Global humanoid robotics and embodied AI Report Interpretation

Overview

This global humanoid-robotics update interprets the 2026 World Humanoid Robot Games and recent industry developments through Moravec's paradox: locomotion has improved substantially, but perception, physical-world reasoning and productive task execution remain the decisive challenges. Morgan Stanley sees commercialization broadening, particularly in China, and frames data, robot intelligence and deployment feedback as the key determinants of progress.

Core views

The report argues that the industry is moving beyond highly visible locomotion demonstrations toward the harder question of whether humanoids can complete useful, variable, multi-step tasks efficiently enough for commercial deployment. The 2026 World Humanoid Robot Games gathered 666 teams and 2,056 robots, versus about 280 teams in 2025. Tiangong ran 100 meters in 8.64 seconds, versus 21.5 seconds in 2025, but Morgan Stanley places greater weight on the Games' 21 real-world scenario competitions across factories, homes, hotels, retail, hospitals and emergency response. These tests required tasks such as interpreting spoken instructions, preparing and delivering food, and performing fine manipulation; fully autonomous execution received twice the scoring weight of remote operation in relevant scenarios. The report's conclusion is that physical capability has advanced, but productivity, adaptability and autonomy are now the relevant commercialization benchmarks. Morgan Stanley identifies robot “brains” and data as the global bottleneck. Recent releases including DYNA-2, Skild AI S1 and Generalist GEN-1.5 seek to reduce the cost of adapting robots to new tasks, using few examples or short demonstrations and, in some cases, tackling longer task sequences. DYNA-2 was trained primarily on more than 1 million hours of egocentric human video and reportedly showed a scaling relationship between human-video training and robotic capability. The report emphasizes that data yield, diversity and feedback from deployments may be more durable advantages than model architecture alone. Companies are expanding model parameters, computing capacity—several operate at roughly thousand-card scale—and data collection, with targets of 1 million usable data hours by end-2026 or 2027. Figure's data initiative illustrates the scale of investment: it reported 260,000 app downloads, more than 44,000 weekly active users and 16 million videos, had paid US$15 million to creators, and planned more than US$1 billion of data and compute spending over the next 12 months. Demand indicators are improving but remain early-stage. Global humanoid shipments reached 19,000 units in 1H26, up 272% year on year from 5,000 in 1H25, with Chinese companies accounting for 97%. About 65% of China's shipments still went to entertainment, education/R&D and data collection, though industrial and commercial use cases were broadening. Half-size bipeds accounted for about 50% of volume, while wheeled robots represented about 29% and were expanding in commercial and industrial settings. Morgan Stanley expects more 1H26 pilots to convert into broader deployments from 2H26 and forecasts China's shipments at 50,000 in 2026e and 446,000 by 2030e. Examples of deployment momentum include Ant Group robots operating pharmacy night shifts, a planned rollout of at least 1,000 Hexagon AEON units in manufacturing, and Hyundai's target for Atlas deployment from 2028 alongside 30,000 units of annual US robot-production capacity. Policy support is another part of the commercialization flywheel, especially in China. The report notes that China's 2026 initiative targets 10,000-level deployment capacity and more than 100 high-value applications by end-2026, with local governments and central SOEs expected to identify operational sites for humanoid training. It also tracks policy activity in South Korea, Japan, the US and the UK. At the same time, the report flags that physical AI's reliance on cameras, LiDAR and force/torque sensors raises data-security and surveillance concerns, while AI-enabled robots have dual-use national-security implications. For long-term market sizing, Morgan Stanley estimates cumulative global adoption of about 28.1 million humanoids by 2036, 138.5 million by 2040, about 430 million by 2044 and roughly 1 billion by 2050. Its 2050 estimate includes about 935.1 million commercial adoptions and 84.2 million household adoptions. Assuming a six-year replacement cycle, the firm estimates annual global humanoid revenue of US$336 billion by 2035, US$2.0 trillion by 2040 and US$7.5 trillion by 2050. The model assumes substantial ASP deflation: high-income-market ASPs fall from US$200,000 in 2024 to about US$75,000 by 2050, while lower-cost markets decline from roughly US$50,000 to roughly US$21,000. The adoption case therefore depends on scaled production, localized supply chains and robots achieving economically useful real-world performance.

Analysis framework

Morgan Stanley combines observations from the 2026 World Humanoid Robot Games, shipment and deployment indicators, company and model-development updates, policy tracking, value-chain screens and a bottom-up adoption model. It tests the transition from demonstrations to commercially productive tasks, then sizes long-term unit demand using regional adoption assumptions, ASP declines and a six-year replacement cycle.

Methodology notes

• Industry AnalysisSupply-demand framework

  Humanoid adoption and market sizing using deployment assumptions, ASP trajectories and replacement demand.

  The report estimates units adopted across country-income groups and end-markets, then converts demand into revenue using declining average selling prices and a six-year replacement cycle.

• Other

  Scaling-law assessment for embodied-AI training data.

  The report uses reported evidence from DYNA-2 and other models to assess whether more human-video data can improve robotic capability and reduce the cost of adapting to new tasks.

Key data

• Global humanoid shipments19,000 units in 1H26Up 272% year on year from 5,000 units in 1H25; Chinese players accounted for 97%.
• China humanoid shipment forecast50,000 units in 2026e; 446,000 units by 2030eMorgan Stanley expects pilots to broaden into deployments from 2H26.
• Cumulative global humanoid adoption~28.1 million by 2036; ~138.5 million by 2040; ~430 million by 2044; ~1 billion by 2050Morgan Stanley estimates.
• Global annual humanoid market revenueUS$336 billion by 2035; US$2.0 trillion by 2040; US$7.5 trillion by 2050Assumes a six-year replacement cycle.
• Humanoid 100 performance+47% since February 6, 2025Equal-weighted, adjusting for additions and deletions; data cited through August 28, 2026.
• China Humanoid Value Chain performance-1.3% MTD and -21.8% YTDEqual-weighted as of August 28, 2026, versus MSCI China at -0.5% MTD and -9.1% YTD.

Impact & implications

The report portrays the next phase of the humanoid industry as a commercialization and intelligence challenge rather than a locomotion race. It expects companies and value-chain participants to be increasingly differentiated by access to scalable data, model capability, deployment feedback, industrial validation and the ability to reduce cost through production scale.

Risks

• Data availability, diversity and quality remain a bottleneck for embodied-AI model development.
• Robots must prove they can perform real tasks efficiently and productively; many deployments remain at tens of units per site.
• Physical-AI sensors and models may raise data-security, surveillance and national-security concerns because of their dual-use characteristics.

What to watch

• Conversion of 1H26 pilots into broader deployments from 2H26.
• Evidence of productive, autonomous execution in real-world industrial and commercial tasks.
• Progress toward 1 million usable data hours and the quality of deployment feedback loops.
• China's implementation of targets for 10,000-level deployment capacity and more than 100 high-value applications by end-2026.
• Whether model updates materially shorten the time and cost required to teach robots new tasks.