Robotics AI has a terminology problem, and a new tutorial aims to fix it.
Researchers have published a concise tutorial that draws a clearer line around what "world models" actually mean in the context of embodied intelligence and generative simulation. The paper defines world models as action-conditioned predictive models — systems that estimate how a scene or task state will evolve given a sequence of actions. It splits existing approaches into two camps: observation-space models, which work directly with visual data, and state-space models, which operate on more abstract representations of the environment. Each trade-off involves visual fidelity on one side and physical interpretability on the other.
The tutorial goes further by introducing the term "world action models" — a layer that bridges predicted futures with actual executable robot commands. Four paradigms are outlined: imagine-then-execute, video-feature-conditioned action prediction, joint video-action modeling, and using auxiliary video prediction to improve policy learning. That last one is worth watching: training a robot to predict video as a side task has emerged as a surprisingly effective way to learn better control policies without additional labeled data.
The robotics field is crowded with overlapping vocabulary borrowed from reinforcement learning, computer vision, and large language model research, and that ambiguity has real costs — teams talking past each other, methods compared on incompatible assumptions. This tutorial does the unglamorous but necessary work of taxonomic cleanup, the kind of paper that rarely gets press but quietly becomes the citation everyone uses.