Reproducing Results: TCP - Trajectory-guided Control Prediction for End-to-end Autonomous Driving: A Simple yet Strong Baseline

I have reproduced the results of the TCP paper and have made the checkpoints public! This model was trained on an Lenovo Legion laptop with an Intel i7-12700H (20 threads) and NVIDIA GeForce RTX 3070 Mobile GPU with 8 GB VRAM. The model was trained upto epoch 20.

Average Score Composed: 40.24
Average Score Penalty: 0.67
Average Score Route: 65.82

TCP

Trajectory-guided Control Prediction for End-to-end Autonomous Driving: A Simple yet Strong Baseline
Penghao Wu*, Xiaosong Jia*, Li Chen*, Junchi Yan, Hongyang Li, Yu Qiao

arXiv Paper, NeurIPS 2022

Blog in Chinese

This repository contains the code for the paper Trajectory-guided Control Prediction for End-to-end Autonomous Driving: A Simple yet Strong Baseline.

TCP is a simple unified framework to combine trajectory and control prediction for end-to-end autonomous driving. By time of release in June 17 2022, our method achieves new state-of-the-art on CARLA AD Leaderboard, in which we rank the first in terms of the Driving Score and Infraction Penalty using only a single camera as input.

Setup

Clone this repo and build the environment

git clone https://github.com/OpenPerceptionX/TCP.git
cd TCP
conda env create -f environment.yml --name TCP
conda activate TCP

export PYTHONPATH=$PYTHONPATH:PATH_TO_TCP

Download and setup CARLA 0.9.10.1

mkdir carla
cd carla
wget -c https://tiny.carla.org/carla-0-9-10-1-linux -O CARLA_0.9.10.1.tar.gz
wget -c https://tiny.carla.org/additional-maps-0-9-10-1-linux -O AdditionalMaps_0.9.10.1.tar.gz
tar -xvzf CARLA_0.9.10.1.tar.gz
mv CARLA_0.9.10.1.tar.gz Import/
./ImportAssets.sh
cd ..
pip install carla==0.9.12

Dataset

Download our dataset through Huggingface (combine the part with command cat tcp_carla_data_part_* > tcp_carla_data.zip) or GoogleDrive or BaiduYun (提取码 8174). The total size of our dataset is around 115G, make sure you have enough space.

Pretrained Checkpoint

You can download the checkpoints from this Google Drive Link and place them at ckpts/.

(TCP)$ tree ckpts/
ckpts/
├── epoch_07.pth
├── epoch_11.pth
├── epoch_12.pth
├── epoch_16.pth
└── epoch_20.pth

0 directories, 5 files

Training

First, set the dataset path in TCP/config.py in root_dir_all. Training:

python -m TCP.train
python -m TCP.train --resume_from_checkpoint ckpts/epoch_20.pth

Data Generation

First, launch the carla server,

cd CARLA_ROOT
./CarlaUE4.sh --world-port=2000 -quality-level=Low

Set the carla path, routes file, scenario file, and data path for data generation in leaderboard/scripts/data_collection.sh.

Start data collection

sh leaderboard/scripts/data_collection.sh

After the data collecting process, run tools/filter_data.py and tools/gen_data.py to filter out invalid data and pack the data for training.

Evaluation

First, launch the carla server,

cd CARLA_ROOT
./CarlaUE4.sh --world-port=2000 -quality-level=Low

Set the carla path, routes file, scenario file, model ckpt, and data path for evaluation in leaderboard/scripts/run_evaluation.sh.

Start the evaluation

sh leaderboard/scripts/run_evaluation.sh

Citation

If you find our repo or our paper useful, please use the following citation:

@inproceedings{wu2022trajectoryguided,
 title={Trajectory-guided Control Prediction for End-to-end Autonomous Driving: A Simple yet Strong Baseline}, 
 author={Penghao Wu and Xiaosong Jia and Li Chen and Junchi Yan and Hongyang Li and Yu Qiao},
 booktitle={NeurIPS},
 year={2022},
}

License

All code within this repository is under Apache License 2.0.

Acknowledgements

Our code is based on several repositories:

AdityaNG/TCP