Closed
Current attitude control architectures are limited by design phase knowledge of the exact dynamics and expected perturbation environment of the spacecraft. As such, most control designs use generous robustness margins to ensure safety in all realistic conditions. This implies a robustness vs. performance trade-off, imposing some limitations to the maximum achievable pointing performance. The static nature of the most commonly used control laws can also lead to performance loss in the case of unforeseen scenarios or time-varying effects. Recent AOCS/GNC research at Airbus has shown that an effective use of data can be key in lifting those limitations, for instance by leveraging AI techniques. Here, we propose our Hybrid Online Policy Adaptation Strategy (HOPAS) to continuously improve pointing performance. This is achieved using reinforcement learning with recurrent neural networks. HOPAS has been initially prototyped by the team using the Solar Orbiter spacecraft simulator as a use-case. In this use-case, performance (Relative Performance Error) is improved by a factor of 10 and the system displays strong robustness to dynamic uncertainty, delays, and time-varying effects. In addition to the previous results, HOPAS holds further benefits: -Low computational footprint solutions thanks to evolution strategies and relatively small networks (a few thousand parameters). -Safety is a core part of the design. A robust controller is part of the solution (hence “Hybrid”), thus the AI solution can easily be turned off if undesired/unsafe behaviours are detected. The goal of this experiment is to quantify performance gains on real flight data while durably improving OPS-SAT pointing performance. This would raise the TRL level of HOPAS to pave the way for novel missions, including in-orbit servicing and high accuracy observation missions.
