We develop a mathematical theory of closed-loop learning in RNNs, showing how it fundamentally differs from conventional open-loop supervise training.

A 3DGS-based Reinforcement Learning training paradigm for end-to-end autonomous driving