| Grasping and manipulation of objects provide important stimulation toward maturing mental representations in human beings during their development. Both are open-ended activities, which can only be fully characterized in situational terms involving the purpose of the activities, the resources deployed, and the role of each effector. This dissertation describes how representations of grasp activity or grasp state may be constructed by an agent as it interacts with a collection of objects. The state representation proposed is based on a set of empirically-derived generative models of the system dynamics. These models are constructed autonomously by the grasping agent as it observes dynamic patterns in the phase portrait induced by a set of feedback grasp controllers. The resulting state representation serves as the basis on which competing grasp strategies are constructed, evaluated, and stored. The results obtained indicate that it is possible to learn grasp policies that maximizes a user-supplied metric for individual objects even if object geometry, pose and identity are not explicit components of the grasp state vector. Contrary to many other approaches described in the literature, the resulting grasp strategy deals with noisy sensors and actuators, multiple objects, and has been implemented in a real hand-arm system. The techniques contributed in this dissertation can also be applied to the development of general sensorimotor control strategies for agents operating in open environments. They also provide insight into the automated construction of situated representations for sensorimotor activities, the scheduling of motor resources, the basis of perceptual categorization, and the uniform integration of continuous and discrete control primitives, control knowledge, control recovery/recognition into context-dependent strategies. |
