World Models Take Over from Language Models: Company Pioneers Physical AGI 'Dual Pyramid' System, Universal Robots Enter the 'Home Era'

On May 20, 2026, at a residential community in Wuhan's Optics Valley, Jijia Vision (Extreme Vision) quietly changed the narrative around humanoid robots. Instead of showcasing backflips or parkour on a brightly lit stage, the company let its robot, Shiguang S1, walk into a real home — where children scatter toys and furniture gets moved around unpredictably.

This launch marked the debut of Jijia Vision's self-developed 'Dual Pyramid' technical framework, a comprehensive system crafted over three years to tackle the fundamental challenges of physical artificial general intelligence (AGI). The framework consists of two pyramids: the Data Pyramid and the Algorithm Pyramid.

The Data Pyramid has five layers: internet video data, human demonstration data, world model simulator, synthetic simulation data, and real robot data. Each layer addresses a specific bottleneck — scale, density, or authenticity — and is supported by proprietary hardware and software, such as the low-cost data-collection device U-01 and the home robot Shiguang S1 itself. This full-stack approach ensures that Jijia Vision controls the entire pipeline from data acquisition to model training.

The Algorithm Pyramid comprises three ascending layers: world simulation, action alignment, and experience reinforcement. On the world simulation layer, the company's GigaWorld-1 model topped the global WorldArena benchmark, surpassing Google and NVIDIA with the first comprehensive score above 60. On the action alignment layer, the GigaBrain-0 series won first place in the RoboChallenge, while GigaWorld-Policy topped the RoboCasa365 benchmark. On the experience reinforcement layer, the company achieved self-evolution of the embodied foundation model by combining the world model with reinforcement learning.

The first product from this pipeline, Shiguang S1, is a wheeled-arm robot designed specifically for household tasks. Its lower body is a wheeled chassis for stability and safety, while the upper body has human-like arms capable of grasping, aligning, folding, and sorting. The robot is powered by the GigaBrain series embodied foundation model, which enables end-to-end perception, understanding, and action.

Jijia Vision announced that Shiguang S1 has already secured orders for 100 units from a real residential community in Wuhan, with deployment starting in the third quarter of 2026. This is one of the largest commitments to household robots globally, surpassing most competitors who are still in pilot phases in industrial settings.

Looking ahead, the company revealed a 12-month roadmap for its foundational models: GigaBrain-1 (Q3 2026), GigaBrain-2, and GigaBrain-3, which will be trained on 10 million hours of video data plus 1 million hours of world-action data, aiming to trigger the 'GPT-3 moment' for physical AGI. The GPT-3 moment refers to the critical point where scaling laws produce emergent capabilities, moving robots from lab demonstrations to true general-purpose utility in any home.

Jijia Vision also has a parallel B-track in industrial manufacturing. Recently, it collaborated with FAW Tooling and Alibaba Cloud to complete the first full-process deployment of an embodied robot in a real factory, compressing the traditionally months-long adaptation cycle to just weeks.

The company's strong team is a key asset. CEO Huang Guan, a Ph.D. candidate at Tsinghua University, previously led robot vision at Horizon Robotics and was algorithm VP at ZhenRobotics. Chief Scientist Zhu Zheng, a Zhiyuan Young Scholar with over 70 top-tier papers and nearly 20,000 citations, has won multiple prestigious awards. Co-founder Sun Shaoyan, former director at Alibaba Cloud, led the first data-closed-loop system for autonomous driving. The engineering VP Mao Jiming has over 16 years of experience in simulation and distributed architecture, including leading the simulation and engineering team at Baidu Apollo.

In terms of funding, Jijia Vision completed two consecutive rounds in March and April 2026, totaling approximately 2.5 billion yuan (about $350 million), making it the first world-model unicorn valued at over 10 billion yuan in China. Backers include Huawei's Hubble investment arm, top financial institutions, and state-backed platforms.

As the company moves forward, three key questions will be watched: Can the 100-unit home deployment generate a truly effective data closed-loop? Will GigaBrain-1 deliver on the promise of the Dual Pyramid system in Q3 2026? And will GigaBrain-3 indeed reach the physical AGI 'GPT-3 moment' within 12 months? Jijia Vision has laid out a clear, testable path to answer these questions, one household at a time.