Object-Centric World Models: Representation Quality and Robustness Under Distribution Shift
A recent paper on arXiv (2608.12078) investigates object-centric world models (OCWMs) in the context of visual model-predictive control, focusing on how representation quality and generalization are affected by distribution shifts. The researchers discovered a positive correlation between planning success and unsupervised slot-quality metrics (FG-ARI, mBO), although improvements level off at high slot quality. By comparing OCWMs with scene-centric models, the study demonstrates that OCWMs provide greater robustness when slots are well-defined. It questions the notion that an object-centric approach is always beneficial for planning, emphasizing the necessity for enhanced slot encoders. This work is pertinent to AI and robotics, indicating that better slot quality could improve world model efficacy. The paper has been released as a cross-type submission on arXiv.
Key facts
- Paper ID: arXiv:2608.12078
- Announce Type: cross
- Study focuses on object-centric world models (OCWMs) for visual model-predictive control
- Examines representation quality and generalization under distribution shift
- Planning success correlates positively with FG-ARI and mBO slot-quality metrics
- Gains from slot quality saturate at high levels
- OCWMs compared to scene-centric models
- Well-bound slots improve robustness under distribution shift
Entities
Institutions
- arXiv