SCIENTIFIC SYSTEMS CO INC — Department of Defense STTR Phase I: AFX20D-TCSO1

Description

We propose a solution framework to the problem of cross-domain collaboration of space and air assets to enable situational awareness and rapid mission planning for eVTOL/UAS logistic missions and disaster response in uncertain/dynamic environments. Additionally, the approach is extensible to observation of other space objects. Autonomous control, coordination and collaboration of LEO satellite constellation to provide ground and space imaging has great potential to rapidly enhance situational awareness. Complex spacecraft operation planning currently relies heavily on expert knowledge. A major challenge to automate the process is the requirements to meet power, fuel and other spacecraft health constraints, besides optimizing mission performance. The SSCI and University of Colorado Boulder team proposes the shielded deep reinforcement learning (SDRL) techniques to manage flight operations safely. Conventional machine learning methods such as reinforcement learning has the potential to automate the planning process. However, they lack guarantee of safety and satisfying operational constraints, which are essential involving space assets. SDRL learns off-line the best policy for coordinated multi-spacecraft flight mode selection under different conditions, while conforming to safety and operation constraints. In Phase 1, we will prove feasibility of SDRL in both improving performance and guaranteeing safety using an open-source simulation environment Basilisk. We will collaborate with stakeholders at AFRL and SMC to identify integration requirements for transition.