Embodied Intelligence

The dominant narrative shift of the first half of the year is "from show floor to factory floor." A Xinhua observation piece in July called 2026 the year one of humanoid mass production: competitive focus has moved from launch events and demo videos to real deployments on production lines.Fact

• Fact Under industry tracking estimates, humanoid robot shipments grew roughly 272% year over year in H1 2026, with Chinese vendors accounting for about 97% of global shipments.
• Fact Unitree became the world's top seller of humanoid robots, scaling quickly on an aggressive price strategy.
• Fact Capital markets show two opposing forces: private funding surged and leading companies entered a "10-billion-RMB club" (~12 firms by valuation), while an IPO wave opened — and the representative listed company (UBTECH) remains loss-making.
• Fact On the model side, VLA (vision-language-action) and "world models" became the standard industry narrative; NVIDIA shipped a full stack of robot foundation models, simulation tools, and edge hardware at CES 2026, publicly positioning itself as the "Android of robotics."
• Fact The industry has started to consolidate: clear differentiation and elimination signals emerged in 2026, and story-driven companies without real delivery capability are struggling to raise.

This layering is not a strict supply-chain taxonomy; it is a way to classify questions when evaluating companies: which layer a company's real moat sits in determines which valuation logic applies.

Fact Market-size forecasts diverge widely: Morgan Stanley, citing a PwC whitepaper, projects China's embodied-intelligence market at roughly $15 billion by 2030; global forecasts range from tens of billions to over a hundred billion depending on when home robots are assumed to scale. My approach: watch shipments and real orders; treat market forecasts as scenario assumptions, not facts.

To verify The U.S. market shows a "six players, three routes" pattern (vertical integration, model platforms, body-plus-ecosystem alliances). None of the three routes has produced a decisive winner by mid-2026 — the key thing to watch in H2.

Others include Kepler, EngineAI, Star Dynasty, and LimX Dynamics, plus model-layer players such as Physical Intelligence (π series) and Skild. For funding and delivery tracking I use the Humanoid Index funding tracker and HumanoidIntel company tracker as public references.Fact

Judgment A counterintuitive pattern: the first real revenues come from the least human-looking machines. Quadrupeds and wheeled mobile manipulators generated real revenue by 2026, while the most human-like general humanoids remain in factory pilots. The form-factor competition is a tradeoff between "how human it looks" and "task density" — industrial scenarios pay for reliability, not human-likeness.

• Fact End-to-end VLA: vision-language-action models fold perception, reasoning, and action generation into one model. Representatives: NVIDIA GR00T series, Physical Intelligence π series, Google Gemini Robotics, Figure Helix.
• Fact World models: predicting the physical world (NVIDIA Cosmos, Genesis) for simulation training, planning, and data synthesis — becoming the second standard component alongside VLA.
• Fact Data pipelines: teleoperation collection, synthetic simulation, and real-world feedback in parallel. Chinese vendors commonly split into a "VLA large model + world model" two-layer architecture, trained in the cloud and deployed on-device.
• Fact Deployment constraints: control loops are latency-sensitive (millisecond to hundred-millisecond budgets); on-device inference budgets and cloud compute costs together define the product boundary — the same question I track in the On-device AI hub.
• To verify Dexterous manipulation, long-horizon tasks, cross-scenario generalization, failure recovery, and safety liability remain open problems without a general solution in 2026.

• Fact Industrial B2B pays first: automotive, 3C manufacturing, and logistics/warehousing are the main sources of real orders in 2026; the "factory floor" narrative maps to pilots and batch purchases in these scenarios.
• Fact Funding and IPOs diverge: private funding surged and leading valuations entered the 10-billion-RMB club; but reporting notes almost no funding flows to consumer products — "not a penny for ordinary people."
• Fact Unit economics are improving: China's supply chain has sharply cut whole-machine BOM cost; leasing (RaaS) and pay-per-task models lower trial barriers; cloud compute adds a new operating-cost line.
• Judgment Home is a long-term market, not a near-term one: long-tail tasks and liability boundaries raise technical difficulty and commercial risk together (see 家庭机器人为什么技术上更难，经济上可能更丰富).
• Judgment The likely order of evolution: industrial first, then services, then home; B2B first, then consumer. When evaluating a company: revenue evidence > order intent > funding amount > launch events.

Update log

• 2026-08-16 Hub created. Baseline is a mid-2026 snapshot: mass-production year one, +272% shipments, consolidation signals, VLA + world models as the main line.

This page is planned to update quarterly; major changes (funding, launches, delivery data) will be appended as short entries in the log.