Galbot joins hands with CSCEC to build a spatial intelligence dataset, winning the national first prize at the 2026 "Data Elements ×" Competition

On October 23, the award ceremony for the national finals of the 2026 "Data Elements ×" competition, jointly organized and implemented by 20 departments including the National Data Administration, the Cyberspace Administration of China, and the Ministry of Industry and Information Technology, was held in Qingdao, Shandong. The Xingzhu Qingchuan Cloud Alliance, composed of China Southwest Architectural Design and Research Institute, Beijing Galbot, Tsinghua University, the Sichuan Provincial Construction Science and Technology Development and Information Center, and China Electronics Cloud Computing Technology Co., Ltd., won the national first prize in the Science and Technology Innovation track with its project "Physical AI Data Foundation — High-Quality Spatial Intelligence Dataset for Embodied Intelligence Training and Building Industry Agent Development." The "Data Elements ×" competition is the highest-level competition platform for implementing the "15th Five-Year Plan" deployment and the "Data Elements ×" Three-Year Action Plan (2024–2026), with participating teams covering 31 provinces (autonomous regions and municipalities directly under the central government) across the country and the Xinjiang Production and Construction Corps, with more than 70,000 teams participating in total. Winning the highest national-level award in the Science and Technology Innovation track this time further demonstrates the team's technical strength and industry-leading position in the field of data infrastructure for embodied intelligence training and building industry agent development. Zhang Zhizheng, co-founder and head of large models at Galbot, said that compared with overseas efforts that focus more on the visual consistency of video data, Galbot focuses more on the physical realism of simulation to support embodied intelligence training and real-world scenario deployment. In the future, Galbot will work with China Southwest Architectural Design and Research Institute through engineering data procurement and data governance cooperation to jointly promote the transformation of building industry data into high-quality physical AI training data. This year's competition reached a new high in popularity, with a total of 29,307 teams registering nationwide and more than 145,000 people participating, and ultimately 330 teams entered the offline final review of the national finals. Among them, the Science and Technology Innovation track was the most competitive, with a total of 19 projects entering the finals, where world-leading research forces such as the Chinese Academy of Sciences, the University of Science and Technology of China, and Tencent Technology competed alongside top enterprises. The Xingzhu Qingchuan Cloud Alliance, in which Galbot participated, stood out in the fierce competition and won the national first prize, fully demonstrating that the company has leading technical strength in embodied intelligence data infrastructure and collaborative model research and development, and has received authoritative recognition from a national-level competition platform. Experience transformation: building a leading "textbook for understanding the physical world" At present, artificial intelligence is accelerating its move from the digital world to the physical world, and embodied intelligence has become an important technical direction for AI to empower the real economy and promote the transformation and upgrading of manufacturing. Unlike large models that mainly process digital information such as text and images, embodied intelligence requires robots to perceive the real environment in real time, understand spatial relationships, judge physical constraints, and complete autonomous decision-making and actual operations accordingly. This requires not only improvements in algorithmic capabilities, but also high-quality data that can accurately describe the physical world and support model learning and action. However, the massive digital information accumulated on the internet can hardly directly meet the needs of robots to understand and operate in real environments. Especially in complex scenarios such as building spaces, spatial structures, geometric relationships, and engineering knowledge have long been accumulated in design drawings, engineering materials, and professional practice, and there is still a lack of systematic data representation for physical AI training. How to transform spatial information in the real physical world into high-quality data that machines can understand and models can learn from has become an important topic in promoting embodied intelligence toward complex real-world scenarios. The high-quality spatial intelligence dataset project in which Galbot participated is precisely aimed at this industry pain point. The dataset transforms the engineering experience accumulated over thousands of years in the building industry and the professional skills of more than 7,000 top engineers at China Southwest Architectural Design and Research Institute into "high-quality textbooks" that physical AI can understand, providing data support for models to understand three-dimensional spatial structures, identify object relationships, and carry out spatial reasoning and action decision-making, and filling the shortcomings of physical AI training data at scale. At the same time, the construction of the dataset will also promote the transformation of knowledge accumulated in engineering archives and professional practice in the building industry into developable and utilizable data assets, opening up a new direction for the development and application of spatial intelligence data in the traditional building industry. This exploration not only provides a new data foundation for the development of physical AI, but also injects new momentum for intelligent transformation into the traditional building industry. AstraData: consolidating the physical AI data foundation with standards and closed-loop validation In this project, Galbot uses real embodied intelligence operation scenarios to reversely define the physical validity standards of spatial intelligence data, and, relying on its self-developed "AstraBrain" embodied large model, carries out full-chain validation from building space data to robot perception, decision-making, and execution, promoting the upgrade of the dataset from "capable of describing space" to "capable of driving embodied behavior" as physical AI training infrastructure. This ability to define data standards starting from real operation needs and to verify the application value of data through real-machine operation stems from Galbot's deep accumulation in embodied intelligence data infrastructure and real-world scenario deployment. The company's self-built tens-of-billions-level embodied intelligence data foundation "AstraData" covers a five-layer data system of internet data, human behavior data, synthetic data, real-machine teleoperation data, and real-machine operation data, providing solid data support for embodied large model training and the improvement of generalization capabilities in complex scenarios. On this basis, the company further applies "AstraBrain" to real business scenarios, promoting the coordinated progress of data construction and model research and development, and forming an iterative closed loop of "data—model—robot—scenario." In the field of industrial manufacturing, Galbot S1 has entered the real production lines of global manufacturing giants such as CATL and Siemens to carry out autonomous operations and conduct normalized autonomous operation; in the field of smart city services, Galbot space capsules have been deployed at 200 sites in nearly 50 cities across the country; in the field of instant retail, smart pharmacies have been deployed in more than 30 cities across the country, with nearly 100 locations operating normally 24/7. The real-machine operation data accumulated in these real scenarios is conducive to providing improvement directions for continuous model iteration and data system optimization, and also provides rich practical experience for the construction and application validation of this high-quality spatial intelligence dataset. Cross-sector collaboration: continuously expanding the boundaries of industrial application of embodied intelligence This award is not only recognition of the project's technological innovation and industrial application value, but also provides a new practical example for cross-domain collaboration between the building industry and the embodied intelligence industry. (At the award ceremony, the team winning the national first prize in the Science and Technology Innovation track came on stage to receive the award) With the continuous accumulation and open application of spatial intelligence data resources, building spaces are expected to become an important training carrier for physical AI to understand three-dimensional environments, learn spatial relationships, and explore autonomous action, providing a new data foundation for embodied intelligence to move from single-task execution to general perception, reasoning, and operation in complex environments, thereby promoting robots to move toward the broader real world. For Galbot, participating in this project is an important practice for the company to expand its embodied intelligence data infrastructure capabilities deeply into the building space field, and is also a key step in promoting the collaborative application of cross-industry data resources and embodied large models and accumulating reusable experience, providing the industry with a replicable paradigm for cross-industry data collaboration. In the future, the company will continue to deepen collaborative innovation in data infrastructure, model research and development, and real-scenario applications, open up the value closed loop from data supply and model capabilities to industrial deployment, and promote the large-scale deployment and sustainable value creation of embodied intelligence technology in more industrial scenarios.