FDL: Mission Support Challenge

Luís F. Simões, Ben Day, Vinutha M. Shreenath, Callum Wilson, Christopher Bridges, Sylvester Kaczmarek, Yarin GalarXiv 2019 · DOI · arXiv

Abstract

The Frontier Development Lab (FDL) is a National Aeronautics and Space Administration (NASA) machine learning program with the stated aim of conducting artificial intelligence research for space exploration and all humankind with support in the European program from the European Space Agency (ESA). Interdisciplinary teams of researchers and data-scientists are brought together to tackle a range of challenging, real-world problems in the space-domain. The program primarily consists of a sprint phase during which teams tackle separate problems in the spirit of 'coopetition'. Teams are given a problem brief by real stakeholders and mentored by a range of experts. With access to exceptional computational resources, we were challenged to make a serious contribution within just eight weeks. Stated simply, our team was tasked with producing a system capable of scheduling downloads from satellites autonomously. Scheduling is a difficult problem in general, of course, complicated further in this scenario by ill-defined objectives & measures of success, the difficulty of communicating tacit knowledge and the standard challenges of real-world data. Taking a broader perspective, spacecraft scheduling is a problem that currently lacks an intelligent solution and, with the advent of mega-constellations, presents a serious operational bottleneck for the missions of tomorrow.

Introduction

Spacecraft scheduling is a high-stakes optimization problem governed by orbital mechanics and operational constraints. For the Cluster-II mission—a constellation of four satellites (Salsa, Samba, Rumba, and Tango) monitoring Earth’s magnetosphere—the primary objective is to download scientific data before it is overwritten in the satellites' "looped" storage systems.

Currently, this remains a manual, labor-intensive process. While automated constraint-solvers exist, they often produce "brittle" schedules that human operators find unusable because they lack the flexibility to handle real-world contingencies or align with human shift patterns. As we transition toward "mega-constellations" comprising hundreds of satellites, this manual approach becomes a critical operational bottleneck. This research explores building a system that is both autonomous and "intelligent" by mimicking the decision-making patterns of human operators.

Methodology

The team developed a multi-stage scheduling framework during an intensive eight-week sprint, focusing on translating human expertise into algorithmic constraints.

Objective Formalization and Simulation

The team translated qualitative operator goals into nine distinct scoring functions, normalized between 0 and 1. These metrics included pass cost-efficiency (minimizing ground station fees), link-budget alignment (ensuring high signal-to-noise ratios), and fragmentation (grouping passes to suit human work shifts). To support these scores, a simulation environment was constructed featuring a link-budget calculator to predict connection quality and a fill-level simulator to track data storage across the constellation.

Visualization of the scheduling environment, showing satellite fill levels, visibility windows, and ground station passes.

Human Decision Modeling

To capture "tacit knowledge," the researchers trained a Gradient Boosting ensemble of 50 shallow decision trees (max-depth 3). This model was designed to predict whether a human operator would book a specific "visibility" window based on the current mission context. This "human-likeness" score was then integrated into a Beam Search algorithm—a heuristic tree search that explores multiple potential schedules simultaneously, pruning less desirable paths based on both formalized scores and the predictive model.

Key Insights and Results

Predictive Performance

The human-modeling ensemble achieved an AUROC of 0.79 in a 10-fold cross-validation. This indicates a strong ability to distinguish between windows a human would book versus those they would ignore, suggesting that while operator knowledge is often described as intuitive, it follows consistent, discoverable patterns.

ROC curve for the human-decision modeling ensemble, achieving an AUROC of 0.79.

Operational Logic

A key finding was the identification of non-linear logic in operator behavior, such as a preference for exactly one-hour passes (driven by ground station billing structures) or passes that fully deplete satellite storage. This "kink" in the decision curve is difficult for traditional solvers to handle but was effectively captured by the machine learning model. Furthermore, the study suggests that "human-like" schedules are inherently more robust, as they often include implicit buffers and patterns that allow for mission recovery in the event of a single ground station pass failure.

Conclusion

The scheduling system was deemed "sufficiently promising" by the Cluster-II operations manager to warrant further development. While the current model is accurate, it remains somewhat agnostic to the broader temporal context, such as the quality of upcoming windows or inter-satellite dependencies. Future work will focus on a "closed-loop" evaluation where expert operators grade the schedules produced by the Beam Search, with the goal of integrating the human-likeness model more deeply into the search's pruning heuristics.

Summary updated

Abstract

The Frontier Development Lab (FDL) is a National Aeronautics and Space Administration (NASA) machine learning program with the stated aim of conducting artificial intelligence research for space exploration and all humankind with support in the European program from the European Space Agency (ESA). Interdisciplinary teams of researchers and data-scientists are brought together to tackle a range of challenging, real-world problems in the space-domain. The program primarily consists of a sprint phase during which teams tackle separate problems in the spirit of 'coopetition'. Teams are given a problem brief by real stakeholders and mentored by a range of experts. With access to exceptional computational resources, we were challenged to make a serious contribution within just eight weeks. Stated simply, our team was tasked with producing a system capable of scheduling downloads from satellites autonomously. Scheduling is a difficult problem in general, of course, complicated further in this scenario by ill-defined objectives & measures of success, the difficulty of communicating tacit knowledge and the standard challenges of real-world data. Taking a broader perspective, spacecraft scheduling is a problem that currently lacks an intelligent solution and, with the advent of mega-constellations, presents a serious operational bottleneck for the missions of tomorrow.
BibTeX
@misc{simoes2019fdl,
  author = {Luís F. Simões and Ben Day and Vinutha M. Shreenath and Callum Wilson and Christopher Bridges and Sylvester Kaczmarek and Yarin Gal},
  title = {{FDL: Mission Support Challenge}},
  year = {2019},
  eprint = {1912.09279},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/1912.09279},
}
Figure