Behavior Trees for Mission Management of High-Altitude Pseudo-Satellites

Abstract

The technology of high-altitude pseudo-satellites (HAPSs) has seen substantial progress in the past decades and is currently approaching operational readiness. However, HAPS aircraft are complex systems and require large crews of highly trained personnel as well as extensive operational safety measures. In order for these platforms to become commercially viable, it is imperative that mission-level tasks are automated in a mission management system, while maintaining flight safety. In this thesis, behavior trees (BTs) as developed for computer game artificial intelligence are investigated for the purpose of such mission management. However, BTs in their original form lack important properties for their application in HAPS mission management. Three new extensions of the framework are thus provided in this thesis in order to close these gaps: (a) Proper initialization and termination of BT tasks is enabled by introducing the new notion of transient and non-transient tasks. (b) A new continuous-time processing enables the method to be efficiently assessed during long-time simulations. (c) Highly modular stateful tasks introduce memory to a BT in a new and consistent way. The new extensions are conceptually developed and practically implemented in object-oriented Modelica libraries. Leveraging the modularity of BTs, the end user is relieved from the most intricate tasks of designing low-level implementations or high-level safety strategies. In addition, the goal-oriented design methodology supports the user in successfully assembling mission plans. The properties of this new approach are evaluated both theoretically and in practical mission simulations with a complex multi-disciplinary HAPS model. The developed extensions provide all necessary capabilities to apply the approach to HAPS mission management, while maintaining the original algorithm’s advantages. Typical mission simulations are accelerated by a factor of 500 compared to sampled implementations, thus enabling the user to test mission plans in extensive simulations before deployment. An upper bound of three event iterations is derived for the computational complexity of arbitrary status updates in the tree, indicating that the algorithm is real-time capable. This upper bound is confirmed in simulations. The modular structure of BTs proves highly suitable as a practical mission management method for HAPSs. The provided extensions close several severe gaps between the simplistic idea of BTs and a mission management methodology readily deployable in relevant HAPS applications. BTs in their extended form will be a powerful tool for future HAPS mission management as well as for similar applications.