World Action Models (WAMs) are designed to jointly predict future video and generate actions, but they remain sensitive to task-irrelevant visual perturbations—changes in appearance that should not alter robot behavior. This sensitivity can lead to inconsistent action outputs when the same scene appears under different lighting, textures, or other surface-level variations.
The paper introduces PICA-WAM, which stands for Visual Perturbation Invariance via Cross-Appearance Self-Alignment. The core idea is to align representations across different appearances so that the model focuses on task-relevant structure rather than superficial visual differences. This is intended to make action generation more robust to appearance shifts.
The work is posted on arXiv and targets a known weakness in current WAMs. By explicitly addressing visual perturbation invariance, PICA-WAM offers a potential path toward more reliable world models for robot decision-making in varied visual environments.