State Representation Learning

State representation learning asks how an agent or model can turn observations and histories into an internal state that preserves what matters for future prediction or action. Methods include predictive objectives such as JEPA and successor representations, inverse-dynamics and bisimulation learning, recurrent belief states, and task- or reward-driven abstractions. They are studied in partially observed control, pixel-based robot planning, navigation, and dynamical-system forecasting, where useful states can improve sample efficiency, transfer, and robustness. A central tension is that predictive accuracy or reconstruction does not guarantee a decision-sufficient state: representations must retain control-relevant distinctions without carrying irrelevant detail, and their adequacy remains difficult to measure.