Arxiv:1804.08229V3 [Cs.AI] 25 Feb 2019 Declarative Language

0 Jiang et al. / Front Inform Technol Electron Eng 2019 ??(?):0-0 Frontiers of Information Technology & Electronic Engineering www.jzus.zju.edu.cn; engineering.cae.cn; www.springerlink.com ISSN 2095-9184 (print); ISSN 2095-9230 (online) E-mail: [email protected] Task Planning in Robotics: an Empirical Comparison of PDDL-based and ASP-based Systems Yuqian Jiang1, Shiqi Zhangy2, Piyush Khandelwal3, Peter Stone1 1Department of Computer Science, The University of Texas at Austin 2Department of Computer Science, SUNY Binghamton 3Cogitai, Inc. E-mail: [email protected]; [email protected]; [email protected]; [email protected] Received mm. dd, yyyy; Revision accepted mm. dd, yyyy; Crosschecked mm. dd, yyyy Abstract: Robots need task planning algorithms to sequence actions toward accomplishing goals that are impossible through individual actions. Off-the-shelf task planners can be used by intelligent robotics practitioners to solve a variety of planning problems. However, many different planners exist, each with different strengths and weaknesses, and there are no general rules for which planner would be best to apply to a given problem. In this article, we empirically compare the performance of state-of-the-art planners that use either the Planning Domain Description Language (PDDL), or Answer Set Programming (ASP) as the underlying action language. PDDL is designed for task planning, and PDDL-based planners are widely used for a variety of planning problems. ASP is designed for knowledge-intensive reasoning, but can also be used for solving task planning problems. Given domain encodings that are as similar as possible, we find that PDDL-based planners perform better on problems with longer solutions, and ASP-based planners are better on tasks with a large number of objects or in which complex reasoning is required to reason about action preconditions and effects. The resulting analysis can inform selection among general purpose planning systems for particular robot task planning domains. Key words: Task planning; Robotics; PDDL & Answer Set Programming https://doi.org/10.1631/FITEE.1000000 CLC number: TP 1 Introduction delwal et al., 2017). It is convenient to achieve all these tasks using knowledge declared in a single de- In a general purpose planning system, task plan- scription of the domain, and general purpose planning problems are tackled by problem-independent ning systems are well suited to the task. solvers based on a description of the domain in a arXiv:1804.08229v3 [cs.AI] 25 Feb 2019 declarative language. Such planning systems are ex- In order to design a general purpose planning tremely useful in application domains where many system, a declarative language for formalizing the different planning goals need to be accomplished, or domain first needs to be selected, followed by the the domain description evolves over time. For in- selection of a suitable solver which supports this lan- stance, in an application domain such as robotics, guage. Many different factors affect this selection a mobile service robot may need to solve plan- process. Every language has its limitations in rep- ning tasks such as collecting documents, making de- resenting task planning problems, and given a par- liveries, or providing navigation assistance to visi- ticular language, specific language-dependent tech- tors (Cambon et al., 2009; Erdem et al., 2012; Khan- niques may need to be employed to succinctly for- malize a particular planning problem. For instance, y Corresponding author c Zhejiang University and Springer-Verlag GmbH Germany, part not all languages support default reasoning, which of Springer Nature 2018 might bring difficulties in formalizing some planning Jiang et al. / Front Inform Technol Electron Eng 2019 ??(?):0-0 1 problems. A solver is also typically tied to a partic- soning about action preconditions and effects than ular language, but may not support all features in problems in the IPC, which focus on generating long that language, requiring careful construction of the plans efficiently. Various properties of the domain domain description using only supported features. or task in these benchmarks are also varied during Additionally, the properties of the domain can af- evaluation to analyze the effect on planning time. fect how quickly a given pair of language and solver The goal of this article is to help a robot planning can solve planning problems. For instance, some practitioner understand the effects of specific domain domains include many objects and their properties, properties to aid the choice of language selection for which can be challenging to some planning systems. general purpose planning. Finally, given a language, a solver and a planning We hypothesize that current state-of-the-art problem, there can be many ways of formalizing the PDDL-based planners perform better on tasks with problem using the language. For these reasons, care- long solutions, and ASP-based planners tend to per- ful consideration needs to be given to the selection form better on shorter tasks with a large number of language and solver. of objects. The hypothesis is confirmed in all three This article aims to help in the language se- benchmark domains. We also hypothesize that ASP- lection process, given a task planning problem at based planners outperform PDDL-based planners in hand. Specifically, we compare two declarative lan- domains where complex reasoning (specified in Sec- guages: the Planning Domain Definition Language tion 3.1) is required to reason about action precondi- (PDDL) (McDermott et al., 1998), the most pop- tions and effects. This is observed in a controlled ex- ular language in the planning community, and An- periment of the Robot Navigation and Blocks World swer Set Programming (ASP) (Gelfond and Kahl, domains. To the best of our knowledge, this is the 2014; Lifschitz, 2008), a popular general knowledge first work on empirical comparisons between PDDL- representation and reasoning (KRR) language that based and ASP-based planning systems, and can has been recently used in a variety of task planning serve as a useful reference to robot planning practi- problems (Lifschitz, 2002; Yang et al., 2014; Erdem tioners. et al., 2016), including robotics (Erdem and Patoglu, 2018). PDDL was created for the explicit purpose 2 Background of solving planning problems, whereas the development of ASP has focused on a broader set of reason- Research in task planning dates back to one of ing tasks, such as inference and diagnosis, as well as the earliest research areas in Artificial Intelligence. planning. Since the development of STRIPS (Fikes and Nils- The main contribution of this article is, within son, 1971) (as part of the Shakey robot project), the context of robotics, a comparison of planning many languages have been developed for represent- time between ASP-based and PDDL-based task ing task planning domains. Such languages typically planners when both are used to model the same need to describe actions’ preconditions and effects, domain. Evaluation is performed across three dif- and are commonly known as action languages.A ferent benchmark problems. While it may be pos- summary of some early action languages is avail- sible to construct ASP and PDDL planners specif- able (Gelfond and Lifschitz, 1998). ically suited to these specific benchmarks, domain- In order to plan for real-world problems (such independent solvers are compared in this article. Al- as robot systems), action languages first need to be though planner performance can be sensitive to do- capable of formally representing complex planning main encoding, we take care, to the extent possible, domains. Some of the recent research in task plan- to encode the domains similarly in each language. ning is focused on developing languages that im- The benchmark problems consist of the Blocks World prove the representation capability of action lan- and Hiking problems from the International Plan- guages (Giunchiglia et al., 2004; Lee et al., 2013; ning Competition (IPC) (Coles et al., 2012), as well Babb and Lee, 2015). Planning using these action as a variant of the Robot Navigation problem (Yang languages usually requires a translation to more gen- et al., 2014; Zhang et al., 2015). The Robot Navi- eral knowledge representation and reasoning (KRR) gation problem typically requires more complex rea- languages such as ASP. A planning paradigm for 2 Jiang et al. / Front Inform Technol Electron Eng 2019 ??(?):0-0 ASP was proposed (Lifschitz, 2002), and has been trajectories, such as to reason about history, non- used in many real-world applications (Chen et al., determinism, or both for diagnosis purposes. From 2010; Erdem et al., 2012; Yang et al., 2014). In this the perspective of design purposes, PDDL is an ac- article, we follow the same planning paradigm when tion description language, and ASP is an action encoding domains in ASP. query language, though their implementations often- In parallel, another line of research in task times support both description and query functional- planning aims at more efficient planning algorithms ities. At the same time, action language systems are and their implementations. PDDL (McDermott usually implemented by compilation. For instance, et al., 1998) was developed as a common formal- Coala (Gebser et al., 2010) is one such compilation ism with the goal of allowing more direct com- system that provides compilation techniques for sev- parison of planning algorithms and implementa- eral action languages. tions. Since then, many efficient search algo- There is existing research on predicting plan- rithms have been developed for task planning prob- ning time using features of domains and problems, such as Fast-Forward (Hoffmann, 2001) and lems (Fawcett et al., 2014), or more generally on Fast-Downward (Helmert, 2006). These algorithms predicting time required to solve a problem (Leyton- have publicly available implementations including Brown et al., 2002). These methods can be used to SAYPHI (Rosa et al., 2007), LAMA (Richter et al., help a planning practitioner to estimate the difficulty 2011) and FDSS (Helmert et al., 2011).

Load more