Learning Gaussian Policies from Corrective Human Feedback