Reward Training Wheels: Adaptive Auxiliary Rewards for Robotics Reinforcement Learning