Title: Enabling Real-Time Physics Simulations via Dynamic Mesh-Gaussian Conversions

Abstract:

While physics simulation relies on fixed mesh topologies to ensure efficient collision detection, integrating dynamic 3D reconstructions presents a challenge because state-of-the-art techniques, such as DG-Mesh, generate meshes with varying topologies that are optimized for geometric accuracy. This study explores whether converting these topologies can facilitate physics integration without compromising reconstruction fidelity. To address this, we introduce a dual-representation framework that pairs fixed-topology meshes, designed for physics calculations, with Gaussian splatting for visual rendering. By updating vertex buffers at runtime, this approach achieves a 4.65$\times$ acceleration compared to baselines utilizing varying topologies. We assessed two conversion methodologies—temporal correspondence tracking and template-based projection—against native fixed-topology methods (MaGS) using the DG-Mesh dataset. The results indicate that both conversion strategies suffer from 65-80% geometric degradation, yielding outcomes that fall short of MaGS performance, even though DG-Mesh starts with higher initial quality. These findings highlight that achieving high-fidelity reconstruction and maintaining a physics-compatible topology are fundamentally conflicting goals that cannot be effectively resolved through post-processing. Our insights guide the future creation of physics-aware reconstruction techniques, while our proposed framework supports real-time simulation for any fixed-topology method.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC