Research map / Probabilistic and scientific modeling / Dynamical systems
K-PWM: Control-Oriented Structured World Models under Partial Observation
Santosh M Rajkumar, Sriram Narayanan, Samuel E Otto, Debdipta Goswami
Read the paperCheck your idea against 137,217 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.
Introduces K-PWM, a Koopman-structured probabilistic world model for control from noisy, partially observed vector measurements of physical systems.
Problem
World models are mostly built from pixel observations, but many physical systems are observed via noisy, partially informative vector measurements, so learning control-ready dynamics under partial observation is harder.
Approach
Uses a Koopman-structured probabilistic latent dynamics model so that the learned world model is structured for control rather than being a generic black-box predictor; details beyond the abstract are not available.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use K-PWM instead of MTS3 | K-PWM · Hopper/Swimmer/Half Cheetah partial obs | prediction RMSE: 0.014 -> 0.0087 (Hopper) | Improved |
| use K-PWM instead of MTS3 | K-PWM · Half Cheetah partial obs | prediction RMSE: 0.14 -> 0.048 | Improved |
| use K-PWM instead of MTS3 | K-PWM · Hopper/Swimmer/Half Cheetah | training time: 7.6-15.1x speedup | Improved |
| use K-PWM instead of MTS3 | K-PWM · MuJoCo Menagerie robots | training time: 1.9-5.6x faster | Improved |
| use K-PWM with MPPI instead of MTS3 model | K-PWM · MuJoCo Menagerie robots | per-step computation time: 4.3-17.0x speedup | Improved |
| use K-PWM instead of DK | K-PWM · MuJoCo Menagerie robots | EE tracking RMSE: Panda 7.3 -> 6.1; ALOHA 2 2.1 -> 2.5 | Mixed |
| use K-PWM instead of DK | K-PWM · MuJoCo Menagerie robots | per-step computation time: 4.9 -> 5.6 (Panda) | Hurt |
| random generator init instead of Koopman-consistent init (ablation) | K-PWM · Swimmer, Half Cheetah | GECM convergence | Hurt |
| use smaller latent dimension than the selected one (ablation) | K-PWM · Half Cheetah | GECM convergence | Hurt |
| GECM instead of joint SGD training (ablation) | K-PWM · MuJoCo prediction | prediction accuracy | No clear effect |
| GECM instead of joint SGD training (ablation) | K-PWM · MuJoCo prediction | training time: several times faster | Improved |
| bilinear instead of linear latent transition (ablation) | K-PWM · MuJoCo prediction | prediction RMSE | Improved |
Closest papers by meaning
- Learning POMDP World Models from Observations with Language-Model Priors (NeurIPS 2026 · similarity 0.87)
- Gated Inference Network: Inference and Learning State-Space Models (NeurIPS 2024 · similarity 0.86)
- Trajectory World Models for Heterogeneous Environments (ICML 2025 · similarity 0.85)
- Efficient Dynamics Modeling in Interactive Environments with Koopman Theory (ICLR 2024 · similarity 0.85)
- Deep-Koopman-KANDy: Dictionary Discovery for Deep-Koopman Operators with Kolmogorov-Arnold Networks for Dynamics (NeurIPS 2026 · similarity 0.85)
- Simplifying Latent Dynamics with Softly State-Invariant World Models (NeurIPS 2024 · similarity 0.85)
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.
