Session
Weekend Session V: Automation - Research and Academia
Location
Utah State University, Logan, UT
Abstract
The rapid development and launch of low-cost satellites have led to constellations of hundreds of satellites, the proliferated Low Earth Orbit (p-LEO) constellation, and has changed the dynamic of space-based earth observation. The availability of massive numbers of satellites creates the opportunity to tackle larger scale problems, such as larger scale monitoring of wildlife, illegal fishing, and climate events. These opportunities must be met with increasing automation, preferably using an approach that executes and adapts on-orbit.
Optimizing distributed resources remains an astronomically challenging problem, even more so when computations are performed on the comparatively compute restricted hardware of satellites. We propose a data-efficient two-part system, modeled after the Actor-Critic architecture, that models the dynamics of p-LEO constellations and other data streams and generates optimized tasking. First, using the Koopman operator theory approach, we model an aggregate representation of a heterogenous (multi-modal sensing) satellite constellation to predict satellite availability, observation capabilities, and resource utilization. An aggregate representation of the constellation enables scalability to model hundreds to thousands of satellites, as well as being agnostic to particular identities (satellites may enter and leave the constellation). Second, we use the Hierarchical Bayesian Program Learning (HBPL) paradigm to formulate and learn a task decomposition and generation ‘program’. Tasks are constructed and defined probabilistically while guided by expert informed structure and bounds, enabling efficient search and optimization of the task space. During execution, the HBPL component proposes sets of tasks (serving as the ‘actor’) which are scored by the Koopman model. The two methods described above are notable due to their low data requirements and speed of model training or updating, making them a stellar pairing for on-orbit applications.
This data-driven learning approach to task generation was explored to solve the task decomposition problem for the BAE Systems Collective Space Tasking and Assimilation Reasoning System (CoSTARS) under the DARPA Oversight program. The full architecture includes a distributed auction mechanism for task assignment, a data assimilation component, and updates on entities or objectives. The task decomposition approach is evaluated under this system architecture for the case of monitoring a set number of entities of interest.
A Data-Efficient Model-Based Task Decomposition Approach for Massive Satellite Constellations
Utah State University, Logan, UT
The rapid development and launch of low-cost satellites have led to constellations of hundreds of satellites, the proliferated Low Earth Orbit (p-LEO) constellation, and has changed the dynamic of space-based earth observation. The availability of massive numbers of satellites creates the opportunity to tackle larger scale problems, such as larger scale monitoring of wildlife, illegal fishing, and climate events. These opportunities must be met with increasing automation, preferably using an approach that executes and adapts on-orbit.
Optimizing distributed resources remains an astronomically challenging problem, even more so when computations are performed on the comparatively compute restricted hardware of satellites. We propose a data-efficient two-part system, modeled after the Actor-Critic architecture, that models the dynamics of p-LEO constellations and other data streams and generates optimized tasking. First, using the Koopman operator theory approach, we model an aggregate representation of a heterogenous (multi-modal sensing) satellite constellation to predict satellite availability, observation capabilities, and resource utilization. An aggregate representation of the constellation enables scalability to model hundreds to thousands of satellites, as well as being agnostic to particular identities (satellites may enter and leave the constellation). Second, we use the Hierarchical Bayesian Program Learning (HBPL) paradigm to formulate and learn a task decomposition and generation ‘program’. Tasks are constructed and defined probabilistically while guided by expert informed structure and bounds, enabling efficient search and optimization of the task space. During execution, the HBPL component proposes sets of tasks (serving as the ‘actor’) which are scored by the Koopman model. The two methods described above are notable due to their low data requirements and speed of model training or updating, making them a stellar pairing for on-orbit applications.
This data-driven learning approach to task generation was explored to solve the task decomposition problem for the BAE Systems Collective Space Tasking and Assimilation Reasoning System (CoSTARS) under the DARPA Oversight program. The full architecture includes a distributed auction mechanism for task assignment, a data assimilation component, and updates on entities or objectives. The task decomposition approach is evaluated under this system architecture for the case of monitoring a set number of entities of interest.
