{"id":{"repo_id":"umn","oai_identifier":"oai:conservancy.umn.edu:11299/280266"},"canonical_url":"https://search.dev.ndltd.org/etd/umn/oai:conservancy.umn.edu:11299/280266","repository":{"repo_id":"umn","name":"University of Minnesota","base_url":"https://conservancy.umn.edu/server/oai/request"},"display":{"title":"Ml Controllers With Memory For Robust Quadrotor Control And Research","abstract":"Over time, Non-Classical controllers are becoming more common and the interest of much research. Neural network controllers are some of the most common. These Neural Network controller offer advantages such as higher efficiencies, ability to learn nonlinear models, and adaptability. Designing such controllers typically relies on optimization and reinforcement learning methods, such as PPO or DDPG learning. These learned controllers require many action-time steps, and performing many of these actions directly while learning can result in undesirable effects on the real system. Undesirable effects may include hardware damage, behaving unsafely or the quadrotor crashing. In order to collect the required data over many action/time steps while also avoiding undesirable system effects, leaned controllers are typically learned on simulated systems. These simulations usually utilize different physics engines such as Bullet, NVIDA PhysX, or Jolt Physics. While these simulations are reasonably good, they are not perfect analogs of our real system; therefore, our controller can learn undesirable behaviors from these imperfect simulations. In this thesis we will explore different control architectures and ML control methods for a quadrotor, and in doing so develop a framework for testing and evaluation of non-classical controllers for future SimToReal research. The system we will control a Crazyflie 2.X quadrotor drone, and its pose will be measured by the VICON motion capture system. We will be controlling the system with a neural network controller which is learned from a pyBullet simulation using the OpenAI Gym framework. Many different training methods and control algorithms are created, tested, and their robustness compared. These methods were all implemented into a real system and their effectiveness compared empirically. The most robust model was a RNN of multi-layer perceptron using Long Short Term Memory (LSTM) units. The resulting system enables users to define different control algorithms and control a quadcopter remotely. Allowing for continued research in neural and non-classical control methods for aerial robotics","abstract_html":"Over time, Non-Classical controllers are becoming more common and the interest of much research. Neural network controllers are some of the most common. These Neural Network controller offer advantages such as higher efficiencies, ability to learn nonlinear models, and adaptability. Designing such controllers typically relies on optimization and reinforcement learning methods, such as PPO or DDPG learning. These learned controllers require many action-time steps, and performing many of these actions directly while learning can result in undesirable effects on the real system. Undesirable effects may include hardware damage, behaving unsafely or the quadrotor crashing. In order to collect the required data over many action/time steps while also avoiding undesirable system effects, leaned controllers are typically learned on simulated systems. These simulations usually utilize different physics engines such as Bullet, NVIDA PhysX, or Jolt Physics. While these simulations are reasonably good, they are not perfect analogs of our real system; therefore, our controller can learn undesirable behaviors from these imperfect simulations. In this thesis we will explore different control architectures and ML control methods for a quadrotor, and in doing so develop a framework for testing and evaluation of non-classical controllers for future SimToReal research. The system we will control a Crazyflie 2.X quadrotor drone, and its pose will be measured by the VICON motion capture system. We will be controlling the system with a neural network controller which is learned from a pyBullet simulation using the OpenAI Gym framework. Many different training methods and control algorithms are created, tested, and their robustness compared. These methods were all implemented into a real system and their effectiveness compared empirically. The most robust model was a RNN of multi-layer perceptron using Long Short Term Memory (LSTM) units. The resulting system enables users to define different control algorithms and control a quadcopter remotely. Allowing for continued research in neural and non-classical control methods for aerial robotics","abstract_has_math":false,"creators":["Navarrete, Nicholas"],"institution":null,"degree_name":null,"degree_level":null,"degree_discipline":null,"degree_department":null,"school":null,"contributors":[],"advisors":[],"committee_chairs":[],"committee_members":[],"year":2026,"date_issued":"2026-02","date_published":"2026-02","updated_at":"2026-07-24T05:20:07Z","subjects":["Control","Machine learning","Neural networks","Proximal Policy Optimization","Quadrotors","Sim2Real"],"languages":["en"],"rights":[],"rights_urls":[],"identifier_entries":[]},"links":{"outbound_url":"https://hdl.handle.net/11299/280266","outbound_label":"Handle","outbound_source":"dc:identifier.uri"},"metadata_groups":[{"id":"people","label":"People","entries":[{"key":"dc:creator","label":"Author","values":["Navarrete, Nicholas"]}]},{"id":"academic_context","label":"Academic Context","entries":[{"key":"dc:date.accessioned","label":"Dc Date Accessioned","values":["2026-05-11T16:32:07Z"]},{"key":"dc:date.issued","label":"Date","values":["2026-02"]},{"key":"dc:type","label":"Dc Type","values":["Thesis or Dissertation"]}]},{"id":"subjects_keywords","label":"Subjects and Keywords","entries":[{"key":"dc:subject","label":"Dc Subject","values":["Control","Machine learning","Neural networks","Proximal Policy Optimization","Quadrotors","Sim2Real"]}]},{"id":"language_rights","label":"Language and Rights","entries":[{"key":"dc:language.iso","label":"Language (ISO)","values":["en"]}]},{"id":"identifiers","label":"Identifiers","entries":[{"key":"dc:identifier.uri","label":"Identifier URI","values":["https://hdl.handle.net/11299/280266"]}]},{"id":"additional","label":"Additional Metadata","entries":[{"key":"dc:description","label":"Description","values":["University of Minnesota M.S.E.C.E. thesis. February 2026. Major: Electrical/Computer Engineering. Advisor: Nicola Elia. 1 computer file (PDF); vi, 37 pages."]},{"key":"dc:description.abstract","label":"Abstract","values":["Over time, Non-Classical controllers are becoming more common and the interest of much research. Neural network controllers are some of the most common. These Neural Network controller offer advantages such as higher efficiencies, ability to learn nonlinear models, and adaptability. Designing such controllers typically relies on optimization and reinforcement learning methods, such as PPO or DDPG learning. These learned controllers require many action-time steps, and performing many of these actions directly while learning can result in undesirable effects on the real system. Undesirable effects may include hardware damage, behaving unsafely or the quadrotor crashing. In order to collect the required data over many action/time steps while also avoiding undesirable system effects, leaned controllers are typically learned on simulated systems. These simulations usually utilize different physics engines such as Bullet, NVIDA PhysX, or Jolt Physics. While these simulations are reasonably good, they are not perfect analogs of our real system; therefore, our controller can learn undesirable behaviors from these imperfect simulations. In this thesis we will explore different control architectures and ML control methods for a quadrotor, and in doing so develop a framework for testing and evaluation of non-classical controllers for future SimToReal research. The system we will control a Crazyflie 2.X quadrotor drone, and its pose will be measured by the VICON motion capture system. We will be controlling the system with a neural network controller which is learned from a pyBullet simulation using the OpenAI Gym framework. Many different training methods and control algorithms are created, tested, and their robustness compared. These methods were all implemented into a real system and their effectiveness compared empirically. The most robust model was a RNN of multi-layer perceptron using Long Short Term Memory (LSTM) units. The resulting system enables users to define different control algorithms and control a quadcopter remotely. Allowing for continued research in neural and non-classical control methods for aerial robotics"]},{"key":"dc:title","label":"Title","values":["Ml Controllers With Memory For Robust Quadrotor Control And Research"]}]}],"canonical_facts":{"dc:creator":["Navarrete, Nicholas"],"dc:date.accessioned":["2026-05-11T16:32:07Z"],"dc:date.issued":["2026-02"],"dc:description":["University of Minnesota M.S.E.C.E. thesis. February 2026. Major: Electrical/Computer Engineering. Advisor: Nicola Elia. 1 computer file (PDF); vi, 37 pages."],"dc:description.abstract":["Over time, Non-Classical controllers are becoming more common and the interest of much research. Neural network controllers are some of the most common. These Neural Network controller offer advantages such as higher efficiencies, ability to learn nonlinear models, and adaptability. Designing such controllers typically relies on optimization and reinforcement learning methods, such as PPO or DDPG learning. These learned controllers require many action-time steps, and performing many of these actions directly while learning can result in undesirable effects on the real system. Undesirable effects may include hardware damage, behaving unsafely or the quadrotor crashing. In order to collect the required data over many action/time steps while also avoiding undesirable system effects, leaned controllers are typically learned on simulated systems. These simulations usually utilize different physics engines such as Bullet, NVIDA PhysX, or Jolt Physics. While these simulations are reasonably good, they are not perfect analogs of our real system; therefore, our controller can learn undesirable behaviors from these imperfect simulations. In this thesis we will explore different control architectures and ML control methods for a quadrotor, and in doing so develop a framework for testing and evaluation of non-classical controllers for future SimToReal research. The system we will control a Crazyflie 2.X quadrotor drone, and its pose will be measured by the VICON motion capture system. We will be controlling the system with a neural network controller which is learned from a pyBullet simulation using the OpenAI Gym framework. Many different training methods and control algorithms are created, tested, and their robustness compared. These methods were all implemented into a real system and their effectiveness compared empirically. The most robust model was a RNN of multi-layer perceptron using Long Short Term Memory (LSTM) units. The resulting system enables users to define different control algorithms and control a quadcopter remotely. Allowing for continued research in neural and non-classical control methods for aerial robotics"],"dc:identifier.uri":["https://hdl.handle.net/11299/280266"],"dc:language.iso":["en"],"dc:subject":["Control","Machine learning","Neural networks","Proximal Policy Optimization","Quadrotors","Sim2Real"],"dc:title":["Ml Controllers With Memory For Robust Quadrotor Control And Research"],"dc:type":["Thesis or Dissertation"]},"updated_at":"2026-07-24T05:20:07Z"}