Planning Mixture of Experts

This repository contains PyTorch implementation of ''Exploring Reflective Limitation of Behavior Cloning in Autonomous Vehicles'' paper.

Dependencies

This repo uses PyTorch and CARLA simulator (0.9.10.1). Please make sure that all dependencies are installed by running:

python3 -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt

Please specify the CARLA root directory at the top of run.sh file.

export CARLA_ROOT=~/Projects/nazeri/CARLA_0.9.10.1          # change to where you installed CARLA

You need at least a middle range GPU to run the experiments with minimum possible batch size.

Training

To run different training stages, please see experiments doc first. You can define the training and model parameters via the respective config file in conf directory, then, use run.sh stage_number to run training where stage_number can be one of stage0, stage1, stage2, stage3. To see more options, and their description execute ./run.sh -h

Evaluation

Sample output of the PMoE. PMoE's predictions are in red whereas the reflective rule-based AI are in green.

PMoE

PU-Net

To evaluate the models there are two options, Carla leaderboard evaluation or nocrash evaluation which can be run via ./run.sh benchmark or ./run.sh nocrash respectively. You can configure the model and benchmark parameters through the conf/benchmark.yaml config file.

Dataset

The dataset was collect using this repo which is fork of LBC leaderboard submission.

Acknowledgements

Thanks for World on Rails, Learning by Cheating, CARLA leaderboard, CARLA scenario_runner, Coiltraine, repos for making their code public.

Reference

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

@INPROCEEDINGS{Nazeri2021,
  author={Nazeri, Mohammad Hossein and Bohlouli, Mahdi},
  booktitle={2021 IEEE International Conference on Data Mining (ICDM)}, 
  title={Exploring Reflective Limitation of Behavior Cloning in Autonomous Vehicles}, 
  year={2021},
  volume={},
  number={},
  pages={1252-1257},
  doi={10.1109/ICDM51629.2021.00153}}

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. The evaluation code is borrowed from World on Rails, and Learning by Cheating which also released under the MIT License.