Robust object-centric decomposition in complex real-world scenes

Develop robust object-centric decomposition methods for richer language instructions and more diverse, visually complex real-world scenes.

Background

SlotDiT uses a fixed slot-based object-centric representation to model task-relevant scene dynamics for text-guided video generation and robot control. The experiments cover simulated and real-world robotic datasets, but the authors identify a remaining challenge when scenes become more diverse and visually complex and when language instructions become richer.

The unresolved issue concerns whether object-centric decomposition can reliably identify and maintain meaningful object representations under the increased visual and linguistic variability of such settings. Addressing this challenge would support scaling SlotDiT and related object-centric diffusion models beyond the comparatively structured environments evaluated in the paper.