Derin pekiştirmeli öğrenme yöntemleri ve geleneksel kontrolcülerin havasal manipulasyon için karşılaştırmalı çalışması

Aerial manipulation with unmanned aerial vehicles is increasingly becoming a necessity in many applications. In this thesis, we analyze the controller approaches for a bi-rotor aerial manipulator for a pick and place operation. First of all, we compare a classical control approach with a minimum snap trajectory generation and Deep Reinforcement actor-critic algorithms for the control of the aerial manipulator. Furthermore, we examine the effects of degrees of freedom of the manipulator for the Deep Reinforcement Learning approaches and analyze how the change of goal position of the object that the aerial manipulator needs to carry affects the training of the learning approaches. Moreover, to obtain a faster convergence for the learning approaches we have added informative states in which the aerial manipulator starts with the object it needs to carry is already grasped. Our results of the 2D simulation environment for the aerial manipulation suggest that all of the actor-critic algorithms yield valuable results with off-policy algorithms being more sample efficient. Still, these algorithms have stability issues that fail in the task for some cases. On the other hand, the classical controller approach does not have this problem but finishes the task slower than the Deep Reinforcement Learning approaches.