China Just Made Physical AI the Center of Its Economy. What's Washington's Answer?
The most consequential paragraphs of the 15th Five-Year Plan weren't about growth targets. They were buried on page 73, and they describe a plan to industrialize robot training on a national scale.
“Control the infrastructure, and you control the possibilities.” — Susan Crawford
When the National People’s Congress adopted China’s 15th Five-Year Plan in March, the Western reaction settled quickly into a familiar groove. Commentary fixated on the GDP target — a cautious 4.5–5 percent range, the lowest since China began publishing five-year plans in earnest — and on the now-routine emphasis on technological self-reliance. The plan was read, summarized, and filed away as continuity rather than rupture.
That was a mistake.
Buried deep in the 141-page document, in an inset box on page 73, is a passage that may prove more consequential than any macroeconomic figure in it. It announces that China will “coordinate the layout of embodied intelligence training grounds, promote virtual-real fusion collaborative training and evolution, develop integrated big-brain/small-brain embodied models and algorithms, tackle key technologies in the body and core components, and accelerate the upgrade and deployment of humanoid robots and other form-factor products.”
This is not the language of aspiration. It is the language of procurement.
The term “artificial intelligence” appears more than fifty times in the plan. “Embodied intelligence” (具身智能) — a term that barely existed in Chinese policy documents before 2023 — now commands its own dedicated inset among the plan’s top ten “new industry tracks,” alongside integrated circuits, biomanufacturing, commercial space, and the C919 jetliner program. In the space of three years, Beijing has moved embodied AI from a niche research interest to a designated engine of the national economy.
The contrast with how Washington is treating the same technology is hard to miss.
What “Embodied Intelligence” Means in Beijing-Speak
To translate the jargon: embodied intelligence is not just humanoids. It is foundation models and AI agents tied to physical systems — robots, drones, autonomous vehicles, agricultural machinery, and any other form factor that interacts with the physical world. The category sits at the intersection of robotics, large models, sensors, and data infrastructure, and Beijing has chosen to treat it as a single integrated industry rather than a collection of adjacent ones.
The plan’s drafting is revealing. Embodied AI sits alongside quantum technology, biomanufacturing, hydrogen and fusion energy, brain-computer interfaces, and 6G as the six designated engines of what the document calls the “intelligent economy.” But its placement is broader than that grouping suggests. Robotics and embodied AI are woven across multiple chapters of the plan — manufacturing, digital transformation, healthcare, social welfare — and treated as a horizontal capability rather than a vertical industry. Several analysts have read this as the more important signal: the plan does not just elevate the sector, it positions it as connective tissue.
The Chapter 13 framing makes this explicit. The plan establishes an “AI+” (人工智能+) action plan as a cross-cutting national program, modeled loosely on Germany’s Industry 4.0 but covering six domains rather than one: science and technology, industrial development, consumer applications, social welfare, governance, and national security. Robots are expected to show up in all six.
When Beijing says “embodied intelligence,” it is not talking about a handful of flashy humanoids at a trade show. It is talking about rewiring how work, services, and social governance operate.
The Training Grounds Clause
The single most underdiscussed item in the plan is Box 3, Item 02 — the directive to build national-scale “training grounds” for embodied AI.
In practical terms, this means warehouse-scale and city-scale testbeds where robots learn through a combination of simulation and real-world deployment, generating proprietary data loops that feed back into national models. The plan frames this as a coordinated buildout across line ministries and provinces, not a research exercise. MERICS, the Berlin-based China research institute, has counted more than forty state-funded robot training centers already operating, generating what the institute estimates as millions of real-world training data entries.
The strategic logic is worth pausing on. The single hardest problem in physical AI today is not algorithms or chips — it is data. Foundation models for language can be trained on scraped internet text. Foundation models for robots need physical experience: contact forces, manipulation trajectories, recovery from failure, performance under wear and friction. That data does not exist on the open web. It has to be generated, and generating it at scale requires deployments at scale.
China’s installed base solves part of that problem by accident. By 2024, China had roughly 2,027,000 industrial robots operating in its factories — about 4.5 times Japan’s stock — and it installed 295,000 new industrial robots in 2024 alone, 54 percent of global deployments that year. The 15th FYP turns that incidental data advantage into an explicit one. Training grounds are not a hedge against the data bottleneck; they are an industrial policy designed to widen it into a moat.
China is not just buying robots. It is industrializing robot training itself — turning data from factories, logistics parks, and cities into a strategic national asset.
The architecture matters. Where Western embodied AI companies are scrambling to assemble proprietary training data from whatever deployments they can negotiate, Chinese firms are being offered access to nationally coordinated data infrastructure as a matter of policy. The advantage is not a head start in any single model. It is a head start in the loop that produces every future one.
Humanoids and Standards as Deliverables
The training-grounds passage sits inside a broader directive that names humanoid robots specifically, calling for accelerated “upgrade and deployment of humanoid robots and other form-factor products” and explicit targeting of “key technologies in bodies and core components.” The wording is not loose. Core components is a reference to harmonic reducers, ball screws, servo motors, and torque sensors — the precision parts where China remains dependent on Japanese and German suppliers, and where 35–45 percent of the country’s high-value robotics components currently come from Japan alone. The plan is essentially a directive to break that dependency on a defined timeline.
In parallel, Beijing has built the standards scaffolding to match. The Ministry of Industry and Information Technology established a dedicated Humanoid Robot and Embodied Intelligence Standardization Technical Committee in December 2025, and by March 2026 had released the first national standard system covering the humanoid robot industry’s entire lifecycle. China is now leading the formulation of global standards for elder-care robots and is actively working to shape international norms for robot safety, interoperability, and data governance.
The pattern is familiar to anyone who watched China’s 5G and high-speed rail playbooks: establish the domestic standard first, build scale around it, then export it as the de facto international norm. IEEE, ASTM International, and ISO committees are all engaged in humanoid robotics, but no single Western jurisdiction has matched the speed or vertical integration of Beijing’s standards push.
Municipalities are racing to localize. Shanghai’s Action Plan for the Development of the Embodied Intelligence Industry (2025–2028) commits to building a complete ecosystem with fifty core partners, at least five national or industry-wide standards, and a roster of flagship applications. Shenzhen launched a dedicated ten billion RMB AI and Robotics Industry Fund in early 2025. Beijing’s Yizhuang development zone has declared its intention to become “a global first-class embodied intelligent robotics industrial new city.” Guangdong has flagged embodied AI, 6G, quantum, and the low-altitude drone economy as its provincial priorities for the FYP period. The central directive is being amplified at every level beneath it.
Where Western regulators are still debating what a safe humanoid is, Chinese cities are being told how many standards and flagship pilots they are expected to deliver.
The Numbers Underneath
It is worth being precise about the scale on which all this is being layered.
China is not entering the robot race from a standing start. By 2024, IFR data put China’s operational stock of industrial robots at roughly two million units, with new installations of 295,000 that year representing more than half of global deployments. In the nascent humanoid segment, Chinese firms shipped approximately 90 percent of the world’s units in 2025, led by AgiBot (5,168 units), Unitree (over 4,200), and UBTech.
The domestic supply share has shifted in parallel. Chinese producers now supply 59 percent of robots installed inside China, and roughly 90 percent in the metal and machinery sector — IFR’s own data, not Chinese government claims. The 14th Five-Year Plan’s stated objective of breaking foreign dependency on the supply side has, by IFR’s reading, substantially worked.
The 15th FYP is not the start of this trajectory. It is an accelerant layered on top of an already enormous base, and a deliberate pivot from what IFR characterizes as “traditional industrial automation” toward “high-end, intelligent robotics integrated with AI.” The shift is from selling more arms to selling smarter ones — and from being the world’s largest robot buyer to being the world’s dominant robot platform.
The question is not whether China catches up in robots. It is whether the rest of the world is ready for a world in which more than half of the world’s industrial robots are plugged into an embodied AI stack shaped in Beijing.
Two Loops: Open AI and Physical Control
The 15th FYP makes a second move that has received even less attention than the training-grounds passage: it commits explicitly to leading the global open-source AI ecosystem.
China’s open-model strategy and its manufacturing base are now being treated as interlocking flywheels. The logic runs as follows. Chinese labs publish capable open-source foundation models at low token cost. Those models get downloaded, fine-tuned, and embedded into robots, logistics systems, and industrial controllers — increasingly Chinese-made ones. The deployments generate real-world data. The data improves the next generation of models. The improved models extend the reach of Chinese platforms globally. Each turn of the loop deepens the dependency.
This is not theoretical. February 2026 data showed Chinese LLM usage surging globally, driven by rising demand for agentic applications and by token pricing substantially below Western equivalents. Energy costs underpin the math. Renewables overcapacity in western China has produced electricity prices lower than in most U.S. states — a structural cost advantage in a domain where inference scales directly with electricity.
The contrast with the U.S. pattern is sharp. Open-model action in the United States remains concentrated in software, where the policy debate is about chip exports and frontier-lab compute. Robots, when they enter the conversation, do so as an afterthought — a downstream beneficiary of progress made elsewhere rather than a strategic priority in their own right. The Trump administration’s AI Action Plan, released in July 2025, mentions robotics in a paragraph. The rumored 2026 executive order on robotics has been “under consideration” for nearly a year. The Hudson Institute and ITIF have both published papers arguing for a CHIPS-style robotics industrial policy. None has been passed.
China is wiring its AI flywheels directly into robots. Washington is still treating robots as an afterthought to software.
The Hudson paper is worth reading in full because it concedes the structural problem: the U.S. theory of victory in AI is to limit Beijing’s compute by restricting GPU exports. That theory may or may not work in software. In embodied AI, where the binding constraint is data and deployment rather than chips, it may not even be the right theory.
What This Means for Boards and Ministries Outside China
For decision-makers outside Beijing — in corporate boardrooms, ministries, standards bodies, and capital allocators — the 15th FYP raises a small number of questions worth taking seriously, regardless of where one sits on geopolitics.
The first is industrial strategy. Most Western jurisdictions today have a robotics policy of sorts, usually consisting of grants, tax credits, and pilot programs. Few have an embodied AI strategy as such — a coordinated framework that treats data infrastructure, deployment environments, standards, and procurement as a single integrated stack. The default state is what one analyst has called “pilot-itis”: dozens of unconnected projects without a national spine.
The second is infrastructure. Beijing is funding physical training grounds for embodied AI directly. In most Western countries, the equivalent infrastructure is expected to emerge from private vendors building their own testbeds or from one-off academic partnerships. Whether that ad hoc model produces national-scale training data within the time window that matters is an open question.
The third is standards. Many of the international standards that will shape how embodied AI systems operate — safety thresholds, behavioral norms, data-sharing rules, certification pathways — are being actively drafted right now in IEC, ISO, and ASTM committees. The Chinese MIIT committee established in December 2025 is contributing to those processes with coordination and resource. Other jurisdictions are participating, but more slowly, and with less domestic alignment between their standards bodies, regulators, and industry. Standards shape markets. Standards drafted under one set of national priorities tend to advantage producers operating under those priorities.
The fourth is procurement and export control. Most existing frameworks were designed for a world in which “AI” meant cloud software and the relevant questions were about data sovereignty, model bias, and computational sovereignty. They are not well-suited to a world in which AI increasingly means physical systems whose design, training data, and behavioral norms are shaped by foreign industrial policy. Updating procurement rules, certification regimes, and export-control thinking for embodied AI is a project most ministries have not started.
None of these questions has an obvious answer. But each is the kind of question that, left unaddressed, tends to be answered for you.
A Note on Time
There is a recurring pattern in technology policy where a new layer of infrastructure is treated as a peripheral concern for one cycle, and as the defining strategic question of the next. Cloud computing went from an IT line item to the backbone of national power in roughly a decade. The same arc is plausible for embodied AI, on a faster clock.
China has decided that the next decade of AI will be embodied, and has organized its industrial policy accordingly. The plan is now public, the funding is committed, the training grounds are being built, and the standards committees are sitting. None of this guarantees success — provincial overcapacity, component dependencies, and uneven execution are real risks for Beijing — but the plan exists and is being implemented.
The rest of the world’s plans, for now, are being written.
Robot News Of The Week
Hello Robot’s latest Stretch 4 is bigger, faster, and stronger than previous versions
Stretch 4 represents something larger than a new robot platform. It reflects a growing realization across robotics that physical AI will not succeed through intelligence alone, but through mobility, sensing, safety, and human-centered design working together. With its sensor-rich architecture, omnidirectional mobility, and focus on real-world assistive use cases, Hello Robot is positioning Stretch 4 as both a research platform and a glimpse into the future of everyday robotics. More importantly, the company’s cautious approach to scaling highlights a lesson the broader robotics industry continues to learn: successful deployment is not about flashy demos, but sustained usefulness in messy, unpredictable human environments.
ABB Robotics launches PickMaster Lite to simplify and accelerate robotic picking
ABB Robotics’ launch of PickMaster Lite reflects a broader shift happening across industrial automation: manufacturers no longer just want powerful robotics software — they want automation that is faster to deploy, easier to integrate, and simpler to scale. By reducing engineering complexity and shortening commissioning timelines, ABB is targeting one of the biggest barriers to robotic adoption for packaging OEMs and system integrators: deployment friction. The platform’s combination of vision-guided picking, digital twin integration through RobotStudio, and simplified workflows highlights how the next phase of industrial robotics growth may depend less on breakthrough hardware and more on making advanced automation accessible to companies that lack deep robotics expertise.
America’s Robotics Industry Just Organized for Washington
Robots for America signals a major shift in U.S. industrial strategy. What began as a robotics coalition may become something much larger: a coordinated effort to modernize American manufacturing through automation, physical AI, workforce development, and industrial policy. The real story isn’t just robots — it’s America reorganizing for the next industrial era.
Robot Research Of The Week
Honeybees teach drones how to navigate
Bee-Nav highlights one of the most important lessons emerging in robotics and physical AI: intelligence does not always require massive compute. By borrowing navigation strategies from honeybees, researchers at Delft University of Technology have demonstrated that lightweight drones can navigate complex environments using tiny neural memories rather than compute-intensive mapping systems. The work points toward a future where smaller, cheaper, and more energy-efficient autonomous robots can operate in places where GPS and large onboard computers are impractical. Beyond its technical achievement, Bee-Nav also reinforces a growing trend in robotics — that some of the most important breakthroughs may come not from endlessly scaling AI models, but from rediscovering the elegant efficiency already present in nature.
The video below shows a small drone learning how to find its way home in a room with obstacles. First, it flies around, taking pictures to learn what the room looks like.
Then the drone is placed in different spots and must return to the home location by itself while avoiding obstacles so it doesn’t crash.
Shipbuilding’s New Co-Pilot: AI and Robots Enter the Drydock
Backed by a $6.2 million grant from Japan’s Ministry of Land, Infrastructure, Transport and Tourism, a new collaboration between American and Japanese researchers is exploring how AI and autonomous robots could help modern shipbuilding adapt to the messy reality of construction. Led by University of Michigan, the project aims to create robotic “co-pilots” that scan ships during construction, compare what was actually built against digital plans, and help workers spot problems before delays spiral out of control. The effort reflects a broader shift in industrial robotics — from simple automation toward systems that can monitor, reason, and help humans make complex decisions in dynamic environments.
Robot Workforce Story Of The Week
The U.S. Small Business Administration has announced a major new push to strengthen domestic manufacturing, unveiling up to $50 million in grants through its Manufacturing in America Empower to Grow initiative. The program will fund up to 10 organizations that can provide training, technical assistance, and hands-on support to small manufacturers operating in critical industries including robotics, shipbuilding, aerospace, advanced manufacturing, and industrial equipment. Beyond the funding itself, the initiative reflects a broader national strategy centered on reshoring supply chains, rebuilding industrial capacity, and helping smaller manufacturers adopt the tools, workforce skills, and operational capabilities needed to compete in an increasingly automated and geopolitically contested manufacturing landscape.
Robot Video Of The Week
For decades, giant piloted mechas belonged firmly to science fiction. Now, companies like Unitree Robotics are beginning to blur the line between fantasy and engineering reality. The debut of the GD01 — a 2.7-meter-tall “manned mecha” capable of transforming between two and four legs — has captured global attention and reignited debate about the future of embodied AI, humanoid robotics, and China’s rapidly accelerating robotics ecosystem. More than a viral spectacle, the GD01 symbolizes China’s growing ability to commercialize futuristic robotics concepts at scale, leveraging deep manufacturing supply chains, aggressive iteration cycles, and national industrial policy to turn science fiction into deployable machines.
Upcoming Robot Events
May 27-28 Robotics Summit & Expo (Boston, MA)
June 1-5 IEEE ICRA (Vienna, Austria)
June 22-25 Automate (Chicago, IL)
June 23-24 Humanoid Robot Forum at Automate Show (Chicago, IL)
July 7 Machina Summit (Paris)
Aug. 17-21 IEEE Intl Conf. on Automation Science & Engineering (Shenyang, China)
Sept. 14-19 International Manufacturing Technology Show (Chicago, IL)
Sept. 15-17 AI Infra Summit (Santa Clara, CA)
Sept. 22-24 ROSCon Global (Toronto, Ont.)
Sept. 27-Oct. 1 IROS (Pittsburgh, PA)
Oct. 6-8 Motek (Stuttgart, Germany)
Oct. 14-15 Natural Disaster Expo (Houston, TX)
Nov. 3-5 International Robot Safety Conference (Detroit, MI)
Nov. 9-13 International Conference on Telepresence (Bristol, UK)
Mar. 2-4 European Robotics Forum (Birmingham, UK)