Unified framework for path-planning and task-planning for autonomous robots
Introduction
Nowadays, most of the robotic systems that are being designed have to move and perform tasks in a variety of environments. In order to do that, they have to avoid obstacles, find a collision free trajectory, and plan tasks. Some of these systems aim to help with ordinary, but tedious tasks. Generally, they move in known surfaces and the number of decisions is usually limited such as night surveillance tasks.
However, new science missions such as the two main missions ExoMars and MARS Sample Return (part of the Aurora program of the European Union Council of Research) will require more capable systems to achieve the goals that they are designed for. Mobility and science capabilities of the rovers are increasing, which require more powerful tools to assist on-ground operators and autonomous capabilities for the rovers. In previous missions, optimality in the path followed by the rover was not a difficult task due to the short distances they could travel per day. However, a critical aspect in future missions is to optimally (or sub-optimally) plan the path that a robot should follow while performing scientific tasks. In addition, improvements for such domains can also be incorporated into commercial robotic applications, e.g. performing inventory tasks in a large warehouse or autonomous logistics domains [1].
In this direction, the paper presents some of the results obtained within the Ph.D. program on the topic of Cooperative Systems for Autonomous Exploration Missions supported by ESA. Particularly, we are pursuing autonomous navigation with a rover for an exploration domain with several scientific targets along a known terrain. To face this problem, we initially had a complex deliberative model expressed in Planning Domain Definition Language (PDDL) [2] in which path-planning and task-planning were solved with a PDDL planner. This solution was very limited in (i) the map size because the planner only solved small grids in a reasonable time, and (ii) the quality of path generated because the planner generated suboptimal paths that were impossible to improve by using recent research advantages on path-planning algorithms. A second approach was to partially detach path-planning and task-planning. Before the search process starts, it is possible to include in the PDDL model a visibility graph with connections between tasks, which is generated using a path-planning algorithm. Then, the PDDL planner can obtain a feasible solution to achieve all tasks. Finally, the path-planning algorithm generates paths for the movements between each pair of locations in the plan. However, in our experiments we observed that solutions were not optimal. This was because of the domain independent heuristic employed by the task-planner. Considering a movement in the same way as other actions (e.g. take picture) leads to not properly consider the distance between tasks.
For this reason, we have developed an Artificial Intelligence (AI) planner that integrates capabilities of path-planning and task-planning. The main idea is to take advantage of path-planning heuristics and merge them with domain independent heuristics to generate better solutions in robotic domains. The proposed planner, called Unified Path-Planning and Task-Planning Architecture (up2ta), is able to plan paths considering the shortest path while performing scientific tasks in an efficient ordered way. A PDDL planner is responsible of ordering the tasks while a path-planning algorithm searches for the route between tasks. In up2ta, these planners are highly coupled, allowing to merge the heuristics of both planners in order to provide better solutions for mobile robotics domains. We perform an experimental evaluation of our planner on two classical exploration domains: pictures acquisition and samples delivering. In addition, up2ta is currently deployed as the deliberative layer of the Model-Based Autonomous Controller (MoBAr) [3], [4], which has been also used in our evaluation.
The next section reviews those heuristic algorithms for path-planning that are used within our planner. Section 3 presents a brief description of heuristic planners in the state-of-the-art, with a special focus on the ff planner that we use in our integration. Section 4 defines approaches that interleave task-planning and motion/path-planning. Section 5 describes our initial attempt to solve problems without a heuristic integration schema. Section 6 introduces two application domains that are then used in an experimental evaluation, which is shown in Section 8. Section 9 presents some conclusions.
Section snippets
Path-planning review
In this section, we review the most common path-planning algorithms and their features. In particular, we focus on heuristic search algorithms applied to path-planning since we use those in our integration.
Heuristic search planners
Heuristic search planners use heuristic functions H() to find a sequence of actions that reaches the goal from the initial state [14]. Heuristic functions are commonly generated by simplifying the original problem. This is known as relaxation. A common relaxation-based heuristic in planning is the delete relaxation, which consists of ignoring delete effects. Considering a delete relaxation problem, the estimated cost from any state to the goal state can be seen as a lower bound of the
Integration of task-planning and path-planning
Traditionally, task-planning and motion/path-planning problems were covered by separating both problems, i.e., high-level task modelling ignores low-level constraints. This simplification results in inefficient or even infeasible solutions. Then, works that integrate task-planning and motion/path-planning to solve problems such as object manipulation or mobile robotics domains have been reported in the literature, providing better solutions. We can mention work done by Zacharias et al. [28],
Unified path-planning and task-planning architecture
The idea behind the up2ta planner comes from our experience using a PDDL planner as the deliberative layer for the control of a mobile robot in exploration domains. We wanted to develop a planner that gets a closer to the optimal ordering of multiple tasks placed in a grid. In order to do that, we combine a modified ff planner and a path-planning algorithm to work together in a coordinated way. The resulting system, called up2ta, takes the benefits of (i) a PDDL planner for task-planning using
Mobile robot exploration domains
In this section, we describe a scenario that we use in our experimental evaluation. It consists of a mobile robot that must achieve a set of exploration tasks in different locations. As explained in Section 5.1, we must provide three files to up2ta: the PDDL domain and problem, and the terrain information (DTM). The PDDL domain was partially presented in Fig. 5. It describes robot operations to manage the different subsystems (power on/off), move between points (Move_To), and perform
Experimental results
In the following, we present an experimental evaluation on the mobile robot exploration domains described in the previous section. The test consists of running the up2ta planner with different path-planning algorithms (the algorithms use the same methods and structures to manage the DTM) over 100 uniforms flat surfaces of 500 × 500 nodes with 40% of blocked cells (the map generation algorithm can be found in Muñoz et al. [37]) and randomly placed tasks. In addition, we try different heuristic
Robotic platform deployment tests
In this section, we present a demonstration of up2ta as the deliberative layer of an autonomous controller called MoBAr [3], [4]. This controller is a 3-layer architecture formed by a functional layer supported on ROS [38], an executive based on PLan EXecution Interchange Language (PLEXIL) [39], and a PDDL based deliberative, which is up2ta in this case. up2ta is in charge of generating a plan to achieve a set of goals (e.g., pictures or samples acquisition for the domains used in our
Conclusions
In this paper, we presented the up2ta planner that integrates task-planning using a PDDL based planner and a path-planning algorithm to obtain feasible paths. The motivation of this new planner is to create a deliberative layer for autonomous robots that can interleave task-planning and path-planning. The aim is to obtain a closer optimal ordering of tasks taking into consideration the path between them. A PDDL model could manage both objectives, but only with small grids and spending several
Acknowledgements
Pablo Muñoz is supported by the European Space Agency under the Networking and Partnering Initiative Cooperative systems for autonomous exploration missions project 4000106544/12/NL/PA. This work was partially supported by UAH project 2015/00297/001, MINECO EphemeCH TIN2014-56494-C4-4-P and Junta de Comunidades de Castilla-La Mancha project PEII2014-015-A. Authors want to thank Héctor Franco Gregorio and Diego López Pajares for their contributions on the up2ta and TurtleBot implementation and
Pablo Muñoz is Ph.D. student at Universidad de Alcalá and funded by the European Space Agency under the Networking/Partnering Initiative project entitled “Cooperative Systems for Autonomous Exploration Missions”. Engineer in Computer Science (2009) and masters in Space Technology and Science (2010, best qualifications), his research is focused on autonomous controllers, applying Artificial Intelligence techniques such as planning, scheduling and intelligent execution to space robotics. He has
References (39)
- et al.
Self-driving vehicles in logistics, Tech. report
(2014) - D. McDermott, The PDDL planning domain definition language, The AIPS-98 Planning Competition Committee,...
- P. Muñoz, M.D. R-Moreno, A. Martínez, A first approach for the autonomy of the Exomars rover using a 3-Tier...
- P. Muñoz, M.D. R-Moreno, Model-based architecture on the ESA 3DROV simulator, in: Procs. of the Application Showcase,...
Grid-based path-finding
- et al.
A formal basis for the heuristic determination of minimum cost paths
IEEE Trans. Syst. Sci. Cybern.
(1968) - et al.
Theta*: Any-angle path planning on grids
J. Artificial Intelligence Res.
(2010) - et al.
Near optimal hierarchical path-finding
J. Game Dev.
(2004) - D. Ferguson, A. Stentz, Field D*: An interpolation-based path planner and replanner, in: Procs. of the 12th...
- A. Nash, S. Koenig, M. Likhachev, Incremental Phi*: Incremental any-angle path planning on grids, in: Procs. of the...
The ff planning system: Fast plan generation through heuristic search
J. Artificial Intelligence Res.
The fast downward planning system
J. Artificial Intelligence Res.
Cited by (35)
Human–robot collaboration for on-site construction
2023, Automation in ConstructionOGATE: A framework for autonomous controllers assessment
2023, Robotics and Autonomous SystemsCitation Excerpt :The architecture follows a sense-(re)plan-act cycle in which the goals and the current state of the world are stored in the PDDL problem to be accessed by the deliberative layer, while the domain (also encoded in PDDL) provides the robot capabilities and the environmental constraints. Currently, we are employing the optic [44] planner (but others can be used [45]), that accepts PDDL 2.1 [46]. Using the functions introduced in PDDL 2.1 enables us to create a temporal model of the actions, thus allowing to decide whether a plan is feasible or not in terms of the communication opportunities.
A chaotic path planning generator enhanced by a memory technique
2021, Robotics and Autonomous SystemsCitation Excerpt :In this work, the application that will be studied is a chaotic path planning for area exploration [9,20–24]. Path planning in general refers to designing a navigation strategy for anautonomous robot exploring an area, usually taking into account some efficiency measures, like fast coverage of the given space [25–28]. In chaotic path planning, apart from area exploration, the additional task is to design a navigation strategy that will look random and unpredictable.
MPTP: Motion-planning-aware task planning for navigation in belief space
2021, Robotics and Autonomous SystemsCitation Excerpt :In this paper, we employ a PDDL-based task planner. UP2TA [34] develops a unified path planning and task planning framework for mobile robot navigation. In this approach, the robot is required to perform a series of tasks at different locations before returning back to the initial location.
Task assignment of multiple agricultural machinery cooperation based on improved ant colony algorithm
2021, Computers and Electronics in AgricultureCitation Excerpt :The task assignment of multi-machine cooperation is a key link in the cooperative navigation system. A reasonable task assignment can improve the efficiency of the entire system and reduce execution costs (Cai, 2015; Muñoz et al., 2016; Hanheide et al., 2017). In order to reduce the production cost to the greatest extent, Giordani et al. (2013) proposed a two-tier distributed multi-agent system (MAS) framework solution.
Pablo Muñoz is Ph.D. student at Universidad de Alcalá and funded by the European Space Agency under the Networking/Partnering Initiative project entitled “Cooperative Systems for Autonomous Exploration Missions”. Engineer in Computer Science (2009) and masters in Space Technology and Science (2010, best qualifications), his research is focused on autonomous controllers, applying Artificial Intelligence techniques such as planning, scheduling and intelligent execution to space robotics. He has published several conference papers on autonomous controllers and others specialized in path planning. Currently, his work aims to better understand the interaction between planning and scheduling with the underlying execution system, and to analyse and measure key factors that affect to the performance of autonomous controllers using a space based scenario as reference.
María Dolores R-Moreno received a Ph.D. in Computer Science from Universidad de Alcalá (UAH) in Madrid with the distinction of European Ph.D., and Associate Professor in the same university since 2008. In 2007 Dr. R-Moreno spent one year at NASA Ames Research Center as a postdoc, and since then, she has visited several research center such as the European Space Agency (ESA-ESTEC), NASA Ames and the Jet Propulsion Lab (JPL). Her research focuses on Automated AI Planning & Scheduling, Intelligent Execution applied to real applications (i.e. aerospace, robotics, e-learning or the web) and Evolutionary Computation. She has served in the program committee of several international AI conferences and reviewer of international journals. She has published over 100 journals, books and conference papers.
David F. Barrero is a senior lecturer in the Computer Engineering Department at the Universidad de Alcalá, Spain. His research interests include genetic programming and evolutionary algorithms, educational data mining, lightweight cryptography, and electronic government. Barrero has a Ph.D. in computing from the Universidad de Alcalá.