This repository contains the code for running experiments and generating results.
The code for training and testing is based on the work in following repository:
https://github.com/bitsauce/Carla-ppo
##Requirements
https://github.com/carla-simulator/carla/releases
We used Windows version 0.9.5 and python 3.5.
This CARLA version comes with PythonAPI for python 3.5.
In order to install the PythonAPI package in your python 3.5 environment, please refer to:
carla-simulator/carla#1466
conda install tf-gpu
conda install pandas
pip install pygame
pip install onnx
pip install onnx-tf
pip install tensorflow-probability==0.7
pip install gym
pip install scikit-image==0.16.2
Pip install opencv-python
Run Carla with 'Town04' command line argument.
Run train.py and pass the command line arguments specified by the script.
Run train.py with -test command line argument.
Pretrained models for agent 1,2 and 3 are located inside 'models' folder.
To generate paper results, run plot_training_results.py and plot_histograms.py
We used windows desktop machines with core i7-8700K 3.7GHz CPU, 32GB of RAM, and a TITAN-V GPU. Training for 6000 episodes took around 5 days for each agent.