FGGS-LiDAR is a geometry-first, plug-and-play framework that bridges 3D Gaussian Splatting (3DGS) and high-performance LiDAR simulation. It converts any pretrained 3DGS asset into a watertight mesh — directly from Gaussian parameters, with no LiDAR-specific supervision or architectural changes — and then performs BVH-accelerated ray-casting that simulates LiDAR returns at 500+ FPS across up to 4096 parallel environments. In large-scale parallel settings, FGGS-LiDAR achieves an order-of-magnitude lower LiDAR-simulation latency than Isaac Sim.
- Geometry-first, pose-free. 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.
- Efficient 3DGS → geometry. 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.
- Massively parallel 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.
- High fidelity. Centimeter-scale Chamfer Distance and F-score > 0.98 against ground-truth LiDAR across indoor and outdoor scenes, with lower LiDAR-simulation error than existing 3DGS-to-mesh baselines.
| 3DGS → Voxelized Mesh → Simulated LiDAR | Denoised Watertight Mesh |
|---|---|
 |
 |
| LiDAR FPS vs Mesh Complexity | LiDAR Throughput vs Mesh Complexity |
|---|---|
 |
 |
- 3DGS → voxel → TSDF mesh conversion pipeline
- GPU-accelerated LiDAR simulation & evaluation
- Isaac / Gym integration (README_ISAAC.md)
- Example datasets & evaluation reproduction packages (evaluation/)
- Pretrained scene assets for more scenes
- More LiDAR sensor models and presets
- Documentation & tutorials
- Linux with an NVIDIA GPU (CUDA 11.8 recommended)
- Python 3.9
- PyTorch 2.3.0 (+cu118), CuPy (cuda11x), Taichi 1.7.3
We recommend a clean conda environment:
conda create -n fggs python=3.9 -y
conda activate fggsStep 1 — GPU-enabled PyTorch and CuPy. For CUDA 11.8:
pip install torch==2.3.0+cu118 torchvision==0.18.0+cu118 torchaudio==2.3.0 \
--index-url https://download.pytorch.org/whl/cu118
pip install "cupy-cuda11x>=11.0"Step 2 — Base dependencies and local submodules.
pip install -r requirements.txtVerify the install:
python -c "import torch, cupy, taichi; print('Torch:', torch.__version__, 'CuPy:', cupy.__version__, 'Taichi:', taichi.__version__)"Expected output (example):
Torch: 2.3.0+cu118
CuPy: 13.6.0
Taichi: 1.7.3
We provide ready-to-use bash scripts for the full pipeline.
bash convert.shThis runs the demo config (gs2mesh/voxel2mesh/configs/pure.yaml, 0.05 m voxels,
post-processing off) and writes output/mesh/out_recon.obj — a quick, coarse
reconstruction meant to exercise the pipeline end-to-end. It is not the
paper-grade mesh; for that see the note below.
bash evaluate.shevaluate.sh scans output/mesh/reconstruction.obj (shipped in data.zip) —
our best mesh, produced with higher-quality settings than the demo config (finer
voxels plus decimation and Laplacian smoothing post-processing), which reproduces
the paper's Outdoor/VLP-32 numbers (CD 0.025, F-score 0.979). The coarse
out_recon.obj from convert.sh above is a demo output and scores well below this.
In addition to the core modules, the project ships LiDAR simulation extensions for physics-based environments:
- Gym Demo (
examples/gym_demo.py) — a lightweight demonstration based on IsaacGym's Python bindings, useful for quick visualization and testing. - Isaac Demo (
examples/isaac_demo.py) — a more complete IsaacGym integration with multi-sensor setups and larger simulation scenes.
Both demos share the same LiDAR models and API as the main package. See README_ISAAC.md for environment setup.
Testing examples are provided as a GitHub release asset: Download data.zip
Contents of data.zip:
data/gt.urdf— URDF model used for Gym simulationdata/outdoor_gt.obj— ground-truth mesh of the outdoor scenedata/outdoor_scene.ply— ground-truth 3DGS representationoutput/mesh/reconstruction.obj— our best reconstructed mesh, produced with higher-quality settings than the demo config (finer voxels plus decimation and Laplacian smoothing); this is the meshevaluate.shscans. (convert.sh'sout_recon.objis a separate, coarser demo output.)
Unzip the archive from the repository root — it preserves the data/ and output/mesh/ paths, so the files land in place automatically:
unzip data.zipThen run convert.sh and evaluate.sh.
evaluation/ holds reproduction packages for the paper's benchmarks and
method comparisons:
evaluation/speed/— batched multi-environment LiDAR rendering latency (256–4096 envs × 1k–32k beams).evaluation/accuracy/— full-scene LiDAR-simulation accuracy (Chamfer Distance, F-score, VLP-32) of Ours vs. GS2Mesh and MILo on two indoor scenes.evaluation/method_comparison/— object-level geometry comparison (HDL-64) — FGGS (unsupervised) vs the LiDAR-supervised methods GS-LiDAR and LiDAR-RT, showing FGGS matches 64-frame LiDAR-supervised methods with zero LiDAR. Bundled third-party outputs reproduce their columns without installing or retraining the baselines.
- DISCOVERSE — Efficient robot simulation in complex high-fidelity environments.
- GS-Playground — A high-throughput photorealistic simulator for vision-informed robot learning.
Released under the MIT License.
If you find FGGS-LiDAR useful, please cite:
@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}
}