CAREER: Motion Planning and Active Vision Strategies for Optimizing Visual Feedback in Robot Control

This project investigates two complementary ways of improving computer vision based robot control particularly in uncertain and changing environments. First, by developing motion planning techniques that account for properties of the visual data -- exploiting the available visual feedback and overcoming sensing limitations. This requires extending the notion of robot configuration space to include image feature space and using topology-preserving learning schemes to handle modeling uncertainties. Second, by developing active vision strategies for optimizing visual feedback in robot control -- systematically extending the range of robot operation beyond that possible by static cameras. Defining invariances under camera motion would allow the decoupling of active camera control from robot control. Both deterministic and stochastic strategies will be explored for exploiting the camera motion in cooperation with the robot motion plan. This research will be interleaved with supporting educational activities which include developing new graduate courses in robotics and human-computer interfaces, revising undergraduate courses in related areas, and developing two new educational tools. The first tool will be for collaborative group projects on the web especially for projects involving video images. The second tool will be for training and visualization using an augmented reality interface that allows interactive mixing of virtual and real objects.