Overview
Embodied intelligence is crossing from lab demos into a scale-up window. At an October 19 State Council press conference, MIIT’s science and technology director Wei Wei said China’s humanoid-robot whole-machine products surpassed 400 models in H1 with output above 40,000 units, and full-year output is expected to exceed 100,000. MIIT has pooled industry-academic forces to tackle motion control, simulation and tactile sensing, and is pushing real-scene embodied training.
International and industrial momentum is building. OpenAI CEO Sam Altman disclosed a humanoid-robot program centered on the “AI brain,” with a demo prototype expected in 2027. China’s CAICT, with Tsinghua and others, released the 2026 Embodied Intelligence Development Report, arguing the field is moving from “performance” to “application,” that model self-evolution remains “bounded,” and that agent engineering has converged on “models handle reasoning, Harness handles orchestration.” On the consumer front, Chery unveiled and opened pre-orders for “Mobao,” a consumer home-companion humanoid, while its “Mojia” robots already work routinely in 100+ education and healthcare scenarios.
From policy to industry, embodied intelligence is positioned as AI’s key bridge from the digital to the physical world and a pillar of “new quality productive forces.” The path—enter via one category (humanoids), then diffuse across the chain—means landing pace will be driven by whole machines pulling parts and software ecosystems.
Background & Interpretation
1. Technical Background and Evolution
The embodied stack is stratifying fast: a base of motion control, tactile sensing and simulation; a middle of “world model + spatial understanding” embodied foundation models (e.g., Daxiao’s ACE-Brain-0.6, fine-tuned on Qwen3.5-9B, beating GPT-6 Astra on six spatial metrics); and a top of reasoning-orchestration Harness. CAICT’s report clarifies physical-AI vs. embodied concepts, proposes a four-element framework, and stresses real-scene training along “scale data—scenario docking—deployment.”
2. Core Drivers and Underlying Mechanisms
The volume inflection rests on “clear demand + falling cost + software-defined.” Over 300 key scenarios—warehousing, retail, home services—are being validated; hardware costs have dropped sharply (capable arms from luxury-car to high-end-laptop prices), exposing a “hardware proliferation, intelligence lacking” gap that creates room for general robot “brain” models. OpenAI’s choice to enter via the brain, not the body, redefines the competitive dimension.
3. Market Structure and Industrial Chain Impact
Humanoids will lift reducers, servos, sensors and dexterous hands, plus embodied large models, simulation testing and data services. Capital has voted: Faux Robotics closed nearly ¥1 billion in C+/D rounds; MemSilicon raised ¥400M+ for LPU inference chips. Whole-machine expansion and cross-over entries shift competition from “demo show” to “delivery power.”
Implications & Outlook
Near-to-medium term, 2026 could be the “batch-delivery first year” for humanoids, though true scale needs time. Watch real-scene data accumulation, dexterous-operation generalization, and whether unit cost falls into consumer range. Maturation of world-model benchmarks (WorldOracle, X2Real) will speed convergence of the “sim-to-real gap.”
Takeaways for Industry Participants
Whole-machine and body firms should hone reliability and cost around real scenarios, not just demos. AI and software companies can enter via the “brain,” simulation benchmarking and data services. Component suppliers must scale delivery and consistency with OEM expansion. Investors should weigh order-binding, production cadence and unit economics, guarding against concept overheating.
This column compiles industry information and shares technical perspectives; it does not constitute investment advice.