JEPA architectures learn by predicting in latent space. This makes them naturally invariant to pixel-level noise (static, occlusion) and irrelevant visual properties (like color shifting) that don't affect the underlying world physics.