Slot-MPC: Object-Centric World Model for Goal-Conditioned Planning
A new framework called Slot-MPC uses object-centric representations and model predictive control to enable agents to plan actions adaptively at inference time, overcoming the limitations of fixed reactive policies in reinforcement learning.
Key facts
- Slot-MPC is an object-centric world modeling framework.
- It leverages vision encoders to learn slot-based representations encoding individual objects.
- It uses these representations to learn an action-conditioned object-centric dynamics model.
- At inference time, the learned dynamics model enables action planning via MPC.
- It allows agents to adapt to novel situations not seen during training.
- The approach is inspired by human perception of scenes as objects.
- It addresses the limitation of most object-centric world models and RL approaches that learn fixed reactive policies.
- The paper is published on arXiv with ID 2605.14937.
Entities
Institutions
- arXiv