This Week in Robotic Learning #11: We're not going to protectionism our way through this
How to navigate the pesky reality that robotics is dependent on a foreign adversary
American tech dominance is in large part, an extension of American cultural and economic dominance rather than the result of superior tech in and of itself. Anyone who has used Salesforce can tell you this. Competition in tech has, generally, been almost the exclusive province of American firms. (Blah blah, Adyen, Spotify, Shopify, fine, whatever, but come on. The lazy, sweeping statement is like 95% right here).
Since Japan’s slide into irrelevance decades ago, it’s been US technological hegemony.
Robotics is… not developing in this way. American tech dominance has been built on bits, where the moat is lock-in–network effects and distribution flywheels. Robotics is built on atoms–neodymium, actuators and manufacturing capacity–atoms do not care who already uses your product. And so, it is not disputable that Chinese hardware is ahead of the United States.
Some evidence in support and reasons why:
Per SemiAnalysis, Unitree may ship its 10,000th humanoid robot in the next few weeks; American companies combined probably haven’t shipped 1,000
China has produced over 7,000 patents related to humanoid design in the past five years; American companies are around 1/5th of that.
Chinese firms have undercut international competitors on price in the key places that robots cost: actuators and screws, with leading humanoid companies such as Unitree more or less fully vertically integrated.
The materials bottleneck is on neodymium rare earths, which are necessary for the magnets in the motors. 60% of that is mined and 90% is refined within China. Everyone needs those to make their robots work, and in April ‘25, China imposed export licensing, leading to delays for US and non-Chinese firms.
There are political elements–the latest Five-Year Plan (yep, they’re still doing those) has robotics and “embodied intelligence” in the top strategic tier, alongside quantum, brain-computer interfaces, 6G, and fusion; that is currently supported by a ¥1 trillion ($138B) national VC “guidance” fund for AI/hard tech and other subtler ways that the state throws its heft around to get what it wants.
So yeah, Chinese firms clearly have a strong position in robot hardware.
SemiAnalysis’ post, China’s Unitree Will Dominate Global Robotics has dominated the robotics conversation this week, with responses to the core thesis ranging from “cool” to “uh oh, letting a Chinese company dominate seems bad.” I think it’s indisputable that Unitree is going to be a humanoid and robotics hardware giant for years to come. But I don’t think Unitree domination is necessarily the only path forward.
In its clumsy way, the United States government is already (potentially) pushing back this week, with the DoD adding Unitree, among a wide range of huge Chinese tech firms to a list of companies said to be aiding the Chinese military, laying the groundwork to eventually ban them from doing any contract work with the US government. Separately, the GUARD Act, which would effectively ban Chinese humanoids from being imported to the US. for even commercial purposes, was introduced to the House and referred to the House Committee on Energy and Commerce. (An important tidbit for anyone inexperienced with legislation names is to see what the backronym is. In this case it’s Guarding the U.S. Against Adversarial Robotics Dominance. I give that a D-. You can’t have an acronym word in the acronym!)
While the bill itself is unlikely to become a law, there is a growing bipartisan consensus on some restrictions on Chinese robotics imports, and I would expect them to be folded into either the broader budget or the Defense budget bill. The key piece of evidence here is that Tom Cotton (R-AR, Head of Intelligence Committee) and Chuck Schumer (D-NY, Senate Minority Leader) introduced a bill in March that is similar, but narrower–banning executive government agencies from importing Chinese robotics. That’s the stronger signal on what’s likely to come, given the powerful position of its bipartisan authors, but the GUARD Act’s harsher provisions warrant notice.
Balancing security, freedom and science
There’s only one Chinese tech company that ever truly won the ongoing battle for Americans’ attention: ByteDance, whose TikTok dropped its slop all over the US, whose citizens went right to the trough. Political debate erupted immediately, centering data privacy and national security concerns. While my general preference would be to avoid government intervention, I have to concede that I was fine with the government stepping in here.
There is a legitimate national security issue here. Chinese corporations can be compelled to hand the CCP whatever it asks for for whatever reason, and Americans’ behavioral data belonging to such an organization isn’t great. The Equifax hack showed that this isn’t just a boogeyman.
The TikTok saga ending in a forced divestment is a useful prelude to robotics. Taking a heavy hand with TikTok is easy because it costs very little and even an outright ban would just mean saying goodbye to a very replaceable social media slop app. Banning robots would cost us the hardware our researchers rely on.
The GUARD Act directs U.S. national security agencies to review humanoid and quadruped robots as well as the software and communications systems that control them produced by foreign adversaries, explicitly China, Russia, Iran, and North Korea (but really, China). If, after review, any platform determined to pose an “unacceptable” national security risk is referred to the FCC for placement on its Covered List, which bans its importation and operation inside the United States.
Now, you may ask: what happens if those agencies do not review the robots within a year? Does that mean they get to enter? It does not! In fact, it means the opposite. Any qualifying robot that is not reviewed within that window is automatically added to the ban list by default.
The rationale comes in three parts. First, espionage/security. This is a legitimate concern when dealing with the CCP. The present concern would be more about preventing IP leakage than protecting consumer data. Autonomous robots are not really in the hands of consumers. Longer-term, as autonomous robots become more widely used, the concern set is like the TikTok use case, only amplified. If we’re assuming household humanoids are widespread, the notion of the Chinese government tapping into those cameras, sensors and everything else is indeed off-putting. So whether the concern is about current-day IP or future-day consumer protection, this is a legitimate worry.
The Select Committee’s May 2025 letter alleged an undocumented remote-access tunnel dubbed pre-installed on Unitree platforms that silently phoned home to PRC servers and could expose sensor data or root access1 So, this is serious, and real and warrants consideration.
I’m a touch less impressed by the second concern over our economic/industrial base. This is the idea that China is flooding the market with artificially cheap hardware that could bankrupt nascent American robotics companies before they reach scale. My initial response to this is: good? I mean, if China wants to use their hard-earned tax yuan to make American researchers’ hardware cheaper… I think that’s great. As with other times when China has been accused of “dumping,” if they’re selling us something below cost in an unsustainable way because their government is compelling them to do it… fine. More to the point: everyone is operating at a loss in robotics. Figure AI and Physical Intelligence are approaching $50 billion in combined valuations on the backing of almost no real-world business outcomes. We’re all losing money until we aren’t. I like our chances here better.
The more legitimate leg of this argument is that there are network effects to particular pieces of hardware, for establishing benchmarks, and allowing researchers to build off of each others’ work, and so it’s important that there’s a balanced field and American firms should compete on all levels of the supply chain, where possible. There’s something to that. That said: I don’t want to eat higher costs so that less competitive American companies can continue to exist. Earn it, fellas.
The third concern, which I think is pretty much wholly illegitimate and, in fact, is directly contradicted by the second, is the military-industrial complex sludge, in which these robotics firms are co-developing the People’s Liberation Army’s modernization and using revenue to develop weaponized robots. I suppose this is possible, but as I mentioned, there are no profits, so if the CCP wanted to develop weaponized robots, a more straightforward approach would be to… pay people to develop weaponized robots rather than do some weird bank shot with a commercial business.
Who is right on the merits?
Supporters have a reasonable case about security concerns. And there is an industrial policy case to be made that this is critical infrastructure. But opponents are right as well that slowing US-based research could be very, very costly.
The good news is that we’re almost certainly ahead in the “robot brain” VLA-type robotic intelligence layer, which is likely the most important. If history is a guide, software tends to eat the hardware layer–think Microsoft v. PC manufacturers, Android/iOS v. assemblers, NVIDIA v. server OEMs and on and on. But still, we’re pretty clearly behind on everything related to hardware, in terms of both manufacturing and design. And the brain only stays ahead if it has bodies to train and deploy on.
This is not necessarily a case where: Chinese ban -> everyone falls in line behind the US-based alternative. It may well just be a case where Chinese ban -> US falls behind because we have worse hardware. So, if we take the industrial policy concern seriously (and we should), the better approach would probably be subsidies to any emerging national champions rather than ineffectually banning Chinese hardware and slowing our own research.
I get that concerns about research progress feel vague and fuzzy. But that’s the thing about the frontier: it’s vague and fuzzy until it isn’t. The benefits are hard to define now, but over the long-term, if we all are excited about robotics’ future, then, well, this will hinder our path there.
Chris Paxton, a widely respected roboticist, has said this legislation would be “a serious blow to American academic robotics research, outside of big corporate labs.” His CEO at Agility Robotics, Peggy Johnson, on the other hand, offered unequivocal support, which to me mostly reads as “I don’t want my company to have to compete with Chinese robots.” That is special pleading, and not something our government policies should support.
What would a better regime look like?
Whether we think we can compete directly with Chinese firms on manufacturing (maybe?), cost (seemingly not), frontier modeling (yes!) or design (possibly), we do not have the rare earths. Robots need the rare earths. There is no viable workaround. And so, we don’t hold the cards. To the extent there’s a robotics industry, for the foreseeable future, it will be a Chinese one, at least in part.
A trade war is basically the worst way to get us a robotics industry that advances American interests. If we enter a trade war, we’ll lose. China can take its ball of rare earths and go home, so we’d need to find other pressure points to get the CCP to back off. In either case, this is zero sum thinking that opens us up to all sorts of retaliatory tariffs instituted by both China and the United States, which, not to get all Econ 101 in here, are bad. This approach is just a bad time for everyone involved.
Even if you assume that there are no second order effects and that we’re comfortable destroying the robotics industry… that’s not going to work. General purpose autonomous robotics is going to happen with or without us, and forming the best, most collaborative relationship with China on this matter will be beneficial for both nations; if we’re going to have a shot to be the preeminent one in this space, it’ll have to be at the research layer, and that research layer depends on Chinese hardware. The more our interests are intertwined, the more smoothly everything can go. As Lyndon Johnson said: better to have them inside the tent pissing out than outside the tent pissing in.
The import restriction we would most benefit from is a trusted-control-stack mandate, meaning that any robot operated in the US must run a certified, US/allied control layer–OS, networking, safety–and any imported bodies must be open at that layer. The data layer produced in deployments, nonnegotiably, must not be exfiltrated to the CCP.
That means no locked proprietary firmware and no CloudSail tunnels. In this case, cheap bodies stay legal, the value/control layer is forced into US and allied hands, and backdoors are neutralized at the network boundary, to eliminate espionage risk, or, I guess, a scenario in which the CCP can either brick US-based machines or use them for all kinds of nefarious purposes. This is a very serious concern that would grant China offensive capabilities against all American civilians–consider what Mossad was able to do with access to pagers and apply that to a device with cameras in every home.
The Cotton-Schumer American Security Robotics Act’s provisions on banning these companies from the US government procurement is eminently reasonable. If there’s one buyer who should be extra sure they’re getting the most secure tech, it’s the US military. The provisions stating that a university lab on a federal grant couldn’t buy or operate covered “unmanned ground vehicles” with those dollars, would limit all research to the private sector. That would be a total own goal that accomplishes very little other than handicapping our universities.
Within this regime, for the time being, we can certainly default to allowing imports rather than defaulting to banning them. Research firms with proprietary IP are aware of the risks and can choose to take them or not. Generally, if they’re at the frontier I expect they’ll want to avoid anything that could result in their most valuable asset–their IP stolen, and being on the ground, are more likely to be able to exercise good judgment on that than government officials.
On industrial policy I hold the loosest views. Maybe a CHIPS-style nudge to stand up domestic hardware pencils out, maybe it doesn’t; I’m not convinced the economics or the strategic case is there. If we must do something, make it tax credits over open-ended grants, and investment over import bans.
What’s clearly strategically important is lessening our dependency on Chinese rare earths. So, the Pentagon taking a 15% stake in US-based rare earth miners, MP Materials? Sure–a drop in the bucket overall, but the bucket is full of drops. To the extent there are opportunities for the public sector to grow our rare earth refining capacity, those should be pursued too.
We won the last era because the moat was made of bits we owned, buoyed by our cultural and economic power. This era’s moat is much more dependent on atoms that we don’t own. We cannot ban our way into owning something we cannot make and we cannot ban rare earths into existence.
So let’s stop trying to do that. Let the cheap bodies in, force the brain and the data into trusted hands, and compete where we’re actually ahead. By opening trade, we can make this a contest that our real advantages–research ingenuity, deep capital, a huge consumer market–are built to win. We’re not going to protectionism our way through this.
Let’s look at some robots
I suppose after all this, we should look at some Chinese robots, yeah? The video above has a Unitree humanoid scoring some goals, autonomously. It’s pretty impressive, though if you’re looking to see soccer played at the highest level, I’d consider watching the World Cup instead.
Continuing along our Unitree theme, I’m interested in the Humanoid Ikea Challenge for later this fall. We’ve seen robots autonomously accomplish harder tasks, and you can watch a Unitree humanoid accomplish the task via teleoperation. But it goes to show that autonomous robots’ usefulness is bottlenecked by intelligence more than physical capability.