Session
Advanced Technologies 2
Location
Salt Palace Convention Center, Salt Lake City, UT
Abstract
UK space industry partners Craft Prospect Ltd (CPL) and GMV deliver an advanced avionics demonstrator which incorporates cutting-edge machine learning (ML) techniques to enhance on-board autonomy for small satellite space missions. The work is funded by the European Space Agency (ESA), and the delivery partners combine expertise in embedded Artificial Intelligence (AI) and space systems engineering to target use cases for on-board autonomy which offer the best gains for a diverse range of mission types including institutional and commercial Earth observation (EO), navigation, scientific, and lunar exploration missions.
Common technical target applications have emerged as underpinning technologies which enable the adoption of on-board autonomous operations, namely: Responsive Mission Planning and Scheduling, and Fault Detection, Isolation and Recovery (FDIR).
Planning and scheduling are core capabilities to enable on-board autonomous operations. Tasks historically performed by ground-based operators are now replaced in large part by on-board activities which can perform short-term task planning and execution for targets of opportunity and tip and cue activities, longer-term goal based planning, and dynamic system reconfiguration.
In every mission scenario, a large emphasis is placed on the detection and mitigation of faults by the FDIR function, in conjunction with monitoring of the spacecraft telemetry. Modern missions fuse a combination of on-board and on-ground FDIR functions to meet the needs of the mission in terms of reactiveness and computational loads.
Two on-board processing enablers are developed to carry out autonomous mission planning and FDIR, designed to meet the requirements of a real upcoming mission scenario: ESA OPS-SAT VOLT (Versatile Optical Laboratory for Telecommunications) led by CPL. This mission requires the handling of multiple payloads and services, including activities in optical communications, quantum key distribution and hyperspectral imaging. Solutions for the VOLT mission must also meet requirements on explainability, trustworthiness and assurance of the decision-making, planning and scheduling pipeline. This demands a multi-modal approach which involves: consideration of mission-level assurance requirements on low-level AI components, a clear specification of test cases for full coverage with respect to these requirements, utilisation of relevant explainable AI (XAI) techniques and deployment of a real-time and verifiable autonomy supervisor on-board to act as a check on the AI planning solution.
This paper describes the development of these on-board processing enablers and the cutting-edge AI technology that underpins them: reinforcement learning for responsive planning, and transformer models for anomaly detection.
Document Type
Event
AI for Autonomous Operation of Small Satellite Missions
Salt Palace Convention Center, Salt Lake City, UT
UK space industry partners Craft Prospect Ltd (CPL) and GMV deliver an advanced avionics demonstrator which incorporates cutting-edge machine learning (ML) techniques to enhance on-board autonomy for small satellite space missions. The work is funded by the European Space Agency (ESA), and the delivery partners combine expertise in embedded Artificial Intelligence (AI) and space systems engineering to target use cases for on-board autonomy which offer the best gains for a diverse range of mission types including institutional and commercial Earth observation (EO), navigation, scientific, and lunar exploration missions.
Common technical target applications have emerged as underpinning technologies which enable the adoption of on-board autonomous operations, namely: Responsive Mission Planning and Scheduling, and Fault Detection, Isolation and Recovery (FDIR).
Planning and scheduling are core capabilities to enable on-board autonomous operations. Tasks historically performed by ground-based operators are now replaced in large part by on-board activities which can perform short-term task planning and execution for targets of opportunity and tip and cue activities, longer-term goal based planning, and dynamic system reconfiguration.
In every mission scenario, a large emphasis is placed on the detection and mitigation of faults by the FDIR function, in conjunction with monitoring of the spacecraft telemetry. Modern missions fuse a combination of on-board and on-ground FDIR functions to meet the needs of the mission in terms of reactiveness and computational loads.
Two on-board processing enablers are developed to carry out autonomous mission planning and FDIR, designed to meet the requirements of a real upcoming mission scenario: ESA OPS-SAT VOLT (Versatile Optical Laboratory for Telecommunications) led by CPL. This mission requires the handling of multiple payloads and services, including activities in optical communications, quantum key distribution and hyperspectral imaging. Solutions for the VOLT mission must also meet requirements on explainability, trustworthiness and assurance of the decision-making, planning and scheduling pipeline. This demands a multi-modal approach which involves: consideration of mission-level assurance requirements on low-level AI components, a clear specification of test cases for full coverage with respect to these requirements, utilisation of relevant explainable AI (XAI) techniques and deployment of a real-time and verifiable autonomy supervisor on-board to act as a check on the AI planning solution.
This paper describes the development of these on-board processing enablers and the cutting-edge AI technology that underpins them: reinforcement learning for responsive planning, and transformer models for anomaly detection.
