In early August 2026, Wujie Dongli (Beijing) signed a 500 million yuan order with Envision Technology — the first hundred-million-yuan-level overseas order in China's "embodied intelligence manipulation" industry. The first batch of robots destined for a battery gigafactory in France has already departed. Wujie Dongli simultaneously announced 700 million yuan in total global orders, with partners including automotive supply chain giants such as ZF LIFETEC and Omway. How did a company founded less than two years ago land such large international contracts?
Data Infrastructure: Shijingshan and Haidian
The answer lies in two Beijing locations. The Shijingshan Humanoid Robot Data Training Center is among China's first and currently largest data training base for full-size bipedal humanoid robots, with three phases built, producing nearly 20 million high-quality data records per year across close to 700 fine-grained scenarios. It uses three collection channels: first-person perspective (VR headset + controller teleoperation), remote teleoperation (data gloves directly controlling dexterous hands), and wearable haptic sensors. The data feeds into large models for training "perception" and "manipulation precision."
In Haidian's Embodied Intelligence Innovation Industrial Park, Wujie Dongli's pilot platform has established a full closed-loop process from raw material inspection to finished-machine testing. Its second-generation in-house robot, the K15, has also obtained EU CE industrial-grade full-directive certification — the "passport" into European industrial production lines.
The Data Gap and the MWATM Bet
Zhu Kai, general manager of the Shijingshan training center, offered a striking comparison: "Currently, the total amount of real-robot data collected worldwide does not exceed 500,000 hours. However, training a general-purpose model that lets robots adapt to complex environments requires at least tens of millions of hours of high-quality data." That gap — 500,000 hours versus 10 million hours — is the most straightforward supply-demand mismatch in embodied AI data today.
Wujie Dongli's "latent space world model" architecture (MWATM) differs from the industry's common VLA (Vision-Language-Action) route, emphasizing unified representation and reasoning of multimodal information within latent space — a key technical bet on fitting "perception + physics + causality" into a single model. Orders from both energy-sector production lines (power batteries) and the automotive supply chain (chassis and electronics) mean their "general-purpose brain" is receiving real-world feedback from at least two industrial scenarios.
C-End Validation: Swancor and Zhiyuan
A parallel line of consumer validation comes from Swancor. Its H1 2026 report, disclosed on August 14, showed roughly 160 million yuan of R&D spending on consumer embodied intelligence robots, two launched products (Qiyuan Q1 and T1), 210 million yuan in advance payments, an R&D team expanded to 227 people, and 273 cumulative patent applications. Swancor is still at the "business validation" stage, but the financial structure — 160 million in R&D against 210 million in pre-orders — indicates preparation for mass production.
Add Zhiyuan Robotics, which mass-produced 15,000 units in H1 2026 (with humanoid share the highest), and China's embodied intelligence manipulation sector now shows simultaneous leaps across three directions: B-end order-data loops, C-end product development, and manufacturing-side production-line data.
Policy and Standards
Supporting policies landed in sync. The "Humanoid Robot and Embodied Intelligence Standards System (2026 Edition)," released in February 2026, covers the full industrial chain, full technology stack, and all application scenarios. MIIT Deputy Minister Ke Jixin recently revealed that China produced roughly 20,000 humanoid robots last year, over 40,000 in H1 this year, with full-year output projected to exceed 100,000. At the August 2026 CIFTIS fair, Shijingshan will launch a "technology service platform for cross-border circulation of embodied intelligence data elements under a compliance framework" — the interface design for data elements moving from factories to overseas markets. Combined with seven national standards for humanoid robot testing methods already underway, the industry is building international influence from "hardware bodies" to "software brains" to "data elements."
Conclusion
Data, algorithms, orders, C-end products, policy, and cross-border channels — five layers advanced simultaneously in a single summer for Chinese embodied intelligence. Each item alone is industry news; stacked together, it amounts to an "infrastructure week" in which China's embodied intelligence sector, starting from data elements, connected industrial orders, overseas orders, consumer pre-orders, policy legislation, and national standards in one sweep.