GARAD-SLAM: 3D GAussian splatting for Real-time Anti Dynamic SLAM

ICRA 2025

^†Mingrui Li¹, ^†Weijian Chen², Na Cheng¹, Jingyuan Xu¹, Dong Li³ and Hongyu Wang^1∗

Paper | Video

intro.mp4

Abstract

GARAD-SLAM is a real-time 3DGS-based SLAM system tailored for dynamic scenes. It directly performs dynamic segmentation on Gaussians and map them back to the front-end to obtain dynamic point labels through a Gaussian pyramid network, achieving precise dynamic removal and robust tracking. It imposes rendering penalties on dynamically labeled Gaussians updated through the network to avoid irreversible erroneous removal caused by simple pruning. We utilize TUM RGB-D and BONN RGB-D datasets to evaluate the performance of our algorithm.

Build

Source code will be released soon.

Prerequisites

We build the project on Ubuntu 20.04 LTS, and the compiling is similar to Photo-SLAM and LC-CRF-SLAM, please refer to them for more details. We extend our sincere gratitude for their outstanding contributions.

Dependencies	Tested with
OS	Ubuntu 20.04 LTS
gcc	3.27.5
CUDA	10.1
Torch	2.0.1
cuDNN	8.9.3

Dependencies

sudo apt install libeigen3-dev libboost-all-dev libjsoncpp-dev libopengl-dev mesa-utils libglfw3-dev libglm-dev

Installation

git clone https://github.com/DrLi-Ming/GARAD-SLAM
cd GARAD-SLAM
chmod +x ./build.sh
./build.sh

License

GARAD-SLAM is released under the GNU General Public License v3.0.

Citation

If you find our work useful, please kindly cite us:

@INPROCEEDINGS{11128757,
  author={Li, Mingrui and Chen, Weijian and Cheng, Na and Xu, Jingyuan and Li, Dong and Wang, Hongyu},
  booktitle={2025 IEEE International Conference on Robotics and Automation (ICRA)}, 
  title={GARAD-SLAM: 3D Gaussian Splatting for Real-Time Anti Dynamic SLAM}, 
  year={2025},
  pages={11047-11053}
}