Research map / 3D vision / 3D generation
Gaussian-Augmented Physics Simulation and System Identification with Complex Colliders
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 AS-DiffMPM, a differentiable MPM with CPIC-style particle-wise collision handling for arbitrarily shaped rigid colliders (meshes or 2D Gaussians), integrated with DVGO, 2DGS, and MDyn-3DGS for system identification from video.
Problem
System identification of continuum material physical parameters from visual observations when objects collide with arbitrarily shaped (non-planar) rigid bodies, which existing differentiable MPM pipelines do not support.
Approach
Extends differentiable MPM with Compatible Particle-in-Cell (CPIC) collision handling that applies particle-wise velocity corrections using a collision grid built from mesh faces or 2D Gaussian disks, allowing end-to-end gradient-based parameter optimization with complex colliders, and plugs into NeRF- and 3DGS-based renderers.
Key results
- On particle-trajectory system identification, AS-DiffMPM achieves the best or second-best errors on most materials; e.g., Armadillo Newtonian log10(mu) error 0.4 vs 5.67 for RP-DiffMPM.
- With visual supervision, MDyn-3DGS reaches PSNR 41.3 on Newtonian Box versus 35.3 for DVGO and 35.7 for 2DGS.
- Real dough experiment yields estimated parameters mu=57,033.1, kappa=139,650.4, tau_Y=7,919.1, eta=25.1, with PSNR dropping from 33.2 at 1 s to 25.3 at 3 s after collision.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use particle-wise AS-DiffMPM collision handling instead of rigid-particle collision handling | System identification from particle trajectories · Complex colliders across materials | physical-parameter absolute error | Helped |
| Use AS-DiffMPM instead of GOP-DiffMPM | System identification from particle trajectories · Complex colliders across materials | physical-parameter absolute error | Mixed |
| Use MDyn-3DGS rendering with dynamic reconstruction supervision | Visual system identification · Newtonian fluid, colliders overall | physical-parameter absolute error | Helped |
| Use 2DGS rendering rather than MDyn-3DGS | Visual system identification · Granular media | physical-parameter absolute error | Helped |
| Replace a mesh collider with a 2DGS collider (ablation) | AS-DiffMPM · Particle trajectories, Newtonian, Box | physical-parameter absolute error: 5.08 $\pm$ 2.73 -> 4.36 $\pm$ 2.66 | Mixed |
| Replace a mesh collider with a 2DGS collider (ablation) | AS-DiffMPM · Particle trajectories, Newtonian, Armadillo | physical-parameter absolute error: 0.4 $\pm$ 0.76 -> 8.75 $\pm$ 6.17 | Hurt |
| Replace a mesh collider with a 2DGS collider (ablation) | AS-DiffMPM · Particle trajectories, Granular, Bunny | physical-parameter absolute error: 0.06 $\pm$ 0.07 -> 0.91 $\pm$ 1.15 | Hurt |
| Use analytical S-MPM planar boundary conditions instead of AS-DiffMPM | Planar-collider system identification · Newtonian, log10(mu) | absolute error: 7.98 -> 7.51 | Helped |
| Use AS-DiffMPM instead of GOP-DiffMPM | Bowl-collider system identification · Newtonian, log10(mu) | absolute error: 7.29 -> 6.01 | Helped |
| Evaluate reconstruction later after collision | AS-DiffMPM dough experiment · Dough reconstruction, 1 s to 3 s after collision | PSNR: 33.2 $\pm$ 4.1 -> 25.3 $\pm$ 7.8 | Hurt |
Surprising findings
- For planar colliders, the standard S-MPM boundary strategy achieves slightly lower errors than AS-DiffMPM for some parameters.
- The 2DGS-based collider performs worse than the mesh collider for system identification from trajectories on most materials.
Limitations
- Colliders must be reconstructed and imported before optimization, so only static rigid bodies are supported.
- Gaussian attributes are not time-dependent, causing reconstruction degradation for large deformations (dough experiment).
- Non-Newtonian materials remain hardest, with large parameter errors across colliders.
Experimental setup
Datasets: simulated MPM rollouts with cross-shaped object and Box, Bunny, Armadillo colliders real dough capture (five cameras)
Benchmarks: simulated particle-trajectory system identification benchmark (Box, Bunny, Armadillo colliders) visual system identification benchmark with DVGO, 2DGS, MDyn-3DGS real-world dough experiment
Metrics: mean absolute error of physical parameters (scaled by 100) Chamfer Distance Earth Mover's Distance PSNR SSIM LPIPS
Baselines: GOP-DiffMPM RP-DiffMPM S-MPM PAC-NeRF
Builds on: PAC-NeRF Compatible Particle-in-Cell 2D Gaussian Splatting DVGO PhysGaussian
Closest papers by meaning
- GIC: Gaussian-Informed Continuum for Physical Property Identification and Simulation (NeurIPS 2024 · similarity 0.84)
- MonoPhysics: Estimating Geometry, Appearance, and Physical Parameters from Monocular Videos (NeurIPS 2026 · similarity 0.84)
- DecoupledGaussian: Object-Scene Decoupling for Physics-Based Interaction (CVPR 2025 · similarity 0.84)
- OmniPhysGS: 3D Constitutive Gaussians for General Physics-Based Dynamics Generation (ICLR 2025 · similarity 0.83)
- Vid2Sim: Generalizable, Video-based Reconstruction of Appearance, Geometry and Physics for Mesh-free Simulation (CVPR 2025 · similarity 0.83)
- Structure from Collision (CVPR 2025 · similarity 0.83)
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.