Session
Poster Session 1
Location
Salt Palace Convention Center, Salt Lake City, UT
Abstract
Autonomously determining fuel-optimal, formation-keeping trajectories for small spacecraft operating in formation flight is often limited by available processing power. Because SmallSat missions are constrained by size, weight, and power (SWaP), field-programmable gate arrays (FPGAs) are an attractive option for onboard computing. Traditional formation-keeping relies on convex optimization to solve for fuel-optimal burns and burn times; however, implementing these solvers on FPGAs is challenging due to limited board resources.
To address the need for fast, reliable trajectory planning onboard space vehicles, we investigate leveraging machine learning to reduce computational cost for real-time, onboard execution. We propose using a deep neural network (DNN) trained offline to predict burn times, replacing heavy online optimization. To ensure robustness, the model is coupled with a post-processing algorithm that solves a linear least-squares problem to determine the exact velocity impulses required to return satellites to their nominal state.
This hybrid approach is evaluated for onboard FPGA execution on the ITASAT-2 mission, consisting of three Low Earth Orbit (LEO) satellites operating in both "string-of pearls" and "cross-meridian" formations². The performance of the post-processing algorithms is benchmarked against traditional convex solvers with respect to computational cost, accuracy, and fuel cost.
Document Type
Event
Paper
Included in
Promoting Assured Formation Flight Autonomy Onboard Resource Constrained Systems
Salt Palace Convention Center, Salt Lake City, UT
Autonomously determining fuel-optimal, formation-keeping trajectories for small spacecraft operating in formation flight is often limited by available processing power. Because SmallSat missions are constrained by size, weight, and power (SWaP), field-programmable gate arrays (FPGAs) are an attractive option for onboard computing. Traditional formation-keeping relies on convex optimization to solve for fuel-optimal burns and burn times; however, implementing these solvers on FPGAs is challenging due to limited board resources.
To address the need for fast, reliable trajectory planning onboard space vehicles, we investigate leveraging machine learning to reduce computational cost for real-time, onboard execution. We propose using a deep neural network (DNN) trained offline to predict burn times, replacing heavy online optimization. To ensure robustness, the model is coupled with a post-processing algorithm that solves a linear least-squares problem to determine the exact velocity impulses required to return satellites to their nominal state.
This hybrid approach is evaluated for onboard FPGA execution on the ITASAT-2 mission, consisting of three Low Earth Orbit (LEO) satellites operating in both "string-of pearls" and "cross-meridian" formations². The performance of the post-processing algorithms is benchmarked against traditional convex solvers with respect to computational cost, accuracy, and fuel cost.
