Learning to See for Robotic Manipulation with Limited Labeled Training Data