DeformMaster: An Interactive Physics-Neural World Model for Deformable Objects from Videos
TLDR
DeformMaster learns an interactive physics-neural world model for deformable objects from real videos, enabling dynamics rollout and novel-view synthesis.
Reasoning
The paper presents a novel unified framework combining physics-based dynamics with neural residuals for high-fidelity modeling of deformable objects from real-world videos. Strengths include real-world experiments and outperforming baselines; weaknesses are the lack of explicit reinforcement learning context and potential complexity in scaling.
Read-first score
Read-first score 61.4, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 51.
Field roles
Rank sensitivity
Stability: volatile; rank range: 396.
Keyword Scores
Deep Analysis
Innovations
- Video-derived interactive physics-neural world model for deformable objects
- Preserves structured physical rollout with neural residual compensation for unmodeled effects
- Grounds sparse hand motion as a distributed compliant actuator for hand-continuum interaction
- Represents material response with spatially varying constitutive experts
- Drives high-fidelity 4D appearance from predicted physical evolution
- Unified dynamics-and-appearance framework
Methodology
DeformMaster is a physics-neural world model that combines structured physical rollout with a neural residual to compensate for unmodeled effects. It uses sparse hand motion as a distributed compliant actuator for hand-continuum interaction and represents material response with spatially varying constitutive experts. The model predicts physical evolution and renders high-fidelity 4D appearance from it, trained on real-world deformable-object videos.
Key Results
Experiments on real-world deformable-object sequences show DeformMaster outperforms state-of-the-art baselines in rolling out future dynamics and rendering dynamic appearance, while supporting novel action rollout, material-parameter variation, and dynamic novel-view synthesis.