Simulation of a Quadcopter Stabilization System Using a Neural Reinforcement Learning Algorithm

Setting up classical PID controllers works for systems with a priori known properties and parameters. After changing the characteristics of an object, such regulators may not be able to stabilize its control variables and break the specified limits e.g. when the angular velocity is exceeded as a result of an engine malfunction. This paper presents an algorithm for the neural stabilization of a quadcopter. We use reinforcement learning method to enable adaptation of the agent during the control process. The quadcopter is simulated in a virtual environment and trained in a series of episodes with sequential assessment of its movement efficiency and tuning neural network parameters. A comparison of the transient characteristics of the quadcopter in various conditions showed that the neural network controller surpasses the classical PID controller by several significant criteria.