Cherry-Picking with Reinforcement Learning : Robust Dynamic Grasping in Unstable Conditions