karthik-ragunath/car_racer_sac

Deep reinforcement learning in autonomous driving

Latent space reinforcement learning for autonomous driving (CarRacing-v0)

This repository contains code to set up a reinforcement learning agent using the Soft Actor-Critic algorithm (SAC) [https://arxiv.org/pdf/1801.01290.pdf]. As perception module we choose a $\beta$-VAE [https://openreview.net/pdf?id=Sy2fzU9gl].
Please note that while we provide a working implementation of Twin Delayed Deep Deterministic Policy Gradient (TD3) this algorithm has only been used for experimental reasons and is thus neither documented nor supported.

Getting started

Clone the repo and install the dependencies from environment.yml (or requirements.txt). We recommend using miniconda to run our code as this project was developed using conda. train_sac.py is the main training script for our algorithm and display_sac.py shows the trained agent.

Project structure:

.
├── models    # Folder for autoencoder and SAC models
|
├── perception   # Vision module
|   ├── autoencoders.py    # Contains our encoder models
|   ├── generate_AE_data.py    # Script to generate encoder training data
|   └── utils.py    # helper Functions for encoder data handling
|
├── sac    # Soft Actor-Critic implementation
|   ├── model.py   # SAC neural net models
|   ├── replay_memory.py   # Replay buffer
|   ├── sac.py   # Main SAC implementation and training class
|   └── utils.py   # Helper functions for SAC training
|    
├── docs    # Documentation for the SAC+perception modules
|   ├── perception    # Docs for the perception module
│   |   └── xxx.html    # Html files
|   |
|   ├── sac    # Docs for the SAC algorithm
|   |   └── xxx.html    # Html files
|   |
|   ├── displaymodel_sac.html    # SAC evaluation script
|   ├── train_sac.html    # SAC training script
|   └── train_vae.html    # VAE training script
|
├── runs    # Example tensorboard training log 
│   └── log folders    # Folders containing a training run's logs
|       └── train logs    # tensorboard log files
|
├── td3    # TD3 implementation (not supported)
|   ├── td3.py    # main TD3 implementation (not supported)
|   └── utils.py    # TD3 replay buffer (not supported)
|
├── displaymodel_sac.py    # Evaluation script for a trained agent
├── train_sac_baselines.py    # Training script for the baselines agent
├── train_sac.py    # Training script for our SAC implementation
├── train_td3.py    # Training script for our SAC implementation
├── train_vae.py    # Training script for our VAE
├── environment.yml    # Conda environment
├── requirements.txt    # Pip requirements
└── README.md     # This file

Weights

The weights of our trained encoder models can be found in models and are named "weights.py" for the deterministic encoder and "VAE_weights.py" for the VAE.
In addition to that we also provide some pretrained SAC agents. The best performing of those is "klein_6_24_18".

Results

We provide a writeup containing the detailed results and further information about our setup here: https://www.dropbox.com/s/ilkrxnwvfe493py/Latent_RL_paper.pdf?dl=0.

Use this work

Feel free to use our source code and results.

@article{LatReinforcementCARLA2019, 
	title={Latent Space Reinforcement Learning for Continous Control: An Evaluation in the Context of Autonomous Driving}, 
	author={Klein, Timo and Sigloch, Joan and Michel, Marius}, 
	year={2019}, 
}