Optimal Motion Planning in 3D Workspaces: Integrating a
Panel-Method-Based Motion Planner with Continuous
Deep Reinforcement Learning

Μαλλιαρόπουλος Κατσίμης, Μάριος; Malliaropoulos Katsimis, Marios

dc.contributor.author	Μαλλιαρόπουλος Κατσίμης, Μάριος	el
dc.contributor.author	Malliaropoulos Katsimis, Marios	en
dc.date.accessioned	2023-08-23T07:22:13Z
dc.date.available	2023-08-23T07:22:13Z
dc.identifier.uri	https://dspace.lib.ntua.gr/xmlui/handle/123456789/57923
dc.identifier.uri	http://dx.doi.org/10.26240/heal.ntua.25620
dc.rights	Αναφορά Δημιουργού - Μη Εμπορική Χρήση - Παρόμοια Διανομή 3.0 Ελλάδα	*
dc.subject	Robotics	en
dc.subject	Optimal Control Systems	en
dc.subject	3D Motion Planning	en
dc.subject	Deep Reinforcement Learning	en
dc.subject	Fluid Mechanics	en
dc.subject	Ρομποτική	el
dc.subject	3D Σχεδιασμός Πορείας	el
dc.subject	Βέλτιστος Έλεγχος Συστημάτων	el
dc.subject	Ενισχυτική Μάθηση	el
dc.subject	Μηχανική των Ρευστών	el
dc.title	Optimal Motion Planning in 3D Workspaces: Integrating a Panel-Method-Based Motion Planner with Continuous Deep Reinforcement Learning	en
dc.contributor.department	Control Systems Lab	el
heal.type	bachelorThesis
heal.classification	Mechanical Engineering, Robotics	en
heal.language	en
heal.access	free
heal.recordProvider	ntua	el
heal.publicationDate	2023-07-01
heal.abstract	This diploma thesis proposes a novel and proven correct reactive method for planning three-dimensional optimal motion in complex environments. By combining fluid flow equations, optimal control theory, and deep reinforcement learning techniques, this study offers an interdisciplinary and unique approach, effectively merging positive attributes from different scientific fields. The method models the 3D motion planning problem by solving streamlines of the potential fluid flow, enabling the proper handling of various terrain types. This is achieved through the discretization of the geometry into surface panels, while the safety criteria are ensured via a set of von-Neumann boundary conditions. The proposed fluid-based planner guarantees a continuous-time, natural-looking, stable and safe solution for the motion planning problem with Artificial Harmonic Potential Fields (AHPFs). Furthermore, this thesis presents a model-based reinforcement learning algorithm for learning the optimal non-linear control in continuous time and action space with respect to an infinite horizon cost function. The algorithm utilizes an actor-critic scheme based on policy iteration, to successively approximate the optimal solution of the Hamilton-Jacobi-Bellman equation. This way, the optimal robot motion is obtained by iteratively updating the fluid flow parameters (i.e., the controller parameters) in a deterministic manner. The proposed method demonstrates fast convergence and outperforms widely used methods such as the RRT*, highlighting its contribution to the field of 3D optimal motion planning.	en
heal.advisorName	Κυριακόπουλος, Κώστας	el
heal.advisorName	Kyriakopoulos, Kostas	en
heal.committeeMemberName	Παπαδόπουλος, Ευάγγελος	el
heal.committeeMemberName	Αντωνιάδης, Ιωάννης	el
heal.academicPublisher	Εθνικό Μετσόβιο Πολυτεχνείο. Σχολή Μηχανολόγων Μηχανικών	el
heal.academicPublisherID	ntua
heal.fullTextAvailability	false
heal.fullTextAvailability	false