Evaluating Model-free Reinforcement Learning toward Safety-critical Tasks