Learning by Cheating

This repo is the implemention of paper Learning by Cheating in CARLA 0.9.6.

Learning by Cheating
Dian Chen, Brady Zhou, Vladlen Koltun, Philipp Krähenbühl,
Conference on Robot Learning (CoRL 2019)
arXiv 1912.12294

Code in CARLA 0.9.5 is coming soon.

Reference

If you find this repo to be useful in your research, please consider citing our work

@inproceedings{chen2019lbc,
  author    = {Dian Chen and Brady Zhou and Vladlen Koltun and Philipp Kr\"ahenb\"uhl},
  title     = {Learning by Cheating},
  booktitle = {Conference on Robot Learning (CoRL)},
  year      = {2019},
}

CARLA Challenge

Checkout our submission to the 2020 CARLA Challenge!

Video

For a summarization of this project, please checkout our video.

Installation

Please refer to INSTALL.md for setup instructions.

Quick start

We provide a quick script here in case you would like to skip compiling and directly use the official binary release:

# Download CARLA 0.9.6
wget http://carla-assets-internal.s3.amazonaws.com/Releases/Linux/CARLA_0.9.6.tar.gz
mkdir carla_lbc
tar -xvzf CARLA_0.9.6.tar.gz -C carla_lbc
cd carla_lbc

# Download LBC
git init
git remote add origin https://github.com/dianchen96/LearningByCheating.git
# rename the LICENSE file to avoid conflicts during the pull
mv LICENSE CARLA_LICENSE 
git pull origin release-0.9.6
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

# 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
cd ../../..
mkdir -p ckpts/image
cd ckpts/image
wget http://www.cs.utexas.edu/~dchen/lbc_release/ckpts/image/model-10.th
wget http://www.cs.utexas.edu/~dchen/lbc_release/ckpts/image/config.json
cd ../..
mkdir -p ckpts/priveleged
cd ckpts/priveleged
wget http://www.cs.utexas.edu/~dchen/lbc_release/ckpts/privileged/model-128.th
wget http://www.cs.utexas.edu/~dchen/lbc_release/ckpts/privileged/config.json
cd ../..

Once you are done with that, you need to start the Carla Server and the LbC agent.

Running the Carla Server

• Open up a terminal.
• Inside the carla directory run ./CarlaUE4.sh -fps=10 -benchmark.

Running the LbC Agent

• Open up another terminal to run the LbC agent.

• To run the LbC agent, your PYTHONPATH needs to be set correctly. Make sure [CARLA PATH]/PythonAPI is in your PYTHONPATH If you are inside the carla_lbc directory (created above), you can run the following command.

export PYTHONPATH="`pwd`/PythonAPI:$PYTHONPATH" 

• After ensuring your PYTHONPATH is set correctly, run this:

CUDA_VISIBLE_DEVICES="0" python benchmark_agent.py --suite=town2 --model-path=ckpts/image/model-10.th --show

• Now you can see the the image model drive in the testing town!

Benchmark Results (0.9.6 w/ pedestrians fix)

Since CARLA does not have an official 0.9+ version that supports pedestrian crossing, we modified the most up-to-date CARLA (0.9.6) to support pedestrian crossing to compare to the original benchmark.

All our changes are only on the client side, and summarized below:

1. Modified navigation mesh, such that pedestrians walk and "cross" the streets.
2. Modified navigation, such that pedestrians avoid cars.
3. Added pedestrian teleport functionality, such that they are teleported if stuck and causing a traffic jam.

Our CARLA 0.9.6

We are cleaning-up our CARLA 0.9.5 implementation, and the code is coming soon.

Autopilot

Priviledged(cheating) Agent

Stage 1(purely offline) Agent

Stage 2(online fine-tuned) Agent

Training Models

Data collection

python data_collector.py --dataset_path=[PATH]

Use --n_episodes to select the number of trajectories you want to collect. Make sure [CARLA PATH]/PythonAPI is in your python path, or add PYTHONPATH=[CARLA PATH]/PythonAPI before the call.

Train a privileged agent

cd training
python train_birdview.py --dataset_path=[DATA PATH] --log_dir=[LOG DIR]

• --dataset_path expects a folder that contains a train and a val subdirectory, where each of these should contain .lmdb trajectory files collected from the data collection script.

• --log_dir will store the model checkpoints, the hyperparameter configurations, and the training losses and visualizations. You can track your model using tensorboard --log_dir [LOG_DIR] to monitor the progress. You should expect a validation loss smaller than 5e-3 for a well trained a privileged model.

Train an image agent

Stage 0 (warm-up)

cd training
python train_image_phase0.py --dataset_path=[DATA PATH] --log_dir=[LOG DIR] --pretrained --teacher_path=[TEACHER PATH]

• --teacher_path expects the path to a privileged agent .th checkpoint. Make sure config.json from priveleged agent trainig lies in the same directory as the checkpoint

Stage 1

cd training
python train_image_phase1.py --dataset_path=[DATA PATH] --log_dir=[LOG DIR] --teacher_path=[TEACHER PATH] --ckpt=[CKPT PATH]

• --ckpt expects the path to the stage 0 .th checkpoint.

Stage 2 (dagger)

cd training
python train_image_phase2.py --teacher_path=[TEACHER PATH] --ckpt=[CKPT PATH] --log_dir=[LOG DIR]

• --ckpt expects the path to the stage 1 .th checkpoint.

For all stages you can track your model using tensorboard --log_dir [LOG_DIR] to monitor the progress.

Note

Due to randomness, the retrained model will not be the same as the published, and you will likely need to retune the controller parameters.

Benchmarking models

1. Start a CARLA server instance ./Carla.sh -fps=10 -benchmark -world-port=[PORT NUM]
2. Run python benchmark_agent.py --suite=[SUITE NAME] --port=[PORT NUM] --model_path=[MODEL PATH]. This will create a summary.csv in /benchmark and benchmarking videos in /benchmark/[SUITE NAME].
3. Once benchmarking is done, use python view_benchmark_results.py [MODEL_PATH]/benchmark/[MODEL NAME] to print a results table like the ones shown below.

Note that CARLA is non-deterministic, since currently we cannot control the random seeds in the server. Our client-side random seed makes sure the other vehicles have deterministic initial positions, but the decision of whether to turn left or right at intersections is non-deterministic.

Detailed Benchmark Results

Autopilot

╔Performance of autopilot════════════╦═════════╦═══════╗
║ Suite Name          ║ Success Rate ║ Total   ║ Seeds ║
╠═════════════════════╬══════════════╬═════════╬═══════╣
║ FullTown01-v1       ║ 100          ║ 100/100 ║ 0     ║
║ FullTown01-v2       ║ 100          ║ 50/50   ║ 0     ║
║ FullTown01-v3       ║ 100          ║ 100/100 ║ 0     ║
║ FullTown01-v4       ║ 100          ║ 50/50   ║ 0     ║
║ FullTown02-v1       ║ 100          ║ 100/100 ║ 0     ║
║ FullTown02-v2       ║ 100          ║ 50/50   ║ 0     ║
║ FullTown02-v3       ║ 100          ║ 100/100 ║ 0     ║
║ FullTown02-v4       ║ 100          ║ 50/50   ║ 0     ║
║ NoCrashTown01-v1    ║ 100.0 ± 0.0  ║ 300/300 ║ 0,1,2 ║
║ NoCrashTown01-v2    ║ 100.0 ± 0.0  ║ 150/150 ║ 0,1,2 ║
║ NoCrashTown01-v3    ║ 98.7 ± 0.6   ║ 296/300 ║ 0,1,2 ║
║ NoCrashTown01-v4    ║ 99.3 ± 1.2   ║ 149/150 ║ 0,1,2 ║
║ NoCrashTown01-v5    ║ 86.3 ± 3.2   ║ 259/300 ║ 0,1,2 ║
║ NoCrashTown01-v6    ║ 82.7 ± 6.1   ║ 124/150 ║ 0,1,2 ║
║ NoCrashTown02-v1    ║ 100.0 ± 0.0  ║ 300/300 ║ 0,1,2 ║
║ NoCrashTown02-v2    ║ 100.0 ± 0.0  ║ 150/150 ║ 0,1,2 ║
║ NoCrashTown02-v3    ║ 99.0 ± 1.0   ║ 297/300 ║ 0,1,2 ║
║ NoCrashTown02-v4    ║ 98.0 ± 2.0   ║ 147/150 ║ 0,1,2 ║
║ NoCrashTown02-v5    ║ 60.0 ± 2.6   ║ 180/300 ║ 0,1,2 ║
║ NoCrashTown02-v6    ║ 58.7 ± 7.6   ║ 88/150  ║ 0,1,2 ║
║ StraightTown01-v1   ║ 100          ║ 100/100 ║ 0     ║
║ StraightTown01-v2   ║ 100          ║ 50/50   ║ 0     ║
║ StraightTown02-v1   ║ 100          ║ 100/100 ║ 0     ║
║ StraightTown02-v2   ║ 100          ║ 50/50   ║ 0     ║
║ TurnTown01-v1       ║ 100          ║ 100/100 ║ 0     ║
║ TurnTown01-v2       ║ 100          ║ 50/50   ║ 0     ║
║ TurnTown02-v1       ║ 100          ║ 100/100 ║ 0     ║
║ TurnTown02-v2       ║ 100          ║ 50/50   ║ 0     ║
╚═════════════════════╩══════════════╩═════════╩═══════╝

Priviledged(cheating) Agent

Model checkpoints

╔Performance of model-512════════════╦═════════╦═══════╗
║ Suite Name          ║ Success Rate ║ Total   ║ Seeds ║
╠═════════════════════╬══════════════╬═════════╬═══════╣
║ FullTown01-v1       ║ 100          ║ 100/100 ║ 0     ║
║ FullTown01-v2       ║ 100          ║ 50/50   ║ 0     ║
║ FullTown01-v3       ║ 100          ║ 100/100 ║ 0     ║
║ FullTown01-v4       ║ 100          ║ 50/50   ║ 0     ║
║ FullTown02-v1       ║ 100          ║ 100/100 ║ 0     ║
║ FullTown02-v2       ║ 100          ║ 50/50   ║ 0     ║
║ FullTown02-v3       ║ 99           ║ 99/100  ║ 0     ║
║ FullTown02-v4       ║ 100          ║ 50/50   ║ 0     ║
║ NoCrashTown01-v1    ║ 100.0 ± 0.0  ║ 300/300 ║ 0,1,2 ║
║ NoCrashTown01-v2    ║ 100.0 ± 0.0  ║ 150/150 ║ 0,1,2 ║
║ NoCrashTown01-v3    ║ 96.3 ± 3.1   ║ 289/300 ║ 0,1,2 ║
║ NoCrashTown01-v4    ║ 97.3 ± 3.1   ║ 146/150 ║ 0,1,2 ║
║ NoCrashTown01-v5    ║ 80.2 ± 4.9   ║ 239/298 ║ 0,1,2 ║
║ NoCrashTown01-v6    ║ 81.3 ± 5.8   ║ 122/150 ║ 0,1,2 ║
║ NoCrashTown02-v1    ║ 100.0 ± 0.0  ║ 300/300 ║ 0,1,2 ║
║ NoCrashTown02-v2    ║ 100.0 ± 0.0  ║ 150/150 ║ 0,1,2 ║
║ NoCrashTown02-v3    ║ 95.0 ± 1.0   ║ 285/300 ║ 0,1,2 ║
║ NoCrashTown02-v4    ║ 93.3 ± 2.3   ║ 140/150 ║ 0,1,2 ║
║ NoCrashTown02-v5    ║ 45.5 ± 8.4   ║ 135/297 ║ 0,1,2 ║
║ NoCrashTown02-v6    ║ 45.3 ± 9.5   ║ 68/150  ║ 0,1,2 ║
║ StraightTown02-v1   ║ 100          ║ 100/100 ║ 0     ║
║ StraightTown02-v2   ║ 100          ║ 50/50   ║ 0     ║
║ TurnTown02-v1       ║ 100          ║ 100/100 ║ 0     ║
║ TurnTown02-v2       ║ 100          ║ 50/50   ║ 0     ║
╚═════════════════════╩══════════════╩═════════╩═══════╝

Stage 1(purely offline) Agent

╔Performance of model-32═══════════╦═════════╦═══════╗
║ Suite Name        ║ Success Rate ║ Total   ║ Seeds ║
╠═══════════════════╬══════════════╬═════════╬═══════╣
║ FullTown01-v1     ║ 93           ║ 93/100  ║ 0     ║
║ FullTown01-v2     ║ 98           ║ 49/50   ║ 0     ║
║ FullTown01-v3     ║ 94           ║ 94/100  ║ 0     ║
║ FullTown01-v4     ║ 96           ║ 48/50   ║ 0     ║
║ FullTown02-v1     ║ 94           ║ 94/100  ║ 0     ║
║ FullTown02-v2     ║ 96           ║ 48/50   ║ 0     ║
║ FullTown02-v3     ║ 92           ║ 92/100  ║ 0     ║
║ FullTown02-v4     ║ 94           ║ 47/50   ║ 0     ║
║ NoCrashTown01-v1  ║ 88.3 ± 1.5   ║ 265/300 ║ 0,1,2 ║
║ NoCrashTown01-v2  ║ 91.3 ± 3.1   ║ 137/150 ║ 0,1,2 ║
║ NoCrashTown01-v3  ║ 74.3 ± 3.8   ║ 223/300 ║ 0,1,2 ║
║ NoCrashTown01-v4  ║ 71.3 ± 4.6   ║ 107/150 ║ 0,1,2 ║
║ NoCrashTown01-v5  ║ 27.7 ± 3.5   ║ 83/300  ║ 0,1,2 ║
║ NoCrashTown01-v6  ║ 24.7 ± 2.3   ║ 37/150  ║ 0,1,2 ║
║ NoCrashTown02-v1  ║ 85.0 ± 2.6   ║ 255/300 ║ 0,1,2 ║
║ NoCrashTown02-v2  ║ 80.7 ± 2.3   ║ 121/150 ║ 0,1,2 ║
║ NoCrashTown02-v3  ║ 64.3 ± 4.0   ║ 193/300 ║ 0,1,2 ║
║ NoCrashTown02-v4  ║ 60.0 ± 4.0   ║ 90/150  ║ 0,1,2 ║
║ NoCrashTown02-v5  ║ 11.7 ± 2.3   ║ 35/300  ║ 0,1,2 ║
║ NoCrashTown02-v6  ║ 11.3 ± 3.1   ║ 17/150  ║ 0,1,2 ║
║ StraightTown01-v1 ║ 100          ║ 100/100 ║ 0     ║
║ StraightTown01-v2 ║ 100          ║ 50/50   ║ 0     ║
║ StraightTown02-v1 ║ 100          ║ 100/100 ║ 0     ║
║ StraightTown02-v2 ║ 100          ║ 50/50   ║ 0     ║
║ TurnTown01-v1     ║ 96           ║ 96/100  ║ 0     ║
║ TurnTown01-v2     ║ 100          ║ 50/50   ║ 0     ║
║ TurnTown02-v1     ║ 97           ║ 97/100  ║ 0     ║
║ TurnTown02-v2     ║ 100          ║ 50/50   ║ 0     ║
╚═══════════════════╩══════════════╩═════════╩═══════╝

Note that this is with different controller parameters from what we submitted for CoRL2019, yielding slightly different numbers.

The original raw numbers are shown below

╔Performance of model-32═══════════╦═════════╦═══════╗
║ Suite Name        ║ Success Rate ║ Total   ║ Seeds ║
╠═══════════════════╬══════════════╬═════════╬═══════╣
║ FullTown01-v1     ║ 94           ║ 94/100  ║ 0     ║
║ FullTown01-v2     ║ 98           ║ 49/50   ║ 0     ║
║ FullTown01-v3     ║ 95           ║ 95/100  ║ 0     ║
║ FullTown01-v4     ║ 92           ║ 46/50   ║ 0     ║
║ FullTown02-v1     ║ 94           ║ 94/100  ║ 0     ║
║ FullTown02-v2     ║ 98           ║ 49/50   ║ 0     ║
║ FullTown02-v3     ║ 88           ║ 88/100  ║ 0     ║
║ FullTown02-v4     ║ 90           ║ 45/50   ║ 0     ║
║ StraightTown01-v1 ║ 100          ║ 100/100 ║ 0     ║
║ StraightTown01-v2 ║ 100          ║ 50/50   ║ 0     ║
║ StraightTown02-v1 ║ 100          ║ 100/100 ║ 0     ║
║ StraightTown02-v2 ║ 100          ║ 50/50   ║ 0     ║
║ TurnTown01-v1     ║ 96           ║ 96/100  ║ 0     ║
║ TurnTown01-v2     ║ 100          ║ 50/50   ║ 0     ║
║ TurnTown02-v1     ║ 95           ║ 95/100  ║ 0     ║
║ TurnTown02-v2     ║ 98           ║ 49/50   ║ 0     ║
╚═══════════════════╩══════════════╩═════════╩═══════╝

Stage 2(online fine-tuned) Agent

Model checkpoints

╔Performance of model-10══════════╦═════════╦═══════╗
║ Suite Name       ║ Success Rate ║ Total   ║ Seeds ║
╠══════════════════╬══════════════╬═════════╬═══════╣
║ FullTown01-v1    ║ 100          ║ 100/100 ║ 0     ║
║ FullTown01-v2    ║ 100          ║ 50/50   ║ 0     ║
║ FullTown01-v3    ║ 100          ║ 100/100 ║ 0     ║
║ FullTown01-v4    ║ 100          ║ 50/50   ║ 0     ║
║ FullTown02-v1    ║ 100          ║ 100/100 ║ 0     ║
║ FullTown02-v2    ║ 100          ║ 50/50   ║ 0     ║
║ FullTown02-v3    ║ 98           ║ 98/100  ║ 0     ║
║ FullTown02-v4    ║ 98           ║ 49/50   ║ 0     ║
║ NoCrashTown01-v1 ║ 99.0 ± 0.0   ║ 297/300 ║ 0,1,2 ║
║ NoCrashTown01-v2 ║ 89.3 ± 3.1   ║ 134/150 ║ 0,1,2 ║
║ NoCrashTown01-v3 ║ 97.3 ± 1.2   ║ 292/300 ║ 0,1,2 ║
║ NoCrashTown01-v4 ║ 95.3 ± 3.1   ║ 143/150 ║ 0,1,2 ║
║ NoCrashTown01-v5 ║ 70.0 ± 4.8   ║ 196/280 ║ 0,1,2 ║
║ NoCrashTown01-v6 ║ 62.7 ± 2.3   ║ 94/150  ║ 0,1,2 ║
║ NoCrashTown02-v1 ║ 99.0 ± 0.0   ║ 297/300 ║ 0,1,2 ║
║ NoCrashTown02-v2 ║ 84.7 ± 3.1   ║ 127/150 ║ 0,1,2 ║
║ NoCrashTown02-v3 ║ 93.3 ± 2.1   ║ 280/300 ║ 0,1,2 ║
║ NoCrashTown02-v4 ║ 70.0 ± 2.0   ║ 105/150 ║ 0,1,2 ║
║ NoCrashTown02-v5 ║ 46.1 ± 4.0   ║ 128/278 ║ 0,1,2 ║
║ NoCrashTown02-v6 ║ 32.7 ± 9.5   ║ 49/150  ║ 0,1,2 ║
╚══════════════════╩══════════════╩═════════╩═══════╝

Note that this is run with different controller parameters from what we submitted for CoRL2019, yielding slightly better numbers.

The original raw numbers are shown below

Benchmark results/videos

╔Performance of model-10═══════════╦═════════╦═══════╗
║ Suite Name        ║ Success Rate ║ Total   ║ Seeds ║
╠═══════════════════╬══════════════╬═════════╬═══════╣
║ FullTown01-v1     ║ 100          ║ 100/100 ║ 0     ║
║ FullTown01-v2     ║ 100          ║ 50/50   ║ 0     ║
║ FullTown01-v3     ║ 100          ║ 100/100 ║ 0     ║
║ FullTown01-v4     ║ 96           ║ 48/50   ║ 0     ║
║ FullTown02-v1     ║ 98           ║ 98/100  ║ 0     ║
║ FullTown02-v2     ║ 100          ║ 50/50   ║ 0     ║
║ FullTown02-v3     ║ 99           ║ 99/100  ║ 0     ║
║ FullTown02-v4     ║ 100          ║ 50/50   ║ 0     ║
║ NoCrashTown01-v1  ║ 97.0 ± 1.0   ║ 291/300 ║ 0,1,2 ║
║ NoCrashTown01-v2  ║ 86.7 ± 4.2   ║ 130/150 ║ 0,1,2 ║
║ NoCrashTown01-v3  ║ 93.3 ± 0.6   ║ 280/300 ║ 0,1,2 ║
║ NoCrashTown01-v4  ║ 87.3 ± 3.1   ║ 131/150 ║ 0,1,2 ║
║ NoCrashTown01-v5  ║ 70.7 ± 4.5   ║ 212/300 ║ 0,1,2 ║
║ NoCrashTown01-v6  ║ 63.3 ± 3.1   ║ 95/150  ║ 0,1,2 ║
║ NoCrashTown02-v1  ║ 99.7 ± 0.6   ║ 299/300 ║ 0,1,2 ║
║ NoCrashTown02-v2  ║ 70.0 ± 4.0   ║ 105/150 ║ 0,1,2 ║
║ NoCrashTown02-v3  ║ 94.0 ± 3.0   ║ 281/299 ║ 0,1,2 ║
║ NoCrashTown02-v4  ║ 62.0 ± 2.0   ║ 93/150  ║ 0,1,2 ║
║ NoCrashTown02-v5  ║ 51.3 ± 3.1   ║ 154/300 ║ 0,1,2 ║
║ NoCrashTown02-v6  ║ 38.7 ± 6.4   ║ 58/150  ║ 0,1,2 ║
║ StraightTown01-v1 ║ 100          ║ 100/100 ║ 0     ║
║ StraightTown01-v2 ║ 100          ║ 50/50   ║ 0     ║
║ StraightTown02-v1 ║ 100          ║ 100/100 ║ 0     ║
║ StraightTown02-v2 ║ 100          ║ 50/50   ║ 0     ║
║ TurnTown01-v1     ║ 100          ║ 100/100 ║ 0     ║
║ TurnTown01-v2     ║ 96           ║ 48/50   ║ 0     ║
║ TurnTown02-v1     ║ 100          ║ 100/100 ║ 0     ║
║ TurnTown02-v2     ║ 100          ║ 50/50   ║ 0     ║
╚═══════════════════╩══════════════╩═════════╩═══════╝

License

This repo is released under the MIT License (please refer to the LICENSE file for details). Part of the PythonAPI and the map rendering code is borrowed from the official CARLA repo, which is under MIT license. The image augmentation code is borrowed from Coiltraine which is released under MIT license.