Title: Autonomous Methods for Learning and Pruning Motion Primitives for Navigation and Adversarial Tasks

 

Date: Wednesday, June 28, 2023

Time: 10:00AM EST

Location: GTMI 114

Virtual Link: https://teams.microsoft.com/l/meetup-join/19%3ameeting_MDViNjgyMTUtYTM1Mi00ZjNkLTlhNWMtODlkOTc4MzRjNjE2%40thread.v2/0?context=%7b%22Tid%22%3a%22482198bb-ae7b-4b25-8b7a-6d7f32faa083%22%2c%22Oid%22%3a%223e22115c-de4b-45a4-bc01-9eb934d29d35%22%7d

 

Zachary Goddard

Robotics PhD Student

School of Mechanical Engineering

Georgia Institute of Technology

 

Committee:

Dr. Anirban Mazumdar (Advisor) – School of Mechanical Engineering, Georgia Institute of Technology

Dr. Jonathan Rogers – School of Aerospace Engineering, Georgia Institute of Technology

Dr. Panagiotis Tsiotras – School of Aerospace Engineering, Georgia Institute of Technology

Dr. Seth Hutchinson – School of Interactive Computing, Georgia Institute of Technology

Dr. Kyle Williams – Pathfinder Technologies, Sandia National Laboratories

 

Abstract: 

Motion primitives provide a powerful means of rapid kinodynamic planning; however, the design of an effective primitive library for complex systems or tasks requires substantial expert knowledge. This work proposes an autonomous framework for learning motion primitives with minimal human input and demonstrates the process on simulated F-16 dynamics for navigation with and without obstacles. The framework combines deep reinforcement learning with our own contributions in the form of algorithms and shaping rewards to generate and select motion primitives for a maneuver automaton. Additionally, we contribute our own heuristics and post-processing algorithm to improve planning time with a state-of-the-art search algorithm, Hybrid A*. The demonstrated examples show significant improvement in the time to reach the goal on navigation tasks. This work then extends the framework to adversarial domains through the development of a primitive-based Monte Carlo Tree Search and a beam search modification guided by a learned heuristic model. We demonstrate the framework's ability to improve performance with primitives learned in the adversarial environment and demonstrate the benefits of motion primitives compared with forward simulated game tree search methods from existing literature.