First-person human video for pretraining.
3,600 hours of egocentric human video used to pretrain manipulation models.
