PICA-WAM: Visual Perturbation Invariance via Cross-Appearance Self-Alignment in World Action Models

Paper recorded by Signals 4 on 2026-10-19 in cs.CV. Abstract reproduced from arXiv; link to the original below.

Published 2026-10-19 on arXiv · recorded by Signals 4 on 2026-10-20

Category: cs.CV · 计算机视觉 · first seen 2026-10-20

Abstract

World Action Models (WAMs) jointly predict future video and generate actions, yet remain sensitive to task-irrelevant visual perturbations, i.e., changes in visual appearance that should not affect robot behavior. We identify a gap in conventional WAM training: video prediction encourages representations to preserve appearance information, while action generation does not explicitly enforce invari

Read on arXiv →