This repository contains a software package to solve motion planning problems on CommonRoad using Reinforcement Learning algorithms. We currently use the implementation for the RL algorithms from Stable Baselines, but the package can be run with any standard (OpenAI Gym compatible) RL implementations.
import gym
import commonroad_rl.gym_commonroad
# kwargs overwrites configs defined in commonroad_rl/gym_commonroad/configs.yaml
env = gym.make("commonroad-v1",
action_configs={"action_type": "continuous"},
goal_configs={"observe_distance_goal_long": True, "observe_distance_goal_lat": True},
surrounding_configs={"observe_lane_circ_surrounding": True, "lane_circ_sensor_range_radius": 100.},
reward_type="sparse_reward",
reward_configs_sparse={"reward_goal_reached": 50., "reward_collision": -100})
observation = env.reset()
for _ in range(500):
# env.render() # rendered images with be saved under ./img
action = env.action_space.sample() # your agent here (this takes random actions)
observation, reward, done, info = env.step(action)
if done:
observation = env.reset()
env.close()commonroad-rl
├─ commonroad_rl
│ ├─ doc # Folder for documentation
│ ├─ gym_commonroad # Gym environment for CommonRoad scenarios
| ├─ action # Action and Vehicle modules
| ├─ observation # Observation modules
| ├─ reward # Reward and Termination modules
| ├─ utils # Utility functions for gym_commonroad
│ ├─ configs.yaml # Default config file for actions, observations, rewards, and termination conditions,
as well as for observation space optimization and reward coefficient optimization
│ ├─ commonroad_env.py # CommonRoadEnv(gym.Env) class
│ └─ constants.py # Script to define path, vehicle, and draw parameters
│ ├─ hyperparams # Config files for default hyperparameters for various RL algorithms
│ ├─ tests # Test system of commmonroad-rl.
│ ├─ tools # Tools to validate, visualize and analyze CommonRoad .xml files, as well as preprocess and convert to .pickle files.
│ ├─ utils_run # Utility functions to run training, tuning and evaluating files
│ ├─ README.md
│ ├─ evaluate_model.py # Script to evaluate a trained RL model on specific recorded scenarios and visualize the scenario
│ ├─ evaluate_model_sumo.py # Script to evaluate a trained RL model on interactive scenarios and visualize the scenario
│ ├─ generate_solution.py # Script to genearte CommonRoad solution files from trained RL models.
│ ├─ train_model.py # Script to train RL model or optimize hyperparameters or environment configurations
│ ├─ sensitivity_analysis.py # Script to run sensitivity analysis for a trained model
│ └─ plot_learning_curves.py # Plot learning curves with provided training log files.
├─ scripts # Bash scripts to install all dependencies, train and evaluate RL models, as well as generate CommonRoad solution files from trained RL models.
├─ README.md
├─ environment.yml
└─ setup.py
Detailed instructions under ./commonroad_rl/install_docker/readme_docker.md
This project should be run with conda. Make sure it is installed before proceeding with the installation.
- download & install conda, and init anaconda to work from the terminal. tested on conda 4.5; 4.9, 4.10
~/anaconda3/bin/conda init
# for minconda
~/miniconda3/bin/conda init
- clone this repository
git clone https://gitlab.lrz.de/tum-cps/commonroad-rl.git
- install build packages
sudo apt-get update
sudo apt-get install build-essential make cmake
- setup a new conda env (or install packages to an existing conda env e.g. myenv
conda env update --name myenv --file environment.yml)
conda env create -n cr37 -f environment.yml
git submodule init
git submodule update --recursive || exit_with_error "Update submodules failed"
- (optional) install pip packages for the docs. If you want to use the jupyter notebook, also install jupyter.
source activate cr37
pip install -r commonroad_rl/doc/requirements_doc.txt
conda install jupyter
Simply run
bash scripts/install.sh -e cr37
cr37 to be replaced by the name of your conda environment if needed.
If you have the necessary libraries installed already, you can also install without sudo rights. Simply run
bash scripts/install.sh -e cr37 --no-root
cr37 to be replaced by the name of your conda environment if needed. Press ctrl+c to skip when asked for sudo password.
This script will build all software in your home folder.
Please ask your admin to install necessary libraries for you if they are missing. The command to install all necessary libraries for mpi, building packages, and the drivability checker is:
sudo apt-get install -y libopenmpi-dev build-essential unzip cmake libboost-dev libboost-thread-dev libboost-test-dev libboost-filesystem-dev libeigen3-dev libcgal-dev xutils-dev libfcl-dev libomp-dev libgmp-dev libglu1-mesa-dev
Further details of our test system refer to ./commonroad_rl/tests.
source activate cr37
pytest commonroad_rl/tests --scope unit module -m "not slow"
To get to know the package, please check ./commonroad_rl/tutorials for further details.
The commonroad_rl folder contains the source files. There are Python scripts for training, evaluating, and visualizing models. The most important scrips are explained in ./commonroad_rl/README.md and can be run with your Python executable. They are especially useful if you are developing a new feature or want to debug a specific part of the training.
If you tested your codes already and everything runs smoothly on your computer and you now want to run the real experiments on larger dataset, the bash scripts help you with that. The are located in ./scripts. They can be used for training with PPO and TD3 and testing an agent. Always adapt the specific paths in the scripts to the corresponding paths on your machine and check the comments in the file to determine which arguments have to be provided.
Please use development branch or open a new feature_xxx branch to make contribution.
- OpenAI Stable Baselines: the implementation of RL algorithms used in our project.
- OpenAI Spinning Up: we do not use their implementations in our project. But they provide quite nice explanations of RL concepts.
- OpenAI Gym: general interface.
- OpenAI Safety Gym: a special collection of Gyms for safe RL. Configurable as our project.
If you use CommonRoad-RL in your paper, please cite:
@inproceedings{Wang2021,
author = {Xiao Wang and Hanna Krasowski and Matthias Althoff},
title = {{CommonRoad-RL}: A Configurable Reinforcement Learning Environment for Motion Planning of Autonomous Vehicles},
booktitle = {Proc. of the IEEE International Conference on Intelligent Transportation Systems (ITSC)},
year = {2021},
}