Transferring Experience from Simulation to the Real World for Precise Pick-And-Place Tasks in Highly Cluttered Scenes

In this paper, we introduce a novel learning-based approach for grasping\nknown rigid objects in highly cluttered scenes and precisely placing them based\non depth images. Our Placement Quality Network (PQ-Net) estimates the object\npose and the quality for each automatically generated grasp pose for multiple\nobjects simultaneously at 92 fps in a single forward pass of a neural network.\nAll grasping and placement trials are executed in a physics simulation and the\ngained experience is transferred to the real world using domain randomization.\nWe demonstrate that our policy successfully transfers to the real world. PQ-Net\noutperforms other model-free approaches in terms of grasping success rate and\nautomatically scales to new objects of arbitrary symmetry without any human\nintervention.\n