arjunnwick/ai-rocket-landing-ppo

AI Rocket Landing — PPO Reinforcement Learning Model 🚀

Overview

AI Rocket Landing is a reinforcement learning project that trains an AI agent to control a simulated lunar lander. The agent learns through trial and error to manage its movement and attempt safe landings in the Gymnasium LunarLander-v3 environment.

The project uses Proximal Policy Optimization (PPO), a reinforcement learning algorithm, implemented with Stable-Baselines3 and PyTorch.

Model Details

Property	Value
Algorithm	Proximal Policy Optimization (PPO)
Policy	MlpPolicy
Environment	Gymnasium LunarLander-v3
Action space	Discrete (4 actions)
Observation space	8 values
Training duration	Approximately 500,000 timesteps
Model format	Stable-Baselines3 ZIP checkpoint

Training

The agent was trained in the LunarLander-v3 simulation using reward feedback from the environment. PPO updates the policy to improve the agent's decisions over repeated episodes.

The model checkpoint is provided as ppo_lunarlander_v2.zip.

Evaluation Results

In a recorded evaluation of 10 episodes:

• Average reward: 207.68
• Best episode reward: 264.36
• Worst episode reward: 0.57
• Episodes with reward of at least 200: 8 out of 10

These results represent one evaluation run. Performance may vary between runs, and a high reward does not guarantee a safe landing in every episode.

How to Load the Model

Install the required libraries:

pip install "gymnasium[box2d]" stable-baselines3

Download ppo_lunarlander_v2.zip from this repository, then run:

import gymnasium as gym
from stable_baselines3 import PPO

env = gym.make("LunarLander-v3")
model = PPO.load("ppo_lunarlander_v2.zip", env=env)

obs, info = env.reset()

for _ in range(1000):
    action, _ = model.predict(obs, deterministic=True)
    obs, reward, terminated, truncated, info = env.step(action)

    if terminated or truncated:
        obs, info = env.reset()

env.close()

Intended Use

This model is intended for educational purposes, reinforcement learning experiments, and simulated lunar landing research.

Safety note: This is a simulation-only model. It has not been validated for controlling real spacecraft or physical landing systems.

Project Links

• GitHub source code: https://github.com/arjunnbharath/AI-Rocket-Landing
• Model repository: Available in this Hugging Face repository

Author

Created as part of an AI engineering and reinforcement learning project.