The FCC’s public notice reads like a technical specification, not a salvo. Published on 28 July 2026, DA 26-786 adds “advanced robotic devices” produced abroad to the Covered List of the Secure Networks framework, and the definition it provides contains not a word of anthropomorphism. No bipedalism, no hands, no face, no dynamic balance. Four cumulative conditions are sufficient to bring a device within scope — as reported by There’s a Robot For That, the independent publication providing the most detailed reading. The text of the notice itself, a public document, has not been read here: everything that follows regarding the definition relies solely on this relay.
- ground locomotion with obstacle avoidance and navigation;
- remote operation by a human operator, via command or sensor data;
- a mass greater than 2 kg, dock included;
- the simultaneous presence of an environment sensor, a wired or wireless connectivity of at least 200 kbit/s, and navigation or decision-making autonomy software.
Put together, these conditions do not outline a humanoid: they outline a connected mobile device weighing more than a toy. Washington does not regulate legs. It regulates a data path.
What the text catches, and what it spares in writing
The scope is also read in the negative. The International Federation of Robotics highlights two points: the attachment criterion is the place of hardware production and not the manufacturer’s nationality, and industrial robots — articulated, Cartesian, SCARA, parallel — are expressly excluded. Industrial robotics exits the scope through the back door, while the remotely piloted mobile device enters through the window.
The measure is also narrower than announced: it blocks the certification of new models only. Devices already certified and already purchased remain usable, as noted by The Robot Report and heise; no installed fleet is disconnected. This detail changes the reading of the timeline: again according to There’s a Robot For That, and without overlap at this stage, Unitree had obtained FCC certification for its R1 on 22 June, then for the H2 and the A2 quadruped on 30 June, a few weeks before the listing. A parallel action targets foreign-made connected uninterruptible power supplies, those already installed in US homes not being concerned, reports TechCrunch. The motive repeats: the target object is a box that talks to the network from a home.
A factory rule, then an intelligence questionnaire
If attachment is by factory, compliance is obtained by moving a factory. The US content threshold applicable is 65% of the cost of components for deliveries from 2024 to 2028, under the Buy American standard of 48 CFR 25.101(a), and it rises to 75% from 2029 onwards — figures reported by There’s a Robot For That, without overlap either, and which derive from a third-party standard, not the letter of the notice. We are no longer talking about robotics; we are talking about BOM and supply chain.
The exit route confirms the nature of the exercise. The Conditional Approval, granted on a case-by-case basis by the Department of War for robots and by the Department of Homeland Security for uninterruptible power supplies alone, requires disclosure of shareholding, investors, supply chain, software and microsoftware architecture, and cybersecurity controls. This is not a customs formality; it is an intelligence questionnaire — and that it applies to microsoftware is enough to say that the object of concern is the link, not the limb. Forbes recalls the invoked basis, the Secure and Trusted Communications Networks Act, and reports that the agency’s chairman, Brendan Carr, says he is acting in concert with national security agencies to secure critical supply chains. Failing to trace back to the FCC press release, which would settle the matter, this statement is rendered in indirect speech, based on this single media source.
The machines providing a service this week are at the end of a wire
The most troubling aspect is not the text: it is that the state of the art validates it word for word. On 29 July, the startup Tau Robotics makes public in San Francisco a home cleaning service provided by humanoid robots, billed at $30 an hour and reserved for selected candidates. All we know comes from a single report, that of ABC7, not corroborated to date. According to this report, the machines are teleoperated from a central site, with AI assistance: they have no decision-making autonomy, and the AI cannot pilot them alone. Still according to ABC7, co-founder Alex Koch himself limits already functional tasks to vacuuming and wiping down a kitchen worktop, while roboticist Ken Goldberg judges it is too early and says he would be very surprised if these machines performed truly useful cleaning work. A distant body, an operator, a link: the FCC definition in commercial version.
The data factory obeys the same constraint. At Neura Robotics, in Metzingen, each Neura Gym cell associates a humanoid with a teleoperation equipment, the “Motion Suit”, worn by an employee; tracking cameras record the gesture while the human executes the task, and the robot learns the sequence through imitation learning, describes the report from automationspraxis. The human is not in the loop out of caution: he is the loop.
As for the VLA (Vision-Language-Action) model published this week, it documents its own limit. On the Apollo 2 humanoid from Apptronik equipped with SharpaWave grippers — five-fingered hands with 22 degrees of freedom (DOF) each, specifies marktechpost — Google DeepMind measures, with Gemini Robotics 2, a 92% success rate for unscrewing a light bulb and 36% for screwing it in. The same publication gives, on the dual-arm platform with Franka Duo grippers, 89.6% in precision insertion and 78.9% in preparation of varied tool kits (kitting): above all tasks measured on multi-fingered hands. Dexterous manipulation remains the bottleneck — on everything measured, the two arms perform better than the 22 DOF hand. The only real data saving, the embedded model Gemini Robotics On-Device 2 adapts to a new dual-arm platform with fewer than 200 demonstration examples collected in a few hours.
Scale remains. Approximately 15,000 humanoid robots have been shipped worldwide over the past year, according to the Taipei Times, which credits Unitree and Agibot with more than 5,000 units each, while Tesla and Figure AI ship at best a few hundred. A competing count picked up by Forbes attributes 87% of shipments to Chinese companies and details 5,168 units for Agibot, 4,200 for Unitree and 1,000 for UBTech — figures that do not overlap with each other or with the previous ones. On totals, the sector we are regulating does not yet know how to count itself. On one point, however, the counters agree: the Chinese share hovers around 85% — the estimate from Barclays, that of a Japanese study for 2025, the count from Leaderobot placing it above 80% on some 13,000 units delivered, and the 87% of the count picked up by Forbes. This is the only figure in this dossier on which four independent estimates converge.
Beijing’s counter-shot targets the pipe itself
The day after the notice, seven ministries led by the MIIT — the Ministry of Industry and Information Technology — launch a national “揭榜挂帅” programme dedicated to humanoid robots and embodied AI, relayed on 30 July by the People’s Daily on page 11. All we know comes from a single report, that of the aggregator Xintouyan, not corroborated at this stage: neither the MIIT notice nor the People’s Daily edition has been read directly. According to this report, the 2028 target is a robotic foundation model multimodal coordinating at least four modalities — vision, language, force and touch feedback, robot body state — a perception and recognition accuracy rate in multi-scenarios greater than or equal to 96%, and embedded edge world models subject to national intellectual property. Nothing documents the development schedule of this programme; the reading we can offer is less that of a counterattack than of a convergence of diagnosis — one regulates the link, the other plans its removal.
The counterattack, however, is verbal. The Chinese Ministry of Commerce demands on 30 July the immediate removal of the listing, denounces discriminatory treatment and warns it will take “resolute countermeasures” if Washington persists, reports the Global Times; the day before, Mao Ning, spokesperson for the Ministry of Foreign Affairs, responded to a Bloomberg journalist that Beijing firmly opposes the extension of the national security concept to suppress Chinese companies and will take “necessary measures”, according to the Beijing Daily report. In the background, a meeting between Donald Trump and Xi Jinping is expected in September, and the Hong Kong-listed UBTech title fell by more than 6%, notes TNW.
A final safeguard, because the week ended on an image: on 30 July, Lei Jun relays a factory video, reported by Sina, where a Xiaomi robot folds and stacks boxes while a second films the scene and gestures a victory. The manufacturer presents the sequence as falling under decision-making autonomy enabled by an embodied AI model; no independent verification of the level of autonomy is provided, and the footage itself could not be viewed at its source. A filmed demonstration does not say who holds the controls.
Three things will be verified quickly: who files a Conditional Approval dossier, and how far the required software architecture disclosure really goes; the IPO of Unitree, whose schedule published by the Shanghai Stock Exchange sets the preliminary price consultation for 5 August and the opening of subscriptions for 10 August, the first market verdict rendered after the listing; and the first physical prototype of a BYD humanoid, which the group confirmed to the China Securities Journal will present in August in its “Di Space” spaces. Each time, the same question, now also a question of US law: does the machine decide, or does it receive?