Research map / Probabilistic and scientific modeling / Dynamical systems
Learning Flexible Body Collision Dynamics with Hierarchical Contact Mesh Transformer
Read the paperCheck your idea against 108,967 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.
Proposes HCMT, a hierarchical mesh graph Transformer with contact-mesh and hierarchical-mesh attention blocks for learning flexible body collision dynamics, and introduces the Impact Plate dataset.
Problem
Mesh-based GNN simulators struggle to propagate collision effects over long distances in flexible body dynamics, where instantaneous contacts occur in short timeframes and stress must be predicted.
Approach
Combines a contact mesh Transformer that attends over separate mesh and contact edges (with edge features) with a hierarchical mesh Transformer that pools nodes via BFS and Delaunay remeshing so collision effects propagate quickly across levels.
Key results
- On Impact Plate, HCMT reaches position RMSE 20.71 (x10^3) vs 40.73 for MGN, a 49.2% improvement, and stress RMSE 14,742 vs 35,871.
- HCMT is reported to be about 176 times faster than ANSYS per step on Impact Plate (81 ms vs 14,280 ms).
- HCMT achieves the best RMSE on all four datasets, e.g. Sphere Simple position 30.41 vs 33.26 for MGN and Deformable Plate 7.67 vs 10.78.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use HCMT instead of the strongest baseline, MGN | Impact Plate · Impact Plate | Position RMSE: 40.73\scriptsize{±}{2.94} -> 20.71\scriptsize{±}{0.57} | Helped |
| use HCMT instead of the strongest baseline, MGN | Deforming Plate · Deforming Plate | Position RMSE: 7.83\scriptsize{±}{0.16} -> 7.49\scriptsize{±}{0.07} | Helped |
| use HCMT instead of the strongest baseline, MGN | Sphere Simple · Sphere Simple | Position RMSE: 33.26\scriptsize{±}{6.33} -> 30.41\scriptsize{±}{1.71} | Helped |
| use HCMT instead of the strongest baseline, MGN | Deformable Plate · Deformable Plate | Position RMSE: 10.78\scriptsize{±}{0.54} -> 7.67\scriptsize{±}{0.42} | Helped |
| apply HMT before CMT instead of contact-then-propagate order (ablation) | HCMT · Impact Plate | Position RMSE: 20.34 -> 42.90 | Hurt |
| remove HMT (ablation) | HCMT · Impact Plate | Position RMSE: 20.34 -> 55.63 | Hurt |
| remove CMT (ablation) | HCMT · Impact Plate | Position RMSE: 20.34 -> 46.27 | Hurt |
| add Laplacian positional encoding (ablation) | HCMT · Impact Plate | Position RMSE: 20.34 -> 21.11 | Hurt |
| replace dual-branch CMT with single-branch CMT (ablation) | HCMT · Impact Plate | Stress RMSE: 14,447 -> 16,046 | Hurt |
| replace dual-branch CMT with single-branch CMT (ablation) | HCMT · Deformable Plate | Position RMSE: 7.67 -> 8.05 | Hurt |
| remove remeshing (ablation) | HCMT · Impact Plate | Stress RMSE: 14,447 -> 17,482 | Hurt |
| remove remeshing (ablation) | HCMT · Impact Plate | Position RMSE: 20.34 -> 19.21 | Helped |
Surprising findings
- The Late-Contact ordering (HMT before CMT) and Only-HMT variants degrade badly, and Only-CMT beats Only-HMT on every dataset.
- Remeshing with Delaunay triangulation improves stress RMSE but the position RMSE is slightly worse with remeshing (19.21 without vs 20.34 with).
Limitations
- Pooling is fixed (BFS sampling plus remeshing) rather than learned, as the authors note.
- Evaluated only on 2D Impact Plate for flexible collisions; Deforming Plate and Sphere Simple are not flexible-collision benchmarks, and the Impact Plate dataset is proprietary-style industry data with 2D meshes.
- Two stress-singularity trajectories had to be excluded from Deforming Plate.
Experimental setup
Datasets: Impact Plate Deforming Plate Sphere Simple Deformable Plate
Benchmarks: Impact Plate Deforming Plate Sphere Simple Deformable Plate
Metrics: RMSE (rollout-all) position RMSE stress RMSE training time per step inference time per step
Baselines: MGN GT (Graph Transformer) HCMT ablations: Late-Contact, Only CMT, Only HMT, HCMT+LPE
Builds on: MeshGraphNets Graph Transformer GNS BSMS-GNN EAGLE
Closest papers by meaning
- EGODE: An Event-attended Graph ODE Framework for Modeling Rigid Dynamics (NeurIPS 2024 · similarity 0.85)
- Towards Universal Mesh Movement Networks (NeurIPS 2024 · similarity 0.85)
- UGM2N: An Unsupervised and Generalizable Mesh Movement Network via M-Uniform Loss (NeurIPS 2025 · similarity 0.84)
- Object Dynamics Modeling with Hierarchical Point Cloud-based Representations (CVPR 2024 · similarity 0.84)
- HOI-Dyn: Learning Interaction Dynamics for Human-Object Motion Diffusion (NeurIPS 2025 · similarity 0.84)
- Geometric Contact Flows: Contactomorphisms for Dynamics and Control (ICML 2025 · similarity 0.84)
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.