Randomization in robot tasks

Abstract

It is argued that randomization is a useful primitive for the solution of robot tasks under uncertainty. The author demonstrates a possible application using a standard peg-in-hole problem, focusing on randomization within simple feedback strategies. More generally, randomization may be thought of as an operator that randomly selects between possible knowledge states of actions. In order to synthesize randomized strategies, it is possible to use this operator within the dynamic programming methodology essentially as one would any other operator. It is further asserted that randomization has three important properties: (1) increases the class of solvable tasks; (2) reduces plan brittleness; and (3) simplifies the planning process.