The Path to General Physical AI: Insights from the Eighth Intelligent Assisted Driving Conference

On July 22, 2025, at the Eighth Intelligent Assisted Driving Conference, Mao Jiming, partner and vice president of Jiga Technology, proposed that the goal of building a world model is to achieve General Physical AI. He noted that humans possess a biologically common intuitive cognitive layer and unique advanced thinking and reasoning abilities, while current machine intelligence falls short in the foundational intuitive cognition. The world model, serving as the underlying operating system of Physical AI, can bridge this gap by facilitating spatial understanding, perception, and operational capabilities, allowing logical intelligence to act upon the physical world. Mao further elaborated on the dual-driven strategy of the world model platform and the embodied foundational model, emphasizing the joint optimization of the world model and intelligent agents. He stated that the world model not only provides data support for intelligent agents but also continually enhances accuracy and generalization capabilities as the agents evolve. Jiga Technology has designed a hierarchical world model, dividing the world into an appearance layer, a geometry layer, and a physical attributes layer, utilizing neural networks and generative technology to model and render each layer, significantly reducing the Sim2Real gap. Looking ahead, Mao believes the world model will become the core carrier of intelligence, requiring deep integration of cloud and edge world models. He suggested that the cloud-side world model should enhance counterfactual reasoning capabilities, while the edge-side world model needs to address slow reasoning speeds in pursuit of instantaneous reasoning. Additionally, a continuous learning mechanism is crucial for the evolution of the world model, which corrects cognitive biases through real-world observation, marking a key step in model evolution. The overarching goal of building a world model is to enable Physical AI to reach the level of General Physical AI, akin to general artificial intelligence. By exploring suitable knowledge injection methods, the model can effectively integrate physical laws without solely relying on video data for learning. The future of the world model is envisioned as a central component of intelligence, necessitating continued efforts to solve current challenges in reasoning speed and enhance the model's capabilities for real-time applications.