Tsinghua & UC Berkeley Introduce ODEWorld, a Continuous-Time Embodied World Model

新智元 · wechat · 2026-08-05

A joint research team from Tsinghua AIR and UC Berkeley BAIR introduced ODEWorld, the first continuous-time embodied world model. While traditional visual world models rely on discrete frame prediction and struggle with transient physical changes between frames, ODEWorld directly learns the continuous rate of change in real physical time.

Core Technical Designs:

Static-Dynamic Decoupling: Uses a frozen DINOv2 encoder to separate static backgrounds from dynamic changes. Dynamic latent variables are aggressively compressed into a single 1×768 token, avoiding redundant calculations of background information.
First-Order Velocity Supervision: Directly supervises the first-order time derivative of the latent state via Jacobian-vector products (JVP), effectively preventing representation collapse and maintaining a high effective rank in the latent space.
Savitzky–Golay Filter: Applied to smooth high-frequency jitter in adjacent frame features, extracting stable physical velocities for model training.

Capabilities & Performance:

By integrating using ODE solvers, ODEWorld unifies previously independent capabilities—arbitrary frame rate generation, time completion (frame interpolation), and backward generation—into a single continuous-time prediction paradigm (PT-Flow).

In LIBERO video prediction experiments, ODEWorld not only outperformed LDP and V-JEPA2 in 64-frame long-horizon prediction quality (PSNR/LPIPS) but also demonstrated extreme efficiency: generating 64 frames took only 0.072 seconds, roughly 1/55th the time of LDP and 1/8.6th that of V-JEPA2.

Original post →

Tsinghua & UC Berkeley Introduce ODEWorld, a Continuous-Time Embodied World Model

More from Embodied