To appear in KR '96 Representing Sensing Actions: The Middle Ground Revisited

To build e ective planning systems, it is crucial to nd the right level of representation: too impoverished, and important actions and goals are impossible to express; too expressive, and planning becomes intractable. Within the classical framework, Pednault's adl [ 24 ] provided a happy compromise between the impoverished strips representation and the expensive situation calculus. Among languages handling sensing actions and information goals, there is a similar spectrum of expressiveness. uwl, an extension of strips, can't express goals like \Rename the le paper.tex to kr.tex." Nor can it represent universally quanti ed goals or e ects. At the other extreme are elegant languages [ 22, 21, 17 ] for which e ective planners do not exist. In this paper, we combine elements of uwl and adl, to de ne sadl: a middle-ground representation for sensing actions. Underlying our language are two insights, missing from uwl: 1) Knowledge goals are inherently temporal. 2) Knowledge preconditions are unnecessary for an important class of domains (those obeying a Markov property). sadl is expressive enough to encode the rich domain theory of the Internet Softbot, including hundreds of UNIX and Internet operators; yet it supports tractable inference by planners such as xii [ 11, 10 ] . Many thanks to Mark Boddy, Bob Doorenbos, Oren Etzioni, Marc Friedman, Robert Goldman, Neal Lesh, Greg Linden, Mike Perkowitz, Rich Segal, Jonathan Shakes and Ellen Spertus for helpful comments. This research was funded in part by O ce of Naval Research Grant N00014-94-1-0060, by National Science Foundation Grant IRI-9303461, by ARPA / Rome Labs grant F30602-95-10024, by a gift from Rockwell International Palo Alto Research, and by a Microsoft Graduate Fellowship