Optimal data augmentation for out-of-distribution generalization in embodied tasks

Determine the optimal data augmentation strategy for a specified embodied perception-cognition-action task to maximize robustness and generalization under out-of-distribution shifts.

Background

Within the discussion of limitations in current perception-cognition-action systems, the authors note that out-of-distribution shifts significantly degrade performance and that data augmentation is a common mitigation approach.

However, it remains unresolved which augmentation strategies are best for a given task, motivating a focused investigation into principled selection or optimization of augmentations to improve real-world robustness.