TorchGRL is the source code for our paper Graph Convolution-Based Deep Reinforcement Learning for Multi-Agent Decision-Making in Mixed Traffic Environments for IV 2022.TorchGRL is a modular simulation framework that integrates different GRL algorithms and SUMO simulation platform to realize the simulation of multi-agents decision-making algorithms in mixed traffic environment. You can adjust the test scenarios and the implemented GRL algorithm according to your needs.


Before starting to carry out some relevant works on our framework, some preparations are required to be done.


Our framework is developed based on a laptop, and the specific configuration is as follows:

  • Operating system: Ubuntu 20.04
  • RAM: 32 GB
  • CPU: Intel (R) Core (TM) i9-10980HK CPU @ 2.40GHz
  • GPU: RTX 2070

It should be noted that our program must be reproduced under the Ubuntu 20.04 operating system, and we strongly recommend using GPU for training.

Development Environment

Before compiling the code of our framework, you need to install the following development environment:

  • Ubuntu 20.04 with latest GPU driver
  • Pycharm
  • Anaconda
  • CUDA 11.1
  • cudnn-11.1,


Please download our GRL framework repository first:

git clone

Then enter the root directory of TorchGRL:

cd TorchGRL

and please be sure to run the below commands from /path/to/TorchGRL.

Installation of FLOW

The FLOW library will be firstly installed.

Firstly, enter the flow directory:

cd flow

Then, create a conda environment from flow library:

conda env create -f environment.yml

Activate conda environment:

conda activate TorchGCQ

Install flow from source code:

python develop

Installation of SUMO

SUMO simulation platform will be installed. Please make sure to run the below commands in the “TorchGRL” virtual environment.

Install via pip:

pip install eclipse-sumo

Setting in Pycharm:

In order to adopt SUMO correctly, you need to define the environment variable of SUMO_HOME in Pycharm. The specific directory is:


Setting in Ubuntu:

At first, run:

gedit ~/.bashrc

then copy the path name of SUMO_HOME to “~/.bashrc”:

export SUMO_HOME=“/home/…/.conda/envs/TorchGCQ/lib/python3.7/site-packages/sumo”

Finally, run:

source ~/.bashrc

Installation of Pytorch and related libraries

Please make sure to run the below commands in the “TorchGRL” virtual environment.

Installation of Pytorch:

We use Pytorch version 1.9.0 for development under a specific version of CUDA and cudnn.

pip3 install torch==1.9.0+cu111 torchvision==0.10.0+cu111 torchaudio==0.9.0 -f

Installation of pytorch geometric:

Pytorch geometric is a Graph Neural Network (GNN) library upon Pytorch

pip install torch-scatter torch-sparse torch-cluster torch-spline-conv torch-geometric -f

Installation of pfrl library

Please make sure to run the below commands in the “TorchGRL” virtual environment.

pfrl is a deep reinforcement learning library that implements various algorithms in Python using PyTorch.

Firstly, enter the pfrl directory:

cd pfrl

Then install from source code:

python develop


flow folder

The flow folder is the root directory of the library after the FLOW library is installed through source code, including interface-related programs between DRL algorithms and SUMO platform.

Flow_Test folder

The Flow_Test folder includes the related programs of the test environment configuration; specifically, is the core python program. If the program runs successfully, the environment configuration is successful.

pfrl folder

The pfrl folder is the root directory of the library after the deep reinforcement learning pfrl library is installed through source code, including all DRL related programs. The source program can be modified as needed.

GRLNet folder

The GRLNet folder contains the GRL neural network built in the Pytorch environment. You can modify the source code as needed or add your own neural network.

  • constructs the fundamental neural network of GRL algorithms
  • constructs the dueling network of GRL algorithms

GRL_utils folder

The GRL_utils folder contains basic functions such as model training and testing, data storage, and curve drawing.

  • contains the training and testing functions for the GRL model.
  • is the function to plot the training data curve.
  • is the function to process the test data.
  • Fig folder stores the training data curve.
  • Logging_Training folder stores the training data generated by different GRL algorithms.
  • Logging_Test folder stores the testing data generated by different GRL algorithms.

GRL_Simulation folder

The GRL_Simulation folder is the core of our framework, which contains the core simulation program and some related functional programs.

  • is the main program, containing the definition of FLOW parameters, as well as the controlling (start and end) of the simulation.
  • is the definition of vehicle control model based on FLOW library.
  • is the core program to build and initialize the simulation environment of SUMO.
  • defines the road network.
  • registers the simulation environment of SUMO to the gym library to realize the connection with GRL algorithms.
  • defines the elements in MDPs, including state representation, action space and reward function.
  • Experiment folder is the core program of co-simulation under different GRL algorithms, including the initialization of the simulation environment, the initialization of the neural network, the training and testing of GRL algorithms, and the preservation of the training and testing results.
  • GRL_Trained_Models folder stores the trained GRL model when the training process ends.


You can simply run “” in Pycharm to simulate the GRL algorithm, and observe the simulation process in SUMO platform. You can generate training plot such as Reward curve:

Verification of other algorithms

If you want to verify other algorithms, you can develop the source code as needed under the “Experiment folder”, and don’t forget to change the imported python script in “”. In addition, you can also construct your own network in GRLNet folder.

Verification of other traffic scenario

If you want to verify other traffic scenario, you can define a new scenario in “”. You can refer to the documentation of SUMO for more details .


View Github