Using Reinforcement Learning to Achieve Two Wheeled Self Balancing Control

The non-linear, unstable system of the two wheeled self-balancing robot has made it a popular research subject within the past decade. This paper outlines the design of a two wheeled robot with self balancing control systems using Reinforcement Learning. The BeagleBone Black platform was used to design the two wheeled robot. Along with the motor, the robot was also equipped with an accelerometer and gyroscope. Using the Q-Learning method, adjustments to the motor were made according to the dip angle and the angular velocity at that given time to return the robot to balance. The experimental results show that using this reinforcement learning method, the robot has the ability to quickly return to a balanced state under any dip angle.