Building a Practical Taxonomy for AI World Models

Short summary

A report proposes a practical taxonomy for AI world models, covering domains from RL and robotics to video generation and autonomous driving. Instead of a single global score, the framework uses fields like domain, input/output modalities, action-conditioning, representation type, temporal horizon, and evaluation type to differentiate models. The key insight is that not every generative model qualifies as a world model—usefulness for acting inside an environment matters more than visual realism.

• •Proposes a multi-field taxonomy instead of a single ranking score for world models
• •Covers domains: RL, robotics, embodied AI, video, autonomous driving, games, industrial simulation, spatial intelligence, software agents
• •Distinguishes world models from generative models by emphasizing action-conditioning and functional evaluation

Generated with AI, which can make mistakes.