This repo contains the inference code and the weights of our paper accepted at CVPR 2020. It's a fork of the repository Learning by Cheating from which we just kept all the code related to the evaluation on the standard CARLA benchmark and on the new released No-Crash benchmark.
We provide a script to install every dependencies needed and download our weights.
# Download CARLA 0.9.6
wget http://carla-assets-internal.s3.amazonaws.com/Releases/Linux/CARLA_0.9.6.tar.gz
mkdir carla_RL_IAs
tar -xvzf CARLA_0.9.6.tar.gz -C carla_RL_IAs
cd carla_RL_IAs
# Download LBC
mv LICENSE LICENSE_CARLA # Conflict with LICENSE from CARLA
git init
git remote add origin https://github.com/marintoro/LearningByCheating.git
git pull origin master
wget http://www.cs.utexas.edu/~dchen/lbc_release/navmesh/Town01.bin
wget http://www.cs.utexas.edu/~dchen/lbc_release/navmesh/Town02.bin
mv Town*.bin CarlaUE4/Content/Carla/Maps/Nav/
# Create conda environment
conda env create -f environment.yml
conda activate carla_RL_IAs
# BE CAREFUL: you need to install pytorch according to your cuda version
conda install pytorch torchvision cudatoolkit=10.1 -c pytorch
#conda install pytorch torchvision cudatoolkit=10.0 -c pytorch
#conda install pytorch==1.1.0 torchvision==0.3.0 cudatoolkit=9.0 -c pytorch
pip install --upgrade pip
pip install pygame
# Install carla client
cd PythonAPI/carla/dist
rm carla-0.9.6-py3.5-linux-x86_64.egg
wget http://www.cs.utexas.edu/~dchen/lbc_release/egg/carla-0.9.6-py3.5-linux-x86_64.egg
easy_install carla-0.9.6-py3.5-linux-x86_64.egg
# Download model checkpoints trained only on Town01/training weathers
wget https://github.com/marintoro/LearningByCheating/releases/download/v1.0/model_RL_IAs_only_town01_train_weather.zip
unzip model_RL_IAs_only_town01_train_weather.zip
# Download model checkpoints used for CARLA challenge
cd ../../..
wget https://github.com/marintoro/LearningByCheating/releases/download/v1.0/model_RL_IAs_CARLA_Challenge.zip
unzip model_RL_IAs_CARLA_Challenge.zip
Then, open up a terminal, inside the carla directory run ./CarlaUE4.sh -fps=10 -benchmark -quality-level=Low
.
Open another terminal and run python benchmark_agent.py --suite=town2 --max-run 100 --path-folder-model model_RL_IAs_only_town01_train_weather/ --render --crop-sky
to see our model driving on test town!
If you want to see our model used for the CARLA challenge you need to run instead
python benchmark_agent.py --suite=town2 --max-run 100 --path-folder-model model_RL_IAs_CARLA_Challenge/ --render
Note that the model we used for the CARLA challenge was trained on a way harder task and on another version of CARLA so the results on the benchmark are lower. On the other hand it handles all towns, including Town03, Town04 and Town05 with US traffic lights!
╔═══════════════════╦══════════════╦═════════╦═══════╗
║ Suite Name ║ Success Rate ║ Total ║ Seeds ║
╠═══════════════════╬══════════════╬═════════╬═══════╣
║ FullTown01-v1 ║ 100 ║ 100/100 ║ 2020 ║
║ FullTown01-v2 ║ 100 ║ 50/50 ║ 2020 ║
║ FullTown01-v3 ║ 100 ║ 100/100 ║ 2020 ║
║ FullTown01-v4 ║ 100 ║ 50/50 ║ 2020 ║
║ FullTown02-v1 ║ 100 ║ 100/100 ║ 2020 ║
║ FullTown02-v2 ║ 100 ║ 50/50 ║ 2020 ║
║ FullTown02-v3 ║ 98 ║ 98/100 ║ 2020 ║
║ FullTown02-v4 ║ 100 ║ 50/50 ║ 2020 ║
║ NoCrashTown01-v1 ║ 100 ║ 100/100 ║ 2020 ║
║ NoCrashTown01-v2 ║ 36 ║ 18/50 ║ 2020 ║
║ NoCrashTown01-v3 ║ 96 ║ 96/100 ║ 2020 ║
║ NoCrashTown01-v4 ║ 34 ║ 17/50 ║ 2020 ║
║ NoCrashTown01-v5 ║ 70 ║ 70/100 ║ 2020 ║
║ NoCrashTown01-v6 ║ 26 ║ 13/50 ║ 2020 ║
║ NoCrashTown02-v1 ║ 99 ║ 99/100 ║ 2020 ║
║ NoCrashTown02-v2 ║ 24 ║ 12/50 ║ 2020 ║
║ NoCrashTown02-v3 ║ 87 ║ 87/100 ║ 2020 ║
║ NoCrashTown02-v4 ║ 34 ║ 17/50 ║ 2020 ║
║ NoCrashTown02-v5 ║ 42 ║ 42/100 ║ 2020 ║
║ NoCrashTown02-v6 ║ 18 ║ 9/50 ║ 2020 ║
║ StraightTown01-v1 ║ 100 ║ 100/100 ║ 2020 ║
║ StraightTown01-v2 ║ 100 ║ 50/50 ║ 2020 ║
║ StraightTown02-v1 ║ 100 ║ 100/100 ║ 2020 ║
║ StraightTown02-v2 ║ 100 ║ 50/50 ║ 2020 ║
║ TurnTown01-v1 ║ 100 ║ 100/100 ║ 2020 ║
║ TurnTown01-v2 ║ 100 ║ 50/50 ║ 2020 ║
║ TurnTown02-v1 ║ 100 ║ 100/100 ║ 2020 ║
║ TurnTown02-v2 ║ 100 ║ 50/50 ║ 2020 ║
╚═══════════════════╩══════════════╩═════════╩═══════╝
The results there are way below the one above, but note that it was trained on much harder tasks and that it works also on US towns, i.e. Town03, Town04 and Town05.
COMING SOON
This repo is released under the MIT License (please refer to the LICENSE file for details). Most of the code come from the repository Learning by Cheating which is under MIT license. Part of the PythonAPI and the map rendering code is borrowed from the official CARLA repo, which is under MIT license.