PhysiFormer: Learning to Simulate Mechanics in World Space
TLDR
PhysiFormer uses diffusion transformer in world coordinates to simulate physically-plausible 3D object motion, outperforming autoregressive baselines.
Reasoning
Strengths: novel coordinate-space diffusion approach, no inductive biases, factorized attention, strong results on simulated data. Weaknesses: limited to simulated training, no real-world validation, unclear scalability to complex scenes.
Read-first score
Read-first score 56.6, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 37.
Field roles
Rank sensitivity
Stability: volatile; rank range: 400.
Keyword Scores
Deep Analysis
Innovations
- Casting vertex trajectory prediction as a single denoising diffusion process directly in world coordinates without explicit inductive biases like rigidity or causality
- Attention factorised over time, space, and objects for efficiency and permutation-invariant multi-object reasoning
- Probabilistic formulation capturing uncertainty in learned dynamics, enabling diverse plausible futures
Methodology
PhysiFormer is a diffusion transformer that takes initial vertex positions, velocities, and object material type (rigid or elastic) as input and samples future vertex trajectories via a denoising diffusion process in world coordinates. It uses attention factorised over time, space, and objects for efficiency and permutation-invariant multi-object reasoning. The model is trained on over 100k simulated trajectories.
Key Results
PhysiFormer substantially outperforms autoregressive baselines in trajectory accuracy, rigidity preservation, and momentum-based physical consistency, and generalizes to mixed-material settings, unseen real-world geometries, and larger object counts.