Gasgoo Munich- On September 15, Mao Jiming, partner and vice president at GIGAAI, laid out a systematic vision for general world models at the Gasgoo Embodied Perception Fusion and Multimodal Large Model Innovation Seminar.
Image Credit: Gasgoo
Mao Jiming divides the general world model into two components: world generation and world action. The former predicts changes in the world based on specific actions, while the latter predicts actions based on observations—creating a closed loop between the two.
Regarding the evolution of world models, Mao outlines an L1-to-L5 roadmap. L1 leverages massive video data to learn the diversity of structural changes and is ready for immediate commercialization. L2 incorporates action to master causality. L3 introduces geometry for precise spatial representation, effectively raising the dimension by one level. L4 tackles physics by handling unobservable implicit states and interacting with specific objects and surfaces. Finally, L5 focuses on operation—engaging with the real physical world for continuous learning to refine modeling accuracy and uncover new patterns.
Mao distills recent progress into an "algorithm pyramid" and a "data pyramid." The algorithm pyramid comprises three layers: world simulation, world and action alignment, and experience reinforcement. The data pyramid maps to the classification of diverse embodied data.
Data serves as fuel for algorithm evolution, while algorithms conversely dictate how data is acquired.
Over the next 5 to 10 years, general world models will unleash infinite possibilities. We will not just be witnesses to this shift, but architects of a brand-new world.
