No more showboating. At the ninth World Artificial Intelligence Conference, which opened in Shanghai on 17 July, humanoid robots stopped performing and started working. Past shows leaned on backflips and boxing; this year robots served as both product and staff, guiding visitors and minding unmanned retail points, the first time full-size humanoids were deployed at scale in a major exhibition’s public areas.
AGIBOT: “deployment era, year one”
AGIBOT set the tone with “deployment era, year one”. Its Yuanzheng A3 Ultra, the only humanoid among the ten exhibit treasures, stands 174 centimetres, carries over 50 degrees of freedom and a 38-kilogram arm load, and runs an end-to-end VLA model. Its G2 robot autonomously handled chip tray loading in a semiconductor line; G2 also ran metro security guidance copied from Guangzhou’s real stations. A sub-brand showed four-legged robots with all-terrain and navigation engines now used in power, chemical and campus security patrols. AGIBOT’s GO-2 model is the decision brain; its GE-2 world model topped the 2026 World Arena leaderboard.
D-Robotics: compute as the utility layer
While others built bodies, D-Robotics built the substrate. Its Sunrise S600 is a robotics SoC with 560 TOPS INT8, an 18-core A78AE CPU and 256-bit LPDDR5x at 204.8 GB/s, aimed at cutting robot development from months to weeks via a three-layer toolchain. It already has 20-plus top customers and 100-plus ecosystem partners.
MagicLab, Leju, Qianjue and Qiyuan
MagicLab launched three models: the 180-centimetre MagicBot X1 with 31 active degrees of freedom and 450 N·m joints, the wheeled D1, and the light-industrial MagicDog T1. Its Magic-VLA K02 passed 90 per cent on long-horizon tasks. Leju built a 1:1 factory zone where robots depalletise with no human in the loop, its Kuafu series now over 95 per cent localised. Qianjue targeted the missing sense, touch: its X-TouchMind V1 and TacVerse dataset let a robot authenticate shoes in five seconds with over 99.9999 per cent agreement with human experts. Qiyuan showed the transformable Q1, a wheeled-and-quadruped personal robot with open-source structure.
VLA versus world model
The loudest debate was the robot brain. VLA reads natural language and adapts fast but struggles with physics like friction and deformation. World models simulate physical interaction but read instructions poorly. Most agree the two will run in parallel until real operation data reaches the 100-million-hour scale. AGIBOT hedges with both GO-2 and GE-2 plus a real-robot reinforcement loop. One executive’s test: when nobody mentions “world model” anymore, the field has arrived, just as nobody debates which tech a self-driving car uses.
The “CATL moment” for supply chains
A clear shift: upstream suppliers stepped to the front. Humanoid core-part localisation passed 75 per cent; joints, long near 50 per cent of body cost and foreign-locked, now have domestic modules at million-unit capacity. Linker Hand and Critical Point showed high-DOF hands with hundreds of tactile points. Tashan’s emerald E10A dynamic tactile chip claims microsecond resolution, 0.9 microamps standby, and over 80 per cent of the tactile hardware market.
No firm tries to own the whole chain anymore. Chips, joints, touch and complete machines are now separate specialisms, much like smartphones. The division is not retreat; it is maturity. WAIC 2026 drew the line: humanoids are employed, creating real value, and the scalable commercial window has opened.
Editor’s note: This is an adapted translation of the original OFweek report. It has been trimmed and restructured for readability for an international business audience.