Image Control of the Inverted Pendulum

There are many different techniques for the measurement and control of dynamic real-world systems. The traditional approach to motor control problems is to find angular position/velocity with one or more rotary encoders. This information is then typical fed into a control algorithm to administer negative feedback on the observed state. In applications where encoders are not feasible, other techniques must be used in order to observe the system. This restriction is the motivation for our project – we aim to control the inverted pendulum solely through images of the system. The model will learn how the images are associated with the angular states of the system, and then it will apply different RL techniques to decide the optimal action for the motor in order to keep the pendulum balanced without any prior information about the physical dynamics. This project compares three different methods to control the inverted pendulum. The first uses a Neural Network to determine the angle of the system from the input images, and then subsequently uses Reinforcement Learning to control and stabilize the system. The second design is the traditional classical control approach that serves as our baseline. These two methods were done successfully in simulation, so we moved onto our third model physical implementation. For this, we experimented with an adapted form of the above method as well as a new addition with Deep Q Learning [1].