At the 2026 World Robot Conference forum, held in Beijing from August 19 to 22, Hu Luhui, founder and CEO of ZhiCheng AI, stated that world models have two primary application directions: digital spaces and the physical world. In the digital realm, world models can enable 3D spatial cognition and interactive Q&A, representing a deeper capability upgrade over language and multimodal large models. Current multimodal models still struggle to genuinely grasp the laws of physical operation, which is precisely the focus of world models.

The other major direction lies in the real physical world, specifically through integration with humanoid robots. Robots are among the best carriers for closing the AI physical loop and achieving real-world deployment. The concept of "understanding the physical world" may sound abstract, but it can be broken down into three layers: cognition of an object's physical properties, understanding of dynamic environmental changes, and task-oriented behavioral planning capabilities. From this perspective, world models serve as an essential underlying infrastructure for humanoid robots. For robots to observe their surroundings and generate next-step actions, they must rely on physical rules, dynamic change patterns, and causal reasoning to make judgments. World models have the potential to address the shortcomings that reinforcement learning and VRA-based approaches face in humanoid robot tasks.

Hu noted that many people assume industrial and specialized scenarios are easier to deploy, while home environments present higher complexity. However, in terms of the fundamental difficulty of AI physical understanding and causal reasoning, no scenario is inherently easier or harder than another. The truly distinctive challenge of home use cases lies not in the action tasks themselves, but in the combined constraints of safety, privacy, and human-machine coexistence. If world models represent the AI 3.0 stage, then one of the core topics AI 4.0 must tackle is safety, privacy, and human-machine cohabitation. The hefty privacy fine imposed on Meta by the EU in the past offers valuable insight: relying solely on regulations and rules is far from sufficient. Solutions must be built at the foundational AI technology layer. To drive the truly large-scale deployment of world models, technical safeguards for security and privacy are indispensable.