Multi Agent LMPC Examples

This repository collects the multi-agent LMPC examples shown in the folloing publications

Zhu, E., Stürz, Y., Rosolia, R., Borrelli, F., "Trajectory Optimization for Nonlinear Multi-Agent Systems using Decentralized Learning Model Predictive Control", Presented at Conference on Decision and Control 2020. [arXiv]

The authors are with the Model Predictive Control Lab at UC Berkeley

Edward Zhu (edward.zhu@berkeley.edu)
Yvonne Stürz (y.stuerz@berkeley.edu)

Directory Structre

decentralized_LMPC collects examples relating to the first paper above where a decentralized LMPC policy is implemented in multiple multi-agent navigation tasks. In this approach, synthesis of the decoupled MPC policies is done in a centralized manner between iterations of task execution. Online execution of the decoupled policies can be done completely in parallel with no communication between agents

3_agent_nl_demo: A simple navigation task for 3 car-like agents navigating an intersection
3_agent_nl_centralized_demo: The above task with centralized policy synthesis and control (for comparison with the decentralized approach)
multi_agent_nl_demo: A more complex task with 10 agents performing collision free manuevers in a shared space. Agents are initialized in a circle
multi_agent_rand_nl_demo: A more complex task with 10 agents performing collision free manuevers in a shared space. Agents are initialized in arbitrary positions

Native/VirtualEnv Installation

The code is written in Python 3 and the following packages are required

sudo apt install coinor-libipopt-dev libblas-dev ffmpeg
sudo python3 -m pip install cvxpy ipopt casadi scikit-learn

Docker Installation

We also provide Dockerfiles and scripts to build a Docker image and instantiate containers to run the code in this repository. To do so, first we need to install Docker CE. Instructions for installing via the apt repository for Ubuntu are included below. Instructions for MacOS can be found here.

Ubuntu

# Update apt package index
sudo apt update

# Install packages to allow apt to use a repository over HTTPS
sudo apt install -y apt-transport-https \
  ca-certificates \
  curl \
  software-properties-common

# Add Docker's official GPG key
curl -fsSL https://download.docker.com/linux/ubuntu/gpg | sudo apt-key add -

# Add Docker repository
sudo add-apt-repository \
  "deb [arch=amd64] https://download.docker.com/linux/ubuntu \
  $(lsb_release -cs) \
  stable"

# Update apt package index
sudo apt update

# Install Docker CE
sudo apt install docker-ce

# Add user to docker group
export user=$(whoami)
sudo usermod -aG docker $user

Docker related files are included in the docker directory. The bash script start_LMPC_docker.sh builds the image, which extends the continuumio/anaconda image by adding the cvxpy and ipopt Python libraries.

Running with Docker

In order to (re)build the image and start the container, run the following command from the base directory of the repository:

bash docker/start_LMPC_docker.sh -r

The -r flag triggers a rebuild of the image. There is no harm in always including this flag to ensure that changes to the Dockerfile are implemented. If the -r flag is not specified, a container will be instantiated using the most recent build of the image.

Once the container is running, open up a new terminal and use the following command to start a bash session in the container (you can get the name of the container using docker container ls):

docker exec -it ${container_name} bash

We can then run our code within this environment.

Important Notes

Upon container instantiation the base directory of the repository is mounted to /home/LMPC within the container
python-tk is installed and TkAgg needs to be selected as the matplotlib backend as opposed to the default Qt5Agg (which doesn't seem to work). This can be done by including the following code once before the first time matplotlib.pyplot is imported:

import matplotlib
matplotlib.use('TkAgg')

A Jupyter notebook server can be started within the container using the command:

/opt/conda/bin/jupyter notebook --notebook-dir=/home/LMPC --ip=0.0.0.0 --port=8888 --allow-root

Once it is up and running, enter the following URL into a web browser:

localhost:8888/?token=[server_token]