration rather than a funded program of record, and this manual declines to present it as established.
Quadrupeds, not bipeds, are the mature category
If any legged robot is militarily mature today, it is the four-legged kind. Ghost Robotics’ Vision 60 already performs base security at Tyndall Air Force Base, and an armed “SPUR” variant — a 6.5mm weapon fired by a human operator through a 30x electro-optical scope — drew public backlash when unveiled at AUSA in 2021. Since then, testing has accelerated: the Army’s 10th Mountain Division ran “Operation Hard Kill” at Fort Drum against aerial drones, and U.S. Army Central tested an armed Vision 60 at Red Sands, Saudi Arabia, in September 2024 that, per an Army Central spokesperson, engaged several static ground targets. Marine units separately tested Vision 60 paired with Onyx Industries’ SENTRY system — AI-assisted detection with mandatory human authorization before firing — and, in a cruder field improvisation, test-fired an M72 rocket launcher mounted on a commercial Chinese-made Unitree Go1. Ghost Robotics itself was acquired in 2024 by South Korean defense firm LIG Nex1, which took a 60 percent stake for roughly $240 million.
China: procurement surge, no confirmed deployment
A Reuters review of more than 100 Chinese military procurement notices, patents, and official records found accelerating PLA interest in humanoid robots — including a 2025 tender for a humanoid robot with installation and training, and a 2026 tender for an “embodied humanoid robot intelligent perception and dexterous operation system” — but found no evidence that China has fielded an armed humanoid with an operational unit. That gap between procurement interest and confirmed capability matters. What has been publicly demonstrated is quadrupedal, not humanoid: a robot dog carrying a QBZ-95 rifle appeared at the 2024 “Golden Dragon” exercise with Cambodia, and at the World Defense Show 2026 in Riyadh, a Chinese-linked manufacturer displayed a quadruped carrying four anti-tank guided missiles with a 2–4 kilometer range — while retaining a human operator for weapons release.
Russia: hardware improvised, claims outrunning it
Russia’s record leans further toward aspiration. The humanoid FEDOR flew to the International Space Station in 2019 and was retired the same year, its developers saying there was nothing more for it to do; it had shown firing pistols in demonstrations but no post-2019 military development, and reportedly relied on roughly 80 percent imported components — a real vulnerability under sanctions. The tracked Marker UGV, which can mount a Kornet anti-tank missile, numbered only five units as of early 2023, a handful reportedly sent toward Donbas; defense analyst Samuel Bendett called the claimed combat use more public relations than a realistic tactic. At the Army-2022 expo, an “M-81” armed robot dog shown was visually identical to a roughly £350 commercial Unitree quadruped fitted with an RPG-26 — evidence that Russia’s quadruped base has been improvised from cheap Chinese commercial hardware rather than built indigenously.
Israel’s semi-autonomous Jaguar, and Ukraine’s actual scale
Israel Aerospace Industries’ six-wheeled “Jaguar” UGV, deployed along the Gaza border, autonomously plans routes and avoids obstacles while a human operator engages targets with its remote 7.62mm machine gun — the human-gated pattern that recurs across every fielded system in this article. Ukraine remains the actor with genuinely operational scale: its Ministry of Defense reports exceeding its 2025 ground-robot supply targets and plans to field 25,000 ground robotic systems by mid-2026, overwhelmingly for frontline logistics rather than combat.
Policy is catching up, unevenly
Congress has moved to restrict the category before it grows further: an NDAA provision would bar the Pentagon from procuring, leasing, or operating “covered humanoid robotic systems” tied to China, Russia, or Iran. The irony is pointed: in June 2026, the Pentagon added 65 entities to its Section 1260H list of Chinese military companies, including Unitree — a signatory to the 2022 industry pledge against weaponizing general-purpose robots, discussed alongside the rest of that pledge’s limits on the manual’s companion governance page.
What the evidence actually supports
Strip away the viral footage and the pattern is consistent: bipedal humanoids remain experimental and largely unarmed, with Foundation’s Ukraine deployment as the narrow exception that proves the rule; quadrupeds are the mature, occasionally armed category, fielded by at least four militaries; and in every documented case where a ground robot carries a weapon, a human — not the machine — decides when it fires. That last fact is a policy choice, not a technical ceiling, and it is the subject this manual turns to next.