From Prediction to Control: A Survey of Closed-Loop World Models for Robot Learning

World models are now used throughout robot learning: to shape policy representations, generate training data, score behavior, plan actions and correct execution. Yet the field has not converged on what such a model should predict, what it should be conditioned on, what data it should learn from, or how it should compute. We argue that this lack of convergence is expected. A world model helps a robot only through the component that uses its predictions, such as a learner, a planner, an evaluator or a controller, so the quality of a model is relative to that consumer: the same predictive error can be harmless when generating offline training data and decisive when a planner ranks actions under a control deadline. Starting from a first-order account of how model errors reach the robot’s real-world return, we organize the literature along the path from prediction to control. On the model side, we characterize world models by four design dials (prediction space, conditioning, knowledge source and computation), each embodying a fundamental trade-off. On the consumer side, we identify five pathways through which predictions change behavior: as features, as generated experience, as scores, as decision-time futures of candidate actions, and as residuals between prediction and observation. We then relate these pathways to the learning regimes and policies that consume them, to the requirements of different embodiments, and to evaluation protocols that measure utility on real robots. Finally, we review how deployment data close the outer loop by refitting models and policies, and we cast open problems as unresolved trade-offs on the design dials. We maintain a companion repository that organizes the surveyed works by section.