An AI pilot that taught itself to land on the Moon through 500,000 steps of trial and error.
Above: The fully trained "Ace Pilot" agent executing a perfect suicide burn and soft touchdown.
Project Selene is a Deep Reinforcement Learning project that solves the classic LunarLander-v2 environment. Unlike traditional PID controllers that use hard-coded physics rules, this agent started with zero knowledge of gravity or thrust.
Using Proximal Policy Optimization (PPO), the agent learned to:
- Manage fuel efficiency.
- Stabilize flight trajectory against random initial forces.
- Execute a soft landing (Velocity
$\approx$ 0) between the flags.
After 500,000 training steps, the agent achieved "Ace" status, consistently outperforming the environment's solving threshold of 200 points.
| Metric | Result | Notes |
|---|---|---|
| Success Rate | 87.0% | Over 100 test episodes |
| Average Reward | +222 | "Solved" status is >200 |
| Algorithm | PPO | MlpPolicy, Ent_coef=0.01 |
To visualize the thinking process, I built a custom telemetry pipeline using mplcyberpunk to track altitude vs. velocity in real-time.
A smooth, confident descent with a final velocity flare right before touchdown.
- Core: Python 3.10
- Environment: Gymnasium (
LunarLander-v2) - RL Algorithm: Stable-Baselines3 (PPO)
- Visualization: Matplotlib, Mplcyberpunk, MoviePy
- Physics Engine: Box2D
Want to see it fly on your own machine?
-
Clone the Repository
git clone [https://github.com/syd311/project-selene.git](https://github.com/syd311/project-selene.git) cd project-selene -
Install Dependencies
pip install gymnasium[box2d] stable-baselines3 shimmy matplotlib mplcyberpunk moviepy
-
Run the Simulation
python record_show.py
This will load the pre-trained
neural_lander_FINAL.zipand generate a new landing video.
This project is licensed under the MIT License - see the LICENSE file for details.
Created by syd311