Standard Issue: Who is writing the humanoid rules?

Google DeepMind, Meta and OpenAI are all moving into physical AI, including humanoids. But they aren't all competing to build the best robot. They're competing to build the intelligence that could power different robots. For DeepMind and Meta, the ambition is to create robotics foundation models that can understand instructions, recognise objects, plan tasks and adapt to unfamiliar situations, regardless of what machine it is running on. The challenge is that robots need more than internet knowledge. They need to learn from physical experience: handling objects, planning paths, moving through spaces and interacting with people. That means collecting huge amounts of real-world data and developing models that can turn it into useful actions. The real prize is not necessarily selling the most robots, but supplying the intelligence that makes them work.

All three want to own the brain; but differ in how much of the body they want to own. Google DeepMind is the most partner-focused, developing its AI with robot makers such as Boston Dynamics and Apptronik, while relying on those companies to evolve the physical platforms. Meta also wants to supply software to other manufacturers, but is getting closer to the hardware: it has acquired robotics-AI startup Assured Robot Intelligence and developed touch-sensing technology. According to The Information, Meta has built a humanoid prototype, though Meta told the outlet the work is for research, not sale. OpenAI is taking a more integrated approach. It is developing its own humanoid while hiring engineers for motors and actuators, while seeking external manufacturing partners to supply critical components. Sam Altman has said he expects a “cool demo” in 2027.

There's a gap between what the companies say and what they are doing. Meta CTO Andrew Bosworth says software, not hardware, is the bottleneck. Yet The Information reports that Meta built a humanoid prototype with China's Fourier, is developing a second version and is hiring supplier-development and product staff in China that hints at broader robotics manufacturing capability. Meta says the Fourier work is for internal research, not a product it plans to sell. OpenAI's hiring tells a similar story: when I checked its robotics careers page, five of the 15 listed job titles explicitly mentioned motors or actuators. I also looked for a design, a specification or a named manufacturing partner for OpenAI's robot and found none.

If intelligence is the prize, robot makers risk becoming suppliers of the less differentiated part: the body. If hardware is the hard part, a lab that borrows a body finds out on someone else's robot. Humanoids are already shipping, but most are still being used for research, entertainment or narrowly defined industrial tasks. The real test of general-purpose capability has yet to arrive. Even Elon Musk has described the hands as one of Optimus's hardest engineering challenges. His comments are a reminder that the physical engineering remains difficult. Either way, the balance of power could shift away from the companies making the machines. Humanoids are shipping, but general-purpose intelligence remains unproven.

China published its first six national standards specifically for humanoid robots on 28 September 2026. The National Standards Information Public Service Platform classifies them as recommended. The standards come into effect on 1 April 2027. Together, they cover general rules, how a robot senses its surroundings, how it plans what to do, how it moves, how it handles tasks, and simulation testing, where a robot can be tried out in a virtual setting.

The six standards’ records are on the National Standards Information Public Service Platform. The database lists Unitree, UBTECH and Beijing Xiaomi Robotics as main drafting organisations on all six. Three research bodies are listed on all six too: the Beijing Research Institute of Automation for Machinery Industry, Jianghuai Laboratory and the Shanghai Artificial Intelligence Research Institute. I compared the names of AGIBOT, Galbot, Fourier and Leju against the six lists and found none of them. But being named as a drafter shows participation, not influence. The records do not show who proposed which requirements.

A separate safety series is still in draft. The database lists five parts as recommended: general safety, industrial use, personal and household services, public services and special applications. As of 2 November, its records show Parts 1 to 3 open for public comment and Parts 4 and 5 in drafting, with no comment deadline given. Part 1 lists Unitree, UBTECH and Beijing Xiaomi Robotics among its drafting organisations, alongside carmakers, a battery maker, telecom firms, research institutes and testing bodies. Galbot is listed on four of the five safety parts. The database lists a draft text and an explanatory note for each of the first three. The records give no effective date.

The records give the safety standard for public service applications a 15-month drafting period from 22 May 2026, which runs to August 2027. The first six take effect in April. On that schedule, the months in between have no finished humanoid standard titled for public service use.

Two separate stories this week. The Signal is about the gap between what three AI labs say and what their hiring shows, and Below the Fold is about how the humanoid standards are taking shape in China and who is drafting them.