Past Event: Center for Autonomy Seminar

Practical Challenges for Compositional RL

Dr. Jared Culbertson, Research Mathematician, U.S. Air Force Research Laboratory's Autonomous Capabilities Team (ACT3)

Abstract

In this talk, rather than presenting any specific novel results, we will focus on the practical problems and technical challenges associated with our current work in applying reinforcement learning to the air combat and spacecraft inspection problems. Framed by previous work on sequential, parallel, and hierarchical composition and informed by the specific constraints of our ongoing system development for these real-world applications, we will discuss the primary problems that we are currently facing in realizing flexible agent training and transfer across different environments and tasks. We will also briefly describe a new highly configurable RL library (CoRL) built to enable rapid experimentation and transfer from simulation-based training to real-world execution. 

Biography

Dr. Jared Culbertson is a research mathematician with the U.S. Air Force Research Laboratory's Autonomous Capabilities Team (ACT3), a research group focused on the development and deployment of flexible AI solutions across a diverse set of air and space mission areas. Jared's research primarily deals with fundamental aspects of representational structures, recently involving compositional approaches for hybrid dynamical systems and now focused on behavior acquisition, diversity, and composition in reinforcement learning problems.