Dual Action Policy for Robust Sim-to-Real Reinforcement Learning