Swarm-Gen: Fast Generation of Diverse Feasible Swarm Behaviors

This repository contains the source code to reproduce the experiments in our paper: "Swarm-Gen: Fast Generation of Diverse Feasible Swarm Behaviors" where we combine a generative model (CVAE/VQ-VAE) with a safty filter (SF) to learn to generate diverse swarm behavior trajectories given a fixed start and goal points.

Getting Started

1. Clone this repository:

git clone https://github.com/cisimon7/SwarmGen.git
cd SwarmGen

2. Create a conda environment and install the dependencies:

conda create -n venv python=3.8
conda activate venv
pip install -r requirements.txt

3. Create a directory resources

4. Download the weights for our trained Models and save to the resources directory

5. Download the train and test dataset and also save to the resources directory.

6. You should be able to run any of the train and/or test norebooks. Note that for num_agent=4 or num_agent=8, the variables lwidth and lheight should both be set to 5, while for num_agent=16, they are set to 10. This applies for all the notebooks.

Visualization

Visualization of diverse trajectories generated using the VQ-VAE model for a given fixed start and goal points.