DG-TrajGen

The official repository for paper ”Domain Generalization for Vision-based Driving Trajectory Generation’‘ submitted to ICRA 2022.

Our Method

• Trajectory representation:
  • Model: ./learning/model.py/Generator
• Latent Action Space Learning:
  • Generator model: ./learning/model.py/Generator
  • Discriminator model: ./learning/model.py/Discriminator
  • Training: ./scripts/Ours/stage1_train_GAN.py
• Encoder Pre-training:
  • Training: ./scripts/Ours/stage2_pretrain_encoder.py
• End-to-End Training the Encoder:
  • Training: ./scripts/Ours/stage3_train_e2e.py

Comparative Study

• RIP:
  • Training: ./scripts/RIP/train.py
  • Referenced official code: github
  • Paper: arxiv
• MixStyle:
  • Training: ./scripts/MixStyle/train.py
  • Referenced official code: github
  • Paper: arxiv
• DIVA:
  • Training: ./scripts/DIVA/train.py
  • Referenced official code: github
  • Paper: arxiv
• DAL:
  • Training: ./scripts/DAL/train.py
• E2E NT:
  • Training: ./scripts/E2ENT/train.py
  • Referenced official code: github
  • Paper: arxiv

Closed-loop Experiments:

We train the model on the Oxford RobotCar dataset and directly generalize it to the CARLA simulation.

• Run: ./scripts/CARLA/run_ours.py

Citation

If you use our source code, please consider citing the following:

@article{wang2021domain,
  title={Domain Generalization for Vision-based Driving Trajectory Generation},
  author={Wang, Yunkai and Zhang, Dongkun and Cui, Yuxiang and Chen, Zexi and Jing, Wei and Chen, Junbo and Xiong, Rong and Wang, Yue},
  journal={arXiv preprint arXiv:2109.13858},
  year={2021}
}

GitHub

https://github.com/IamWangYunKai/DG-TrajGen