Aquaplanning: An Educational Framework for Automated Planning Tom ´ aˇ s Balyo, Dominik Schreiber, Patrick Hegemann, Jeremias Trautmann Karlsruhe Institute of Technology, Germany {tomas.balyo,dominik.schreiber}@kit.edu Abstract We present Aquaplanning, a new framework for PDDL and automated planning. The framework has been de- veloped for educational and experimental purposes and may also serve as an efficient and versatile planner with a multitude of different strategies and techniques. For the Sparkle Planning Challenge 2019, we employ a se- quential portfolio consisting of two different forward search strategies and a third planning approach based on incremental SAT solving. Introduction In the following, we describe the framework Aquaplan- ning 1 and its configuration that we submit to the Sparkle Planning Challenge 2019. Aquaplanning is a framework written in Java, the name being a recursive acronym for “Aquaplanning Quick Automated Planning”. It has been originally developed for educational purposes, serving as a modular and integrated platform with which students and re- searchers alike can easily understand, implement and evalu- ate ideas related to automated planning. Note that Aquaplan- ning is self-contained: it does not depend on any planning- related frameworks. We focus on keeping the planning in- stance as an object-oriented and comprehensible structure throughout the entire planning pipeline. As such, modifica- tions and extensions to the employed algorithms are possible in a comparably easy manner. The document is structured as follows: First, we present the general features and internal workings of Aquaplanning. Then, we describe the specific solving techniques we em- ploy in the Sparkle Planning Challenge 2019. Features Aquaplanning fundamentally operates on PDDL as an input format. All PDDL features related to basic classical plan- ning are supported, as well as a range of advanced features such as conditional effects, quantifications, derived predi- cates (axioms), general disjunctive logic, and numerical con- ditions. We use an original PDDL parser written on top of the Antlr parser generator (Parr 2013). Copyright c  2019, Association for the Advancement of Artificial Intelligence (www.aaai.org). All rights reserved. 1 https://github.com/domschrei/aquaplanning Our grounding algorithm features a planning graph traver- sal (Blum and Furst 1997) in order to find all relevant facts and actions in the provided problem and to simplify the problem accordingly. Actions with disjunctive conditions can be split into multiple actions by constructing the con- dition’s Disjunctive Normal Form (DNF), or they can retain their complex logical structure for the remaining procedure. The grounding is constructed in a way such that operators can have a large number of arguments and still be instanti- ated in a reasonably efficient way: We order an operator’s ar- guments descendingly by the number of occurrences within the operator’s preconditions. Then, using this ordering, we employ a depth-first search on the operator’s possible partial assignments to its arguments, immediately discarding any partial assignment which leads to an unfulfilled precondition in the planning graph’s current layer. For planning purposes, we have implemented several no- table kinds of planning approaches: in addition to a general interface for heuristic forward search planning as well as a more specialized greedy forward search procedure, we also included a number of Satisfiability-based approaches, en- coding the problem into propositional logic and solving it with the help of an incremental SAT solver. For this pur- pose, either the integrated SAT4j library (Le Berre and Par- rain 2010) or any other SAT solver over the incremental SAT interface IPASIR (Balyo et al. 2016) can be used. In the con- figuration we submit to the challenge, we use MiniSAT (E´ en and S ¨ orensson 2003). Some of the further features of Aquaplanning include a plan validation tool as well as general interfaces for parallel portfolio planning and plan optimization techniques. Sparkle Planning Challenge 2019 For the upcoming planning challenges, we have constructed a sequential portfolio consisting of three different planners. We explain each of these planners in some more detail in the following sections. We have found that Aquaplanning’s planning approaches have strong capabilities in different kinds of planning do- mains and, as a consequence, that we can achieve better re- sults with a sequential portfolio than when just employing one of the planners. The approaches are arranged in in the following way: First, we try try to solve the problem using a greedy and very fast planner. If we do not find a plan after