Research map / Probabilistic and scientific modeling / Dynamical systems
Time-Aware World Model for Adaptive Prediction and Control
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 TAWM, a TD-MPC2-based world model conditioned on the time-step size Δt and trained on log-uniformly sampled Δt, so one model handles varying observation rates without extra samples or training steps.
Problem
World models are trained at a single fixed time-step size, so their performance degrades when deployed at different observation rates; training at multiple rates also raises questions about sample efficiency.
Approach
Condition the latent dynamics, reward, value and policy on the time-step size Δt, model the latent derivative with Euler or RK4 integration scaled by a log-based τ(Δt), and train on a mixture of Δt values sampled per episode (log-uniform or uniform) at the same training budget as the baseline.
Key results
- TAWM outperforms fixed-Δt TD-MPC2 baselines across evaluation Δt values on Meta-World tasks, while matching or exceeding baseline success at the default Δt = 2.5 ms.
- Baselines trained only at Δt ≥ 10 ms fail on all Meta-World tasks, whereas TAWM trained on a Δt mixture consistently surpasses any single-Δt baseline.
- On Meta-World Faucet-Open, TAWM keeps about 90% success at Δt = 50 ms, while MTS3 degrades rapidly as evaluation Δt increases.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add Δt conditioning and mixture-of-time-step training instead of training only at a fixed Δt | TD-MPC2 world model · Meta-World control tasks across varying inference Δt | success rate | Helped |
| Add Δt conditioning and mixture-of-time-step training instead of training only at a fixed Δt | TD-MPC2 world model · PDE-control tasks across varying inference Δt | episode reward | Helped |
| Train a non-time-aware baseline only at a fixed Δt ≥ 10 ms | TD-MPC2 baseline · Meta-World tasks | success rate: zero success | Hurt |
| Use Euler integration instead of RK4 (ablation) | TAWM · Most Meta-World tasks | success rate | Helped |
| Use RK4 integration instead of Euler (ablation) | TAWM · PDE-control tasks | episode reward | Helped |
| Use uniform rather than log-uniform Δt sampling (ablation) | TAWM with RK4 · Meta-World tasks at inference Δt ≥ 30 ms | success rate | Helped |
| Use uniform rather than log-uniform Δt sampling (ablation) | TAWM with RK4 · Meta-World Assembly at smaller inference Δt | success rate | Hurt |
| Use uniform rather than log-uniform Δt sampling (ablation) | TAWM with RK4 · Meta-World Lever Pull at smaller inference Δt | success rate | Hurt |
| Replace the TAWM dynamics model with MTS3 | TAWM control setup · Meta-World tasks across different evaluation Δt | success rate | Hurt |
Surprising findings
- Euler integration outperforms RK4 on the majority of Meta-World tasks, suggesting RK4 is not uniformly beneficial.
- Uniform Δt sampling performs better than log-uniform in many environments, especially at large evaluation Δt, though it is worse at small Δt in some tasks.
Limitations
- The upper bound Δt_max is chosen empirically, with no systematic method for the highest task frequency.
- The dynamics model is deterministic rather than probabilistic.
- Theoretical lemmas rely on strong assumptions such as shared interpolation learning and model capacity.
Experimental setup
Datasets: Meta-World control-gym PDE environments (Burgers, Allen-Cahn, Wave)
Benchmarks: Meta-World control-gym PDE control
Metrics: success rate episode reward (LQ error)
Baselines: TD-MPC2 MTS3
Builds on: TD-MPC2 MTS3 Runge-Kutta integration MPPI
Closest papers by meaning
- PWM: Policy Learning with Multi-Task World Models (ICLR 2025 · similarity 0.87)
- Trajectory World Models for Heterogeneous Environments (ICML 2025 · similarity 0.86)
- World Models as Reference Trajectories for Rapid Motor Adaptation (NeurIPS 2025 · similarity 0.86)
- Learning Hybrid Dynamics Models with Simulator-Informed Latent States (AAAI 2024 · similarity 0.85)
- From Forecasting to Planning: Policy World Model for Collaborative State-Action Prediction (NeurIPS 2025 · similarity 0.85)
- Overcoming Slow Decision Frequencies in Continuous Control: Model-Based Sequence Reinforcement Learning for Model-Free Control (ICLR 2025 · similarity 0.84)
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.