Heteroscedastic Uncertainty Decomposition for Learned Gyro Bias Correction in Spacecraft Attitude Estimation

arXiv 2607.24608·low signalXBlueskyLinkedInCopy link

This work (arXiv 2607.24608, July 27) trains a 1-D convolutional network to predict residual angular-rate corrections from multi-sensor inputs including gyroscope and star-tracker measurements, feeding the corrections into a flight-representative Gyro-Stellar Estimator. The network outputs both mean corrections and input-dependent heteroscedastic aleatoric uncertainty, with epistemic uncertainty handled separately, and the paper investigates how attribution and uncertainty behave together. It is a narrow aerospace application, but the pattern — a learned residual correction feeding a classical estimator, with explicit aleatoric/epistemic separation so the downstream filter knows how much to trust it — is the standard template for safely inserting ML into an existing deterministic control loop.