“The size of the policy network was found to be important; halving the model size for the tabletop manipulation task prevented it from handling all the holdout tasks.”