Deep reinforcement learning for ball balancer control

Table of Contents

Objective
[Technologies used](#Technologies used) [Control architecture](#Control architecture)

Objective

The goal of this project is to create a DDPG agent able to control a ball balancer in order to move the ball with the desired trajectory.

Technologies used

• DDPG (deep deterministic policy gradient)

  • Ball to plateau angle controller

• OpenAi gym

  • Environement / System simulation

• Genetic optimisation

  • Motor position controller

• SGD feed forward neural network regression

  • Motor and ball simulation

• Neural networks

  • Pytorch

Control architecture

Here is the heigh level control loop.

Given this choice of architecture, the project can be devided in the main parts :

• Plateau control and simulation Dedicated page
• Ball control and simualtion Dedicated page

A simulation is needed for every component of the system because the controllers training need that.

The plateau controller is a ffw neural network trained thanks to a genetic optimisation. The ball controller is also a ffw neural network, but this time trained with DDPG.

Detailed in its dedicated page.