End-to-end deep reinforcement learning and control with multimodal perception for planetary robotic dual peg-in-hole assembly