Skip to content
Bud's Wiki

FreeForm: Reduced-Order Deformable Simulation from Particle-Based Skinning Eigenmodes

TL;DR

Section titled “TL;DR”

FreeForm is a mesh-free, reduced-order elastodynamics method for simulating deformable hyperelastic objects across arbitrary geometric representations (meshes, point clouds, 3D Gaussian splats). Instead of learning skinning weights via per-shape neural-field optimization (Simplicits), FreeForm builds the reduced deformation basis by solving a generalized eigensystem on the Hessian of the elastic energy using a Reproducing Kernel Particle Method (RKPM) discretization. The result is a ~40× speedup in the “training” (basis-construction) stage and lower simulation error vs. FEM ground truth. NVIDIA demonstrates it inside a robot-arm interaction pipeline where the deformable objects are 3DGS assets, with Cosmos-Transfer rendering the post-processed video.

Key claims

Section titled “Key claims”
  • Reduced-order elastodynamic simulation can be made mesh-free by replacing per-shape neural-field optimization of skinning weights with a generalized eigensystem solve on the elastic-energy Hessian under an RKPM discretization [Method Overview].
  • Basis construction runs at 3.19 ± 2.48 s (Thingi10K) and 3.49 ± 2.39 s (Simready) with RKPM vs. 121.44 ± 10.15 s and 117.45 ± 1.13 s for Simplicits — roughly 40× faster training time [Comparison Table].
  • FreeForm produces lower simulation error than Simplicits under three boundary-condition regimes (Fix Side, Pull Farthest, Pull Boundary) on both Thingi10K and Simready, evaluated against FEM converged results [Comparison Table].
  • In beam bending and twisting benchmarks against FEM, RKPM outperforms both Simplicits and full-order MPM / SPH baselines given enough degrees of freedom [Standard Beam Test].
  • On heterogeneous materials (a sphere with four alternating hard/soft layers), FreeForm captures distinct per-layer deformation while Simplicits collapses to global stiff-body motion [Additional Results].
  • Limitations: high-frequency details (wrinkles) and sharp contact are difficult under any reduced-order basis; topology changes (fracture) are not modeled; self-collision is not explicitly handled; kernel-radius / sampling-density parameter tuning matters for RKPM basis quality [Limitation].

Method

Section titled “Method”

Traditional FEM requires volumetric meshing, which is difficult on arbitrary shapes and undefined for point-based representations like 3DGS — the motivation for going mesh-free. FreeForm discretizes the object with Reproducing Kernel Particle Method (RKPM) particles, then constructs a reduced-order deformation subspace of skinning eigenmodes by solving a generalized eigensystem on the Hessian of the elastic energy. This replaces Simplicits’ iterative per-shape neural-field optimization with a one-shot linear-algebra solve using efficient LAPACK-style routines. Once the skinning weights are obtained, elastic simulation runs on the reduced subspace with time-stepping. The reduced basis is what drops runtime cost — the simulation only touches subspace coordinates, not every particle. Robot-object interaction is demonstrated by controlling a manually authored robot arm mesh against pasta bags, fruit bowls, and plants simulated with FreeForm; the 3DGS renders are optionally post-processed by Cosmos-Transfer for realistic lighting/reflections while dynamics stay physically grounded.

Results

Section titled “Results”

Training-stage timings show ~38× speedup on Thingi10K (3.19 s vs. 121.44 s) and ~34× on Simready (3.49 s vs. 117.45 s), each with lower error across all three boundary-condition benchmarks. On the Simready “Pull Farthest” test, RKPM error is 4.75e-04 vs. Simplicits’ 9.38e-04 (roughly 2× lower). In the standard beam bending/twisting comparison against FEM ground truth, FreeForm beats Simplicits, MPM, and SPH at equivalent DOF counts. The heterogeneous-sphere qualitative test is the sharpest demonstration: FreeForm reproduces layer-wise deformation gradients that Simplicits smooths away.

Why it’s interesting

Section titled “Why it’s interesting”

For Luma’s video and world-model work, this matters as a simulation backend for deformable 3DGS assets that plausibly slots into pipelines like Agentic Real2Sim: Physics-based World Modeling with Vision-Language Agents (physics-simulator + VLM-agent orchestration) and 3D as code (neural-graphics-as-spatial-programming). The 40× training-time drop makes reduced-order elastodynamics practical inside per-scene 3DGS pipelines, not just offline research demos. Related in spirit to Genesis World — Simulation platform for general-purpose robotics & embodied AI learning‘s unified multi-physics stack, but where Genesis targets FEM/MPM/SPH breadth, FreeForm goes narrow-and-fast on hyperelastic reduced-order simulation across representations. The Cosmos-Transfer post-processing step is also a nice example of decoupling physically-grounded dynamics from photorealistic appearance — a factorization that keeps showing up (compare A Functional Taxonomy of World Models‘s Marble simulator category).

See also

Section titled “See also”