Research map / 3D vision / Gaussian splatting
EnliveningGS: Active Locomotion of 3DGS
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 EnliveningGS, a pipeline that makes 3DGS objects move actively by inversely solving muscle activations under a hybrid LCP-penalty contact model, with local-embedding-based Gaussian splitting and plane-prior inpainting.
Problem
Active, muscle-driven locomotion of objects represented by 3D Gaussian splatting, which requires modeling frictional contacts with the environment in an inverse physics problem; prior 3DGS physics methods only simulate passive forward dynamics.
Approach
Embed muscle fibers in a tetrahedral mesh bound to Gaussians, combine a penalty method with LCP-based contact so the active set is reduced, and solve a two-stage QP for muscle activations that track user trajectories; bidirectional local embeddings guide Gaussian deformation and splitting.
Key results
- On bending, twisting and falling deformations of NeRF synthetic objects, Ours reaches PSNR 28.79/26.89/25.67 versus 28.63/26.53/25.31 for VR-GS.
- The hybrid LCP-penalty locomotion solver averages 17.01 iterations versus 46.19 for QPCC [41] on a 50-frame, 60-variable sequence, with lower average objective value (0.83 vs 1.29).
- Plane-prior inpainting fine-tunes in about 20 seconds (about 100 iterations), reported as 60x faster than Gaussian Grouping's 20 minutes.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use EnliveningGS instead of VR-GS | 3DGS deformation · NeRF synthetic (bending) | PSNR: 28.63 -> 28.79 | Helped |
| Use EnliveningGS instead of VR-GS | 3DGS deformation · NeRF synthetic (twisting) | LPIPS: 0.066 -> 0.041 | Helped |
| Use EnliveningGS instead of VR-GS | 3DGS deformation · NeRF synthetic (falling) | SSIM: 0.938 -> 0.951 | Helped |
| Add Gaussian splitting (ablation) | EnliveningGS · NeRF synthetic (bending) | PSNR: 28.62 -> 28.79 | Helped |
| Add Gaussian splitting (ablation) | EnliveningGS · NeRF synthetic (twisting) | LPIPS: 0.062 -> 0.041 | Helped |
| Add Gaussian splitting (ablation) | EnliveningGS · NeRF synthetic (falling) | SSIM: 0.939 -> 0.951 | Helped |
| Use the two-stage penalty-guided solver instead of QPCC | EnliveningGS locomotion solver · Locomotion solver (average) | iterations: 46.19 -> 17.01 | Helped |
| Use the two-stage penalty-guided solver instead of QPCC | EnliveningGS locomotion solver · Locomotion solver (average) | object value: 1.29 -> 0.83 | Helped |
| Add non-penetration regularization Lseg (ablation) | EnliveningGS segmentation · Segmentation plane (object upper half) | PSNR: 25.917 -> 35.291 | Helped |
| Add non-penetration regularization Lseg (ablation) | EnliveningGS segmentation · Segmentation plane (object upper half) | SSIM: 0.827 -> 0.945 | Helped |
| Add non-penetration regularization Lseg (ablation) | EnliveningGS segmentation · Segmentation plane (object upper half) | LPIPS: 0.211 -> 0.034 | Helped |
| Add non-penetration regularization Lseg (ablation) | EnliveningGS segmentation · Segmentation plane (lower-half residual) | residual: 0.058 -> 0.002 | Helped |
Limitations
- Relies on plane-based simplification of the environment; complex surroundings need more sophisticated contact approximation.
- Triangulated surfaces would add computational overhead to the contact solver.
- Locomotion is evaluated on hand-designed trajectories and a small set of scenes.
Experimental setup
Datasets: NeRF synthetic Mip-NeRF360 garden DTU truck
Benchmarks: NeRF synthetic Mip-NeRF360 DTU
Metrics: PSNR SSIM LPIPS iterations objective value residual
Baselines: PhysGaussian VR-GS SuGaR QP solver QPCC Gaussian Grouping
Builds on: PhysGaussian Linear Complementarity Problem contact model (Stewart and Trinkle) Neo-Hookean elasticity 2DGS Gaussian Grouping
Closest papers by meaning
- MotionGS: Exploring Explicit Motion Guidance for Deformable 3D Gaussian Splatting (NeurIPS 2024 · similarity 0.86)
- FlexGS: Train Once, Deploy Everywhere with Many-in-One Flexible 3D Gaussian Splatting (CVPR 2025 · similarity 0.85)
- Mani-GS: Gaussian Splatting Manipulation with Triangular Mesh (CVPR 2025 · similarity 0.85)
- H3D-DGS: Exploring Heterogeneous 3D Motion Representation for Deformable 3D Gaussian Splatting (NeurIPS 2025 · similarity 0.84)
- OmniPhysGS: 3D Constitutive Gaussians for General Physics-Based Dynamics Generation (ICLR 2025 · similarity 0.84)
- DecoupledGaussian: Object-Scene Decoupling for Physics-Based Interaction (CVPR 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.