DigiNews

Summary

A detailed exploration of training a world model to play Pokémon using JEPA/LeWorldModel. The post covers how the model learns embeddings from pixel screenshots, how planning is performed with imagined rollouts and the cross-entropy method, and the rollout fine-tuning approach. It also includes data on the training set, results showing the model acquiring a starter in several runs, and references an open-source implementation (lePokeRed).