Jürgen Schmidhuber from Technische Universität München outlined an algorithm using self-supervised recurrent neural networks for reinforcement learning and planning in non-stationary environments in his 1990 paper

KLKangwook Lee@Kangwook_Lee

And guess what ;-) "The algorithm may concurrently learn the model and learn to pursue the main goal."

KLKangwook Lee@Kangwook_Lee

World-action models seem to be emerging across different domains, which is super cool! Action model = learn to control World model = learn what control does World-action model = learn both together Robotics: https://arxiv.org/pdf/2602.15922 Terminal agents: ↓