In this paper, we propose a computation-aware learning method for stable control with Gaussian process by accounting computational uncertainity. The task of the quadrotor is to carry a bottle of water while following a trajectory. The unknown disturbances, originating from complex slosh dynamics and the asymmetric centroid caused by the swaying water in the bottle, are learned online using GP.

For every set of 20 data points, we perform online learning for both the computation-aware and computation-agnostic GP models using different conjugate gradient (CG).  We use a zero mean prior and the product of a linear and a Matern kernel to construct the composite kernel of the GP model.