A narrower question than the general autonomous-weapons debate
The manual’s autonomous weapons and military escalation page covers the broad policy fight over AI-directed targeting, drones, and command-and-control. This page asks a narrower question raised specifically by the military humanoid and ground robots already fielded: what governs a physical, weight-bearing robot body that can carry a weapon, and how well is that governance actually holding up. The short answer, drawn from one pledge, one broken deal, and one stalled treaty process, is: unevenly.
This page describes public statements, corporate policy, and the state of international negotiations. It is not legal advice, and it does not resolve open disputes among states or companies — several of which, described below, remain genuinely unresolved as of September 2026.
The 2022 pledge, and what it does not cover
In October 2022, six robotics makers — Boston Dynamics, Agility Robotics, ANYbotics, Clearpath Robotics, Open Robotics, and Unitree Robotics — published an open letter pledging not to weaponize their “advanced-mobility general-purpose robots” and not to help others do so, a commitment NPR covered as a notable industry first. Read closely, the pledge is narrower than the headlines suggested. It covers only general-purpose robots, leaving companies free to design a robot for military use from the start and restricting only after-the-fact conversion. Open Robotics’ own text admits it cannot legally stop this, because the permissive open-source licensing behind ROS and Gazebo lets anyone build on the software regardless of the pledge. And one major fielded-quadruped maker never signed at all: Ghost Robotics published a competing code of conduct instead, with then-CEO Jiren Parikh saying the company was happy to let government customers mount whatever weapon their job required — a direct rejection of the pledge’s premise, not merely an omission.
When the pledge met a real sale
A 2026 case shows exactly where the pledge’s authority actually comes from. Boston Dynamics’ own ethics committee approved selling Spot robots fitted with flash-bang grenades to a law-enforcement equipment supplier. Employees objected within days, invoking the company’s anti-weaponization pledge and a well-known prior case in which a flash-bang device thrown during a no-knock raid critically injured a toddler; Semafor reported that Boston Dynamics reversed the deal shortly after. The pledge’s text did not block the sale — an internal committee approved it. Workforce pressure did. That is a meaningful enforcement mechanism, but it is not one a reader should mistake for a rule with teeth: it depends on employees noticing, caring, and having enough leverage to act, none of which is guaranteed at every company or in every country.
Unitree’s open contradiction
The pledge’s most uncomfortable fact concerns a signatory, not a holdout. Unitree signed the 2022 letter, and its robots have since appeared armed at Chinese military exercises; in June 2026 the Pentagon added Unitree to its Section 1260H list of Chinese military companies. No public reconciliation statement from Unitree addressing the tension between its 2022 pledge and its Section 1260H designation has surfaced in reporting reviewed for this manual. That is presented here as an open, unresolved contradiction — not a case this manual can adjudicate on the company’s behalf.
The line that actually matters: who pulls the trigger
Across every fielded example described on this manual’s companion page — the Marine Corps’ SENTRY-equipped Vision 60, the Chinese missile-carrying quadruped shown at World Defense Show 2026, Israel’s Jaguar UGV — the same design choice recurs: a human, not the machine, authorizes the final shot, even where AI assists detection. That distinction is the one international law actually turns on. A weapon where a human selects and engages the target — even remotely, even with AI-assisted detection — is a semi-autonomous weapon system, a category that already includes Predator drones, Phalanx close-in weapon systems, and Iron Dome interceptors. A lethal autonomous weapon system (LAWS) is one where the machine itself selects and engages without a human in that final loop. Legal analysts describe removing the human gate as technologically straightforward once a detection pipeline already exists — meaning every human-authorized ground robot fielded today sits deliberately, not technically, on the permitted side of that line. That is precisely the “gray zone” the international policy fight below concentrates on.
Where the treaty fight stands, this month
The relevant venue is the UN Convention on Certain Conventional Weapons (CCW), which operates by consensus — meaning a single opposing state can block any outcome. Its Group of Governmental Experts on lethal autonomous weapons concluded a three-year mandate in Geneva on September 5, 2026, without full agreement: the United States and Russia forced the removal of language on design, development, ethics, predictability, reliability, explainability, and traceability. Despite that, Human Rights Watch reported that 76 states — a majority of the Convention’s parties — now support moving to binding treaty negotiations, calling it the strongest political momentum yet for a CCW treaty process. At a September 4, 2026 session with 128 states present, the U.S. and Russia specifically disputed draft language on “lethality” and on preserving human judgment and control; one anonymous diplomat told Defense News the window is getting smaller to reach an agreement. A November 2025 UN General Assembly resolution had already passed 156-5-8, urging CCW parties to finish the “elements” text. The decisive moment is now the Seventh CCW Review Conference, scheduled for November 2026, where states will decide whether to mandate binding-treaty negotiations, extend the discussion-only mandate again, or let the three-year process lapse.
A measured reading
None of this supports either extreme claim sometimes made about armed ground robots: that treaties have already tamed the technology, or that no constraints exist at all. A real industry pledge exists, and it has real gaps — general-purpose scope only, unenforceable licensing, a non-signatory building armed quadrupeds anyway, and a signatory now on a Pentagon sanctions list with no explanation offered. A real design norm exists — human authorization before lethal engagement — but it is a policy choice every fielded program has made, not a wall the technology cannot cross. And a real treaty process exists, closer to binding negotiations than at any point since the CCW began discussing autonomous weapons, but still one veto away from failing again in November. Readers tracking this space should watch the Seventh Review Conference outcome, not the marketing footage, for what actually changes next.