Understanding Representations Pretrained with Auxiliary Losses for Embodied Agent Planning