An Active Vision System for Detecting, Fixating and Manipulating Objects in the Real World
This report discusses the construction of adaptive control structures for teams of semi-autonomous robots. A exible architecture for adaptive control was developed that forms control policies for allocating resources in open environments and tasks. This architecture is designed to identify the control context and to specify coordinated robot behavior. To identify the control context we form categories of distinct dynamic responses with which to augment state information | a kind of hidden state removal. To evaluate the feasibility of the approach, we have applied it to grasping and manipulation tasks with multiingered robot hands. In this class of tasks, the dynamic plant that we wish to control depends directly on the geometry of the object. We have constructed a working platform that uses the dynamic response of the controller to classify types of plants and uses this information to improve the grasp/manipulation strategy. Since resource alternatives in this platform involve various combinations of ngers, the control context is used to learn resource schedules that coordinate the allocation and de-allocation of ngers.