VisualPatchWorld: Code World Models for Efficient Planning
HKBU-KnowComp · hf · 2026-07-29
To bridge the gap between neural predictors and physics engines, the research introduces VisualPatchWorld (VPW), representing world dynamics as code to balance scalability and inspectability.
- Mechanism: VPW selects a qualitative dynamical form via short active probes, then fits the form's free parameters by minimizing multi-step prediction error. The resulting programs can be rolled forward like a simulator and directly applied in model-predictive control (MPC).
- Performance: Compared to prior code-based world models, VPW achieves a 69.0% mean planning success rate, exceeding the strongest baseline by 23.5 points.
- Use Cases: VPW approaches ground-truth engine success on navigation and grasp-rich control tasks. A residual gap remains for contact-rich pushing, but checking a shortlist of promising plans in the engine closes most of it.
More from Research
- Coding agents can speed up projects by quietly locking in the wrong decisions — bibryam · 2026-07-29
- ICLR’s 2026 deadline joke lands with a NeurIPS timing twist — delliott · 2026-07-29
- Microsoft removes Mage Flow and points to a more efficient Qwen-VL-class encoder — Dante_77A · 2026-07-29
- Profluentbio is building protein foundation models for AI-designed therapeutics — nathanbenaich · 2026-07-29
- A staged diagnosis finds most short-text generation loss comes from the codec — ITMO · 2026-07-29
- Nature paper measures non-Gaussian order-parameter statistics across a phase transition — burny_tech · 2026-07-29