janapavlasek/multi-robot-svbp

Code for the multi-robot SVBP paper.

Stein Variational Belief Propagation for Multi-Robot Coordination

This repository is an application using torch-bp. The code contains Python scripts that run experiments for multi-robot collision tests in different environments using the belief propagation algorithms in torch-bp. This code accompanies the paper Stein variational belief propagation for multi-robot coordination. Please cite us if you use this code for your research.

Visit Project Webpage for more details about the project.

Installation

prerequisite:

• Python >=3.8 (tested with 3.8.18)
• PyTorch >=2.0 (tested with up to v2.2 with CUDA 12.1)
• torch-bp (follow link to install)

After clone the repo, under /multi-robot-svbp run:

pip install -e .

Recommend installation:

# create folder for svbp project
mkdir svbp_workspace
cd svbp_workspace

# create python virtual environment
python3 -m venv svbp_venv
source svbp_venv/bin/activate
######### TODO: install suitable pytorch version #########
# for 12.1 eg. pip3 install torch torchvision torchaudio #

# clone the necessary repo
git clone https://github.com/janapavlasek/torch-bp.git
git clone https://github.com/janapavlasek/multi-robot-svbp.git

# install torch-bp
pip install -e torch-bp/
pip install -e multi-robot-svbp/

# may need additional package
pip install matplotlib

Usage

The main experiment scripts are under scripts folder.

Basic use of the scripts:

# in the folder /multi-robo-svbp
# run stein mpc controller for point robot in simple scene
python scripts/point_robot_stein_mpc.py

# recommend to use cuda when running experiments with gabp or svbp
# use save flag to save images of the experiments during each step
python scripts/run_collision_avoid_gabp.py --cuda --save
python scripts/run_collision_avoid_svbp.py --cuda --save

All experiment outputs (data, visualization) will be sotred under ouput folder.

To use a different scene/map of the experiment, use the --scene flag and specify a .yml file in data/scenes folder. For example,

# run the experiment using gabp in the map 'swap_rooms'
python scripts/run_collision_avoid_gabp.py --cuda --save --scene multi-robot-svbp/data/scenes/swap_rooms.yml

Citation

This code accompanies the paper Stein variational belief propagation for multi-robot coordination (Robotics and Automation Letters, 2024). If you use it in your research, please cite:

@inproceedings{pavlasek2024stein,
  title={Stein Variational Belief Propagation for Multi-Robot Coordination},
  author={Pavlasek, Jana and Mah, Joshua Jing Zhi and Xu, Ruihan and Jenkins, Odest Chadwicke and Ramos, Fabio},
  booktitle={Robotics and Automation Letters (RA-L)},
  year={2024}
}

multi-robot-svbp's Issues

HIgh memory usage from `jacrev` operations

• run_collision_avoid_svbp.py example uses alot of memory
  

• proposed problem linked to jacrev mapping every single output to every single input which is not required for our usage in BatchKernel

Proposed fix:

• implement manual gradients that uses significantly less memory