NUMERICA CORPORATION — Department of Defense SBIR Phase I: AF151-058

Description

ABSTRACT:The United States Air Force must increasingly provide supplies and munitions via airdrop to ground forces spread diffusely over large areas amid hostile enemy-held terrain. The calculation of the airdrop release point is crucial for ensuring safe and accurate delivery to designated landing sites. Wind and parafoil models have insufficient fidelity to accurately predict the landing location from a given release point. We propose a machine learning approach using Gaussian process regression to assimilate historical data of previous trajectories to more accurately learn the predicted landing locations. Having learned more accurately the mapping from release point to landing site, the optimal release point is then selected. Our approach then recalculates small adjustments to the aircraft navigation to achieve the optimal aircraft approach trajectory subject to constraints of a pre-specified flight corridor. The Gaussian process approach provides a unified holistic approach rigorous mathematical framework for probabilistic machine learning that allows the data to speak for itself while imposing very few assumptions. Moreover, we propose a Cramer-Rao bound analysis to determine if an improvement in air drop accuracy can indeed be supported by the data. BENEFIT:The proposed machine learning approach will allow the Air Force to utilize historical trajectory data to improve the accuracy of predicted airdrop landing locations. The process will ensure that supplies and munitions are delivered more accurately and with greater certainty leading to less waste and fewer payloads going astray into enemy held terrain. The approach also provides for optimal navigation updates for the delivery aircraft subject to constraints on terrain and designated safety corridors.The benefits are not restricted to military applications only the same approach will provide improved delivery of humanitarian aid and supplies throughout the world. Numericas intended commercialization is to integrate the approach into the existing Joint Program Airdrop System. Numerica will work with prime contractors and suppliers of airdrop equipment to commercialize the machine learning algorithms to improve commercial airdrop applications. Beyond airdrop applications, improved model predictions via machine learning meets an emerging need in many industries including cyber-security, air traffic safety, and civilian policing.