PhysWorld: From Real Videos to World Models of Deformable Objects via Physics-Aware Demonstration Synthesis

arXiv 25.10 2025 58.8 method, application

TLDR

PhysWorld synthesizes physics-aware demonstrations from real videos to train fast, accurate world models for deformable objects.

Reasoning

The paper presents a novel framework that addresses data scarcity by creating digital twins and generating diverse demonstrations, achieving significant speedup over prior work. Strengths include physics-consistent modeling and real video refinement; weaknesses are the reliance on a specific simulator (MPM) and focus on deformable objects, limiting generality.

Read-first score

Read-first score 58.8, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 53.

Recency 8%

86.7

Uses a gentle age decay so recent papers surface without erasing older foundations. 2025

Topical relevance 42%

75.7

Uses existing LLM keyword relevance scores normalized to 0-100. world model,world simulator,generative world model,interactive world model,video world model,world dynamics prediction,model-based reinforcement learning world model

Methodology quality 25%

Screens visible abstract and analysis fields for experiment, dataset, baseline, metric, and limitation evidence. markers=experiment

Reproducibility 25%

Screens links and visible text for paper, code, dataset, artifact, and repository signals. pdf=True; code=False; dataset=False; markers=none

Field roles

Frontier

Rank sensitivity

Stability: volatile; rank range: 467.

Keyword Scores

world model

video world model

world dynamics prediction

interactive world model

generative world model

world simulator

model-based reinforcement learning world model

Deep Analysis

Innovations

Physics-aware demonstration synthesis using a simulator to overcome data scarcity for deformable objects
Construction of a physics-consistent digital twin via constitutive model selection and global-to-local optimization of physical properties
Part-aware perturbations to physical properties for generating diverse motion patterns and extensive demonstrations
Lightweight GNN-based world model embedded with physical properties for fast and accurate future prediction
Real video refinement of physical properties to improve model accuracy

Methodology

PhysWorld first constructs a physics-consistent digital twin of a deformable object within an MPM simulator by selecting a constitutive model and optimizing physical properties globally and locally. It then applies part-aware perturbations to these properties to generate diverse motion patterns, synthesizing extensive training demonstrations. A lightweight GNN-based world model is trained on these demonstrations, with physical properties embedded, and real video can be used to further refine the properties.

Key Results

PhysWorld achieves accurate future predictions for various deformable objects and generalizes well to novel interactions, with inference speeds 47 times faster than the state-of-the-art method PhysTwin.

Resources

Paper