TOWARD A GOAL-BASED MISSION PLANNING CAPABILITY

Using PDDL Based Automated Planners

John Bookless and Glenn Callow

BAE Systems, Advanced Technology Centre, Bristol, U.K.

Keywords: Automated Planning, Distributed Task Assignment, PDDL.

Abstract: This paper proposes a generic goal-based mission planning framework which provides an integration

environment to support evaluation of existing planning and task assignment technologies. The framework

facilitates planning across a team of heterogeneous assets with a distributed capability for generating plans

to collaboratively achieve goals. A human operator assigns a team with a top-level goal which the

framework then decomposes into a list of tasks that can either be tackled by an individual asset or

collectively by a sub-team of assets with the appropriate capabilities. Each asset can generate individual

plans with knowledge of the current world state and a goal state. A selection of candidate planners are

investigated using the framework including a Hierarchical Task Network (HTN) Planner for goal

decomposition and a Partial Ordered PDDL (Planning Domain Definition Language) Planner for action-

based plan generation. The developed framework is applied to a search-and-rescue scenario requiring a team

of UAVs (Unmanned Aerial Vehicle) to search a specified area of operation.

1 INTRODUCTION

Successful planning for large scale missions can be a

difficult process requiring good understanding of the

overall mission objectives, knowledge of the

capabilities of available assets and ability to update

the top-level plan as new information becomes

available. Converting the top-level plan into a list of

tasks which are then assigned to individual assets,

both manned and unmanned, represents a significant

proportion of the effort in the planning problem.

Development of a goal-based mission planning

framework aims to automate part of that process

making it easier for operators to manage a team of

assets. The planning solution supports the following

features:

 Handling planning within ad-hoc teams of

assets which dynamically change over time

 No central point of failure within the system

 Decentralised task allocation and mission

planning

 Changes to state requiring regular re-planning

A search and rescue problem involving a team of

simulated UAV assets has been defined to evaluate

the mission planning framework. This scenario has

similarities with the open vehicle routing problem

(Li, Golden and Wasil, 2007), and the travelling

salesman problem. At each location in the search

area a number of tasks may be required to be

executed depending on whether a survivor has been

found or if the location has previously been

searched.

2 PLANNING FRAMEWORK

A generic mission planning framework has been

proposed, with the key components illustrated in

Figure 1. An instance of this framework will run

independently onboard every asset in a team. The

components are responsible for the following:

Task Manager: Maintains a task-stack and is

responsible for requesting the next task from the list

when the current task has been completed.

Coordinator: A central module which supports

interfacing between other modules in the framework.

Goal Decomposition Module: Interfaces to an

underlying HTN planner which decomposes top-

level goals into manageable tasks.

Task Assignment Module: Computes the next

available task to be completed by the asset based on

a team utility value.

Automated Planner Module: Interfaces to an

underlying PDDL planner which computes a list of

actions given an initial state and a goal state.

481

Bookless J. and Callow G..

TOWARD A GOAL-BASED MISSION PLANNING CAPABILITY - Using PDDL Based Automated Planners.

DOI: 10.5220/0003718104810484

In Proceedings of the 4th International Conference on Agents and Artiﬁcial Intelligence (ICAART-2012), pages 481-484

ISBN: 978-989-8425-95-9

 2012 SCITEPRESS (Science and Technology Publications, Lda.)

Figure 1: Goal-based planning framework.

State Estimator: Stores current world-state

information for assets in a team.

Deconflictor Module: Stores future world-state

information determined by assets sharing plan steps

and used to detect potential location/resource

conflicts in generated plans.

Plan Executive: Responsible for executing

deconflicted plans generated by the Automated

planner module. Also provides a low-level collision

avoidance mechanism to protect against

deconfliction errors.

Communication Module: Supports UDP

communication of state information and decisions

between assets

The modular nature of the framework supports

stand-alone implementation of the outlined

components with a common interface defined

between modules to facilitate easy integration. One

novel aspect of the planning framework is the

integration of a PDDL based planner, normally

applied to deterministic, offline planning problems

such as those set at the biennial International

Planning Competition (Long et al, 2000). This

technology supports on-the-fly goal-based planning

and re-planning to complete assigned tasks. This is

in contrast to developing a multi-agent framework

based on the BDI (Belief, Desire & Intent) paradigm

which commonly utilise pre-compiled plan fragment

libraries to construct plans (Bellifemine, Poggi and

Rimassa, 1999; Howden et al, 2001; d’Iverno et al,

2004).

The Goal Decomposition module uses the

JSHOP2 HTN planner to determine the required

tasks to complete an assigned goal. The Task

Assignment module applies a brute-force method,

but future research will consider integrating auction-

based approaches such as CBBA (Brunet, Choi and

How, 2009), or meta-heuristic approaches such as

simulated annealing (Osman, 1993). Plan

deconfliction currently assigns priorities based on

the assigned task, but future work will consider

spatial and temporal deconfliction to locally repair

plans and resolve conflicts.

3 SCENARIO

A Search-and-Rescue scenario has been defined to

provide a test case for the prototype framework. This

is based on a team of helicopter UAVs which can be

tasked with searching for survivors in an area of

operation, perhaps following the occurrence of a

natural disaster. The problem was simplified by

considering only four possible moves for each UAV

- north, south, east and west. A turning circle was

modelled in the problem such that a vehicle could

not directly move in the opposite direction to which

it is facing. Additionally, it was assumed that once a

survivor was identified, they would remain fixed at

that location.

To tailor the framework to handle the scenario

required a set of tasks and their relationship to top-

level goals to be represented as a HTN domain, and

the set of low-level actions and their relationships to

tasks to be represented as a PDDL domain. The

output list from the Goal decomposition module will

be composed of a combination of the low-level tasks

which are required to achieve the assigned top-level

goal. The output plan from the Automated Planner

module will be composed of a list of the low-level

actions which achieve the assigned task.

The following processes are performed whenever

the distributed team is assigned a new top-level goal:

1) A Top-level goal is assigned to all team

members

2) Each team member decomposes the goal into

a list of sub-tasks which are stored in a task

stack

3) The Task Assignment module is invoked to

select the next task from the stack which the

asset can complete

4) If a plan is required, the Automated Planner

module generates a problem definition file

and the PDDL planner is invoked

5) The output plan is checked for potential

conflicts by the Deconflictor module and the

plan steps are shared with other assets via the

Communication Interface. If conflicts exist a

replan may be required

6) Once conflicts have been resolved, the plan is

passed to the Plan Executive which executes

the actions

Steps 3-6 are repeated until the task stack is empty.

Updates to the state data which may require

ICAART 2012 - International Conference on Agents and Artificial Intelligence

482

additional tasks will result in Step 2 being executed

again to update the task stack (such as discovery of a

survivor). If a new goal is assigned to the team then

Steps 1 and 2 are repeated.

4 COMPARISON BETWEEN

PDDL PLANNERS

To demonstrate how the framework can be used to

evaluate planning technology, some candidate

planners were selected for comparison. Due to the

modular architecture of the developed framework, it

is easy to switch between planners. This is an

advantage of using the generic PDDL planning

language to express the scenario (Fox and Long,

2003). The performance of two candidate PDDL

planners were compared – POPF and SGPlan.

The POPF planner was selected following

collaboration with SciSys UK Ltd and the

Strathclyde University Automated Planning group

(Coles, Coles, Fox and Long, 2010). SGPlan was

selected as an alternate planner as it was a winner in

the Deterministic Planning track of the 2006

International Planning Competition (Hsu and Wah,

2008).

A mission scenario was defined where 6 assets

with sensor capability are tasked with searching an

area with no hidden survivors. However a range of

search area sizes were tested varying from a 5x5 to

an 8x8 grid. Results for the number of required team

moves and plan generation times were recorded

using both the SGPlan and POPF planners.

Figure 2 demonstrates a comparison of the plan

generation times between the two planners executed

on a 2.2GHz processor. For the 5x5 case POPF and

SGPlan have comparable times, 0.58 secs and 2.5

secs respectively. However, the POPF planner takes

a lot longer to generate a valid plan in the 8x8 case,

78.6 secs compared with 3.7 secs for SGPlan.

Comparison of Planner: Plan Generation Time

0.58

1.42

3.7

2.5

19.4

39.8

78.6

5x5 6x6 7x7 8x8

Search Area Sizes

Plan Generation Time (secs)

SGPLAN

POPF

Figure 2: Plan generation time.

Comparison of Planners: Number of Moves

100

5x5 6x6 7x7 8x8

Search Area Sizes

Number of Team Moves

SGPLAN

POPF

Figure 3: Number of team moves.

Figure 3 shows that the required number of plan

moves produced by POPF is better than that

produced by SGPlan. In the 8x8 case, SGPlan

requires 91 moves to complete the search where as

POPF requires 74 moves. This was found to be due

to a number of inefficient steps in the plan produced

by SGPlan whereby assets would transition over

unsearched grid locations and not perform a search

operation, requiring an asset to return later in the

plan. The observed performance difference between

the planners is comparable to those highlighted by

the Strathclyde Automated Planning Group (Coles,

Long and Rendell, 2010).

Although both planners found valid solutions in

all cases, platform utilisation in the plans generated

by SGPlan was not evenly distributed across the

team, compared with the results produced by POPF.

For the case of an 8x8 search area, Figure 4 and

Figure 5 demonstrates the area search coverage for

each asset executing plans produced by SGPlan and

POPF respectively. Most of the search actions in the

plan produced by SGPlan are performed by UAV5

and it can be seen that UAV1, 2 and 3 only search

the squares they initially occupy with no further

movement. Not visible in Figure 4, there are also a

number of unnecessary moves in the generated plan

whereby an asset does not perform a search as it

passes over an unsearched location, requiring a

transition back to that location later in the plan.

The results for the POPF planner, illustrated in

Figure 5, demonstrate improved asset utilisation. The

number of moves performed per asset varies

between 11 and 14 in this case. The generated plan

is not optimal but significantly improves upon the

number of unnecessary moves observed with

SGPlan.

This highlights that there is a trade-off between

the solutions produced by the two planners. SGPlan

can generate a valid plan quicker than POPF and

was found to be able to handle slightly larger search

areas. However, this is at the expense of plan quality

TOWARD A GOAL-BASED MISSION PLANNING CAPABILITY - Using PDDL Based Automated Planners

483

Figure 4: Asset coverage for SGPlan software.

Figure 5: Asset coverage for POPF software.

for this particular scenario where asset utilisation is

not evenly distributed and there are a number of

inefficient steps inserted in the plans. These are the

factors which should be considered when selecting

an appropriate planning solution.

5 CONCLUSIONS

Tailoring the prototype framework to a simple

Search-and-Rescue scenario has enabled a proof-of-

concept evaluation to be performed. This has

demonstrated that it is feasible to construct a

decentralised mission planning system which is

capable of performing goal-decomposition, task

allocation, automated planning and plan

deconfliction.

The framework’s modular architecture facilitates

integration of algorithms such that it could be used

as a test-bed to evaluate and compare planning

technology. Future work will consider the following

framework updates:

- Updates to the interface between the framework

and the automated planners to support scalability

to larger search spaces

- Handling of dynamic environments investigating

extensions to PDDL, such as PDDL+ which

enables modelling of external events and

processes (Fox and Long, 2002)

REFERENCES

Li, F., Golden, B. L., Wasil, E. A., 2007. The open vehicle

routing problem: Algorithms, large-scale test

problems, and computational results, Computers &

OR, 34(10): 2918–2930.

Long, D., et al, 2000. The AIPS-98 Planning Competition,

Artificial Intelligence Magazine, 21(2).

Bellifemine, F., Poggi, A., Rimassa, G., 1999. JADE-A

FIPA-compliant agent framework, Proceedings of

PAAM, 97-108.

Howden, N. et al, 2001. JACK intelligent agents-summary

of an agent infrastructure, 5

International Conference

on Autonomous Agents.

d’Iverno, M. et al, 2004. The DMars Architecture: A

specification of the distributed multi-agent reasoning

system, Autonomous Agents and Multi-Agent

Systems, 9: 5-53.

Fox, M., Long, D., 2003. PDDL2.1: An extension to

PDDL for expressing temporal planning domains,

Journal of Artificial Intelligence Research, 20: 61–

124.

Coles, A. J., Coles, A. I., Fox, M., Long, D., 2010.

Forward-Chaining Partial-Order Planning,

Proceedings of ICAPS-10.

Hsu, C. W., Wah, B. W., 2008. The SGPlan planning

system in IPC-6, Artificial Intelligence.

Brunet, L., Choi, H. L., How, J., 2009. Consensus-based

auction approaches for decentralised task assignment,

In AIAA Guidance, Navigation and Control

Conference and Exhibit.

Osman, I., 1993. Metastrategy simulated annealing and

tabu search algorithms for the vehicle routing

problem, Annals of Operations Research, 41: 421-451.

Coles, A. I., Long, D., Rendell, P., 2010. Experiences with

Temporal Planning, Proceedings of the Workshop of

the UK Planning and Scheduling Special Interest

Group

Fox, M., Long, D., 2002. PDDL+: Modelling continuous

time dependent effects, Proceedings of the 3rd

International NASA Workshop on Planning and

Scheduling for Space.

ICAART 2012 - International Conference on Agents and Artificial Intelligence

484