RGD-SLAM: Robust Gaussian Splatting SLAM for Dynamic Environments

Overview of RGD-SLAM: our system is designed to estimate camera pose in dynamic environments and reconstruct static scenes from sequences of RGB-D frames. It consists of two main components: a front-end tracking and a back-end mapping. The frontend generates a motion mask for each frame and uses the adaptive weight to optimize the camera pose. The backend uses a visibility-aware keyframing strategy and maintains a sliding window, optimizing the static 3DGS scene representation comprehensively.

Installation

You can create an anaconda environment called ismap. Please install libopenexr-dev before creating the environment.

conda env create -f environment.yaml

We recommend following the MonoGS method for SLAM environment configuration．

Then you will then need to install OneFormer to use the segmentation network. We recommend installing it from here.

Download Dataset

You can download the data as below.

bash scripts/download_tum.sh

Run

After downloading the dataset, you can run RGD-SLAM:

python slam.py --config configs/rgbd/tum/fr3_walking_halfsphere.yaml

The system defaults to performing single-threaded tracking and mapping. Dual-threaded tracking and mapping is not currently supported and is planned to be implemented in the next version.

Evaluation

To evaluate the average trajectory error. Run the command below with the corresponding config file:

python slam.py --config configs/rgbd/tum/fr3_walking_halfsphere.yaml --eval

This flag will automatically run system, and log the results including the rendering metrics.

Acknowledgement

Thanks to previous open-sourced repo: MonoGS, DG-SLAM, OneFormer, dotmask

Citing

If you find our work useful, please consider citing:

@article{WANG2026113071,
title = {RGD-SLAM: Robust Gaussian splatting SLAM for dynamic environments},
journal = {Pattern Recognition},
volume = {175},
pages = {113071},
year = {2026},
issn = {0031-3203}
}