Artificial Intelligence (AI) and Machine Learning (ML) are an increasing area of emphasis for the Department of Defense (DoD). In their 2022 budget the Pentagon requested $4.3 billion dollars for AI/ML related efforts, an over 50% jump in a two-year span. Unfortunately, much of the driving technology is being developed for commercial applications due to the relatively small size and challenges associated with the defense market. Common commercial assumptions like ample compute power, internet connectivity, and stable power are not valid in DoD applications due to deployment in what is referred to as the edge on low size, weight, and power (SWaP) platforms. Edge AI/ML refers to the process of processing data and running algorithms on the deployed device rather centrally. This concept allows AI/ML to run without connectivity, an important feature for military platforms that can experience degraded communication. In addition, DoD edge computing often happens in low SWaP environments due to mission and cost requirements. This can be challenging for commercially developed technology that often is designed to live in large data centers with access to ample compute resources. As a result, to make AI/ML technology deployable and tactically relevant, edge computing and low SWaP deployment problems need to be addressed.
This paper describes the process of deploying AI/ML models in low SWaP environments at the edge. Specifically, the paper looks at developing a vision based automatic take off and landing deep learning system for an unmanned aerial vehicle (UAV). The paper details the model development and optimization process for low SWaP deployment. In addition, the paper covers testing of the model using different deployment optimization strategies and the trade-offs associated with each. Ultimately, the paper will provide the community with an example of how to address a pressing problem associated with deploying AI/ML for military applications.
Keywords
DEEP LEARNING,MACHINE LEARNING,UAV
Additional Keywords