Edge Computing in Space: Design Optimization and Reinforcement Learning Scheduling of Onboard Computing Satellites
Name
ravishankar_rashmir_phd_aeroastro_2026_thesis.pdf
Size
7.38 MB
Format
Adobe PDF
Checksum (MD5)
740940ae1a4ffece93e8fb75daf6d988
Author(s)
Ravishankar, Rashmi
Advisor(s)
de Weck, Olivier
Date Issued
February 2026
Publisher
Massachusetts Institute of Technology
Abstract
As satellite missions heighten in complexity, there is a need for spacecraft equipped with greater intelligence and autonomy. This calls for onboard, real-time decision-making abilities on weight, power, and radiation-constrained hardware and processors, one of the most challenging of all "edge computing" problems. Although an existing body of literature has studied satellite design and scheduling optimization, few have considered an environment where space systems make decisions of their own, and none have combined the challenges of edge computing with the multidisciplinary optimization problem of architecting and scheduling a computationally heavy satellite system. This thesis evaluates the potential of edge computing and prescribes an optimal strategy for computer operations. First, a hardware survey is conducted and certain selected terrestrial and space-based processors are benchmarked on speed and thermal performance using fundamental computing algorithms. An edge computing satellite model, "EdgeSat", is defined, and its competing subsystems are optimized for a satellite mission where onboard computing and autonomy are needed. Features of the model include data and computing models of space-qualified hardware, thermal models of computing in space, and a dynamic data management protocol. Next, the value of edge computing is defined and quantified, a pareto optimization of design points is conducted and a utopia point is identified. Finally, reinforcement learning is used to prescribe optimal mission profiles in various scenarios using appropriate rewards and visibility windows. The rewards consider the utility of downlinked data, battery preservation, thermal constraints, against eclipse conditions and downlink windows. Edge computing and design optimization were found to add successive value to the base case, enhancing downlinked data utility by 88% and 115% over the baseline respectively, and RL scheduling of the onboard computer adds further value amounting to between 8% and 23.6% over simple control. Ultimately, the satellite model enhanced with edge computing and a combination of pareto design optimization and reinforcement learning control is shown to bring about a 166% (2.7x) improvement to the original counterpart.
MIT Department
Massachusetts Institute of Technology. Department of Aeronautics and Astronautics
Massachusetts Institute of Technology. Center for Computational Science and Engineering
Terms of Use
In Copyright - Educational Use Permitted
Copyright retained by author(s)
Persistent DSpace Link
