LASER: SCRIPT EXECUTION BY AUTONOMOUS AGENTS FOR ON-DEMAND TRAFFIC SIMULATION

To make a great film, you need three things–the script, the script and the script. – Alfred Hitchcock

LASER.mp4

Framework Overview

Experiment Results

Experiments show that LASER can generate scripts effectively based on user requirements, with only 3.18% of the final script's characters input by the user. It executes the scripts efficiently, achieving a 90.48% success rate and averaging 1606.09 tokens per simulation second per agent.

Experiment records are available in the experiment_result folder.

Getting Started

Environment

clone the repo

git clone https://github.com/CXYyp5SkNg/CXYyp5SkNg.github.io.git
git submodule update --init --recursive

create conda environment

conda create -n laser python=3.8
conda activate laser
pip install -r requirements.txt 

setup and run CARLA 0.9.15

setup Docker

setup NVIDIA Container Toolkit

docker build -t carla_amap:0.9.15 .
docker run --privileged --gpus all --net=host -e DISPLAY=$DISPLAY carla_amap:0.9.15 /bin/bash ./CarlaUE4.sh

exit CARLA

docker stop $(docker ps -q)

set up openai key

export OPENAI_API_KEY=sk-your-apikey

set up Interfuser as test subject

conda activate laser
cd interfuser
pip install -r requirements.txt
python setup.py develop # this will install the customized timm where Interfuser lies in

export PYTHONPATH=$PYTHONPATH:./scenario_runner
export PYTHONPATH=$PYTHONPATH:./PythonAPI/carla/
export PYTHONPATH=$PYTHONPATH:./leaderboard

Download model weights at here. Then move the model to leaderboard/team_code/interfuser.pth.tar.

Generate Script with LASER

mkdir laser_scenes/your_scene_name
cp laser_scenes/Swerve/UserPrompt.txt laser_scenes/your_scene_name
vim laser_scenes/your_scene_name/UserPrompt.txt
python3 laser_sg.py -s "your_scene_name"

Execute Script with LASER

python3 laser_se.py -r T04Highway -s "laser_scenes/Swerve/script.json" -t 10

After the execution, check se_records folder for generated video.