JournalofArtificial Intelligence Research 20 (2003) 379-404 Submitted 10/02; published 12/03
SHOP2: An HTN Planning System
Dana Nau nau@cs.umd.edu
Dept. of Computer Science, and Institute for Systems Research
University of Maryland, College Park, MD 20742 USA
Tsz-Chiu Au chiu@cs.umd.edu
Dept. of Computer Science
University of Maryland, College Park, MD 20742 USA
Okhtay Ilghami okhtay@cs.umd.edu
Dept. of Computer Science
University of Maryland, College Park, MD 20742 USA
Ugur Kuter ukuter@cs.umd.edu
Dept. of Computer Science
University of Maryland, College Park, MD 20742 USA
J. William Murdock murdockj@us.ibm.com
IBM Watson Research Center
19 Skyline Dr.
Hawthorne, NY 10532 USA
Dan Wu dandan@cs.umd.edu
Dept. of Computer Science
University of Maryland, College Park, MD 20742 USA
Fusun Yaman fusun@cs.umd.edu
Dept. of Computer Science
University of Maryland, College Park, MD 20742 USA
Abstract
The SHOP2 planning system received one of theawards for distinguished performance
in the 2002 International Planning Competition. This paper describes the features of
SHOP2 which enabled it to excel in the competition, especially those aspects of SHOP2
that deal with temporal and metric planning domains.
1. Introduction
SHOP2, Simple Hierarchical Ordered Planner 2 (Nau, Mun˜oz-Avila, Cao, Lotem, &Mitchell,
2001), is a domain-independent planning system based on Hierarchical Task Network (HTN)
planning. In the 2002 International Planning Competition, SHOP2 received one of the top
four awards, one of the two awards for distinguished performance. This paper describes
some of the characteristics of SHOP2 that enabled it to excel in the competition.
Like its predecessor SHOP (Nau, Cao, & Mun˜oz-Avila, 1999), SHOP2 generates the
steps of each plan in the same order that those steps will later be executed, so it knows
the current state at each stepoftheplanning process. This reduces the complexity of
reasoning by eliminating a great deal of uncertainty about the world, thereby making it
easy to incorporate substantial expressive power into the planning system. Like SHOP,
c©2003 AI Access Foundation. All rights reserved.

Nau, Au, Ilghami, Kuter, Murdock, Wu, & Yaman
SHOP2 can do axiomatic inference, mixed symbolic/numeric computations, and calls to
external programs.
SHOP2also has capabilities that go significantly beyond those of SHOP:
• SHOP2 allows tasks and subtasks to be partially ordered; thus plans may interleave
subtasks from different tasks. This often makes it possible to specify domain knowl-
edge in a more intuitive manner than was possible in SHOP.
• SHOP2 incorporates many features from PDDL, such as quantifiers and conditional
effects.
• If there are alternative ways to satisfy a method’sprecondition, SHOP2 can sort the
alternatives according to a criterion specified in the definition of the method. This
gives a convenient way for theauthor of a planning domain to tell SHOP2 which parts
of the search space to explore first. In principle, such a technique could be used with
any planner that plans forward from the initial state.
• So that SHOP2 can handle temporal planning domains, we have a way to translate
temporal PDDL operators into SHOP2 operators that maintain bookkeeping infor-
mation for multiple timelines within the current state. In principle, this technique
could be used withanynon-temporal planner that has sufficient expressive power.
The rest of this paper is organized as follows. Section 2 gives some background on HTN
planning, and Section 3 describes SHOP2’s features and planning algorithm. Section 4
describes how to write domain descriptions for SHOP2: in particular, Section 4.1 discusses
basic problem-solving strategies, and Sections 4.2 and 4.3 describe aspects of SHOP2 that
are specific to handling temporal and metric domain features. Section 5 discusses SHOP2’s
performance in the competition, Section 6 discussesrelated work, and Section 7 gives a
summary and conclusion. Appendix A contains a SHOP2 domain description for one of the
problem domains in the planning competition.
2. HTN Planning
HTN planning is like classical AI planning in that each state of the world is represented by
asetofatoms, and each action corresponds to a deterministic state transition. However,
HTN planners differ from classical AI planners in what they plan for, and how they plan
for it.
The objective of an HTN planner is to produce a sequence of actions that perform some
activity or task.Thedescription of a planning domain includes a set of operators similar
to those of classical planning, and also a set of methods,eachofwhichisaprescription for
how to decompose a task into subtasks (smaller tasks). Figure 1 gives a simple example.
Given a planning domain, the description of a planning problem will contain an initial state
like that of classical planning—but instead of a goal formula, the problem specification will
contain a partially ordered set of tasks to accomplish.
Planning proceeds by using the methods to decompose tasks recursively into smaller
and smaller subtasks, until the planner reaches primitive tasks that can be performed di-
rectly using the planning operators. For each nonprimitive task, the planner chooses an
380