- Custom CARLA maps that mimic real-world roads and human driver behaviors from NGSim dataset (I80 and US101)
- We've trained and benchmarked policies on real-world lane change maneuvers from NGSim dataset
- We provide the source code for running NGSim-based scenarios in CARLA. Scenario interface is similar to openai gym interface
- Download and extract CARLA (0.9.6 download link). Then, add PythonAPI wheel to your
PYTHONPATH
:export CARLA_ROOT=/path/to/your/carla/release-folder export PYTHONPATH=$CARLA_ROOT/PythonAPI/carla/dist/carla-0.9.6-py3.6-linux-x86_64.egg:$PYTHONPATH
- You also need to install our asset packages with two new CARLA maps
- Download our archive: Google Drive download link
- Move the archive to:
$CARLA_ROOT/Import
- Ingest into CARLA release:
cd Import && ./ImportAssets.sh
./CarlaUE4.sh -benchmark -fps=10
# (wait until server loads)
python example/example_roundabout_scenario_usage.py
# (wait until scenario script connects successfully, map rendering may tak a while)
python example/manual_driving.py --res 900x500
Code tested with CARLA 0.9.6.
- Download dataset: Google Drive download link
- Unpack:
tar xf xy-trajectories.tgz
- Adjust
data_dir
in source code with your path toxy-trajectories
python example/example_scenario_usage.py
- shows how to run scenario in training looppython example/example_replay_ngsim_in_carla.py
- shows how to replay NGSim dataset in CARLA. It was used to generated GIF in this README file
Code for interfacing with NGSim dataset was based on https://github.com/Atcold/pytorch-PPUU