ORTAC+ : A User Friendly Domain Specific Language for Multi-Agent Mission Planning

Full text

A. Ferrando, R.C. Cardoso (Eds.): Third Workshop on
Agents and Robots for reliable Engineered Autonomy 2023 (AREA’23)
EPTCS 391, 2023, pp. 127–133, doi:10.4204/EPTCS.391.14
© C. Bonhomme, J-L. Dufour
This work is licensed under the
Creative Commons Attribution License.
ORTAC+ : A User Friendly Domain Specific Language for
Multi-Agent Mission Planning
Caroline Bonhomme
Safran Electronics and Defense
Massy, France
ONERA
Toulouse, France
caroline.bonhomme@safrangroup.com
Jean-Louis Dufour
Safran Electronics and Defense
Massy, France
jean-louis.dufour@safrangroup.com
A tactical military unit is a complex system composed of many agents such as infantry, robots, or
drones. Given a mission, an automated planner can find an optimal plan. Therefore, the mission itself
must be modeled. The problem is that languages like PDDL are too low-level to be usable by the
end-user: an officer in the field. We present ORTAC+, a language and a planning tool designed for
this end-user. Its main objective is to allow a natural modeling of the mission, to minimize the risk of
bad modeling, and thus obtain reliable plans. The language offers high-level constructs specifically
designed to describe tactical missions, but at the same time has clear semantics allowing a translation
to PDDL, to take advantage of state-of-the-art planners.
1 Introduction
A tactical military ground unit is a complex system composed of many agents, such as infantry sec-
tions, armored units, ground robots, or flying drones. A typical mission lasts one day and is a mix of
displacements on a road network of 20 km between given start and finish points and actions. We deal only
with the reconnaissance action of a route or an area when an enemy presence is suspected. For us, the
reconnaissance action is equivalent to a speed reduction of the progression on a route and exploration of
an area. The progression and exploration should be done more safely, thus it is slower. It could be seen as
the first visit to a location. This brings us to a variant of the classic Multi-Agent Path Finding (MAPF)
problem [
1
]. Of course, if we consider moves as actions, this is also a multi-agent planning problem [
2
],
but classification as a MAPF problem is more accurate.
Given a mission, if it is described in a programming language, an automated planner can find an
optimal plan. Therefore the mission itself must be modeled. The automated planning community has
developed the PDDL language in order to specify high level planning problems [
3
]. However, the
specification of the mission should be done by an officer in the field who is not an expert in automated
planning. Languages like PDDL are too low-level to be usable by the defined end-user who is not an
expert in planning. This is why we present ORTAC+. Its main objective is to allow a natural modeling of
the mission, so as to minimize the risk of errors on the modeling of the desired mission, and thus obtain
reliable plans. The language is composed by high-level constructs specifically designed to facilitate the
specification of tactical missions, with a clear semantic allowing a translation to PDDL, to take advantage
of state-of-the-art planners. Planning is done offline, and the plans are deterministic, with no contingency.
The default assumption is that the enemy, when suspected, is in fact not present. Otherwise, if there is
contact with the enemy, the execution of the plan is de facto interrupted in the field, and the officer must
replan with modified objectives.
We proceed by presenting related work of the paper in section 2. The domain-specific language
ORTAC+ is presented in section 3, the operational context in section 4, and the conclusion in section 5.

128 ORTAC+ : A Domain Specific Language for Multi-Agent Mission Planning
2 Related Work
ORTAC stands for "Optimal Resource and Technical Action Control". It is a multi-agent planning language
and tool combining constraint-solving techniques with advanced search strategy to deconflict single-agent
plans [
4
]. At the language level, ORTAC is innovative in the constraints of coordination between units
introducing the notion of "support" explained later [
4
]. The proper term for that is "cross-schedule
dependencies" [5], typically in the case of "coalition formation."
PDDL is the standard planning language used commonly by the automated planning community.
One of the purposes of PDDL was to benchmark planning algorithms. Since its creation in 1998 for the
International Planning Competition [
6
], the language evolved to model more complex planning problems
as temporal or numerical planning [
3
]. The Competition of Distributed and Multiagent Planners [
7
]
benchmarks some centralized and decentralized algorithms with an extension of PDDL for multiagent
planning as a reference language. ORTAC and PDDL are powerful tools for modeling high-level missions
and finding a corresponding plan, nevertheless, they cannot be used by an end-user.
In recent years, an effort on designing domain specific languages (DSL) for non-expert in robotics
raised in the research community. Indeed, some of these languages lack clear semantics [
8
]. Additionnaly,
users are required to define the behavior of each robot. Consequently, the autonomy of the system is
hindered. Theses systems lack planning capabilities, making it challenging to initiate the integration of
planning tools. The research community uses PDDL as the standard task planing language. However, its
integration into robotics remains limited. To remedy this issue, efforts were done to integrate planning in
robotics [9].
3 ORTAC+
ORTAC+ is a tool that assists a military officer in planning a mission. It is a language that aims to be
expressive with well-defined syntax and semantics. To do that, the language should allow the description
of a given mission and be compatible with automated planners. A military mission is limited in space and
time. The first version of ORTAC+ emphasizes the spatial aspects of the mission. The temporal aspects
will be added in future works.
3.1 Problem Definition
The missions could be seen as a multi-agent planning problem where each agent collaborates to reach
a global goal. However, we only consider agents movements. Thus, the model is a variant of a MAPF
problem. A classical MAPF problem is represented with a set of agents with an initial position and a
target position, in an undirected graph. The solution of a classical MAPF problem is a set of single-agent
paths that does not collide [
1
]. ORTAC+ allows the description of this kind of problem but aims to be
more general with the following variants:
•
Targets are no longer assigned to each agent but to teams of agents, leading to the possibility of
having less target positions than agents. Several agents could be assigned to the same target position
but only one of them should reach it.
•
Spatial constraints are added to the problem. The constraint can be forbidden or enforced vertex for
specific agent.
•
Agents can be heterogeneous. While classical MAPF problems harnesses agents with similar
characteristics, our problem differs in this aspect.
•
The position of an agent is not only on vertices but can be on edges too. The path of an agent is
then a succession of vertices and edges.

C. Bonhomme, J-L. Dufour 129
3.2 Modelisation
Our DSL describes the mission within a three axis representation:
•
A geographical representation, corresponding to an undirected graph
G= (V,E)
, with
V
the space
of vertices which models geographic points and
E
the space of edges that links one vertice to
another. If there is an edge between two vertices, it means that there is a path between these two
points, a road for example.
•
A resource representation, including the agents involved in the mission and their characteristics.
Each element of the description is an object with attributes.
• An operational representation where the constraints and the goal will be defined.
This approach allows a structured and organized representation of critical information, enabling a
comprehensive analysis of a complex mission. While the geographic and resource representations corre-
spond to static meta-data available at the start of the mission, the operational representation corresponds
to flexible operational and tactical elements from the mission, allowing the definition of the goal and
mandatory waypoints. Furthermore, some components of the description are objects with attributes. The
type of mission modeled by ORTAC+ deals with three types of objects: the nodes, the edges, and the
agents. By leveraging the three axes, we enable an in-depth portrayal of mission-critical information with
a more expressive and accessible representation. It is worth noting that the agents involved in the missions
are usually humans, but the tool aims to be compatible with robots.
During the conception phase, the graph is pre-defined or already given. The mission specification
includes the available resources and the operational elements within the system with high-level predicates.
Each predicate has a logical meaning and is a constraint for a planning algorithm.
3.2.1 Resources
The resources encompass the agents, their initial states, and their characteristics. The agent
_
define(init,
characteristics) predicate instantiates the agent, creating the object agent with its initial position init and
its characteristics. It is still possible to add attributes to the agent after its instantiation.
3.2.2 Operational Predicates
This representation is the modular part of the modeling. Indeed, modeling a complex mission can lead to
the absence of solutions. In this part the user can change some elements of the problem and relax some
constraints to find a plan. The planning tool allocates a path for each agent to reach the goal, and the
tactical interpretation of the mission remains to the user.
Table 1 presents the high-level predicates introducing the constraints. Each predicate has a description
and a logical meaning. For example, the predicate node
_
goal() can allocate a node goal to an agent. The
predicates aim to give a declarative specification.
Besides, the predicates can be generalised so that they operate on lists of objects and no longer
individual objects. To facilitate modeling, when a "generalised" predicate is applied to an individual
object, this is automatically replaced by the application on the singleton containing that object. So
the preceding predicate is in fact syntactic sugar for: node_goal([node],[agent]). The general case
node_goal(node_list,agent_list) has the following meaning:
∀n∈node_list,∃a∈agent_l ist,a is in n at t f inal (1)
The
∀n
is in fact a general pattern to which all ’list-generalized’ predicates conform, so as to facilitate
their understanding: the first list argument can always be "and-expanded", which means:

130 ORTAC+ : A Domain Specific Language for Multi-Agent Mission Planning
Predicates Description Logic formula
node_goal(node, agent)
The location of agent should be
node at the final state
agent.loc(tfinal ) = node
node_visit(node, agent)
Agent should visit node during
execution
∃t∈[[tinit ial ;tf inal ]],
agent.loc(t) = node
edge_visit(edge, agent)
Agent should visit edge during
execution
∃t∈[[tinit ial ;tf inal ]],
agent.loc(t) = edge
node_avoid(node, agent) Agent should never visit node ∀t∈[[tinit ial ;tf inal ]],
agent.loc(t)6=node
edge_avoid(edge, agent) Agent should never visit edge ∀t∈[[tinitial ;tf inal ]],
agent.loc(t)6=edge
node
_
supported
_
from(node
1
,
node2)
An agent can visit node
1
only
if another agent supports him in
node2
∀t∈[[tinit ial ;tf inal ]],∀agent,
agent.loc(t) = node1⇒
∃agent26=agent,
agent2.loc(t) = node2
Table 1: Existing predicates in ORTAC+ DSL. Other predicates can be constructed
predicate([o1,o2], list2, ...) ⇔predicate(o1, list2, ...) & predicate(o2, list2, ...)
Because of the multi-agent nature of the model, coordination constraints could be required. We are
interested in the coordination presented in ORTAC [
4
]. The main coordination predicate uses quantitative
time, but if we restrict ourselves to causality, it boils down to: support(unit1,node1,unit2,node2)
which means: "when unit1 goes through node1 (if it goes there), unit2 must be in node2". For this to make
sense, node1 and node2 must be close enough, and from a tactical point of view, it is the standard way to
progress in dangerous areas. However, we are no longer specifying the mission, but have begun designing
the plan. We will call ORTAC’s description style "imperative", and ORTAC+ will add a new style that we
will call "declarative": node_supported_from(node1,node2) which means: "when an agent goes through
node1, another agent must be in node2". This first addition to ORTAC will be called "anonymization",
it is not a generic evolution because it is only suitable for certain predicates like support or goal. And
concerning goals, their "anonymization" corresponds to an interesting variant of MAPF initially called
"permutation-invariant" [10] and today called "anonymous" [11].
3.2.3 Ontology
Considering the specificity of our language for its domain, the proposal entails adding a knowledge base
to augment its capability. This knowledge base is an ontology designed to help the mission specification.
The characteristics of the agents is defined with the declarative predicate presented in the section 3.2.1.
It is possible to specify a constraint involving several agents. For example, considering a mission with
two types of agents, an agent can be a "UAV" for unmanned aerial vehicle or a "UGV" for unmanned
ground vehicle. This characteristics will be define as an attribute of the agents. Instead of writing:
node
_
goal(14, [agent1, agent2]), it is possible to write node
_
goal(14, "UGV") if agent1 and agent2 have
the attribute UGV, meaning at the end of the mission a "UGV" should be at the node 14.
Another example of the combined efficiency of objects and ontologies is: french infantry uses
impressive vehicles called "VBCI" which need some place to maneuver. If the officer wants to ban the
progression of these vehicles on narrow roads, he has to stipulate: edge_avoid("width < 10", "VBCI") in
spite of edge_avoid(list_of_edges_with_width_less_than_10, list_of_units_with_VBCI)
Currently, the ontology is hierarchical. An attribute can be a leaf in knowledge tree. For instance, if

C. Bonhomme, J-L. Dufour 131
agent1 is a UGV on wheels and agent2 is a UGV on caterpillars. There are the relations: a UGV with
wheels is a UGV, and the UGV on caterpillars is a UGV. They have the same parent. The ontology allows
to specify node_goal(14, "UGV") to involve agent1 and agent2 in the constraint.
4 Operational Context
ORTAC+ is built to specify military missions in order to assist an officer planning the mission. In the
following part, an example of mission is described, then its representation in the proposed language.
4.1 Description of the Mission
Figure 1: Cartography of the mission
in Goma, Congo.
Figure 2: Geographical mission mod-
elisation using graphs
Congolese rebels, supported by elements from Rwanda, are located in the eastern region of Goma.
The rebels raise a significant threat to the peace and security of the city. The mission focused on two
specific objectives: securing the United Nation (UN) point at the border, between Congo and Rwanda, and
preparing for the potential evacuation of civilians at the airport. With forces comprising 4 distinct military
sections and 4 motorized companies positioned along the west cost of the city, this mission implies the
recognition of crucial points and axes leading to the destination goals.
Operational constraint. Units follow specific rules of engagements during the execution of their tasks:
• Simultaneous coexistence of two distinct units at the same spatial coordinates is not allowed.
•
Any unit advancing on a main street or an intersection must be supported by another unit recently
deployed to an adjacent position.
Tactical information. The mission encompasses critical data as a foundation for modeling and planning:
•
The airport: the airport is a crucial point and must be controlled during the mission, with a designated
convoy of one company and one unit equipped with night vision glasses will be tasked to evacuate
people from the airport.
•
Securing the UN point: The UN point is identified at the border, requiring strict security measures,
with a priority to maintain control over
• Support team: A support team will be positioned at the southern region of the operation area

132 ORTAC+ : A Domain Specific Language for Multi-Agent Mission Planning
Figure 3: Specification of the "Secure the UN in Goma" mission in ORTAC+
4.2 Mission specification in ORTAC+
An example of the mission specification in ORTAC+ is given in Figure 3. The mission goal is the
securing of the United Nations. The cartography of the mission is given in Figure 1. Friendly forces
are blue rectangles at the west of Goma. Target positions are triangles at the east near to enemies in
red rectangles. The black triangle correspond to the airport and the red one the UN. The 3 axis of
representation (geography, resources and operations) appear clearly combined with the construction
of a hierarchical ontology. Geography elements of the mission are supposed to be already given. It
is represented in Figure 2 The Goma mission involves two types of agents: compagnies and sections.
The predicates "agent
_
define()" is used to instantiate the units. At the initial state, the 4 mechanized
compagnies are at the same node (9). This violates the constraint that only one unit can be at a node at the
same time. That’s why the notion of capacity is introduced for nodes and edges. By default the capacity
of a node or an edge will be 1 but it can be changed for some exceptions. The method "attribute()" allows
the user to add some characteristics to the units. Then, the high level predicates described earlier are used
to add the constraints of the mission. The final state is defined with the predicate "node
_
goal()" with the
help of the knowledge base. A compagny should be at the node 11, corresponding to the airport, at the
final state. Other constraints are defined, for instance, the unit1 should visit the node 9 but avoid the edge
(9,8). ORTAC+ introduces coordination constraints. If an enemy is suspected on a node or an edge, a unit
cannot visit it without support. Therefore, the predicate "node
_
supported
_
from()" model this coordination.
In the presented example, it is specified in order to visit the node 3, there must be a support at the node 18.
5 Conclusion
We have presented ORTAC+, a DSL that hopefully permits a military officer to specify a mission.
Compared to its predecessor, ORTAC, the evolution can be summarized as follows:
1.
predicates are anonymized when suitable, in particular for the declaration of goals or supports and
operate on lists of homogeneous objects (agents, nodes, or edges),
2.
these lists can be described intentionally, with a propositional sub-language constraining the
attributes of objects,
3.
these propositions, when they involve the "or" operator, can be shortened thanks to the declaration
of the adequate ontology.
As a result, ORTAC+ allows more compact and declarative models, which will be more easily written
and checked by end-users. At the same time, we take care to keep semantics to allow a translation to
PDDL, making it possible to rely on the existing planners. Current and future work explores 3 orthogonal
directions: new coordination constraints, quantitative time and resources, and compatibility with flying
drones.

C. Bonhomme, J-L. Dufour 133
References
[1]
Roni Stern, Nathan Sturtevant, Ariel Felner, Sven Koenig, Hang Ma, Thayne Walker, Jiaoyang Li, Dor
Atzmon, Liron Cohen, TK Kumar, et al. Multi-agent pathfinding: Definitions, variants, and benchmarks. In
Proceedings of the International Symposium on Combinatorial Search, volume 10, pages 151–158, 2019.
doi:10.48550/arXiv.1906.08291
.
[2]
Alejandro Torreño, Eva Onaindia, Antonín Komenda, and Michal Štolba. Cooperative multi-agent planning:
A survey. ACM Comput. Surv., 50(6), 2017.
doi:10.1145/3128584
.
[3]
Patrik Haslum, Nir Lipovetzky, Daniele Magazzeni, Christian Muise, Ronald Brachman, Francesca Rossi,
and Peter Stone. An introduction to the planning domain definition language, volume 13. Springer, 2019.
doi:10.1007/978-3-031-01584-7
.
[4]
Christophe Guettier, Willy Lamal, Israel Mayk, and Jacques Yelloz. Design and experiment of a collaborative
planning service for netcentric international brigade command. In Proceedings of the AAAI Conference on
Artificial Intelligence, volume 29, pages 3967–3974, 2015.
doi:10.1609/aaai.v29i2.19055
.
[5]
G Ayorkor Korsah, Anthony Stentz, and M Bernardine Dias. A comprehensive taxonomy for multi-robot
task allocation. The International Journal of Robotics Research, 32(12):1495–1512, 2013.
doi:10.1177/
0278364913496484
.
[6]
Derek Long, Henry Kautz, Bart Selman, Blai Bonet, Hector Geffner, Jana Koehler, Michael Brenner, Joerg
Hoffmann, Frank Rittinger, Corin R. Anderson, Daniel S. Weld, David E. Smith, Maria Fox, and Derek Long.
The aips-98 planning competition. AI Magazine, 21(2):13, 2000.
doi:10.1609/aimag.v21i2.1505
.
[7]
Michal Stolba, Antonin Komenda, and Daniel Kovacs. Competition of distributed and multiagent planners
(codmap). In Proceedings of the AAAI Conference on Artificial Intelligence, volume 30, 2016.
doi:10.1609/
aaai.v30i1.9870
.
[8] Swaib Dragule, Thorsten Berger, Claudio Menghi, and Patrizio Pelliccione. A survey on the design space of
end-user-oriented languages for specifying robotic missions. Software and Systems Modeling, 20(4):1123–
1158, February 2021.
doi:10.1007/s10270-020-00854-x
.
[9] Erez Karpas and Daniele Magazzeni. Automated planning for robotics. Annual Review of Control, Robotics,
and Autonomous Systems, 3:417–439, 2020.
doi:10.1146/annurev-control-082619-100135
.
[10]
S. Kloder and S. Hutchinson. Path planning for permutation-invariant multirobot formations. IEEE Transac-
tions on Robotics, 22(4):650–665, August 2006.
doi:10.1109/tro.2006.878952
.
[11]
Hang Ma and Sven Koenig. Optimal target assignment and path finding for teams of agents. arXiv preprint
arXiv:1612.05693, 2016.
doi:10.48550/arXiv.1612.05693
.
[12]
Jingjin Yu and Steven M. LaValle. Multi-agent path planning and network flow. In Springer Tracts in Advanced
Robotics, pages 157–173. Springer Berlin Heidelberg, 2013.
doi:10.1007/978-3-642-36279-8_10
.
[13]
Bernhard Nebel. On the computational complexity of multi-agent pathfinding on directed graphs. Proceedings
of the International Conference on Automated Planning and Scheduling, 30:212–216, June 2020.
doi:
10.1609/icaps.v30i1.6663
.
[14]
Alexander Kott, Ray Budd, Larry Ground, Lakshmi Rebbapragada, and John Langston. Building a tool
for battle planning: Challenges, tradeoffs, and experimental findings. Applied Intelligence, 23(3):165–189,
December 2005.
doi:10.1007/s10489-005-4606-z
.
[15]
Hang Ma and Sven Koenig. Optimal target assignment and path finding for teams of agents, 2016.
doi:
10.48550/ARXIV.1612.05693
.