• Classified by Topic • Classified by Publication Type • Sorted by Date • Sorted by First Author Last Name • Classified by Funding Source •
Dyn-O: Building Structured World Models with Object-Centric Representations.
Zizhao
Wang, Kaixin Wang, Li Zhao, Peter Stone, and Jiang Bian.
In Annual
Conference on Neural Information Processing Systems, December 2025.
World models aim to capture the dynamics of the environment, enabling agents topredict and plan for future states. In most scenarios of interest, the dynamicsare highly centered on interactions among objects within the environment. Thismotivates the development of world models that operate on object-centric ratherthan monolithic representations, with the goal of more effectively capturingenvironment dynamics and enhancing compositional generalization. However, thedevelopment of object-centric world models has largely been explored inenvironments with limited visual complexity (such as basic geometries). Itremains underexplored whether such models can be effective in more challengingsettings. In this paper, we fill this gap by introducing Dyn-O, an enhancedstructured world model built upon object-centric representations. Compared toprior work in object-centric representations, Dyn-O improves in both learningrepresentations and modeling dynamics. On the challenging Procgen games, wedemonstrate that our method can learn object-centric world models directly frompixel observations, outperforming DreamerV3 in rollout prediction accuracy.Furthermore, by decoupling object centric features into dynamic-agnostic anddynamic-aware components, we enable finer-grained manipulation of these featuresand generate more diverse imagined trajectories. The code of Dyn-O can be foundat: https://github.com/wangzizhao/dyn-O.
@InProceedings{dyno_neurips2025,
author = {Zizhao Wang and Kaixin Wang and Li Zhao and Peter Stone and Jiang Bian},
title = {Dyn-O: Building Structured World Models with Object-Centric Representations},
booktitle = {Annual Conference on Neural Information Processing Systems},
year = {2025},
month = {December},
location = {San Diego, United States},
abstract = {World models aim to capture the dynamics of the environment, enabling agents to
predict and plan for future states. In most scenarios of interest, the dynamics
are highly centered on interactions among objects within the environment. This
motivates the development of world models that operate on object-centric rather
than monolithic representations, with the goal of more effectively capturing
environment dynamics and enhancing compositional generalization. However, the
development of object-centric world models has largely been explored in
environments with limited visual complexity (such as basic geometries). It
remains underexplored whether such models can be effective in more challenging
settings. In this paper, we fill this gap by introducing Dyn-O, an enhanced
structured world model built upon object-centric representations. Compared to
prior work in object-centric representations, Dyn-O improves in both learning
representations and modeling dynamics. On the challenging Procgen games, we
demonstrate that our method can learn object-centric world models directly from
pixel observations, outperforming DreamerV3 in rollout prediction accuracy.
Furthermore, by decoupling object centric features into dynamic-agnostic and
dynamic-aware components, we enable finer-grained manipulation of these features
and generate more diverse imagined trajectories. The code of Dyn-O can be found
at: https://github.com/wangzizhao/dyn-O.
},
}
Generated by bib2html.pl (written by Patrick Riley ) on Thu Oct 23, 2025 16:14:14