Towards Physically Safe Reinforcement Learning under Supervision