Humanoid robots are stuck between two unfinished layers: a body that is still too heavy, too power-hungry and too hard to build at precision, and a brain with no internet-scale body of physical data to learn from. China leads on the layer that matters most right now, which is building and shipping at volume, and Chinese vendors accounted for over 97% of global shipments in the first half of 2026. IDTechEx‘s reports Humanoid Robots 2026-2036: Technologies, Markets, and Opportunities and Materials for Humanoid Robots 2026-2036: Technologies, Players, Forecasts put real commercialization in the 2026 to 2030 window, led by car factories. Catching up runs through magnets, precision screws and training data, and Washington’s response has been a policy wall.

Eighty percent of humanoids are being built for the factory floor

IDTechEx’s analysis shows that 80% of humanoid robots and prototypes to date have been designed for industrial settings, and that choice drives everything about the materials. Industrial humanoids average 78kg, and some wheeled models pass 100kg. Actuators make up more than half that weight, because the gears and screws in hips and knees need hardened, fatigue-resistant steel and aluminum alloys. Every kilogram of metal fights every kilogram of battery, which is why some OEMs have given their robots wheels instead of legs.

Drawing on more than 50 humanoid products and prototypes, IDTechEx finds platforms converging on electric joints, integrated sensing and modular subsystems, with cost and performance still well short of mass-market readiness. Development priorities have moved toward manufacturability, maintainability and deployment feasibility, away from research-lab designs. The demand implication is large: IDTechEx projects total materials demand for humanoids growing at a 47% CAGR from 2026 to 2036, passing 112,000 tonnes by 2036.

The bottleneck you can hold in your hand

The hardest parts to make are the smallest. Planetary roller screws and gears need precision and consistency that current suppliers struggle to hold at scale, and wear and fatigue remain unsolved. IDTechEx’s main humanoid report names battery energy density and thermal limits as constraints on operating time, and says supply chains for screws, bearings and high-performance actuators are not yet built for mass-volume humanoid production. Dexterous hands and tactile sensing remain a major barrier to anything beyond simple industrial work.

Then there are the magnets. NdFeB rare-earth magnets deliver the torque density humanoid actuator motors need, and their supply chain is heavily consolidated in China. That concentration already has a body count in missed targets: Tesla aimed for 5,000 Optimus units for internal use in 2025, actual output was reportedly a few hundred, and China’s April 2025 rare-earth magnet export restrictions were cited as one cause. Those April 2025 restrictions on seven rare earth elements remain in place, alongside a dual-use licensing regime.

The brain has no internet to learn from

Language models trained on text humanity had already written. Training a robot policy requires synchronized action and state streams that do not exist at internet scale; the sources are world models, human teleoperation, which is expensive, simulation, which has a fidelity gap, and pooled open datasets. IDTechEx’s report treats the lack of training data as a core pain point for the AI stack, with synthetic data generation through tools like NVIDIA’s Isaac GR00T as the leading workaround.

World models are the strongest intelligence option for getting around the data bottleneck, though so far they work alongside vision-language-action models (VLAs) rather than replacing them. A world model learns how the physical world changes when something acts on it, so a robot can predict consequences before it moves. They can learn from video, which is abundant, instead of relying only on teleoperated robot data, which is scarce and expensive.

They are showing up in three roles.

As a data engine. They generate synthetic training scenarios. IDTechEx’s humanoid report treats NVIDIA’s Cosmos and Isaac GR00T as the main answer to the training-data gap.

As a planner. Meta’s V-JEPA 2 was trained on just 23,000 robot trajectories. It then ran zero-shot on Franka arms in two labs that weren’t in its training data, using a single uncalibrated low-resolution camera. It plans by imagining future states and picking actions that move toward a goal image. One 2026 paper argues that world action models can serve as zero-shot policies on their own.

Inside the VLA. Researchers are merging the two approaches. One line of 2026 work argues that predictive embeddings from video world models like V-JEPA 2 give robot policies better priors than static visual representations, and uses them as the backbone for existing VLAs.

The limits map onto the weak points of the body. Video carries no force or touch information, and hands and tactile sensing are where humanoids fail most. Simulation fidelity also breaks down on hard cases: visual sim-to-real gaps cause most manipulation failures involving shiny or transparent objects. Running a world model on the robot adds compute and heat to a machine already limited by battery and thermal limits. So world models can sharpen the brain, but they can’t fix a 78kg body with two hours of battery.

Leading intelligence-hardware innovation

NVIDIA is the clear platform leader. The other serious players split into three groups: big labs, humanoid makers building their own world models, and robot-brain startups that use someone else’s. AgiBot, the Chinese shipment leader, has its own.

The platforms: NVIDIA. Cosmos is the default world model for the humanoid industry. It was trained on 20 million hours of real-world data, and early users include 1X, Agility, Figure and Skild AI. At GTC in March, NVIDIA announced Cosmos 3, which it describes as the first world foundation model to combine synthetic world generation, vision reasoning and action simulation. Its GR00T-Dreams pipeline uses Cosmos to generate synthetic training trajectories from a single image and a language instruction, so robots can learn tasks without teleoperation data. One investor analysis calls NVIDIA’s decision to build and open-source the full physical AI stack the most important strategic development in the space.

• 1X released a physics-based world model that learns new tasks for its Neo robots from video and prompts.

• AgiBot released Genie Envisioner 2.0 in April. It is an interactive simulator where robots can be trained and evaluated at scale without real-world trial and error. AgiBot says the model ranked first on the WorldArena benchmark in May, which is a company claim.

• Unitree has an open world-model-action framework, UnifoLM-WMA.

Brain startups and world builders. Skild AI and FieldAI build general robot brains, using Cosmos to generate data and Isaac to validate policies in simulation. World Labs, Fei-Fei Li’s company, shipped Marble, its first commercial world model, and was reportedly raising at a $5 billion valuation. Yann LeCun’s AMI Labs is reportedly seeking a similar valuation.

China’s intelligence edge rests on collecting physical data at scale, which is what its data standards and training grounds are built for. World models trained on human video are the one route that could let the West compete on intelligence without matching China’s deployment volume.

The money pattern is telling; companies using world models have raised more than companies building them. Either the world-model layer is underfunded or the big robotics companies will bring it in-house, as 1X already has.The catch is that contact-rich skills still need real robots doing real work, which brings us back to the factory gap.

China is treating data as infrastructure. Its National Data Administration said it will develop standards for embodied AI training data, ten days after seven companies asked the agency for public data infrastructure and common standards. A national guidance document on real-world embodied intelligence data, GB/Z 218.1-2026, is already published, with a dataset quality standard taking effect 1 November.

How far commercialization actually is

IDTechEx’s forecast is measured. It expects 2026 and 2027 to be transition years, with more players moving from pilots toward production readiness and early rollout, while conceding deployment volumes remain below earlier targets. Automotive manufacturing scales first, on OEM backing, controlled environments and clearer ROI for repetitive tasks; logistics follows but competes with AMRs, AGVs and robotic arms; home robots stay limited through 2036. The overall market reaches roughly US$29.5 billion by 2036, with unit volumes approaching a million around 2030.

On the ground in 2026, the numbers are small. As of mid-2026, no humanoid from any manufacturer had been deployed above the low hundreds of units in a sustained commercial environment. The best-documented Western record is Figure at BMW Spartanburg, where robots run ten-hour shifts loading sheet-metal parts into welding fixtures, with more than 90,000 parts placed at over 99% accuracy. Tesla is behind its own rhetoric: in January 2026 Musk said Optimus was “not in usage in our factories in a material way,” and the first units off the Fremont line are going to an internal training programme, not to factory work or customers.

China ships at scale, and much of it goes to labs

The volume gap is real. Smart Analytics Global counted about 19,100 humanoid shipments globally in the first half of 2026, up 272% year over year, with AgiBot at 44% of the market on roughly 8,400 units. Counterpoint’s count runs higher, at more than 22,000 units, with Chinese companies taking all five top positions. China’s industry ministry projects full-year 2026 output above 100,000 units, a fivefold jump from about 20,000 in 2025. UBTech’s Walker S2 units are working on assembly lines at BYD, Geely, FAW-Volkswagen and others, plus Foxconn factories and SF Express logistics, and its new Liuzhou plant is designed for more than 10,000 robots a year on a 10-minute takt, a planned capacity rather than verified output.

Robots designed for factories are not the same as robots working in factories. Through Q3 2025, more than 70% of Unitree’s humanoid revenue came from research and education customers, with industrial applications at 9%. Unitree’s Q1 2026 net profit fell 52% year over year. The G1 starts at $13,500, but the spec lists roughly two hours of battery and a 2kg maximum arm load on the standard model.

China’s lead, then, sits in the manufacturing learning curve. It comes from a base no one else has: China installed about 276,000 industrial robots in 2025, more than Japan, the US, Germany and South Korea combined. Add magnet dominance, component density and a state that turns industry requests into standards in ten days, and the cost curve bends faster there.

What catching up requires

Magnets first. China’s concentration is already driving alternative supply chains across the US, Europe and Asia, including a $400 million US Department of Defense investment in domestic magnet production.

Precision components second: actuator suppliers that traditionally serve automotive are diversifying into humanoids, which is the natural path for Canada’s parts sector.

Third is data infrastructure built as a public good, which the West has not attempted.

Fourth is deployment hours; the data flywheel only turns when robots are working.

Canadian Sovereignty Check

Canada’s position is thin. Sanctuary AI is Canada’s only homegrown humanoid company, and in June 2026 it pivoted to selling its physical AI software for other robots, with its hydraulic hands still in development. The most visible Canadian deployment is a foreign robot: Agility signed a commercial agreement with Toyota Motor Manufacturing Canada for seven Digit units on RAV4 logistics.

The wall Washington built

In late July the US FCC added foreign-made humanoid and quadruped robots to its Covered List, covering new models only, with federal government sales exempt, and producers can seek a Conditional Approval exemption from the Department of War. Morningstar’s Li Kangyuxiao noted that Chinese manufacturers have been scaling production and cutting costs faster than most overseas competitors.

The implications compound. A US market walled off from 97% of global supply pays more per robot and deploys fewer of them, which means fewer working hours and less training data for the model layer where US firms lead. China keeps building volume, learning the manufacturing curve and accumulating standardized physical data under state coordination. The materials market IDTechEx forecasts will be supplied by whoever qualifies first. For Canada, the gap arrives, as much now does, as a sovereignty question: whose robots, whose magnets, whose training data, and whose rules govern machines that will eventually work beside people in plants from Oshawa to Cambridge.