DDP-WM: Disentangled Dynamics Prediction for Efficient World Models

World models are essential for autonomous robotic planning. However, the substantial computational overhead of existing dense Transformer-based models significantly hinders real-time deployment. To address this efficiency-performance bottleneck, we introduce DDP-WM, a novel world model centered on the principle of Disentangled Dynamics Prediction (DDP). We hypothesize that latent state evolution in observed scenes is heterogeneous and can be decomposed into sparse primary dynamics driven by physical interactions and secondary context-driven background updates. DDP-WM realizes this decomposition through an architecture that integrates efficient historical processing with dynamic localization to isolate primary dynamics. By employing a cross-attention mechanism for background updates, the framework optimizes resource allocation and provides a smooth optimization landscape for planners. Extensive experiments demonstrate that DDP-WM achieves superior efficiency and performance across diverse tasks, including navigation, precise tabletop manipulation, and complex deformable or multi-body interactions. Specifically, on the challenging Push-T task, DDP-WM achieves an approximately 9 times inference speedup and improves the MPC success rate from 90% to 98% compared to state-of-the-art dense models.

For robots to act autonomously in the real world, they need an internal "mental model" that can imagine what will happen next when they take an action — much like how humans mentally rehearse before making a move. Current approaches build such models using powerful but computationally expensive neural networks that process every part of a visual scene equally, regardless of whether anything is actually changing there. This makes them far too slow for real-time robotic control. We observe that in most physical interactions, only a small portion of the scene actually changes — for example, when a robot pushes an object, the object moves but the table and background stay mostly the same. Based on this insight, we propose DDP-WM, a system that separates scene changes into two types: the main physical changes (like an object moving) and the subtle ripple effects those changes have on the rest of the scene. By focusing most computation on the small changing regions and handling the rest cheaply, our approach runs approximately 9 times faster than previous methods while actually improving the robot's planning success rate from 90% to 98% on a challenging task. This brings us closer to deploying intelligent planning systems on real robots that need to make decisions quickly and continuously.