🚁 Drone Stabilization AI

This project is a physics-based simulation where a drone learns to stabilize itself using reinforcement learning. It employs an Actor-Critic neural network to make real-time stabilization decisions, adapting to dynamic conditions through trial and error.

🚀 Features

• Reinforcement Learning: Uses an Proximal Poclicy Optimization (PPO Actor-Critic) to train.
• Physics Simulation: (Somewhat) realistic drone motion with forces and torque calculations.
  • Playable by both the AI and the User
• Visualization: Built with Pygame for to display saved models' performances
  • No visualization during training, but there are generated checkpoints that can be visualized as the model trains

🏃‍♂️ Running The Simulation

python train.py

The simulation creates logs that are visualized via tensorboard (localhost website). When running train.py it will by default make a directory runs/ that will store logs.

To run tensorboard:

tensorboard --logdir=runs

Environment

This project was built and tested around Python 3.11 (so it might not work on other versions).

Dependencies:

pip install torch numpy pygame tensorboard