Machine Learning for Flexible Robotics
Final rept. 1 Oct 1986-31 Dec 1990
ILLINOIS UNIV AT URBANA-CHAMAPIGN
Pagination or Media Count:
Robotic planning, if it is to be successful in real-world situations, must find some way to side-step the now-well-documented obstacles to classical AI planning. These recent results show that the computational complexity of standard planning is unacceptable even with drastic and untenable simplifying assumptions about the world. The source of complexity in real-world robotic domains includes the problems of data uncertainty, large amounts of data to consider, as well as the problem of tractably producing plans according to the given domain rules. Pretending that these complexities do not exist relegates a computer system to a trivialized micro-world with little hope of applications to the real world. The research of this grant has been directed towards dealing with the real-world constraints that artificial intelligence robotics systems must address. We have made significant progress on two fronts. The first investigates an integrated approach to planning wherein a classical a priori planner is augmented with reactive abilities. The second thrust of this grant explores a new approach called permissive planning. We have implemented our ideas in the GRASPER system which has capabilities to monitor execution of its plans and to tune its model of the world on failure through use of explicit approximations.