Editing Object permanence (section)

== In artificial agents ==
It has been shown that artificial intelligent agents can be trained to exhibit object permanence.<ref name=":0">{{Cite journal |last1=Shamsian |first1=Aviv |last2=Kleinfeld |first2=Ofri |last3=Globerson |first3=Amir |last4=Chechik |first4=Gal |date=2020 |title=Learning Object Permanence from Video |url=https://www.ecva.net/papers/eccv_2020/papers_ECCV/papers/123610035.pdf |journal=European Conference on Computer Vision (ECCV)|arxiv=2003.10469 }}</ref><ref name=":1">{{Cite journal |last=Tomakov |first=Pavel |date=2021 |title=Learning to Track with Object Permanence |url=https://openaccess.thecvf.com/content/ICCV2021/papers/Tokmakov_Learning_To_Track_With_Object_Permanence_ICCV_2021_paper.pdf |journal=International Conference in Computer Vision (ICCV)|arxiv=2103.14258 }}</ref> Building such agents revealed an interesting structure. The object permanence task involves several visual and reasoning components, where the most important ones are to detect a visible object, to learn how it moves and to reason about its movement even when it is not visible. Shamsian et al<ref name=":0" /> found that object permanence was achieved when the agent had two separate time-sensitive modules, one that tracks visible objects, and a second that decides "what to track" when one object occludes or contains the target. Object permanence has further been shown to apply to videos "in the wild".<ref name=":1" />