Modern industrial processes (nuclear, chemical industry), public service needs (firefighting, rescuing), and research interests (undersea, outer space exploration) have established a clear need to perform work remotely. Whereas a purely autonomous manipulative capability would solve the problem, its realization is beyond the state of the art in robotics [Stark et al.,1988]. Some of the problems plaguing the development of autonomous systems are: a) anticipation, detection, and correction of the multitude of possible error conditions arising during task execution, b) development of general strategy planning techniques transcending any particular limited task domain, c) providing the robot system with real-time adaptive behavior to accommodate changes in the remote environment, d) allowing for on-line learning and performance improvement through "experience", etc. The classical approach to tackle some of these problems has been to introduce problem solvers and expert systems as part of the remote robot workcell control system. However, such systems tend to be limited in scope (to remain intellectually and implementationally manageable), too slow to be useful in real-time robot task execution, and generally fail to adequately represent and model the complexities of the real world environment. These problems become particularly severe when only partial information about the remote environment is available.