Robotic grasping in unstructured and dynamic environments

Grasping and transporting objects is a fundamental trait that underpins many robotics applications, but existing works in this area are not robust to real-world challenges such as moving objects, human interaction, clutter and occlusion. In this thesis, we combine state-of-the-art computer vision techniques with real-time robotic control to overcome these limitations. We present a number of algorithms that can compute grasps for new items in a fraction of a second, react to dynamic changes in the environment, and intelligently choose improved viewpoints of occluded objects in clutter.