Living Object Grasping Using Two-Stage Graph Reinforcement Learning