We aim to combine real-world motivated RL with easy accessibility within a highly relevant problem: the stochastic longitudinal control of an autonomous vehicle. The LongiControl environment consists of a data based electric vehicle model (the Downloadable Dynamometer Database D3 of the Argonne National Laboratory is used here) and a single-lane track with stochastic speed restrictions. The state of the agent includes the actual speed, previous acceleration, current speed limit and the next speed limit as long as it is within a visual range of 150m. The agent selects the acceleration of the vehicle and receives as a reward a combination of speed, energy consumption, jerk and a measure for speeding. LongiControl could be used to elaborate various challenges within Reinforcement Learning. E.g. MORL due to excplicitly contradictory reward terms (minimize energy consumption, travel time, jerk), SafeRL (comply with speed limits) or Explainable AI (as the problem is quite easy to grasp; also see our study on this here).

Please use this bibtex if you want to cite this repository in your publications:

@conference{icaart21,
  author={Dohmen, Jan and Liessner, Roman and Friebel, Christoph and Bäker, Bernard},
  title={LongiControl: A Reinforcement Learning Environment for Longitudinal Vehicle Control},
  booktitle={Proceedings of the 13th International Conference on Agents and Artificial Intelligence - Volume 2: ICAART,},
  year={2021},
  pages={1030-1037},
  publisher={SciTePress},
  organization={INSTICC},
  doi={10.5220/0010305210301037},
  isbn={978-989-758-484-8},
}

• Python 3

• OpenAI Gym

• Numpy, Scikit-learn, Pandas

• Pytorch to train/use the given RL agent

• Pyglet, Matplotlib, Seaborn for visualization

cd gym-longicontrol
pip install -e .

Instances of the environment can be created and handled similar to other Gym environments:

• gym.make('gym_longicontrol:DeterministicTrack-v0')
• gym.make('gym_longicontrol:StochasticTrack-v0')

cd gym-longicontrol/rl/pytorch
python main.py --save_id 99

A trained agent is given in gym-longicontrol/rl/pytorch/out:

• .tar ... model and weights
• .out ... quick overview of the training results
• .npy ... more detailed information about the course of training (can be used within the jupyter notebook gym-longicontrol/rl/pytorch/monitor.ipynb)

Load the trained model an visualize an example track:

cd gym-longicontrol/rl/pytorch
python main.py --load_id 9 --env_id StochasticTrack-v0 -vis

It is also possible to save it as mp4 video:

cd gym-longicontrol/rl/pytorch
python main.py --load_id 9 --env_id DeterministicTrack-v0 -vis -rec