Learning To Grasp

Providing robots with the ability to grasp objects has, despite decades of research, remained a challenging problem. The problem is approachable in constrained environments where there is ample prior knowledge of the scene and objects that will be manipulated. The challenge is in building systems that scale beyond specific situational instances and gracefully operate in novel conditions. In the past, heuristic and simple rule based strategies were used to accomplish tasks such as scene segmentation or reasoning about occlusion. These heuristic strategies work in constrained environments where a roboticist can make simplifying assumptions about everything from the geometries of the objects to be interacted with, level of clutter, camera position, lighting, and a myriad of other relevant variables. With these assumptions in place, it becomes tractable for a roboticist to hardcode desired behaviour and build a robotic system capable of completing repetitive tasks. These hardcoded behaviours will quickly fail if the assumptions about the environment are invalidated. In this thesis we will demonstrate how a robust grasping system can be built that is capable of operating under a more variable set of conditions without requiring significant engineering of behavior by a roboticist. This robustness is enabled by a new found ability to empower novel machine learning techniques with massive amounts of synthetic training data. The ability of simulators to create realistic sensory data enables the generation of massive corpora of labeled training data for various grasping related tasks. The use of simulation allows for the creation of a wide variety of environments and experiences exposing the robotic system to a large number of scenarios before ever operating in the real world. This thesis demonstrates that it is now possible to build systems that work in the real world trained using deep learning on synthetic data. The sheer volume of data that can be produced via simulation enables the use of powerful deep learning techniques whose performance scales with the amount of data available. This thesis will explore how deep learning and other techniques can be used to encode these massive datasets for efficient runtime use. The ability to train and test on synthetic data allows for quick iterative development of new perception, planning and grasp execution algorithms that work in a large number of environments. Creative applications of machine learning and massive synthetic datasets are allowing robotic systems to learn skills, and move beyond repetitive hardcoded tasks.