Overview
From November 2 to 5, the 2026 World Internet Conference “Light of Internet” expo in Wuzhen put embodied intelligence center stage. Shanghai Securities News reported landmark figures: UBTECH disclosed a manufacturing cadence of “one humanoid robot off the line every 10 minutes”; Xiaomi’s embodied foundation model completed a 30-minute-plus, three-room, 50-plus-task long-horizon chain at a pilot apartment.
Unitree’s founder Wang Xingxing predicted 2027 will likely be the inflection point for the robotics industry, with embodied intelligence approaching its “important breakthrough moment.” Yet experts stressed robots remain separated from truly understanding and manipulating the physical world by their “last few millimeters.”
Background & Interpretation
1. Technical Background and Evolution
China’s Internet Development Report 2026 puts the 2025 embodied-intelligence market at about RMB 915 billion, up 20.4% YoY. The expo launched 1,000-plus new technologies and had 100-plus embodied robots working live—from running and sparring to playing ping-pong and cooking, signaling rapidly rising hardware and model capabilities.
2. Core Drivers and Underlying Mechanisms
The bottleneck is “alignment between model and physical world.” Wang noted top multimodal models already excel at spatial understanding and action planning; the real shortfall is the real-time loop—humans judge a cup’s position, weight and friction by sight and touch and adjust force instantly, while robots must form a closed loop of vision, physical reasoning and motor control. Moqi’s CTO Huang Qingqiu stressed high-quality data over scale: 100k hours of precise manipulation data can beat 10M hours of low quality.
3. Market Structure and Industry Chain Impact
Xiaomi disclosed pretraining on 100k+ hours across 1,700+ scenarios, confirming the “data flywheel” as the new battleground. Synchronized breakthroughs in hardware cadence and model generalization are moving the industry from prototype validation to scale manufacturing—yet fine manipulation (disassembling phones, threading needles) still has low success rates, the key gap before mass production.
Implications & Outlook
Around 2027 the industry may hit a mass-production inflection, but the “last few millimeters” of fine manipulation and real-time loops remain decisive. Watch: the accrual speed of high-quality manipulation datasets, edge deployment cost of VLA models, and scaled real-scenario reliability validation.
Takeaways for Industry Participants
Robot firms should put the “data flywheel” and simulation at the strategic core, not just stack hardware specs; compute/model companies can target lightweight edge-cloud deployment; scenario owners (manufacturing, services, elderly care) should define needs via real task chains and co-build data loops; investment should favor perception-control integration that crosses the “last few millimeters.”
This column compiles industry information and shares technical perspectives; it does not constitute investment advice.