Human-inspired Data-light Cultural Alignment for Cross-regional Deployment of Autonomous Vehicles

# Human-inspired Data-light Cultural Alignment for Cross-regional Deployment of Autonomous Vehicles This repository is the official codebase for the paper "Human-inspired Data-light Cultural Alignment for Cross-regional Deployment of Autonomous Vehicles". ## Overview This project provides scripts to build trajectory datasets from multiple public driving datasets, construct PsiPhi training data, and train SF+PV models for cross-regional cultural alignment. ## Requirements - Python 3.10+ - Install dependencies: ```bash pip install -r requirements.txt ``` ## Dataset Preparation The raw datasets are not included. Please obtain access to each dataset and follow its license terms. ### Build per-recording trajectory files Each script converts a raw dataset into `.pt` files containing per-frame states. Examples (replace paths with your local dataset paths): ```bash python scripts/dji/build_dataset_dji_new.py --data-path /path/to/datasets/DJI --output-dir data/generated python scripts/highd/build_dataset_highd_new.py --data-path /path/to/datasets/highD --output-dir data/generated python scripts/ind/build_dataset_ind_new.py --data-path /path/to/datasets/inD --output-dir data/generated python scripts/ngsim/build_dataset_ngsim_new.py --data-path /path/to/datasets/NGSIM --output-dir data/generated python scripts/sind/build_dataset_sind_new.py --data-path /path/to/datasets/sinD --output-dir data/generated python scripts/citysim/build_dataset_citysim_new.py --data-path /path/to/datasets/CitySim --output-dir data/generated ``` ### Organize train/test splits `trainning/build_dataset.py` expects the following structure: ``` data/generated///*.pt ``` Create your train/test splits by moving or copying the generated `.pt` files into: ``` data/generated/US/train data/generated/US/test ``` Repeat for other regions (e.g., `China`, `Germany`) as needed. ## Build PsiPhi Dataset After arranging the train/test splits, build the PsiPhi dataset: ```bash python trainning/build_dataset.py ``` This script writes outputs to: ``` data/psiphi_datasets/psiphi__.pt ``` If you need different countries, edit the country list in `trainning/build_dataset.py`. ## Training Single run: ```bash python trainning/train_psiphi.py \ --train_data data/psiphi_datasets/psiphi_US_train.pt \ --test_data data/psiphi_datasets/psiphi_US_test.pt \ --save_dir outputs/run_us ``` Fine-tuning with fixed PV: ```bash python trainning/train_psiphi.py \ --mode ft_sf_fix_pv \ --sf_path /path/to/sf_state_xxx.pth \ --pv_path /path/to/pv_xxx.npy \ --train_data data/psiphi_datasets/psiphi_US_train.pt \ --test_data data/psiphi_datasets/psiphi_US_test.pt \ --save_dir outputs/run_us_ft ``` Batch runs: - Configure `trainning/batch_train.py` with your datasets and model paths, then run: ```bash python trainning/batch_train.py ``` ## Outputs Training outputs are written to the chosen `--save_dir`, including: - Model checkpoints - Exported SF state dicts and PV vectors - Metrics JSON logs ## Usage and Data Policy - This repository does not distribute any datasets. - You must have proper rights to use each dataset and follow its license terms. - Replace all dataset/model paths with your own environment-specific paths. ## License and Open-Source Compliance - Copyright (c) 2025 Authors. All rights reserved unless a license file is added. - Third-party dependencies are governed by their respective licenses; review `requirements.txt` and comply accordingly. - If you plan to redistribute or publish derived artifacts, ensure compliance with all dataset and dependency licenses.