Robots Just Joined the FCC's Covered List

On July 29, 2026, the Federal Communications Commission added a device class to its Covered List that had never been there before: advanced robotic devices — humanoids and quadrupeds — together with connected power inverters. New models from covered foreign manufacturers can no longer receive FCC equipment authorization, which is the permission slip a device needs before it can be legally marketed or imported in the United States.

This is not a tariff and not a seizure. Robots already authorized stay legal, already-purchased units keep working, and federal buyers and devices with conditional Department of War or Homeland Security approval are carved out. The order applies to the next model, not the current one.

That distinction is what makes it worth reading carefully. The immediate news is a trade action. The durable story is that embodied AI has now been formally classified as infrastructure risk, and the hardware layer underneath it turns out to be far less domestic than the software layer above it.

Executive summary

The FCC added humanoid robots, quadrupeds, and connected power inverters to its Covered List on July 29, 2026, blocking new equipment authorizations for covered foreign models.
The mechanism is prospective. Existing authorizations and already-purchased units are untouched; federal purchases and conditionally approved defense and homeland-security devices are exempt.
China ships an estimated 80 to 90 percent of humanoid robots to date and accounts for more than half of global industrial robot installations, so the addressable share of the market is large.
The dependency that the order does not address is upstream: actuators, harmonic reducers, batteries, sensors, and above all rare-earth magnets, where China holds over 90 percent of refining and production.
Industry reaction split cleanly. Domestic manufacturers called it overdue; roboticists warned it removes the cheap hardware baseline that academic embodied-AI research runs on, before a Western equivalent exists.
Scope is genuinely uncertain. Researchers have flagged that European and Canadian platforms, and US companies that manufacture in China, could be caught depending on implementation.

What the Covered List actually is

The Covered List is a register maintained under the Secure and Trusted Communications Networks Act. It names communications equipment and services the US government has determined pose an unacceptable risk to national security. Its best-known entries are telecom gear from Huawei and ZTE, later joined by video surveillance equipment and, in 2024 and 2025, several categories of drones and connected vehicle technology.

The enforcement mechanism is subtle and worth understanding, because it explains both the reach and the limits of what happened this week.

Nearly every modern robot contains a radio. Wi-Fi for teleoperation and fleet management, Bluetooth for pairing, often cellular for field deployment. Any device that intentionally radiates radio energy requires an FCC equipment authorization before it can be marketed, imported, or sold in the United States. That authorization is not a formality — it is the legal precondition for a product existing in the American market at all.

Adding a device class from covered entities to the list means the FCC simply stops granting those authorizations. There is no border enforcement action, no confiscation, and no retroactive revocation. The product line just cannot renew itself. For a category on an annual hardware refresh cycle, which describes essentially all of humanoid robotics right now, that is a slow-acting ban rather than a soft one.

Why inverters and robots landed in the same order

The pairing looks strange until you read it as a threat model rather than a product taxonomy.

A connected power inverter is the device that converts DC from solar panels or batteries into grid-compatible AC, and modern ones are networked for telemetry and remote firmware updates. A large installed fleet under a single vendor’s remote update authority is, in principle, a coordinated lever on grid stability. That argument has circulated in energy-security circles for several years and has now been formalized.

A humanoid or quadruped robot is a different device with the same shape of risk. It is a mobile, networked computer with cameras, microphones, and depth sensors, operating inside a facility, running vendor firmware that updates over the network. In a warehouse it sees inventory and staffing. In a substation or a data hall it sees layout and access. The White House interagency task force whose determination underpins the order concluded that foreign-built robots could constitute “a cybersecurity risk that threatens the security of critical infrastructure.”

Whether or not one finds the remedy proportionate, the technical premise is not fabricated. This is the same category of argument previously applied to telecom base stations and consumer drones, extended to a device class that moves under its own power.

The dependency the order does not fix

Here is the part that matters more than the headline, and it is a supply-chain argument rather than a policy one.

Blocking foreign finished robots does not create a domestic robot. It creates demand for one. Whether that demand can be met depends on layers of the stack that the Covered List does not touch:

Layer	Where the capability actually sits
Finished humanoids/quadrupeds	China, est. 80–90% of units shipped to date
Industrial robot installations	China, >50% of global installations
Rare-earth magnets (in every servo)	China, >90% of refining and production
Actuators, harmonic reducers	Concentrated in China and Japan
Cells and battery packs	Concentrated in China
Perception software, VLA models	Broadly distributed; US strong

Major industrial robot manufacturers — ABB, FANUC, Yaskawa, KUKA, Teradyne Robotics — have all expanded US operations. That expansion is real, and it is mostly final assembly. The upstream components remain concentrated abroad, and the magnet layer is the hardest of all to relocate, because it depends on refining capacity and environmental permitting rather than on factory floor space.

The honest summary is that the United States is strong at the layer where embodied AI is a software problem and weak at the layer where it is a mechanical one. This order applies pressure to the mechanical layer without, by itself, supplying the capacity to answer it. Whether domestic capacity actually materializes in the resulting window is the open question, and it will be answered in years rather than quarters.

What the industry said

The reaction split predictably but not uninformatively.

Supporters framed it as overdue. Standard Bots CEO Evan Beard called it “one of the strongest technology-security actions in modern U.S. history.” Former Novanta robotics strategy chief Robert Little put the security case plainly: “These robots are mobile, connected computers with cameras.”

The objections were about sequencing rather than principle. Georg Stieler of STM warned the measure could “slow U.S. physical AI innovation by cutting startups and researchers off from future low-cost Chinese platforms before comparable Western alternatives exist.” Henrik Christensen of UC San Diego cautioned that “the ban is potentially much wider than China. Canadian and European robots could be impacted.” Rueben Scriven of Interact Analysis noted the awkward case of US humanoid makers who manufacture in China and may find themselves in scope of their own government’s order.

That last point is the one most likely to generate news over the next quarter. Country of incorporation and country of manufacture have drifted apart in this industry, and the order’s practical boundary depends on which one implementation guidance keys on.

Why this matters for embodied AI research

The part of this story that will show up in papers rather than press releases is the cost of a lab robot.

Robot learning has a data problem that everyone in the field agrees on: policies need real-world interaction data, simulation transfers imperfectly, and collecting real data requires physical robots operating for many hours. The single biggest change in the last three years was not an architecture. It was that a capable quadruped stopped costing $75,000 and started costing $2,000 to $16,000, almost entirely because of Chinese manufacturers. That price change is what made large, multi-robot academic data collection feasible at all.

Cut off new authorizations for those platforms and the software half of the field is untouched — no restriction here applies to models, datasets, code, or simulators — while the hardware half reverts toward its pre-2023 economics. Existing fleets keep running, which means the effect is delayed and lands on the labs that have not yet bought in: new groups, teaching labs, and smaller institutions.

If you believe real-world data is the bottleneck in embodied AI, this is a policy that acts directly on the bottleneck. Whether the domestic hardware ecosystem fills the gap before the research cost compounds is the thing to watch, and there is no evidence either way yet.

What to actually do about it

For teams that deploy robots rather than study them, the useful questions are narrower than the policy debate:

Check authorization status, not origin. Already-authorized models are unaffected. Whether your deployed platform holds a current FCC equipment authorization determines whether you have an operational problem or only a roadmap problem.
Look at the refresh cycle, not the fleet. Exposure concentrates in the next hardware generation. A three-year refresh plan built on a single foreign platform is where the risk actually sits.
Ask vendors where final assembly happens. A US-headquartered supplier manufacturing abroad may face the same authorization question you are trying to route around.
Price the domestic alternative now. Not necessarily to buy it, but because a procurement plan with no evaluated second source is a plan with a single point of failure that is now regulatory as well as commercial.
Separate the software bet from the hardware bet. Perception stacks, VLA policies, and simulation pipelines are unaffected. Architecture that assumes one specific robot chassis is the part that just got more expensive to be wrong about.

The larger pattern

Three consecutive weeks of AI policy news have pointed in the same direction, and it is not the direction most AI regulation debates assume.

The arguments that dominate public discussion concern model capability, alignment, and content. The actions that actually landed concern hardware: memory allocation and price, chip interconnect funding, data center siting and power, and now the physical robots themselves. The regulatory surface that turned out to be enforceable is the one with serial numbers and customs codes attached.

For engineers, the practical reading is that the AI supply chain is becoming a jurisdictional object. Where a model runs, what memory it runs on, and now what body it inhabits are increasingly decisions with legal constraints attached rather than purely technical ones. That is a new column in a lot of architecture decision records.

Sources

Frequently asked questions

What exactly did the FCC do to humanoid robots?

The FCC added a new category — advanced robotic devices, covering humanoids and quadrupeds, alongside connected power inverters — to its Covered List, the register of equipment the agency considers an unacceptable risk to national security. The practical mechanism is equipment authorization rather than a customs ban. Almost anything with a radio in it, which includes every robot that speaks Wi-Fi, Bluetooth, or cellular, needs an FCC equipment authorization before it can be legally marketed or imported into the United States. Once a device class from a covered entity is on the list, the FCC will not grant new authorizations for it. Nothing is confiscated and nothing already authorized is revoked, so units already in labs and warehouses keep working and models that cleared authorization before the order can still be sold. The effect is entirely forward-looking: it applies to the next model, not the current one.

Does this ban only Chinese robots?

The stated target is China, and the practical target is China, but the mechanism is not written as a country-of-origin tariff and researchers have flagged that the wording reaches further than the headline. Several roboticists have publicly warned that Canadian and European platforms could be caught depending on how the equipment classes are interpreted. There is also a second-order effect that has nothing to do with where a company is headquartered: a US humanoid startup that manufactures in China, which many do because that is where the actuator and battery supply chain lives, may find its own product in scope. Treat the practical boundary as unsettled until the FCC issues implementation guidance.

Why are power inverters on the same list as robots?

Because the security argument is about network-connected devices that sit inside critical infrastructure and take instructions from outside it. A connected inverter routes electricity through solar arrays, battery installations, and data centers, and it typically phones home for firmware updates and telemetry. A large fleet of them under common remote control is, in the threat model the White House task force applied, a lever on the grid. Robots and inverters look unrelated on a product shelf and nearly identical in a threat model: both are physically embedded in something that matters, both are remotely updatable, and both are dominated by manufacturers outside US jurisdiction.

How dependent is the robotics industry on China?

Very, and at more than one layer. Public estimates put China at roughly 80 to 90 percent of humanoid robots shipped to date and more than half of global industrial robot installations. The deeper dependency is upstream of the finished robot: China refines and produces the overwhelming majority of the world’s rare-earth magnets, which sit inside essentially every servo motor, and it holds a dominant position in the actuators, harmonic reducers, cells, and sensors that make up a robot’s bill of materials. A policy that blocks finished foreign robots does not by itself relocate any of that.

What does this mean for AI researchers working on embodied AI?

The immediate risk is to the cheap hardware baseline that a great deal of published robot-learning work quietly depends on. Low-cost Chinese quadrupeds and humanoid platforms became the default lab robot because they cost thousands rather than tens of thousands of dollars, which is what made it feasible for a graduate student to collect real-world data at all. If new models of those platforms cannot be authorized, labs keep their existing fleet and lose the upgrade path, while new labs face a much higher entry price. Nothing here restricts models, datasets, code, or simulation, so the software half of embodied AI is untouched.

Is the security concern legitimate or is this pure protectionism?

The security concern is coherent on its own terms, which does not settle whether the remedy is proportionate. A modern humanoid or quadruped is a mobile networked computer carrying cameras, microphones, and a detailed map of wherever it operates, with a vendor-controlled firmware channel into all of it. Applied to a factory floor, a substation, or a hospital, that is a genuine intelligence and access surface. What is contested is scope and timing: the order lands while comparable Western platforms are not yet available at similar cost, so the near-term effect is a capability gap for US buyers rather than a substitution.

What should a company deploying robots do about this now?

Start with an inventory question rather than a policy question: identify which of your deployed and planned platforms hold current FCC equipment authorizations, because already-authorized models are unaffected. Then look at the refresh cycle, since the exposure is concentrated in the next hardware generation rather than the current fleet, and at spare parts and RMA paths for units you rely on. For anything on a multi-year procurement plan, it is worth pricing a domestic or allied alternative now even if it is more expensive, and worth asking vendors directly where final assembly happens.