Machine Learning and Friends Lunch
Reaching and Grasping Skills
Andrew Fagg
UMass
Abstract
Humanoid-form robots show great promise in performing a wide range of tasks in unstructured, real-world environments. However, to operate in a truly flexible manner, they require tremendous complexity in their design. Computational challenges of managing these complexities include providing a layer of abstraction that 1) enables a programmer to work at a task level, and 2) allows machine learning algorithms to be used in the automatic acquisition of control policies. In this talk, I will describe my work at UMass to develop a set of high-level robotic controllers for reaching, grasping and manipulation. These closed-loop controllers are responsible for satisfying subgoals such as tactile-driven establishment of high-quality grasps, moving the robot's hand into alignment with a target object, and transporting an object held by the hand. These high-level controllers become "primitives" that can be assembled together by a programmer or by a learning algorithm in order to perform a complete task. I will illustrate this approach with a grip selection learning problem that parallels a related learning process in children. In addition, I will show how the controllers may be used in the recognition and explanation of actions taken by a user serving as a remote teleoperator of the robot.