In the application of machine learning to robotics, the quality of data assumes a pivotal role. Many methods use exploration algorithms to select the more informative data points for the model. Nevertheless, these approaches overlook the detrimental influence of measurement errors that invariably impact the data. This paper proposes a novel method to improve the quality of datasets employed in model learning by optimizing metrics based on the combination of exploration and active sensing measures. We use metrics based on a Gaussian Process covariance matrix as exploration metrics, with the aim of letting the system to explore the state space regions where the model uncertainty is higher. We combine it with an active sensing metric based on a norm of the Continuous Riccati Equation optimized to reduce the negative effect of measurement noise on the data. To prove the validity and versatility of our approach, we tested it in a simulation scenario on a unicycle and a quadruped robot.