Roderic Grupen - US grants

The department of Computer and Information Science at the University of Massachusetts at Amherst will continue to develop the infrastructure created with their prior CER award, with the emphasis on supporting their research real-time computing. This research will be concentrated in the following areas: real-time communication, real-time distributed operating systems, real-time artificial intelligence, cooperative/distributed problem solving, robotic manipulation and geometric reasoning computer vision and robot navigation, and special-purpose architectures. The groups involved in the various research areas will have a rich interaction, so that the software developed by the systems group will be used in the AI, robotics and vision application areas, and these areas will in turn give valuable feedback to the systems group. Equipment to be purchased in support of the research includes upgrades to current workstations, an upgrade for the Sequent Symmetry multiprocessor, as well as specialized equipment for robotics and vision.

This is the first of a three year continuing award. There search focuses on planning of multifingered robotic grasping of objects in situations where there is incomplete information. The grasping behavior will be continuously and incremetally refined based on sensed information during task execution and reasoningabout object and manipulator geometry. The grasping operation will combine reflexive behavior based on simple stimulus-action rules and local information only, with reactive behavior which provides robust capabilities for dealing with uncertainties and minimizes the amount of preplanning resourcefully without a priori models of the environment or the task. Such systems will be needed for autonomous robotic operation in unstructured, unfamiliar, or unpredictably changing situations such as space, underwater, or environmental operations, and may also be easier and more economical to deploy in more structured environments such as some manufacturing processes.

This research investigates the origins of knowledge and conceptual structure. An interactionist perspective leads to a computational model in which the earliest knowledge structures are influenced by both native structure in the agent and its exposure to the world. `Physical schemas` are learned that form stable, dynamical relationships between the agent and the world based on closed-loop `physical primitives.` A preimaging technique is used to form context-dependent categories spontaneously in order to improve the quality of sensory and motor behavior. Declarative models are lifted from the sensorimotor substrate in the form of `figurative schema.` This representation includes explicit predictive models of sensation and affect, and thus captures some of the underlying semantics of behavioral policies. An integrated robot system is used as the experimental platform. The implications of this work extend to developmental psychology, epistemology, and psycholinguistics, artificial intelligence and machine learning. Computational agents built on these principles refer ultimately to the physical world for meaning --- the same physical world that shapes human conceptual structure. We foresee interfaces to electronic agents that appeal to semantics based in the physical world to understand the `meaning` of a human request.

This SGER proposal concerns the accumulation and representation of skills and control knowledge by robots that interact with unstructured environments. There has been comparatively little work on representations that capture re-useable knowledge in robotics---an issue that lies at the heart of many future applications. Thus, this SGER represents a potentially transformative technology and addresses significant gaps in the state-of-the-art for which the payoff, despite the risk, is extremely high. We aim our 1 year study on learning techniques that accumulate knowledge related to grasping and manipulation. We shall extend pilot studies and build prototypes for self-motivated learning techniques and generative models for manipulation and multi-body contact relationships. The approach relies on learning to discover and exploit structure over the course of several staged learning episodes; from sensory and motor knowledge concerning the robot itself, to controllable relationships between the robot and external bodies, to multi-body contacts involved in tasks like stacking and insertion. The project has three principal technological goals: to advance the state-of-the-art of robotic manipulation and knowledge representation; to extend machine learning methods toward intrinsically motivated, cumulative, and hierarchical learning; and to advance computational accounts of the longitudinal processes of sensorimotor and cognitive development in humans and machines.