The quaternion Bingham Distribution, 3D object detection, and dynamic manipulation

Abstract

Over the past few years, the field of robotic computer vision has undergone a 3-D revolution. One of the biggest challenges in dealing with 3-D geometry lies in appropriately handling 3-D rotational data. To specify "where" an object is in space, one must provide both a position and an orientation for the object. Noise and ambiguity in the robot's sensory data necessitate a robust model for representing uncertainty on the space of 3-D orientations. This is given by the quaternion Bingham distribution-a maximum entropy probability distribution on the 4-D unit quaternion hypersphere. In this thesis, we apply the quaternion Bingham to two applications: 3-D object instance detection from RGB-D images, and robot ping pong. The Bingham enables our object detection system to achieve state-of-the-art detection rates in highly cluttered scenes, while also enabling the ping pong robot to track the orientation and spin on flying ping pong balls. To enable the robot to actually play ping pong, we also explored a new method for incorporating human advice into a robot's motor control exploration policies.