Autonomous Methods for Learning and Pruning Motion Primitives for Navigation and Adversarial Tasks
Author(s)
Goddard, Zachary Carl
Advisor(s)
Editor(s)
Collections
Supplementary to:
Permanent Link
Abstract
Kinodynamic motion planning presents the difficult problem of finding a trajectory to reach a goal under both kinematic and dynamic constraints. Motion primitives can simplify this problem by solving the dynamics offline and allowing for fast planning via concatenation. A key challenge with these methods is the design of a primitive library which concisely captures the capabilities of the system. This work explores the use of reinforcement learning and genetic algorithms to autonomously generate motion primitive libraries and prune them down to their most effective components. The core contributions of this framework are the design of a shaping reward and primitive extraction procedure to learn new motion primitives, and mutation operations to help optimize the learned library. We evaluate this learning framework on navigation and adversarial tasks to demonstrate its ability to improve planning performance by learning effective motion primitives. Additionally, we develop heuristics and a post-process optimization procedure for the Hybrid A* algorithm with the Maneuver Automaton, and we design a primitive-base Monte Carlo Tree Search algorithm for adversarial applications.
Sponsor
Date
2023-07-25
Extent
Resource Type
Text
Resource Subtype
Dissertation