Object Permanence

In AI, object permanence asks whether a system continues to represent an object's existence, identity, and motion when it is occluded or out of view. Approaches include persistent object slots with temporal memory, latent world-model prediction, self-supervision from reappearance, and physics-based trajectory simulation. Researchers test these abilities in synthetic occlusion tasks, video benchmarks such as MOVi and IntPhys, 360-degree video, and autonomous-driving scenes. A central tension is that tracking an object while visible does not ensure it is remembered when hidden; current systems often lose information quickly, and evaluation must distinguish genuine persistence from plausible-looking predictions.