Aplicação do algoritmo DQN a um robô Delta

Abstract

The present work aims to model the kinematics of a Delta parallel robot and the application of the DQN algorithm in an environment developed based on this model, the agent trained in such environment learns through trial and error, receiving positive rewards when it correctly guesses the next step and negative rewards when it misses. Thus, through the iterative process, the agent develops new policies, that is, new rules on how to take actions given a new state of the environment. A simulation environment whose objective is to correctly position the Delta robot's effector at a given point in space is proposed, this environment consists of an algorithm that uses the mechanism mathematical model to generate renders of the robot every time the agent takes an action. It was possible to obtain 80% of correct answers during training, which can be observed in the rewards graph in the results chapter. With the implementation of the algorithm in the Delta robot environment, it was possible to observe that the implemented solution can solve most of the episodes to which the agent is presented, however, in some specific points, the agent fails to obtain a solution. This usually occurs at singularity points, that is, regions where the mechanism may lose its structural rigidity or fail to obtain a unique solution.