Fast Complementary Dynamics via Skinning Eigenmodes

Otman Benchekroun, Jiayi Eris Zhang, Siddhartha Chaudhuri, Eitan Grinspun, Yi Zhou, Alec JacobsonACM Trans. Graphics 2023 · DOI · OpenReview

TL;DR

This paper introduces "skinning eigenmodes," a novel subspace for simulating secondary dynamics in rigged characters that is both rotationally equivariant and computationally efficient. This allows for real-time, high-quality secondary motion augmentation of character animations, overcoming limitations of previous methods that suffered from artifacts or were too slow for interactive applications.

What Problem Does This Paper Address?

Creating realistic animated characters requires not only the primary motion driven by a rig but also secondary dynamics – the subtle, often physics-driven movements like cloth simulation, hair sway, or muscle jiggle. While "complementary dynamics" methods can generate these secondary motions by simulating physics orthogonal to the rig's motion, they are computationally expensive, making them unsuitable for real-time applications like games or interactive VR. Existing acceleration techniques, such as using displacement modes derived from the Hessian of the elastic energy, fail to capture rotational motion accurately, leading to warping and frame-dependent artifacts. This paper addresses the need for a computationally efficient and rotationally robust method to generate high-quality secondary dynamics for interactive character animation.

Methods

The core innovation is the skinning eigenmodes subspace. Instead of relying on traditional displacement modes, the authors leverage the structure of Linear Blend Skinning (LBS). They derive a set of "skinning weights" that parameterize a deformation subspace. This subspace is constructed by solving a generalized eigenvalue problem on a "weight-space Hessian," which is derived from the elastic energy of the mesh. Crucially, this approach is designed to be rotationally equivariant, meaning the simulation's behavior is consistent regardless of the character's orientation.

To further accelerate the simulation, especially for non-linear elastic models, they employ a clustering approach. This groups tetrahedra based on the derived skinning weights, allowing for per-cluster computations instead of per-tetrahedron, significantly reducing computational cost. Finally, they combine these techniques with a hyper-reduced local-global solver that operates entirely within the reduced subspace and uses the clusters, avoiding expensive full-space operations. This solver iteratively optimizes for primary and secondary degrees of freedom, efficiently handling both linear and non-linear elasticity.

Datasets

Not discussed in the paper. The authors use various 3D character models (e.g., elephant, bulldog, dragon, moray eel) and demonstrate their method on these models.

Key Insights and Ideas

  • Skinning as a Subspace Basis: The paper demonstrates that the structure of Linear Blend Skinning can be directly used to define a deformation subspace that inherently possesses desirable properties like rotational equivariance, which is a significant advantage over traditional displacement modes.
  • Rotation Equivariance is Crucial: The authors highlight that for character animation, where rigs often involve significant rotations, rotation equivariance is not just a nice-to-have but a necessity to avoid artifacts and ensure intuitive behavior.
  • Weight-Space Optimization: By optimizing skinning weights directly in a lower-dimensional "weight-space," they can efficiently derive modes that are sensitive to both material properties and rig constraints.
  • Clustering for Non-Linearity: Using clustering based on skinning weights allows them to efficiently handle non-linear elastic energies without being bottlenecked by mesh resolution.

Key Results

  • Significant Speedup: The method achieves a speedup of over 100x compared to full-space complementary dynamics, enabling real-time performance (e.g., >60 fps on a complex scene).
  • High-Quality Secondary Motion: The skinning eigenmodes subspace captures rich secondary dynamics, including twisting and volume-preserving effects, which are difficult for traditional methods.
  • Robustness to Rotations: The simulations are rotationally equivariant, meaning they produce consistent results regardless of the character's orientation, unlike methods based on displacement modes.
  • Efficient Subspace Construction: The generalized eigenvalue problem for skinning eigenmodes is efficient, especially compared to constructing equivalent displacement mode subspaces.

Limitations and Future Work

  • Linear Rigs: The current method assumes linear rigs. Extending it to non-linear rigs is a direction for future work.
  • Collision Handling: Comprehensive collision handling is left as future work. While rigid-body augmentation offers some basic collision effects, more accurate methods are needed.
  • High-Frequency Motion: The subspace can struggle with very high-frequency motions, requiring more degrees of freedom to capture them accurately.
  • General Elasticity: While the framework is adaptable to different elastic potentials, exploring beyond co-rotational models and investigating more complex material behaviors is suggested.
  • VR Tracking Bottlenecks: For VR applications, the tracking and mapping of user movements to primary rig controls can be a bottleneck, which the secondary dynamics would inherit.

Summary updated

Abstract

We propose a reduced-space elasto-dynamic solver that is well suited for augmenting rigged character animations with secondary motion. At the core of our method is a novel deformation subspace based on Linear Blend Skinning that overcomes many of the shortcomings prior subspace methods face. Our skinning subspace is parameterized entirely by a set of scalar weights, which we can obtain through a small, material-aware and rig-sensitive generalized eigenvalue problem. The resulting subspace can easily capture rotational motion and guarantees that the resulting simulation is rotation equivariant. We further propose a simple local-global solver for linear co-rotational elasticity and propose a clustering method to aggregate per-tetrahedra non-linear energetic quantities. The result is a compact simulation that is fully decoupled from the complexity of the mesh.
BibTeX
@article{benchekroun2023fast,
  author = {Otman Benchekroun and Jiayi Eris Zhang and Siddhartha Chaudhuri and Eitan Grinspun and Yi Zhou and Alec Jacobson},
  title = {{Fast Complementary Dynamics via Skinning Eigenmodes}},
  journal = {ACM Trans. Graphics},
  year = {2023},
  doi = {10.1145/3592404},
}
Figure