STTP (Stackelberg Training-Testing Paradigm) is a novel approach that aligns with the hierarchical nature of interactions between autonomous vehicle (AV) policy training and risky scenario generation, facilitating AV performance improvement in a leader-follower order of play. We empirically demonstrate that AV and background vehicle (BV) agents can both achieve better performance in STTP. Essentially, STTP outperforms those competing baselines with simple adversarial training paradigms, non-game modeling and other game modeling approaches, especially in higher dimensional scenarios.

Before getting started, first you should install SUMO properly following the sumo documentation.

Then, to install the dependencies, run the command:

conda env create -f environment.yml
conda activate STTP
pip install -e .

Then, download our pre-processed dataset from here, and unzip it using

unzip dataset.zip

We run experiments on STTP, Non-Game, SimGM, NSD and I-STTP. To begin with, enter the folder Scripts:

cd Scripts

Then you can run STTP and baselines using following example commands.

python main_spg_ac_inv.py --r_ego stackelberg --r_adv stackelberg --is_save True --save_model True --adv_policy sumo --num_agents 5

If you want to run all ablation experiments (which is relatively time consuming), simply use

bash run.sh

python main_NSG.py --r_adv stackelberg3 --r_ego stackelberg --is_save True --save_model True --pretrain_ego True --adv_policy sumo --num_agents 5

python main_SimGM.py --r_adv stackelberg2 --r_ego stackelberg --is_save True --save_model True --pretrain_ego True --adv_policy sumo --num_agents 5

You can resort to wandb to login your personal account with your wandb API key.

export WANDB_API_KEY=YOUR_WANDB_API_KEY

and using --used_wandb True to turn on the online syncronization.

Our codebase structure mainly derives from (Re)2H2O. Many thanks to them!