Research map / Probabilistic and scientific modeling / Dynamical systems
ODEWorld: A Continuous Predictive Architecture via Physical-Time Flow
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes PT-Flow, a continuous latent ODE world model trained with direct first-order (JVP-based) velocity supervision in physical time, instantiated as ODEWorld over frozen DINOv2 features for video generation and robot policy guidance.
Problem
Existing world models are discrete-time, struggle with irregularly sampled data, are costly at pixel level, and JEPA-style latent predictors suffer representation collapse.
Approach
Model the latent state as a velocity field along physical time with an initial-state-conditioned encoder/decoder (dynamical representation decoupling), and directly supervise the latent velocity with JVP-projected ground-truth velocities from Savitzky-Golay derivative estimates instead of consistency losses, enabling ODE-solver prediction at arbitrary time resolution and reverse direction.
Key results
- On video prediction over 64 frames, ODEWorld reaches 19.46 PSNR and 0.134 LPIPS at 0.072s latency, versus V-JEPA 2 at 16.47 PSNR and 0.619s, and LDP at 16.27 PSNR and 3.953s.
- On LIBERO-LONG (10 tasks, 3 seeds), the sequential-subgoal ODEWorld variant averages 83.6% success, above VPP (81.0%), Seer (78.6%), SuSIE (76.3%) and GLCBC (78.0%).
- In real-world bi-manual experiments, ODEWorld guidance raises X-VLA average success from 55% to 80% across four tasks.
Surprising findings
- Removing stop-gradient on the target velocity improves PSNR/LPIPS slightly (21.18/0.107 vs 20.53/0.109 short horizon), and the authors report PT-Flow remains effective without it.
- Without dynamical representation decoupling, PSNR is marginally higher (20.65 vs 20.53 short horizon) even though LPIPS and FPS are worse.
Limitations
- Small scale: trained on only two robotics datasets with limited compute.
- Uses a frozen DINOv2 encoder not designed for temporal consistency, requiring a Savitzky-Golay filter to smooth velocity targets.
- No action conditioning in the current version, so it focuses on video generation rather than control.
Experimental setup
Datasets: LIBERO AgiBot-World (AgiBotWorldChallenge-2025 subset)
Benchmarks: LIBERO LIBERO-LONG AgiBot-World
Metrics: PSNR LPIPS latency success rate effective rank (RankMe)
Baselines: LDP V-JEPA 2 GLCBC SuSIE Seer VPP X-VLA
Builds on: flow matching Joint Embedding Predictive Architecture RAE Savitzky-Golay filter FiLM
Closest papers by meaning
- Flow Equivariant World Models: Structured Memory for Dynamic Environments (ICML 2026 · similarity 0.89)
- Space and time continuous physics simulation from partial observations (ICLR 2024 · similarity 0.87)
- CALM-PDE: Continuous and Adaptive Convolutions for Latent Space Modeling of Time-dependent PDEs (NeurIPS 2025 · similarity 0.86)
- Generating Physical Dynamics under Priors (ICLR 2025 · similarity 0.86)
- Breaking the Discretization Barrier of Continuous Physics Simulation Learning (NeurIPS 2025 · similarity 0.86)
- StateSpaceDiffuser: Bringing Long Context to Diffusion World Models (NeurIPS 2025 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.