ESA GNC Conference Papers Repository
Enabling AI-in-the-loop AOCS algorithms on in-flight hardware: from conception to in-orbit demonstration in ESA OPS-SAT
There is a current trend and opportunity in technology to apply Artificial Intelligence (AI) and Machine Learning (ML) to improve systems performance. In the space sector, convincing demonstrators of AI/ML technologies applied to space use cases have been developed in simulation environments throughout the industry, showing a great potential for AOCS and GNC applications. However, their adoption on spacecraft flight software has yet to happen in a significant manner. This presents a challenging step: from theory and simulator environment to hardware in the space environment. Among some of the opportunities that exist for in-orbit demonstrations, the ESA OPS-SAT CubeSat enables quick prototyping of promising algorithms and in-orbit testing. OPS-SAT is a 3U CubeSat operated by ESA, that gives the opportunity to run experiments on their platform. The satellites on board computer, Satellite Experimental Processing Platform (SEPP), is the payload on which experiments are run. The SEPP is based on Linux, which makes it easily programmable and it also incorporates a high-bandwidth system for embedded applications. It not only offers computing resources rarely seen on a spacecraft, but it is also able to run more conventional software. Past experiments have demonstrated that the platform does run AI code properly, which makes it an ideal candidate for testing Airbus GNC teams HOPAS algorithm. The Hybrid Online Policy Adaptation Strategy (HOPAS) AI algorithm was created driven by the desire to keep improving the AOCS performance in some existing missions. HOPAS uses a family of AI techniques called Reinforcement Learning (RL) in an online fashion in order to continuously improve the spacecraft attitude control pointing performance. This algorithm is a striking example of an AI algorithm for space systems that improves the pointing error of the nominal control system and can adapt and continuously learn in the context of new space environments or disturbances. After showing very convincing performance on a representative simulator of the Solar Orbiter spacecraft, the algorithm was proposed for ESA OPS-SAT. Implementing such a control algorithm on an experimental platform like OPS-SAT is especially challenging. This is due to a number of factors, including the relatively low computing power available on-board traditional spacecraft and the long development process required in the industry. The main challenges to resolve were: time cycle management, as the algorithm needs to be executed continuously over time at a steady frequency whereas retrieval of attitude telemetry and sending command executes at varying speed; memory budget, as AI applications generally require a larger computing footprint; and adapting the software and the HOPAS algorithm to manage online training on the OPS-SAT specific spacecraft. This paper presents the development of a testing infrastructure for AI algorithms in a spacecraft with a particular use case in the work developed over 2022 to adapt and port the algorithm to OPS-SAT, as well as the results of the simulator and experiments performed.