The objective of this project is to find a generic method for solving non-linear, continuous control problems, using deep learning and an Iterative Linear Quadratic Regulator controller.
During the course of this project, we solved a private control problem, for which we built an algorithm that may be applied to generic control problems. In the private problem, we are given a surface, centered on which is a robotic arm with 2 joints. At random, a ball is generated in a near-by location. The arm is controlled by two control signals which apply a separate force on the distal and proximal parts, with the goal of finding a sequence of actions that will move the tip of the arm from its initial position to a terminal position that is as close as possible to the ball.
In this paper, we present a generic solution that is based on the private solution which we engineered to resolve this problem. It is comprised of a neural network that learns the system’s dynamics and a deterministic controller (ILQR) that utilizes it when calculating the optimal control signals. The given solution can effectively handle control signal constraints.
Solutions to a multitude of problems can be achieved using the algorithm that we present, assuming necessary adjustments for each specific case.
