Robot Grasping Of Novel Objects In Dense Clutter
A key challenge in robot grasping is detecting where in a scene to move the hand in order to grasp novel objects. Typically, this is handled using sliding window methods similar to those used in computer vision for object detection tasks. However, this approach ignores significant geometric structure in the problem. In this talk, I will introduce a new method of encoding grasp detection as a machine learning problem that leverages knowledge about the geometry of grasping. Our key insight is to recognize that if we were given a "perfect" point cloud that contained points on all exposed object surfaces, then it would be possible to detect grasps using geometric reasoning alone. It is only because real-world point clouds contain significant occlusions that it is necessary to use machine learning at all. This perspective enables us to use geometry to prune the hypothesis space and to focus machine learning on predicting the presence of occluded grasp geometries. This method performs well in practice. We get high accuracies on test sets and obtain grasp success rates for novel objects near 90% in real robot experiments. Of particular interest is the fact that we perform particularly well in dense clutter. Our grasping system is available as a ROS package and is among the most effective systems for novel object grasping available.
Robert Platt is an Assistant Professor of Computer Science at Northeastern University. Prior to coming to Northeastern, he was a Research Scientist at MIT and a technical lead at NASA Johnson Space Center, where he helped lead the development of the control and autonomy subsystems of Robonaut 2, the first humanoid robot in space. He is an inventor on more than 18 US patents or patent applications. He earned his PhD in Computer Science from the University of Massachusetts, Amherst. Professor Platt is interested in developing robots that can function robustly alongside people in the uncertain everyday world. He is particularly interested in the planning, control, and perception algorithms that could enable a robot to assist people in the context of manipulation or assembly work. His work has applications in a variety of areas including human-robot interaction, health care, assisted living, personal robotics, manufacturing and package handling, and space robotics.