When Object-Centric World Models Meet Policy Learning: From Pixels to Policies, and Where It Breaks

2025/11/08 by Stefano Ferraro, Ferraro, Stefano, Akihiro Nakano +5
Computer Science · #Artificial Intelligence (cs.AI) #Domain Adaptation and Few-Shot Learning #FOS: Computer and information sciences #Generative Adversarial Networks and Image Synthesis #Reinforcement Learning in Robotics

paper · pdf · doi:10.48550/arxiv.2511.06136

openalex publication_date 2025/11/08 · openalex created_date 2025/11/12 · openalex updated_date 2026/07/28

Abstract

Object-centric world models (OCWM) aim to decompose visual scenes into object-level representations, providing structured abstractions that could improve compositional generalization and data efficiency in reinforcement learning. We hypothesize that explicitly disentangled object-level representations, by localizing task-relevant information, can enhance policy performance across novel feature combinations. To test this hypothesis, we introduce DLPWM, a fully unsupervised, disentangled object-centric world model that learns object-level latents directly from pixels. DLPWM achieves strong reconstruction and prediction performance, including robustness to several out-of-distribution (OOD) visual variations. However, when used for downstream model-based control, policies trained on DLPWM latents underperform compared to DreamerV3. Through latent-trajectory analyses, we identify representation shift during multi-object interactions as a key driver of unstable policy learning. Our results suggest that, although object-centric perception supports robust visual modeling, achieving stable control requires mitigating latent drift.

Citations

• Object-Centric World Models from Few-Shot Annotations for Sample-Efficient Reinforcement Learning
• SOLD: Slot Object-Centric Latent Dynamics Models for Relational Manipulation Learning from Pixels
• Masked Generative Priors Improve World Models Sequence Modelling Capabilities
• Representing Positional Information in Generative World Models for Object Manipulation
• STORM: Efficient Stochastic Transformer based World Models for Reinforcement Learning
• DDLP: Unsupervised Object-Centric Video Prediction with Deep Dynamic Latent Particles
• Transformer-based World Models Are Happy With 100k Interactions
• An Investigation into Pre-Training Object-Centric Representations for Reinforcement Learning
• SlotFormer: Unsupervised Visual Dynamics Simulation with Object-Centric Models
• Transformers are Sample-Efficient World Models
• TransDreamer: Reinforcement Learning with Transformer World Models
• Conditional Object-Centric Learning from Video
• Generalization and Robustness Implications in Object-Centric Learning
• On the Binding Problem in Artificial Neural Networks
• Mastering Atari with Discrete World Models
• Improving Generative Imagination in Object-Centric World Models
• Multi-Stage Learning for Grasp-Constrained Object Manipulation with a Simulated Panda Robot
• Object-Centric Learning with Slot Attention
• Dream to Control: Learning Behaviors by Latent Imagination
• Entity Abstraction in Visual Model-Based Reinforcement Learning
• SCALOR: Generative World Models with Scalable Object Representations
• MONet: Unsupervised Scene Decomposition and Representation
• Learning Latent Dynamics for Planning from Pixels

Related

• Learning Interactive World Model for Object-Centric Reinforcement Learning
• Object-Centric World Models from Few-Shot Annotations for Sample-Efficient Reinforcement Learning
• Relational Object-Centric Actor-Critic
• Efficient Exploration and Discriminative World Model Learning with an Object-Centric Abstraction
• SOLD: Slot Object-Centric Latent Dynamics Models for Relational Manipulation Learning from Pixels