/rl-duckiebot

This repository contains different reinforcement learning algorithms and classic control comparison benchmarked in duckietown.

Primary LanguagePython

Docker Image CI DeepSource

Reinforcement Learning in Duckietown

Outline

Work in progress repository, that contains various rl algorithms and classic control theory benchmarked in duckietown Current algortihms will be:

  1. Classic PID.
  2. Deep Deterministic Policy Gradient Algorithm and it's modifications. Currently implemented DDPG with VAE taken from https://github.com/araffin/learning-to-drive-in-5-minutes

Installation

  1. Install https://github.com/Laboratory-of-Embodied-Intelligence/gym-duckietown with instructions in readme.
  2. conda env create --name MY_ENV_NAME -f environment.yml
  3. cd rl && python3 stable_train.py will train DDPG+VAE, cd rl && python3 train.py will train vanilla version of DDPG taken from https://github.com/philtabor/Deep-Q-Learning-Paper-To-Code

Training parameters

There are a lot of hyperparameters to choose, all list is written in stable_train.py, by tweaking them you can drastically change behaviour of the model.