“The ability of the policy's LSTM memory to predict environment parameters, such as cube size, showed a positive correlation with the diversity of the training distribution.”