ruiyang-061x / lunar-lander

Lunar lander trained by reinforcement learning. Implemented by pytorch.

Lunar Lander

INTRODUCTION

Lunar Lander is about landing a spacecraft on a specific landing area on moon safely.
Gymnasium is a open-source toolkit for reinforcement learning, which contains various reinforcement learning environments and provides a standard development procedure of reinforcement learning. It originates from OpenAI's gym. Refer to their website for more information.

METHOD

The actions of spacecraft contain:

• do nothing
• fire left orientation engine
• fire main engine
• fire right orientation engine

The rules of award:

• increased the closer the lander is to the landing pad
• increased the slower the lander is moving
• increased the more the lander is horizontal
• increased by 10 points for each leg that is in contact with the ground
• decreased by 0.03 points each frame a side engine is firing
• decreased by 0.3 points each frame the main engine is firing

The spacecraft is trained by DDPG(Deep Deterministic Policy Gradient), which is a extension of DQN(Deep Q-Network) aiming to solve the problem of continuous control.

ENVIRONMENT

The core dependencies are listed below.

• python 3.11
• torch 2.0.1+cu117
• gymnasium 0.29.0
• tensorboardX 2.6.1
• cost of GPU is around 1GB

TRAINING

python main.py

TESTING

python load.py

There is a ready-to-use model in this repo. Run the command above to see the result.

DEMO

rl-video25.18.7.57-episode-0.mp4

ACKNOWLEDGEMENT

The code is mainly based on drl-ddpg-lunar-lander. Thanks the aurther for the excellent work.

AND

Feel free to raise an issue if you encounter any problem :).