Towards Learning Autonomous Spacecraft Path-Planning from Demonstrations and Environment Interactions

Abstract

An increasing interest in the robotic exploration of uncharted space environments is warranting a spacecraft path-planning approach that is capable of producing not only autonomous, but also generalized solutions. Though current methods allow for some degree of autonomy in specific phases of the mission timeline, such as proximity operations and "touch-and-go" maneuvers, they are not currently extended. Traditional spacecraft path-planning solutions rely on rigorous baseline strategy design, which inherently requires a priori knowledge of the state dynamics, which can be quantified via relevant dynamical parameters. This assumption may break down in the context of unknown environmental dynamics, as is the case with the exploration of uncharted environments. Alternative to approaches that require a baseline, path-planning methods driven by machine learning methods may offer solutions that are more generalized and autonomous. However, the extent to which the performance of a machine learning model compares to an optimal solution has not been extensively evaluated, and initial implementation provided in this work aims to provide further insight into this domain. Path-planning methods based on optimal control theory, pure machine learning methods, as well as the novel field of cooperative human-AI interaction are explored in order to not only analyze the advantages and inherent drawbacks of each distinct approach, but also how each method may be leveraged to construct a synergistic framework that is capable of producing generalized and autonomous solutions. Additionally, extensive exploratory work regarding the implementation of cooperative human-AI interaction as applicable to spacecraft path-planning may serve as the first, empirical observations of behavioral cloning within higher fidelity dynamics, as representative of more realistic scenarios, as well as provide early identification of challenges in training fully autonomous agents for a multi-body dynamics path planning problem.