ODEWorld: A Continuous Predictive Architecture via Physical-Time Flow
A continuous-time latent world model that learns physical-time dynamics through an ODE-governed latent velocity field.
We introduce Physical-Time Flow, a continuous modeling approach that represents sequential dynamics as a latent velocity field evolving along physical time. Building upon this formulation, ODEWorld performs future prediction through temporal integration in a compact latent space, enabling arbitrary temporal resolution, backward prediction, long-horizon reconstruction, and planning-oriented representations for downstream robotic control.
Overview
Continuous Dynamics
Instead of relying on fixed discrete prediction steps, the model treats temporal evolution as a continuous flow in physical time.
Latent ODE Prediction
Future states are generated by solving an ordinary differential equation in a compressed latent representation space.
Planning-Oriented Space
The learned dynamics-centric latent space provides compact and structured information for downstream policy learning.
Method
Physical-Time Flow
ODEWorld models the time-varying component of sequential observations with a continuous latent velocity field. The transition from one latent state to another is obtained by integrating this velocity field over physical time.
- Dynamical representation decoupling
- Latent velocity field learning
- ODE-based temporal integration
- Continuous and bidirectional prediction
Add pipeline visualization here after anonymization.
Results
Video Prediction
Evaluates visual fidelity and temporal consistency under long-horizon prediction.
Temporal Flexibility
Supports arbitrary timestamp generation beyond fixed-frame discrete prediction.
Robotic Control
Uses learned latent dynamics as planning-oriented representations for downstream control tasks.
Resources
Anonymous Materials
Code, videos, and additional qualitative results will be provided in anonymized form. This page intentionally excludes author names, affiliations, institutional links, and non-anonymous metadata during the review period.