FREA: Feasibility-Guided Generation of Safety-Critical Scenarios with Reasonable Adversariality

✨ News

• 2024-09-05 Exciting news! FREA is accepted by CoRL 2024 for Oral Presentation🎉🎉!
• 2024-08-10 Explore our project page, now live here🔗!
• 2024-08-10 Codes are now release!
• 2024-06-05 Our paper is available on Arxiv📄!

This repository contains the implementation of the paper.

FREA: Feasibility-Guided Generation of Safety-Critical Scenarios with Reasonable Adversariality

Keyu Chen¹ , Yuheng Lei², Hao Cheng¹, Haoran Wu¹, Wenchao Sun¹, Sifa Zheng¹

¹School of Vehicle and Mobility, Tsinghua University ²The University of Hong Kong

If you find our work useful, Please give us a star 🌟!

💫 FREA incorporates feasibility as guidance to generate adversarial yet AV-feasible, safety-critical scenarios.

📃 Outline

• 🔧 Setup
• 📚 Usage
  • 📂 1. Collect Offline Data
  • 🔥 2. Train optimal feasible value function of AV
  • 🔥 3. Train adversarial policy of CBV
  • 🔥 4. Train AV (optional)
  • ❄️ 5. Evaluation
  • 📈 6. Results Analysis
  • 🎬 7. Visualization
• 🔖 Citation
• 📋 Acknowledgement

🔧 Setup

Recommended system: Ubuntu 20.04 or 22.04

Step 1: Install Carla (0.9.13 recommended)

Step 2: Setup conda environment

conda create -n frea python=3.8
conda activate frea

Step 3: Clone this git repo in an appropriate folder

git clone [email protected]:CurryChen77/FREA.git

Step 4: Enter the repo root folder and install the packages:

cd FREA
pip install -r requirements.txt
pip install -e .

📚 Usage

📂 Collect Offline Data

Collect data from specific AV and CBV

# Launch CARLA
./CarlaUE4.sh -prefernvidia -RenderOffScreen -carla-port=2000

# Launch in another terminal
python scripts/run.py --agent_cfg expert.yaml --scenario_cfg standard_train.yaml --mode collect_feasibility_data

Merge all data from different AV and CBV

# Merge data
python frea/feasibility/unify_offline_data.py

To use the offline dataset in our paper download from here (data.zip) put them into the feasibility folder.

🔥 Train optimal feasible value function of AV

# Train optimal feasible value function of AV
python train_feasibility.py

🔥 Train adversarial policy of CBV

Train FREA (need well-trained LFR)

# Launch CARLA
./CarlaUE4.sh -prefernvidia -RenderOffScreen -carla-port=2000

# Train FREA
python scripts/run.py --agent_cfg expert.yaml --scenario_cfg fppo_adv_train.yaml --mode train_scenario

Train FPPO-RS (need well-trained LFR)

# Launch CARLA
./CarlaUE4.sh -prefernvidia -RenderOffScreen -carla-port=2000

# Train FPPO-RS
python scripts/run.py --agent_cfg expert.yaml --scenario_cfg fppo_rs_train.yaml --mode train_scenario

Train PPO

# Launch CARLA
./CarlaUE4.sh -prefernvidia -RenderOffScreen -carla-port=2000

# Train PPO
python scripts/run.py --agent_cfg expert.yaml --scenario_cfg ppo_train.yaml --mode train_scenario

🔥 Train AV (optional)

Train PPO AV based on FREA scenarios

# Launch CARLA
./CarlaUE4.sh -prefernvidia -RenderOffScreen -carla-port=2000

# Train FREA
python scripts/run.py --agent_cfg ppo.yaml --scenario_cfg fppo_adv_train.yaml --mode train_agent

Train PPO AV based on standard scenarios

# Launch CARLA
./CarlaUE4.sh -prefernvidia -RenderOffScreen -carla-port=2000

# Train FREA
python scripts/run.py --agent_cfg ppo.yaml --scenario_cfg standard_train.yaml --mode train_agent

❄️ Evaluation

To use the learning-based method PlanT, you need to extract the checkpoint of PlanT from here (PlanT_medium.zip) and put the them here.

Evaluation for data analysis (recording results)

# Launch CARLA
./CarlaUE4.sh -prefernvidia -windowed -carla-port=2000

# Evaluation FREA
python scripts/run.py --agent_cfg expert.yaml --scenario_cfg FPPO_adv_eval.yaml --mode eval --eval_mode analysis

Evaluation for video visualization

# Launch CARLA
./CarlaUE4.sh -prefernvidia -windowed -carla-port=2000

# Evaluation FREA
python scripts/run.py --agent_cfg expert.yaml --scenario_cfg FPPO_adv_eval.yaml --mode eval --eval_mode render

📈 Results Analysis

Results analysis of the paper

• Evaluation Results
• Learning Curve
• Feasibility Results

Generate your own results analysis

Make sure the Evaluation has finished and the result are saved in folder.

# Process the recorded data
python eval_analysis/process_data/process_all_data.py

# Plot the evaluation result
python eval_analysis/plot_data/plot_evaluation_result.py

🎬 Visualization

World spectator

# Launch CARLA
./CarlaUE4.sh -prefernvidia -windowed -carla-port=2000

# Set world spectator
python scripts/run.py --agent_cfg expert.yaml --scenario_cfg standard_eval.yaml --mode eval -sp

AV route

# Launch CARLA
./CarlaUE4.sh -prefernvidia -windowed -carla-port=2000

# Visualize AV route
python scripts/run.py --agent_cfg expert.yaml --scenario_cfg standard_eval.yaml --mode eval -viz_route

BEV map

# Launch CARLA
./CarlaUE4.sh -prefernvidia -windowed -carla-port=2000

# Visualize BEV map
python scripts/run.py --agent_cfg expert.yaml --scenario_cfg FPPO_adv_eval.yaml --mode eval --eval_mode render

🔖 Citation

If you find our paper useful, please kindly cite us via:

@inproceedings{
  chen2024frea,
  title={{FREA}: Feasibility-Guided Generation of Safety-Critical Scenarios with Reasonable Adversariality},
  author={Keyu Chen and Yuheng Lei and Hao Cheng and Haoran Wu and Wenchao Sun and Sifa Zheng},
  booktitle={8th Annual Conference on Robot Learning},
  year={2024},
  url={https://openreview.net/forum?id=3bcujpPikC}
}

📋 Acknowledgement

This implementation is based on code from several repositories. We sincerely thank the authors for their awesome work.

• SafeBench
• FISOR
• PlanT
• ElegantRL
• distance3d
• King
• Two-Dimensional-Time-To-Collision