README.md

March 29, 2026 · View on GitHub

Compact Keyframe-Optimized Multi-Agent Gaussian Splatting SLAM

Monica Li · Pierre-Yves Lajoie · Jialing Liu · Giovanni Beltrame

⚙️ Setting Things Up

Clone the repo:

git clone https://github.com/lemonci/Coko-SLAM.git

We tested the installation with gcc and g++ of versions 10, 11 and 12. Also, make sure that nvcc --version matches nvidia-smi version.

Run the following commands to set up the environment

conda create -n coko-slam python=3.11
conda activate coko-slam

Install pytorch:

# CUDA 11.8
conda install pytorch==2.5.1 torchvision==0.20.1 pytorch-cuda=11.8 -c pytorch -c nvidia
# CUDA 12.1
conda install pytorch==2.5.1 torchvision==0.20.1 pytorch-cuda=12.1 -c pytorch -c nvidia
# CUDA 12.4
conda install pytorch==2.5.1 torchvision==0.20.1 pytorch-cuda=12.4 -c pytorch -c nvidia

Install other dependencies:

conda install -c conda-forge faiss-gpu=1.8.0
pip install -r requirements.txt

In case you are on a local machine and want to run visualization script run:

pip install rerun-sdk

We tested our code on RTX4070 and RTX H100 GPUs with Ubuntu22 and AlmaLinux 9 respectively.

🔨 Running Coko-SLAM

Here we elaborate on how to load the necessary data, configure Coko-SLAM for your use-case, debug it, and how to reproduce the results mentioned in the paper.

Getting the Data

We tested our code on ReplicaMultiagent and AriaMultiagent datasets. Make sure to install git lfs and hugging face cli before proceeding.

ReplicaMultiagent was created by CP-SLAM authors. However, it is a bit tricky to find it on the web. Therefore, we uploaded it to HF datasets for easier access. Install git lfs and download it by running:
git lfs install
git clone https://huggingface.co/datasets/voviktyl/ReplicaMultiagent

AriaMultiagent consists of clips from AriaDigitalTwin. First, download raw AriaDigitalTwin sequences following instructions here. We recommend using the data tools described here. You can find the names of the raw videos we used in prepare_aria_room_data.py in DATASET_DICT. After you have downloaded the raw videos, process them using prepare_aria_room_data.py. This will create a folder with the format matching our repository.

Running the code

Ideally, our system needs n + 1 GPUs where n is the nubmer of agents. If you want to run the system for debugging purposes set multi_gpu: False and agent_ids: [0]. In this way, you will run a single agent and use the same GPU for the server and the agent. Start the system by running:

python run_slam.py configs/<dataset_name>/<config_name> --input_path <path_to_the_scene> --output_path <output_path>

For example:

python run_slam.py configs/AriaMultiagent/room0.yaml --input_path <path_to>/AriaMultiagent/room0 --output_path output/AriaMultiagent/room0

In addition, you can set up wandb to log the results of the runs.

Reproducing Results

While we tried to make all parts of our code deterministic, differential rasterizer of Gaussian Splatting is not. The metrics can be slightly different from run to run. In the paper we report average metrics that were computed over three seeds: 0, 1, and 2.

You can reproduce the results for a single scene by running:

python run_slam.py configs/<dataset_name>/<config_name> --input_path <path_to_the_scene> --output_path <output_path>

If you are running on a SLURM cluster, you can reproduce the results for all scenes in a dataset by running the script:

./scripts/reproduce_aria_sbatch.sh

Demo

To create the demo we first visualized the run in an offline fashion using Rerun. Further, we manually added the titles to the video. You can create the visualization by downloading the folder with the results and running run_visualization.py on top of it.

🙏 Acknowledgments

This codebase is based on MAGiC-SLAM. We thank Vladimir Yugay, for his detailed explaination and the fruitful discussions.