World models move to the center of robot learning as a comprehensive 2026 survey maps the field
A comprehensive 2026 survey on world models for robot learning marks the shift: rather than training policies on task-specific data, robots increasingly learn inside learned simulators of the world. A model that predicts how the environment will respond to an action is both a multimodal generator and the training ground for control.
A world model is a multimodal prediction engine turned inward. Instead of generating content for a human to watch, it predicts how a scene will change so an agent can plan against it — which makes the same generative machinery that produces video the substrate a robot rehearses in. The survey formalising the field is a sign the approach has consolidated from scattered results into a program.
The practical draw is data efficiency. Real-world robot data is slow and expensive to collect; a world model lets a policy practice thousands of times in a learned simulator for every real trial, which is the only way the data economics of physical AI start to resemble the data economics of language models.
The convergence with generative video is not incidental. The better video-generation models get at predicting realistic dynamics, the better they serve as world models, which is why progress in multimodal generation and progress in robot learning are increasingly the same curve viewed from two directions.
arXiv — World model for robot learning: a comprehensive survey → · NVIDIA — National Robotics Week — latest physical AI research →