Skip to content
Go to file

Latest commit


Git stats


Failed to load latest commit information.

Reinforcement Learning in CARLA

We release a trained RL agent from the CoRL-2017 paper "CARLA: An Open Urban Driving Simulator". This is only the inference code, the training code is not released yet.

The agent was trained with the asynchronous advantage actor-critic (A3C) algorithm by V. Mnih et al. (2016). We build on this open-source Chainer implementation: .


Tested with:

  • CARLA 0.8.2
  • python 3.6
  • chainer 1.24.0
  • cached-property 1.4.2
  • PIL 5.1.0
  • opencv 3.3.1
  • h5py 2.7.1

In Anaconda, you can create and activate an environment with installed dependencies (except for CARLA) by running:

conda create -n carla_rl python=3.6 chainer=1.24.0 cached-property=1.4.2 pillow=5.1.0 opencv=3.3.1 h5py=2.7.1
source activate carla_rl

To start evaluation on the CoRL-2017 benchmark:

  • Start a CARLA server on town TownXX (Town01 or Town02) and port PORT (this is to be executed in the CARLA server folder):
./ /Game/Maps/TownXX -carla-server -benchmark -fps=10 -windowed -ResX=800 -ResY=600 -carla-world-port=PORT
  • Make sure CARLA client is in your python path, e.g. by running:
export PYTHONPATH=/path/to/CARLA/PythonClient:$PYTHONPATH
  • Run the evaluation:
python --city-name TownXX --port PORT --corl-2017

The results will be stored in _benchmarks_results.


If you use this code in your research, please cite our CoRL 2017 paper:

  title = { {CARLA}: {An} Open Urban Driving Simulator},
  author = {Alexey Dosovitskiy and German Ros and Felipe Codevilla and Antonio Lopez and Vladlen Koltun},
  booktitle = {Proceedings of the 1st Annual Conference on Robot Learning},
  pages = {1--16},
  year = {2017}


Reinforcement learning baseline agent trained with the Actor-critic (A3C) algorithm.




No releases published


No packages published


You can’t perform that action at this time.