OPS-SAT VOLT: ESA finances on-board autopilot combining reinforcement and transformer learning

Presented at the Small Satellite Conference on 22 August, a British demonstrator entrusts on board the planning of observations and the detection of anomalies, under the supervision of an autonomy supervisor.

Two British space players, Craft Prospect and GMV NSL, exhibited at the Small Satellite conference in Salt Lake City an avionics demonstrator for the European Space Agency's OPS-SAT VOLT mission, which will carry several payloads, including optical communications and a hyperspectral imager.

The project, funded by ESA, targets two functions that small satellites currently largely delegate to the ground. The first is reactive planning: deciding on board, during orbit, which observations to keep and in what order, a task entrusted to reinforcement learning. The second is fault detection, processed by transformative architecture models applied to telemetry.

An AI whose choices we want to be able to verify

The particularity of the approach lies less in the models than in what surrounds them. The teams insist on the explainability and verifiability of decisions, and provide for an on-board autonomy supervisor responsible for validating the plans produced before they are executed. This is the condition set for an operator to agree to let a satellite arbitrate alone, whereas an aberrant decision taken out of ground visibility can only be made up for on the next pass.