Tool Transformation And Validation Pipeline

via “data transformation and task augmentation pipeline”

Generalist robot policy model from Open X-Embodiment.

Unique: Implements a composable data transformation pipeline that applies observation normalization, image augmentation, and task augmentation (language paraphrasing, goal image transformations) on-the-fly during training. Transformations are applied in a configurable order, enabling efficient augmentation without storing augmented data.

vs others: More efficient than offline augmentation by applying transformations during data loading, and more flexible than fixed augmentation strategies by supporting composition of multiple transformation types (image, language, action space).