Interactive imitation learning for spacecraft path-planning in binary asteroid systems

Abstract

Exploration of small body systems poses the problem of designing path planning strategies for possibly uncharted environments. Traditional methods aimed at developing rigorous trajectory baselines may suffer inefficiencies, or turn infeasible when confronted with unknown dynamics. In strongly non-linear dynamics, mapping point design solutions from one dynamical regime to another may be hindered by underlying chaotic behavior. Rather than relying on baseline driven approaches, more generalized strategies may be found by observing human pilots controlling spacecraft motion within varying dynamical environments; the resultant data can then be utilized to initialize machine learning agents to provide more autonomous solutions. A previous numerical experiment resulted in a technical dataset comprising of human-based path planning strategies across a range of binary asteroid systems. This dataset is now used to train various imitation learning agents, and initiate the creation of a framework that integrates human–machine cooperation into the early training phases of artificial intelligent agents; the current application is for spacecraft guidance in binary asteroid systems, as a prototype of complex, potentially unknown, orbit dynamics. An interactive training architecture, based on the DAgger algorithm, is designed and employed to train both original and interactively coached agents, the latter stemming from both corrective and evaluative feedback by a real time human interactor. All agents were interactively trained for a predefined time period. The results from this investigation may provide the first, empirical observations of behavioral cloning within multi-body dynamics with largely randomized parameters, with some notable contributions including early characterization of training time, initial evidence of an autonomous agent learning meaningful policy features via imitation, and early identification of challenges in training fully autonomous agents for a multi-body dynamics path planning problem of this complexity and high dimensional state space.