WRC2026 Live: AI Models Shift from Parameter Specs to Closed-Loop Capabilities & Home Scenario Implementation | 2026 World Robot Conference

On August 20, the World Robot Conference (WRC) was held at the Yizhuang Guochuang Convention and Exhibition Center in Beijing. The event coincided with the listing of Unitree, whose market value exceeded 400 billion yuan, adding to the industry’s growth. This year’s WRC attracted over 300 exhibitors—a 40% surge from last year—with 3,000 robot products on display. Among the new participants were embodied AI startups and real end customers, including 49 central state-owned enterprises that brought 12 real engineering scenarios (power grid, aerospace, mining, metallurgy) to seek robotic solutions. The WRC is divided into four pavilions (A, B, C, D), one more than last year. Visitors start in Pavilion C with humanoid robots, move through complete landing scenarios in Pavilions A and D, and end in Pavilion B, dominated by upstream industrial chain manufacturers. This year’s focus shifted from model specifications to closed-loop capabilities, with companies addressing whether their models can integrate data and tasks into real-world scenarios. Key developments include: - **World Models:** Companies now prioritize world models over Vision-Language-Actors (VLA), enabling robots to predict outcomes before acting. However, integrating VLA and world models remains a challenge, with some companies embedding world models directly into decision-making frameworks (e.g., LuMing Robotics’ Prime R0), while others use explainable reasoning (e.g., Stardust Intelligence’s Lumo-2). - **Closed-Loop Systems:** Companies like LuMing Robotics and Fourier Robotics emphasize iterative feedback loops, where robots collect real-world data to refine models (e.g., NexCore’s skill evolution engine). This contrasts with traditional training methods, where models are perfected in labs before deployment. - **Industrial Implementation:** The industry has adopted a "fast-slow brain" architecture, where small models handle high-frequency tasks while large models manage deep reasoning. This reduces latency and power consumption, improving efficiency in scenarios like food processing. - **Household Scenarios:** The event highlighted growth in household robotics, with companies like Qiansun demonstrating practical tasks (e.g., tidying a living room with voice commands). The boom is driven by world models and the availability of large-scale first-person data, with industry consensus favoring tens of thousands of hours of pre-training. Despite progress, challenges remain: limited real interactive data, lack of standardized training standards, and testing methods that push models to their limits (e.g., Muka Robotics’ capability boundary tests). While embodied models show promise, scaling to large-scale deployment requires overcoming these hurdles.