EnliveningGS: Active Locomotion of 3DGS

Siyuan Shen, Tianjia Shao, Kun Zhou, Chenfanfu Jiang, Yin YangCVPR 2025 · DOI

Introduction

While 3D Gaussian Splatting (3DGS) has revolutionized real-time novel view synthesis, the resulting models are typically "lifeless" static shapes. Previous attempts to add physics, such as PhysGaussian, focused primarily on passive dynamics—how an object reacts to external forces like gravity. The challenge addressed by EnliveningGS is active locomotion: enabling an object to move itself using internal forces (muscles). This presents an "inverse physics" problem: determining the internal muscle activations required to generate necessary friction and contact forces against the environment to follow a desired trajectory.

Methodology

The authors propose a multi-stage pipeline to bridge the gap between static splats and active characters, utilizing a hybrid approach that maintains visual fidelity during extreme deformations.

Overview of the EnliveningGS pipeline, showing the hybrid mesh-GS representation, the locomotion solver, and the final animated output.

1. Hybrid Mesh-Gaussian Representation

The system embeds the 3DGS model into a coarse tetrahedral mesh. This mesh serves as the structural "skeleton" and "flesh," while the Gaussians provide the high-fidelity visual "skin."

2. Muscle-Driven Dynamics and Solver

Virtual muscle fibers, modeled as springs, are embedded within the mesh. To handle the complex problem of frictional contact, the framework employs a Two-Stage Locomotion Solver. It first uses a penalty method to obtain a rough estimate of contacts, then refines the motion using Linear Complementarity Programming (LCP) for precise, hard-contact physics.

3. Maintaining Visual Fidelity

To prevent "spiky" artifacts during significant stretching, the system monitors Gaussian deformation. If a kernel becomes too distorted, it is dynamically split into two smaller, more accurate kernels. Additionally, as the character moves, the system uses plane-prior inpainting to fill the "holes" left in the background, operating up to 60x faster than general-purpose 3D inpainting methods.

Key Results

EnliveningGS demonstrates superior performance in both physical realism and visual quality compared to existing state-of-the-art methods.

A sequence showing a chess piece performing active locomotion (jumping) over time, demonstrating realistic trajectory and contact.
  • Visual Fidelity: In high-deformation tasks like "Twisting," EnliveningGS achieved a PSNR of 26.89, outperforming PhysGaussian (26.53).
  • Solver Efficiency: The two-stage solver reached lower objective values (0.83) with significantly fewer iterations (17.01 avg) compared to standard QPCC solvers (1.29 value / 46.19 iterations).
  • Inpainting Speed: The plane-prior method took only 20 seconds to fill large background gaps, a massive improvement over the 20 minutes required by Gaussian Grouping.

Conclusion and Future Work

EnliveningGS successfully transforms static 3DGS captures into dynamic, controllable entities. By treating models as biological entities with internal muscle logic, the framework enables complex behaviors like walking and jumping. Future work may involve integrating reinforcement learning to automate more expressive, lifelike gaits and expanding contact models to handle highly irregular 3D geometries.

Summary updated

BibTeX
@inproceedings{shen2025enliveninggs,
  author = {Siyuan Shen and Tianjia Shao and Kun Zhou and Chenfanfu Jiang and Yin Yang},
  title = {{EnliveningGS: Active Locomotion of 3DGS}},
  booktitle = {IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  year = {2025},
  doi = {10.1109/cvpr52734.2025.00092},
}
Figure