A researcher trained a world model based on Yann LeCun's JEPA architecture to play Pokémon Red, specifically to select a starter Pokémon in Professor Oak's lab. World models learn to predict future game states from current observations and actions, enabling AI agents to plan sequences of button presses to achieve objectives in the game.
This paper introduces temporal straightening, a technique for improving representation learning in latent planning with world models. By using a curvature regularizer to encourage straightened latent trajectories in a JEPA architecture, the method makes Euclidean distance a better proxy for geodesic distance and improves planning stability, achieving higher success rates on goal-reaching tasks.