barna22/duckietown

Duckietown reinforcement learning project

duckduck

Duckietown reinforcement learning project

Team name: Syntax Terror

Medium cikket fogunk írni, valamint már van aláírásunk előző félévből, Csapó Tamás Gáborral megbeszéltük, hogy leadhatjuk a vizsgáig a házit.

Members:

• Istenes Dóra PQ08W1
• Kenesei Benjámin MS725L
• Suciu Barnabás ASOG9J

The aim of this project is to train a reinforcement learning agent in a simulated autonomous driving environment. The agent is referred to as a "duckiebot", and it is trained using reinforcement learning in the simulator, then it can be transferred to a real model vehicle in the physical test environment.

The repository contains the simulation environment, and the scripts and configuration files necessary for the trainig.

Setup and simulation

After installing the packages specified in the requirements file, the simulation environment can be started with the following command:

python manual_control.py --env-name Duckietown-udem1-v0

The running application looks like this with the default config:

Training

After fixing several errors related to outdated code, and moving train_reinforcement.py and enjoy_reinforcement.py scripts from /learning/reinforcement/pytorch to /learning, we can run the training with the following command:

python3 train_reinforcement.py

The training should finish like this:

This generates a model in the following folder: learning/reinforcement/pytorch/models.

Evaluation is done by running the following script, which simulates the agent in the environment:

python3 enjoy_reinforcement.py

Metrics can also be found in the results folder, in the rewards.npz file.