| dc.description.abstract | The deployment of drones has become increasingly popular in a variety of new applications. Many of such applications require autonomous and adaptive behaviors, especially when tasks require the need for dynamic object tracking, such as follow-me behavior. In the past decade, many applications have seen drastic benefits from using machine learning methods such as Reinforcement Learning (RL). RL uses an animal conditioning based approach in computational tasks to learn new behaviors in specific domains. In this thesis, the implementation of an RL algorithm is trained and tested inside simulation environments, specifically for the task of a drone following a person. This algorithm, a Deep Q-Network (DQN), is tested using four different approaches. First, two changes to the DQN inputs have been proposed to help improve the training process and performance. These suggestions include
the use of directionality of objects in its camera inputs using stacked image frames and the inclusion of depth information about its surroundings using depth maps. Tests have been run with these additions in two environments, each increasing in obstacle complexity. The results have shown that the use of stacked imaging resulted in improvements in environments where they relay valuable information to the agent about the objects in its view. Meanwhile, in environments where the task can be performed without it, they unnecessarily increase the state-space, resulting in degraded performance. Depth images have shown to be a strong improvements to each agent that used them, further reinforcing their strong simplifying capabilities and reduction of state-space. Second, the benefits of using RL compared to a static preprogrammed baseline have been evaluated. These tests have shown that RL allows for much more adaptive and flexible behavior, which is beneficial in each type of environment. Finally, the ability of RL agents to generalize behavior from simpler environments to a third, more complex environment has also been examined. This showed that the agents who were trained in an environment with obstacles, were able to transfer their knowledge to new similarly designed situations. Meanwhile, agents that had never seen an obstacle could not. These results show that the use of RL in the specific task of follow-me behavior or drones is a successful tool because of their adaptive and generalizable behavior. | |
