CIL++ with Multi-View Attention Learning

---

Publications

This is the official code release of the paper:

Yi Xiao, Felipe Codevilla, Diego Porres and Antonio M. Lopez. Scaling Vision-based End-to-End Driving with Multi-View Attention Learning.

Please cite our paper if you find this work useful (will be soon updated with the IROS citation):

     @misc{xiao2023scaling,
     title={Scaling Vision-based End-to-End Driving with Multi-View Attention Learning},
     author={Yi Xiao and Felipe Codevilla and Diego Porres and Antonio M. Lopez},
     year={2023},
     eprint={2302.03198},
     archivePrefix={arXiv},
     primaryClass={cs.CV}
     }

Video

Please watch our online video for more interesting scenario demonstrations

---

Summary

In this repository, you could find materials in order to:

• Benchmark the trained CIL++ model proposed in our paper
• Collect datasets using Roach RL expert driver from this paper
• Train/evaluate (offline) your own trained CIL++ models
• Test your models on CARLA 0.9.13

---

Environment Setup

Python version: 3.8

Cuda version: 11.6

Required packages: requirements.txt

• Set up the conda environment for the experiments:

    conda create --name CILv2Env python=3.8
    conda activate CILv2Env
  

• Download CARLA 0.9.13 to your root directory and build up CARLA docker:

    export ROOTDIR=<Path to your root directory>
    cd $ROOTDIR
    export CARLAPATH=$ROOTDIR/CARLA_0.9.13/PythonAPI/carla/:$ROOTDIR/CARLA_0.9.13/PythonAPI/carla/dist/carla-0.9.13-py3.7-linux-x86_64.egg
  

• For using CARLA docker, you can either 1) pull or 2) build up the container:

  To pull, run:

    docker pull carlasim/carla:0.9.13
  

  To build up, run:

    docker image build -f $ROOTDIR/CARLA_0.9.13/Dockerfile -t CARLA0913 $ROOTDIR/CARLA_0.9.13/
  

• Download the CIL++ repository in your root directory:

    cd $ROOTDIR
    git clone https://github.com/yixiao1/CILv2_multiview.git
  

• Define environment variables:

    export TRAINING_ROOT=$ROOTDIR/CILv2_multiview
    export DRIVING_TEST_ROOT=$TRAINING_ROOT/run_CARLA_driving/
    export SCENARIO_RUNNER_ROOT=$TRAINING_ROOT/scenario_runner/
    export PYTHONPATH=$CARLAPATH:$TRAINING_ROOT:$DRIVING_TEST_ROOT:$SCENARIO_RUNNER_ROOT
    export TRAINING_RESULTS_ROOT=<Path to the directory where the results to be saved>
    export DATASET_PATH=<Path to the directory where the datasets are stored>
    export SENSOR_SAVE_PATH=<Path to the directory where the driving test frames are stored>
  

• Install the required packages:

    conda install --file requirements.txt
  

---

Benchmark our trained CIL++

• Download our trained CIL++ models _results.tar.gz to your TRAINING_RESULTS_ROOT/_results. The saving pattern should be $TRAINING_RESULTS_ROOT/_results/Ours/TownXX/...:

    mkdir -p $TRAINING_RESULTS_ROOT/_results
    tar -zxvf _results.tar.gz -C $TRAINING_RESULTS_ROOT/_results/
  

• Benchmark our trained CIL++:

    cd $DRIVING_TEST_ROOT
    run ./scripts/run_evaluation/CILv2/nocrash_newweathertown_Town02.sh
  

---

Dataset Collection with Roach RL expert

For training models, you can either

• Download our collected datasets:

  For easy downloading, the folders are divided into several parts and all compressed in zip files. For model training, please unzip them into your DATASET_PATH. The dataloader will access to the full paths of $DATASET_PATH/<dataset_folder_name>, where the <dataset_folder_name> will be defined by TRAIN_DATASET_NAME/VALID_DATASET_NAME in the exp yaml file

  • Small test datasets:

    Training: part 1, part 2

    Offline Evaluation: part 3

  • Single-lane towns:

    Training: part 4, part 5

    Offline Evaluation: part 6

  • Multi-lane towns:

    Training: part 7, part 8, part 9, part 10, part 11, part 12, part 13, part 14

    Offline Evaluation: same as part 6

• Collect new datasets. The RL expert driver we used for data collection is from this work

---

Training & performing offline evaluation on new trained CIL++ models

• You need to define a configuration file for training. Please refer to this file in configs folder as an example

• Run the main.py file:

    python main.py --process-type train_val --gpus 0 --folder CILv2 --exp CILv2_3cam_smalltest
  

  where --process-type defines the process type (could be either train_val or val_only), --gpus defines the gpus to be used, --folder is the configuration folder name, and --exp is the configuration yaml file name. Your results will be stored in $TRAINING_RESULTS_ROOT/_results/<folder_name>/<exp_name>/

---

Test your own trained models on CARLA simulator

• Please make sure that your models are saved in the proper pattern as the downloaded CIL++ model:

    cd $TRAINING_RESULTS_ROOT/_results/<folder_name>/<exp_name>/
  

  where folder_name the the experiment folder name, and exp_name is the configuration file name. Your models are all saved in ./checkpoints/

• Define a config file for the benchmarking:

    cd $TRAINING_RESULTS_ROOT/_results/<folder_name>/<exp_name>
    > config45.json
  

  In the json file, you need to define the model/checkpoint to be tested:

        {
            "agent_name": "CILv2",
            "checkpoint": 45,
            "yaml": "CILv2.yaml"
        }
  

  where checkpoint indicates the checkpoint to be tested, yaml is the training configuration file which was automatically generated during training. Please refer to the json file in the downloaded _results.tar.gz

• Benchmark your model:

  Notice that to benchmark your own trained models, you need to modify the script by changing the --agent-config

    cd $DRIVING_TEST_ROOT
    run ./scripts/run_evaluation/CILv2/nocrash_newweathertown_Town02.sh
  

---

License

The code is released under a CC-BY-NC 4.0 license, which only allows personal and research use. For a commercial license, please contact the authors. Portions of source code taken from external sources are annotated with links to original files and their corresponding licenses.

---

Acknowledgements

This research is supported as a part of the project TED2021-132802B-I00 funded by MCIN/AEI/10.13039/501100011033 and the European Union NextGenerationEU/PRTR.

Yi Xiao acknowledges the support to her PhD study provided by the Chinese Scholarship Council (CSC), Grant No.201808390010. Diego Porres acknowledges the support to his PhD study provided by Grant PRE2018-083417 funded by MCIN/AEI /10.13039/501100011033 and FSE invierte en tu futuro. Antonio M. López acknowledges the financial support to his general research activities given by ICREA under the ICREA Academia Program. Antonio thanks the synergies, in terms of research ideas, arising from the project PID2020-115734RB-C21 funded by MCIN/AEI/10.13039/501100011033.

The authors acknowledge the support of the Generalitat de Catalunya CERCA Program and its ACCIO agency to CVC’s general activities.