Event

Slot-JEPA proposes a structured object prediction interface

1 reports · 1 sources

Researchers propose the Slot-JEPA method, which decomposes JEPA's dense patch representations into stable semantic slots and time-varying spatial masks; after a single pretraining, it can be frozen and used as a reusable structured object interface. The Transition Memory Predictor (TMP) built on this interface stores visual transitions as compact masks, retrieves context using slot statistics, and supervises mask prediction and candidate plan evaluation with unbalanced optimal transport (UOT). The paper claims that on seven visual understanding benchmarks, slot reconstruction features remain transferable, and that on the REALM-JEPA benchmark's six planar environments and four robotic arm planning tasks, it improves goal-conditioned planning.