Transparent Reactive Plans

Abstract

A robot that forestalls flaws in its intended course of action while carrying out its jobs must perform difficult and complex computations at a speed that is fast compared to the time that it takes to execute plans. The computational processes must infer the purpose of subplans, find subplans with a particular purpose, automatically generate a plan that can achieve some goal, determine flaws in the behavior that is caused by subplans, and estimate how good the behavior caused by a subplan is with respect to the robot’s utility model. The computational problems these processes try to solve are, in their general form, unsolvable, or at the very least computationally intractable.