At the 2026 World Robot Conference (WRC), held from August 19 to 23 at the Beijing Yichuang International Convention and Exhibition Center, a special session titled "AI Large Models Empowering Robots and Embodied Intelligence Industry New Paradigms" took place. Xiao Zhongyang, founder, chairman, and CEO of Corona Genesis, attended and delivered a speech.
Xiao Zhongyang stated that Corona Genesis breaks down the scaling problem of embodied foundation models into three levels: the learning paradigm, the representation system, and the closed loop of data and feedback. Regarding "how to learn," he pointed out that industry consensus is converging on world models, which learn by predicting the next state. The core logic is to determine the current state of the world, what actions to take, and what will happen next. He believes the efficiency of this physical intelligence learning paradigm stems from two key factors: the ability to construct large amounts of unsupervised data, and the fact that predictive capability relies on an accurate understanding of underlying physical laws.
On the question of "what to learn," Xiao Zhongyang explained that sensors capture raw observations, and the model needs to encode this complex information into comprehensible internal states to further complete prediction, reasoning, decision-making, and action. Therefore, representation must answer the question of how the model internally understands the world. He argued that paradigm shifts in AI foundation models are often accompanied by innovations in representation methods. The field of computer vision has evolved from handcrafted features to learnable representations, the breakthrough of large language models is built on token and contextual representations, and autonomous driving paradigms have gradually converged from manual rules and features to unified spatial representations like BEV and Occupancy.
Addressing "how to continuously learn," Xiao Zhongyang emphasized the significant differences between physical world AI and digital world AI. Data for digital world AI is relatively easy to obtain; text, images, videos, and code can be reused repeatedly in data centers. In contrast, training data for physical world AI must be authentic, derived from real interactions with the physical world, including long-tail issues such as real lighting, real materials, machining tolerances, friction, sensor noise, and degradation. He noted that digital AI might be created in a laboratory, but physical AI cannot be separated from real-world scenarios. The GPT moment for embodied intelligence cannot be "engineered in isolation" in a lab. If a physical foundation model company's strategy is to first complete model training and then deliver products, it will struggle to continuously obtain high-quality interaction data.
In concluding his presentation, Xiao Zhongyang summarized that Corona Genesis revolves around three questions: how to learn, with the answer being next-state prediction; what to learn, with the answer being embodied-native representation E3; and how to continuously learn, with the answer being that true scaling of physical foundation models will increasingly occur through real-world interactions, where product deployment itself becomes part of the model's learning closed loop. He stated that Corona Genesis's mission is to build a learning closed loop, enabling physical AI to truly achieve scaling.
Sina statement: All conference transcripts are compiled from on-site shorthand and have not been reviewed by the speakers. Sina's publication of this article is for the purpose of disseminating more information and does not imply agreement with its views or confirmation of its descriptions.