FGGS-LiDAR · Ultra-Fast LiDAR Simulation from General 3DGS Models (IROS 2026)

LiDAR simulation demo in Isaac Gym on a large 3DGS scene.在 Isaac Gym 中对大规模 3DGS 场景进行的 LiDAR 仿真演示。

Real-time LiDAR simulation (3DGS → Mesh → LiDAR).实时 LiDAR 仿真（3DGS → 网格 → LiDAR）。

500+ FPS

Real-time LiDAR returns实时 LiDAR 出点

4096

Parallel environments并行仿真环境

~10×

Lower latency vs Isaac Sim相比 Isaac Sim 的时延

>10⁸ pts/s

Peak point throughput峰值点云吞吐

Pose-free无需位姿

No COLMAP / LiDAR supervision无需 COLMAP 位姿或 LiDAR 监督

3DGS → Mesh

Directly from Gaussian params直接由高斯参数转换

Abstract摘要

A geometry-first bridge from 3DGS to LiDAR从 3DGS 到 LiDAR 的几何优先桥梁

While 3D Gaussian Splatting (3DGS) has emerged as a strong representation for photorealistic rendering, its vast ecosystem of assets remains difficult to use for high-performance LiDAR simulation, a critical tool for robotics and autonomous driving. We present FGGS-LiDAR, a geometry-first framework that bridges this gap in a plug-and-play manner. Our method converts pretrained 3DGS assets into watertight meshes directly from Gaussian parameters, without requiring LiDAR-specific supervision or architectural alterations, via volumetric discretization and Truncated Signed Distance Field (TSDF) extraction. We pair this with a GPU-accelerated ray-casting module that simulates LiDAR returns at over 500 FPS and supports batched multi-environment simulation with up to 4096 environments. In large-scale parallel settings, FGGS-LiDAR achieves an order-of-magnitude lower LiDAR-simulation latency than Isaac Sim. We validate FGGS-LiDAR on both indoor and outdoor scenes, demonstrating high LiDAR-simulation fidelity. Furthermore, on COLMAP-posed indoor benchmarks, we compare against existing 3DGS-to-mesh baselines and report lower LiDAR-simulation error.

3D 高斯泼溅（3DGS） 已成为照片级真实感渲染的强大表示，但其庞大的资产生态难以直接用于高性能 LiDAR 仿真——而 LiDAR 仿真是机器人与自动驾驶的关键工具。我们提出 FGGS-LiDAR，一个即插即用、几何优先的框架。该方法通过体素化离散与截断符号距离场（TSDF）提取，直接从高斯参数将预训练的 3DGS 资产转换为水密网格，无需 LiDAR 专用监督，也无需修改原有架构。我们将其与 GPU 加速的光线投射模块结合，可在 500 FPS 以上 模拟 LiDAR 回波，并支持最多 4096 个环境 的批量多环境仿真。在大规模并行场景下，FGGS-LiDAR 相比 Isaac Sim 实现了约一个数量级更低的 LiDAR 仿真时延。我们在室内外场景中验证了方法的高仿真保真度；在带 COLMAP 位姿的室内基准上，与已有的 3DGS-to-mesh 方法相比，取得了更低的 LiDAR 仿真误差。

Highlights亮点

Three contributions三大贡献

01

Geometry-first, pose-free几何优先，无需位姿

A pose-free pipeline that recovers watertight meshes directly from Gaussian parameters — no COLMAP poses, rendered depth maps, or LiDAR supervision. LiDAR can be simulated from arbitrary pretrained 3DGS assets.一条无需位姿的流程，直接从高斯参数恢复水密网格——无需 COLMAP 位姿、渲染深度图或 LiDAR 监督，可对任意预训练 3DGS 资产进行 LiDAR 仿真。

02

Efficient 3DGS → geometry高效的 3DGS → 几何

GPU-accelerated voxelization with Gaussian AABBs and Morton-sorted LBVH indexing, followed by narrow-band TSDF reconstruction that yields topology-consistent watertight meshes while scaling to large scenes.基于高斯 AABB 与 Morton 排序 LBVH 索引的 GPU 加速体素化，配合窄带 TSDF 重建，生成拓扑一致的水密网格，并可扩展到大规模场景。

03

Massively parallel LiDAR大规模并行 LiDAR

A plug-and-play, GPU-batched ray-casting engine supporting thousands of environments (up to 4096) in parallel at over 500 FPS — substantially outperforming general-purpose simulators such as Isaac Sim.即插即用的 GPU 批量光线投射引擎，可并行支持数千个环境（最多 4096 个），帧率超过 500 FPS，显著优于 Isaac Sim 等通用仿真器。

Method方法

From Gaussians to LiDAR returns从高斯到 LiDAR 回波

Framework Overview框架总览

3DGS → Occupancy → TSDF Mesh → BVH Ray-casting

FGGS-LiDAR integrates Gaussian-based scene representation, watertight mesh reconstruction, and GPU-parallel LiDAR simulation. Pretrained 3DGS assets are discretized into occupancy volumes, fused into a narrow-band TSDF, and converted to meshes. A BVH-accelerated ray-casting engine then performs first-hit queries at real-time rates, enabling physically grounded, scalable LiDAR sensing directly from general 3DGS content.FGGS-LiDAR 将高斯场景表示、水密网格重建与 GPU 并行 LiDAR 仿真融为一体。预训练 3DGS 资产先被离散化为占据体，融合为窄带 TSDF，再转换为网格；随后由 BVH 加速的光线投射引擎以实时帧率完成首次命中查询，从而可直接基于通用 3DGS 内容进行物理可信、可扩展的 LiDAR 感知。

3DGS Voxelization3DGS 体素化

LBVH-accelerated Gaussian-to-occupancy

Gaussian primitives are organized by an LBVH for efficient spatial querying, enabling localized voxel-density evaluation instead of global Gaussian–voxel accumulation. Thresholding and surface filtering then produce a compact, surface-supporting occupancy representation that preserves thin structures and high-frequency geometry for the next stage.高斯基元由 LBVH 组织以实现高效空间查询，从而进行局部体素密度评估，而非全局的高斯–体素累加。随后通过阈值化与表面过滤，得到紧凑且支撑表面的占据表示，保留薄结构与高频几何，供下一阶段使用。

Denoised → Mesh去噪 → 网格

Narrow-band TSDF + watertight isosurface

Surface-supporting voxels are denoised and converted into a narrow-band TSDF, from which a watertight isosurface is extracted. Mesh simplification and structure-preserving smoothing then reduce simulation cost while preserving the geometry needed for accurate LiDAR first-hit queries — eliminating local artifacts without shrinking geometry.支撑表面的体素经去噪后转换为窄带 TSDF，并从中提取水密等值面。随后通过网格简化与保结构平滑，在保留 LiDAR 首次命中查询所需几何的同时降低仿真开销——在不收缩几何的前提下消除局部伪影。

LiDAR Across Scenes多场景 LiDAR

3DGS · voxelized mesh · simulated LiDAR

For each scene we show, from left to right, the 3DGS visualization, the voxelized mesh, and the LiDAR point cloud rendered by FGGS-LiDAR. The reconstructed geometry retains global structure and sharp boundaries with reliable normals and visibility, yielding consistent surfaces and physically grounded ray interactions across both indoor and outdoor scenes.对每个场景，从左到右依次为 3DGS 可视化、体素化网格，以及由 FGGS-LiDAR 渲染的 LiDAR 点云。重建几何保留了全局结构与清晰边界，并提供可靠的法向与可见性，从而在室内外场景中获得一致的表面与物理可信的光线交互。

Results结果

Fidelity, ablations & real-time performance保真度、消融与实时性能

LiDAR FPS vs Mesh ComplexityLiDAR 帧率 vs 网格复杂度

frame rate

Thanks to the BVH acceleration structure, across all evaluated 3DGS scenes (up to ~6M Gaussian primitives) we observe no systematic degradation of frame rate with primitive count: performance is effectively decoupled from raw cardinality and is instead governed by spatial distribution. FGGS-LiDAR delivers far beyond real-time performance and ranks among the fastest in peer comparisons.得益于 BVH 加速结构，在所有评测的 3DGS 场景（最多约 600 万高斯基元）中，帧率不随基元数量系统性下降：性能基本与原始数量解耦，主要由空间分布决定。FGGS-LiDAR 提供远超实时的性能，在同类对比中名列前茅。

LiDAR Throughput vs Mesh ComplexityLiDAR 吞吐 vs 网格复杂度

throughput

Point throughput grows with sensor density and stays high across mesh complexities. Denser sensors (e.g., OS128, VLP32) exceed 1×10⁸ points/s on lightweight meshes and sustain >7.5×10⁷ points/s even at multi-million-triangle scales — fully exploiting BVH-accelerated parallelism across diverse LiDAR types.点云吞吐随传感器密度增长，并在不同网格复杂度下保持高位。更密集的传感器（如 OS128、VLP32）在轻量网格上超过 1×10⁸ 点/秒，即便在数百万三角形规模下仍维持 >7.5×10⁷ 点/秒，充分发挥了 BVH 加速并行在多种 LiDAR 上的优势。

Fidelity vs. GT Mesh Simulation与 GT 网格仿真的保真度对比

CD / F-score @ 0.05 m

LiDAR激光雷达	Indoor-RealLiDAR		Outdoor-RealLiDAR
LiDAR激光雷达	CD ↓	F-score ↑	CD ↓	F-score ↑
HDL64	0.0034	0.9950	0.0157	0.9816
OS128	0.0034	0.9950	0.0104	0.9867
VLP32	0.0053	0.9918	0.0250	0.9789
Avg.	0.0041	0.9939	0.0170	0.9824

Our simulation closely matches ground-truth LiDAR (centimeter-scale CD, F-score > 0.98) across three sensors on real indoor and outdoor scenes.在真实室内/室外场景、三种传感器下，我们的仿真与真值 LiDAR 高度吻合（CD 为厘米级，F-score > 0.98）。

vs. 3DGS-to-Mesh Baselines (VLP32)与 3DGS-to-Mesh 基线对比（VLP32）

Indoor-COLMAP-1/2 · same pose & beam layout

Method方法	CD ↓	F-score ↑	T (min) ↓
Scene 1场景 1
Ours	0.0089	0.9957	<1
GS2Mesh	0.0126	0.9902	20
MILo (std)	0.0523	0.8910	30
MILo (highres)	0.1647	0.7625	120
Scene 2场景 2
Ours	0.0182	0.9815	<1
GS2Mesh	0.0816	0.9582	20
MILo (std)	0.8206	0.8026	30
MILo (highres)	0.3783	0.8042	120

T is end-to-end 3DGS→mesh time; pose-dependent baselines follow their default protocol (rendering & fusion of 400 COLMAP-posed views). FGGS-LiDAR attains the lowest error and the shortest reconstruction time.T 为端到端 3DGS→网格时间；位姿相关基线遵循其默认流程（渲染并融合 400 个 COLMAP 视角）。FGGS-LiDAR 取得最低误差与最短重建时间。

best最优

Ablation Study消融实验

A: system components · B: post-processing

Variant变体	VoxPts (K)	Faces (M)	T (s) ↓	GPU Mem (GB) ↓	CD ↓	F-score ↑
Block A — system ablations (post-processing fixed)A 部分 —— 系统消融（后处理固定）
Full pipeline	8192.2	1.478	60	22	0.00688	0.9970
w/o surface extraction	21274.7	1.462	90	26	0.00667	0.9975
w/o narrow-band TSDF	8192.2	1.478	>600	22	0.00688	0.9970
w/o flood sign seeding	8192.2	1.964	80	20.65	0.01390	0.9943
w/o LBVH	—	—	—	—	—	—
Block B — post-processing (pipeline fixed)B 部分 —— 后处理（流程固定）
w/o decim., w/o smooth	8192.2	16.707	55	22	0.01554	0.9929
w/o decim., w/ smooth	8192.2	16.761	80	22	0.01816	0.9922
w/ decim., w/o smooth	8192.2	1.478	57	22	0.00820	0.9959

Narrow-band TSDF is essential for throughput (60 s vs. >600 s); decimation is the main driver of low face count and high fidelity. Removing LBVH is infeasible end-to-end (~10³× slower in a micro-benchmark).窄带 TSDF 对吞吐至关重要（60 s 对比 >600 s）；网格抽取是降低面数、提升保真度的主因。端到端移除 LBVH 不可行（微基准中慢约 10³ 倍）。

best最优2nd best次优

Citation引用

BibTeX

@article{wu2025fggs,
  title={FGGS-LiDAR: Ultra-Fast, GPU-Accelerated Simulation from General 3DGS Models to LiDAR},
  author={Wu, Junzhe and Jia, Yufei and Yan, Yiyi and Chen, Zhixing and Tan, Tiao and Wang, Zifan and Wang, Guangyu and Chen, BoKui and Zhou, Guyue},
  journal={arXiv preprint arXiv:2509.17390},
  year={2025}
}