“The combination of a policy with memory, such as an LSTM, and a diverse training distribution can lead to emergent meta-learning, enabling rapid adaptation to unknown dynamics.”