The world model is entering a critical transition period towards complex intelligent behavior generation, becoming a key infrastructure for promoting the integration of physical AI with virtual worlds, and helping China to take a leading position in global AI competition. Currently, world models are moving from research and development testing to mass production empowerment in the field of autonomous driving. By generating massive amounts of high-fidelity scenarios, they drive autonomous driving systems to continuously learn, autonomously verify, and rapidly iterate and optimize, facilitating the implementation of L3/L4 systems and significantly reducing the cost and time of real-world vehicle testing. In the field of embodied intelligence, world models serve as synthetic data engines, breaking through the bottleneck of scarce physical interaction data, providing efficient and safe virtual training environments for robots, and accelerating their adaptation to real-world tasks. Both applications highlight the core value of world models in promoting the closed-loop evolution of AI from perception to action through simulation and generation.
This white paper focuses on 'World Models', a cutting-edge artificial intelligence technology, analyzing its current development status, technical path, market pattern, and future trends. World models are generative AI models that understand the dynamics of the real world (including its physical and spatial attributes). They use input data such as text, images, videos, and motion to generate videos. Through learning, they can understand the physical characteristics of real-world environments, thereby representing and predicting dynamics such as motion, stress, and spatial relationships in sensory data, accelerating virtual world generation for physical AI, generating scalable augmented data, thereby eliminating data bottlenecks, and achieving more efficient basic model training. The research purpose of this white paper is to comprehensively sort out the development history, current status, core technologies of world models, and their applications in intelligent driving and embodied intelligence. Through comparative analysis of the capabilities of different manufacturers, it explores the future development trends of world models.
