/ai-fip

AI-controlled Flywheel Inverted Pendulum

AI-controlled Flywheel Inverted Pendulum

The goal is to build a flywheel inverted pendulum (FIP) model stabilized by a neural network trained by deep reinforcement learning method. Literally, the network will not know anything about the physics of the phenomena and about it's own physical "body", it will use a method of trial and error in order to figure out how to get to upright position.

Milestones

  • Get enough understanding of a physics model
  • Simulate free-fall system with spontaneous rotation of a wheel
  • Implement a balancing algorithm in a simulation
  • Implement a swing-up algorithm in a simulation
  • Build a physical model with stepper motor for rotating a wheel
  • Try other controller to stabilize the system (TBD)
  • Learn DRL enough to apply to the system
  • Pre-train a NN in a simulation and use it inside a real device
  • Allow NN to tune itself in a real device

References