HAD-Gen: Human-like and Diverse Scenario Generation

This repository contains the PyTorch implementation of our paper titled "HAD-Gen: Human-like and Diverse Agent Behavior Modeling for Controllable Scenario Generation".

HAD-Gen can controllably generate human-like and diverse scenarios for autonomous vehicle testing and validation.

Superior Generalization: Compared with baselines, our trained policies demonstrate the strongest generalization ability and can adapt to new real-world scenarios.

Workflow:

1. Data Clustering: Driving behavior in the dataset is clustered into Aggressive, Normal, and Cautious categories
2. Reward Reconstruction: Each sub-dataset is used to reconstruct a reward function representing human driving behavior
3. Policy Training: Reconstructed rewards enable offline RL and MARL to generate driving policies for each cluster
4. Scenario Generation: Various driving policies are deployed in simulation using a policy selection strategy

This project uses the AUTOMATUM dataset. Install the Python utility to interface with the data:

pip3 install openautomatumdronedata

Follow these steps in order:

1. Clustering Analysis

   python riskclustering.py

   Generates clustering results saved as JSON files in scenarios/configs/

2. Feature Extraction

   python feature_extraction_irl.py

   Extracts feature values from dataset for IRL processing

3. IRL Training

   python training_irl.py

   Trains MaxEnt IRL and saves weights as xxxtraining_log.pkl

4. Feature Normalization

   python feature_normalization.py

   Calculates mean and standard deviation for feature normalization

5. Data Preprocessing

   python preprocessing.py

   Requires: feature_normalization.pkl and XXXtraining_log.pkl

Choose your training approach:

• Offline RL: python td3bc_automatum.py
• SAC (Log-Replay): python sac/model.py
• Online RL (Self-Replay): python train_multicar.py

• /baselines: Implementation of baseline algorithms
• /evaluation: Metrics, unit tests, scoring, evaluation, and plotting code
• /simulator: Lightweight simulator and RL environment implementations

• evaluation/evaluation_methods.py: Evaluate baseline performance
• baselines/bc_baseline.py: Train Behavioral Cloning baseline
• baselines/model.py: Train SAC baseline
• simulator/lightweight_simulator.py: Run lightweight simulations

Set these parameters in simulator/lightweight_simulator.py:

# Install the 'gym_env' package in development mode
pip install -e simulator/gym_env

# Install PyTorch (from official website)
# Install additional requirements
pip install -r requirements.txt

OpenDRIVE Map Issues: If you encounter errors about None header attributes, modify map files:

<!-- Change this line in each map file -->
<OpenDRIVE xmlns="http://www.opendrive.org">
<!-- To this -->
<OpenDRIVE>

If you use this code in your research, please cite our paper:

@article{WANG2025108270,
title = {HAD-Gen: Human-like and diverse driving behavior modeling for controllable scenario generation},
journal = {Accident Analysis & Prevention},
volume = {223},
pages = {108270},
year = {2025},
issn = {0001-4575},
doi = {https://doi.org/10.1016/j.aap.2025.108270},
url = {https://www.sciencedirect.com/science/article/pii/S0001457525003586},
author = {Cheng Wang and Lingxin Kong and Massimiliano Tamborski and Stefano V. Albrecht},
}

We welcome contributions! Please feel free to submit issues, feature requests, or pull requests.

This project is licensed under the terms specified in the LICENSE file.