A GENERIC COGNITIVE SYSTEM ARCHITECTURE APPLIED TO

THE UAV FLIGHT GUIDANCE DOMAIN

Stefan Br¨uggenwirth, Ruben Strenzke, Alexander Matzner and Axel Schulte

Institute of Flight Systems, Bundeswehr University, Munich, Germany

Keywords:

Cognitive automation, Hybrid agent architecture, UAV ﬂight guidance, Graph matching, Soar.

Abstract:

We present an overview of our cognitive system architecture (COSA) with applications in the multi-UAV ﬂight

guidance and mission management areas. Our work is based on a modiﬁed version of the Rasmussen scheme,

which is an established model for human behaviour in ergonomics and cognitive systems. We belief that

modeling in close analogy with categories of human behavior simpliﬁes human-machine interaction as well as

the knowledge engineering process. Our hybrid agent architecture is comprised of a low-level, reactive layer

with prestored procedures and a goal-oriented, deliberative layer that enables inference and dynamic planning.

The ﬁrst, fully functional implementation of our architecture used production rules and the Soar interpreter,

enhanced with syntax extensions such as type-safety and class-inheritance speciﬁc to our modeling approach.

We then developed a specialized inference algorithm based on graph matching, which natively supports these

extensions and resulted in performance improvements over the original Rete algorithm of Soar. A major

weakness of our current implementation still lies in its static planning functionality which is realized by a

means-ends plan library. We discuss a concept that interleaves the planning process with knowledge about

anticipated action outcomes, followed by an interpretation of projected future world states with respect to

current goals. We illustrate this principle with a multi-UAV scenario.

1 COGNITIVE AUTOMATION IN

A MULTI-UAV SCENARIO

In the following, a scenario is considered in which

multiple simulated UAVs (Uninhabited Aerial Vehi-

cles) have to accomplish a time critical and resource

bounded military mission in a dynamic and continu-

ous environment. The mission comprises ﬂight plan-

ning, communication activities and payload operation

tasks. We have implemented and evaluated a conﬁg-

uration in which a human team member steers one

of the aircraft and controls an uncrewed team on a

task-based level. The possible actions of the aircraft

include selecting a ﬂight plan, loitering, performing

evasive maneuvers, payload management and sending

messages to other participants like the request to ful-

ﬁll a task. Experiments (Schulte et al., 2008) showed

that an assistant system is required to aid the human

operator of such a multi-UAV system to avoid human

factors related problems (Rauschert et al., 2008).

In a different experimental setup a manned-

unmanned teaming (MUM-T) mission was ﬂown with

human operators located aboard a simulated airborne

helicopter. The results showed that it generates se-

vere problems to guide the UAVs on a state-of-the-

art waypoint-based level, rather than on an abstract,

task-based level like in the system mentioned before

(Uhrmann et al., 2009). Human decision makers and

operators must trust in such a system of high auton-

omy, a fact that leads to the necessity that the system

is able to inform about its states and future behaviors

(Freed et al., 2005). For the intelligent agent to under-

stand and correctly interpret the task commands and

for the human operator to understand the agent’s feed-

back, we base our software development approach

on cognitive automation, which is described further

in (Onken and Schulte, 2009) and the following sec-

tions.

Designing technical systems which are ergonomic

and understandable for the human operator is a chal-

lenging task. A natural approach to designing such

human-machine systems is to share work with the op-

erator on a communication and cooperation level that

is similar to that between two human operators. This

is especially useful, if the technical system has the

purpose of supporting the human operator as in the

assistant systems mentioned in the previous section.

292

Brüggenwirth S., Strenzke R., Alexander A. and Schulte A. (2010).

A GENERIC COGNITIVE SYSTEM ARCHITECTURE APPLIED TO THE UAV FLIGHT GUIDANCE DOMAIN.

In Proceedings of the 2nd International Conference on Agents and Artiﬁcial Intelligence - Agents, pages 292-298

DOI: 10.5220/0002718202920298

 SciTePress

As explained in (Schulte et al., 2008), these sys-

tems must support the human operator in maintaining

or even improving the operator’s situation awareness

(SA). A well accepted theory of situation of SA in the

human factors community stems from Endsley (End-

sley and Garland, 2000).

Figure 1: Situation awareness according to Endsley (Ends-

ley and Garland, 2000).

As shown in ﬁgure 1, she distinguishes three lev-

els of SA with increasing complexity. While level 1

SA simply represents the elementary building-blocks

of a perceived situation, level 2 SA strives to unify

these elements in order to create a most comprehen-

sible situational understanding. Interestingly, level 3

SA includes both, expected outcomes of own actions

and the anticipated evolution of environment: Knowl-

edge, which will also play a key role in the planning

process. We argue that this human-oriented deﬁni-

tion of situation awareness should be closely linked

to what is commonly known as the belief state of an

agent. Accordingly, an agent that builds up a human-

like high-level SA, e.g. inspired by Endsley’s cate-

gories, has advantages when when interacting with a

human operator. Likewise, information that is used

internally to derive this high-level abstraction should

not be exposed to the operator.

Furthermore the artiﬁcial agents must be able to

pursue goals which are either matching directly to the

cognitive goals of the human operator or are subgoals

that the operator has assigned to them. We summa-

rize these concepts as cognitive automation (Onken

and Schulte, 2009). Cognitive automation enables

technical systems to have a human-oriented situation

representation, situation awareness and an explicit

representation of goals. We belief that, although in

some cases computationally less efﬁcient, cognitive

automation simpliﬁes human-machine interaction as

well as the knowledge engineering and system devel-

opment processes.

Although our ﬁrst, Soar-based (Laird et al., 1987)

implementation of the multi-UAV system has a rela-

tively simple means-ends plan-library that is statically

deﬁned at developmenttime, the knowledgebased ap-

proach already shows emergent behavior not prede-

ﬁned by the developer. This behavior is not generated

by directed reasoning or global optimization, instead

it can be seen as a compromise of multiple actions

related to multiple goals that arise sequentially.

Figure 2: Emergent behavior due the explicit goals: avoid

threat and return to base (Meitinger and Schulte, 2009).

An example is the generation of a ﬂight trajectory

shown in ﬁgure 2. In the depicted situation, two goals

are relevant - one is the shortest possible ﬂight path to

the home base and the other is not to come too close

to any SAM (Surface to Air Missile) site. The lat-

ter goal is (reactively) triggered as soon as a minimal

distance is under-run and then replaces the ﬁrst goal

as the actively pursued one (Meitinger and Schulte,

2009). Even though the resulting ﬂight path resem-

bles the optimal curved trajectory, the emergent be-

havior described above is the result of a greedy algo-

rithm, which follows a locally optimal heuristic (al-

ternating between the shortest ﬂight path and SAM

site avoidance goal). Currently, no projection of fu-

ture environment states is made and therefore loops

in the behavior of the agent may occur.

2 THE COGNITIVE PROCESS

According to the principle of cognitive automation

described in section 1, we have derived an infor-

mation processing scheme, which we call the cog-

nitive process. It is structured in close analogy to

Rasmussen’s model of human information processing

(Rasmussen, 1983), often cited among cognitive psy-

chologists and within the human factors engineering

community.

In ﬁgure 3, we present an interpreted version of

the traditional Rasmussen model, tailored for applica-

tion in intelligent agent environments. Our model is

comprised of a low-level, subsymbolic feedback con-

trol layer (skill based), a symbolic, reactive layer with

prestored procedures (procedure-based) and a delib-

erative, goal-oriented layer (concept-based) compara-

ble to a BDI-agent design. From an agent-theoretical

point of view, the approach resembles a hybrid agent

architecture with horizontal layering, comparable to

Fergusons TouringMachines (Ferguson, 1992). Per-

A GENERIC COGNITIVE SYSTEM ARCHITECTURE APPLIED TO THE UAV FLIGHT GUIDANCE DOMAIN

293

ceived data (feature formation, often called the see-

function) is available to the procedure based and con-

cept based layer, but the control policy is currently

straightforward: An exact match to a reactive rule al-

ways takes priority over the deliberative layer. We

currently do not partition our model in hierarchical

layers, e.g. competency-based or multi-agent-wise.

Communication mechanisms between agents have to

be explicitly encoded.

Figure 3: Interpretation of Rasmussen’s model of human

performance incorporating an information technology ap-

proach (Onken and Schulte, 2009).

The cognitive process separates static, a-priori

knowledge in the form of production rules (i.e. the

design-time knowledge: concepts, motivational con-

texts or procedures) from instantiated and parameter-

ized situational knowledge (i.e. the run-time data:

matching concepts, goals & constraints or action in-

structions). In practice, the cognitive subfunctions

(e.g. identiﬁcation, planning, etc.) may use all the

knowledge that is available to the system. The con-

ceptional processing of information, however, is se-

quential, that is, the output of one subfunction serves

as the input to the subsequent one.

2.1 Identiﬁcation

The ﬁrst processing step for concept-based behav-

ior matches identiﬁcation relevant cues with known

concepts. This inference step is done using Soar

style production system syntax with extensions for

an object-oriented representation of the scene, such

as class encapsulation, inheritance and type-safety as

shown in ﬁgure 5. After the forward chaining pro-

cess has come to quiescence, all identiﬁcation rele-

vant cues have been translated into matching high-

level concepts.

2.2 Goal Determination

Matching concepts represent the high-level abstrac-

tion that is meant to resemble human-like situation

awareness. It also serves as input data to the goal de-

termination module and may hence trigger the activa-

tion of certain goals or a subgoal hierarchy as shown

in ﬁgure 4. A goal is met if the associated instance

of a matched concept satisﬁes certain constraints, e.g.

usually an environment attribute lies within a certain

numerical interval. The agent typically strives to ful-

ﬁll all currently active goals - In case it detects a failed

goal, it will activate its planning module to establish

the unmet conditions of the goal.

Figure 4: And-Or tree representing a goal-hierarchy exam-

ple used for aircraft health monitoring

2.3 Planning

The planning module has the purpose of generating a

sequence of tasks to bring the system from the current

situation into a desired state - in our case to fulﬁll all

its currently active goals. Similar to Strips (Fikes and

Nilsson, 1971) planners we associate preconditions

and expected effects (postconditions) with each pos-

sible action. To enable backward chaining, we also

require that every production involved in the planning

process is function free. We furthermore adhere to

a framing axiom, that is, facts which are not explic-

itly mentioned as postconditions remain unchanged.

Unlike a classical planner however, we do not add or

remove literals which are checked for a goal state di-

rectly, but project the expected identiﬁcation relevant

cue of every feasible task candidate (as shown in ﬁg-

ure 3). This introduces additional inference steps to

identify matching concepts before the update and test-

ing of the goal-hierarchy is performed, but has the ad-

vantage of reusing the knowledge for situation com-

prehension (concepts) and motivational contexts. The

planning process announces failure, if it does not en-

counter a state that satisﬁes all active goals within a

ICAART 2010 - 2nd International Conference on Agents and Artificial Intelligence

294

class

<belief> danger

// class ’danger’ is added to namespace ’belief’

{

attributes:

link aircraft;

// link to an object of class ’aircraft’

string threat;

behavior:

{

create // Soar-like production

(instance[belief::aircraft::*] <own> ˆownAC true)

// IF WM contains an ’aircraft’ object that has attribute ’ownAC’ set to ’true’

(instance[belief::aircraft::*] <other> ˆotherAC true)

// and an aircraft object that has attribute ’otherAC’ with value ’true’

(instance[belief::distance::*] <dist> ˆbetween <ownAC>

// and an object of class ’distance’ the two aircraft as ’between’ parameters

ˆbetween <otherAC> ˆdistance small)

// and its ’distance’ attribute with value ’small’

-->

(elaborate <i> ˆaircraft <otherAC> ˆthreat high)

// THEN create object of this class (’danger’) ref. to other A/C, set ’threat’:=’high’

}

Figure 5: Code-Example used in the cognitive process.

certain timespan. Currently, we use a PDDL 2.2 plan-

ner to encode inferred facts as actions and derived

predicates. The re-implemented system may offer the

option to perform planning by direct manipulation of

the working memory graph

2.4 Monitoring

The outcome of a successful planning process is a

task agenda. The items on the agenda are tuples,

composed of tasks and associated task-relevant cues

which trigger the execution of the task. A task-

relevant cue might trivially be a simple time tag or

any other indicator which initiates the task, depend-

ing on the ontology. For execution monitoring pur-

poses, the agenda also contains the anticipated world-

state generated from the pre- and postconditions of a

task during the planning phase. This usually happens

whenever the next task is toggled or at intermediate

discretization points deﬁned at design time. Should

the system detect a deviation from the task agenda,

it triggers a re-planning processes from the current

world state. Alternatively it may ask the operator for

instructions to alter the system knowledge, e.g. if a

hardware failure renders a task option infeasible.

2.5 Procedure-based Behavior

We follow the general convention that an exact

match of a procedure-based behavior rule immedi-

ately ﬁres its associated reactive task, while suspend-

ing concept-based behavior during execution. Once

the procedure-based task has completed, the moni-

toring function will recognize a deviation from the

agenda and initiate re-planning from the current world

state.

Experiments and surveys among pilots and UAV-

operator personnel showed that it is this procedu-

ral knowledge, extracted directly out of manuals,

regulations, combat operation procedures or gained

from experience, that frees pilots from spending cog-

nitive resources on routine tasks and enables them

to direct their attention on high-level mission goals.

Procedure-based behavior therefore poses an impor-

tant feature for knowledge extraction and engineering

in the UAV guidance domain.

Exact matching of procedure-based rules however

implies a closed and consistent world, and is hence

prone to reproducing typical problems encountered in

conventional automation, such as literalism or brittle-

ness (Billings, 1997). We therefore envision as future

work a concept-based, high-level supervision mecha-

nism, which concurrently projects the expected out-

come of the procedure-based behavior and may ex-

plicitly interrupt it, in case it interferes with certain

high-level goals. Ideally, procedure-based behavior

can increase performance, since only the effects of the

associated task need to be projected against the goals,

instead of going through the complete planning phase.

3 IMPLEMENTATIONS OF THE

COGNITIVE PROCESS

Both our implementations of the cognitive process

use a production system that holds in its working

memory the current situation representation including

the system’s beliefs, desires and intents and contains a

knowledge base consisting of two types of production

rules that are applied to the working memory to infer

new facts or make decisions based on existing facts.

The two rule types are (i) inference rules that cause

A GENERIC COGNITIVE SYSTEM ARCHITECTURE APPLIED TO THE UAV FLIGHT GUIDANCE DOMAIN

295

a reversible modiﬁcation which is retracted, once the

rule’s left hand side does not match anymore and (ii)

operator rules that cause persistent modiﬁcations of

the state. The application of the production rules is

directed by a simple automaton as depicted in ﬁgure

6 that operates on the working memory.

Figure 6: Decision cycle controlling the Cognitive Process.

After reading the input the automaton updates the

inference rules by ﬁring the newly applicable ones

and retracting the invalidated ones. In this phase all

ﬁrings occur in parallel until no inference rule is ap-

plicable. This state is referred to as quiescence. In-

ference rules can propose the creation operator nodes

that enable graph transformation rules by generating

operator proposal nodes in working memory. Dur-

ing operator selection the automaton selects one of

the operator proposal nodes and creates an operator

node from it. This operator node serves as a precon-

dition for the operator rules that are then applied. In

the next step the inference rules are again updated to

incorporatethe modiﬁcations of the operator rules and

the operator node created during operator selection is

deleted. Finally the automaton generates the output

and starts over again.

This way the automaton provides the ﬂow control

required by the cognitive process that (i) separates the

parallel execution of the inference rules from the se-

quential execution of the operator rules and (ii) allows

priority based decisions on the order of the applica-

tion sequence of the latter (Laird et al., 1987).

The described model of the cognitive process was

initially implemented in our framework COSA. It

used the well established production system Soar as

knowledge processor but extends the functionality

and the Soar syntax to enable the object oriented ele-

ments of our programming language. The framework

has been extensively used in various applications in

the ﬁeld of multiple co-operating semi-autonomous

aerial vehicles (Schulte et al., 2008; Uhrmann et al.,

2009). Respective prototypeshave been demonstrated

in a collaborative simulation environment with great

success, but with respect to forthcoming ﬁeld exper-

iments including the use on embedded platforms and

increasing complexity of the encoded knowledge the

observed run-time performance of the existing system

is not sufﬁcient.

Our experimental results showed that the pattern

matching phase within the graph transformation is the

bottleneck in the Soar-based implementation of the

cognitive process – more than 90% of the run-time is

spent in this phase (Matzner et al., 2008). Therefore

an efﬁcient graph pattern matching for the left hand

side of the rule is the key to performance of such an

implementation. To achieve this, we re-implemented

the system by transforming the situation representa-

tion from untyped facts in the Soar-based implemen-

tation into directed, typed multigraphs. We also de-

veloped an incremental pattern matching algorithm

that on the one hand incorporates techniques from

the ﬁeld of graph transformations to increase perfor-

mance and on the other hand extends the expressive-

ness our modeling language by directly supporting in-

heritance. As the performance of a pattern matching

algorithm very much depends on the quality of the

search plan which again depends on the situation spe-

ciﬁc cardinality of objects in the host graph we further

augmented our algorithm with an evolutionary search

plan optimizer.

Figure 7: Re-implemented system shows signiﬁcant in-

crease in performance.

Examination of the re-implemented system in a

benchmark multi-UAV application in which an in-

creasing amount of rules were imposed on the sys-

tem, showed a signiﬁcant decrease of overall mem-

ory consumption which enabled the system to han-

dle scenarios with over 200 instantiated SAM site

(threat) objects, whereas the existing system reached

the memory limits at 50 objects. Also, the time used

to complete the pattern matching in each cycle was

decreased by around an order of magnitude (see ﬁg-

ure 7).

4 APPLICATION FOR

MANNED-UNMANNED

TEAMING

The manned-unmanned teaming scenario regards a

manned transport helicopter with a crew consisting

ICAART 2010 - 2nd International Conference on Agents and Artificial Intelligence

296

of a pilot ﬂying and a commander, who is respon-

sible for the helicopter and may delegate tasks to a

self-organizing pool of UAVs. Figure 8 shows the

helicopter cockpit plus multi-UAV simulator in use.

The transport helicopter has the mission goal to de-

liver troops to a speciﬁed object which is reachable

via a predeﬁned route. The way home is also pre-

deﬁned. Both routes include a corridor which may

only be used during speciﬁc time intervals. For the

protection of the transport helicopter there is an addi-

tional attack helicopter, that ﬂies in front of the trans-

port helicopter and may also leave the route in or-

der to provide protection. This is necessary because

there are previously unknown vehicles distributed in

the mission area, which are either posing a threat or

not. To enable the helicopter commanders to assess

the situation, these vehicles have to be photographed

by the UAVs early enough before any manned heli-

copter passes the corresponding area. So both he-

licopter crews are in need of the capabilities of the

UAV pool. The intelligent agent software on-board

the UAVs has to cope with payload management (i.e.

set camera angles, take picture) and ﬂight manage-

ment (i.e. trajectory planning) at the same time to

create the photos the commanders need in order to

identify the vehicles. Furthermore, the tasks given to

the UAV pool have to be sorted temporally and dis-

tributed among its participants, which leads to the ne-

cessity of a mission management capability. This ca-

pability also enables the UAVs to receive, distribute

and execute tasks that originate from various sources

in addition to the two manned helicopters mentioned

above. From the view point of the human operators

the UAVs represent capabilities (like reconnaissance)

which is the reason why we speak of capability man-

agement instead of vehicle management or the like.

Figure 8: Helicopter plus multi-UAV simulator with the

commander viewing UAV sensor data (Uhrmann et al.,

2009).

4.1 Planning and Resource Problems

As our work is evolving towards a non-deterministic,

real-world application (which will culminate in the

deployment of our physical UAVs, see ﬁgure 9), we

have to deal with much more complex, open envi-

ronments. Another recent trend is the focus on the

manned-unmanned teaming application, leading to a

more seamless integration of manned aircraft and/or

human operators into the missions. In military mis-

sions which include manned components, predeﬁned

time schedules have to be kept to reduce risk of life.

This can only be ensured, if the artiﬁcial agents them-

selves have a concept of time and are able to generate

plans that consider certain temporal constraints. In

addition, the human operator in such a dangerous en-

vironment expects the artiﬁcial agents to act on a high

optimization level.

Figure 9: Unmanned helicopter (left) / plane (right) as

MUM-T testbed.

In the scenario described above, time is the most

critical resource, but one may easily think of other

bounded resources that could play a role if realism is

raised, e.g. fuel or payload.

In the example presented in ﬁgure 10 two heli-

copters are on their way to the operation area. There

are two UAVs to support them ﬂying in front, tasked

to reconnoiter the route, then move towards the oper-

ation area and eventually build a communications re-

lay (necessary as soon as the transport helicopter has

landed). Only one UAV is needed to work as relay,

but the reconnaissance task can be covered by mul-

tiple UAVs to either achieve higher quality results or

faster accomplishment. Therefore a suitable plan for

the UAVs is to cooperate concerning the reconnais-

sance tasks until one UAV takes over the role as relay,

while the other one is free to work on tasks that are

out of the scope of this example.

We now consider the unforeseen event that a ve-

hicle is discovered along the route and the comman-

der of the attack helicopter generates additional tasks

for the UAV pool: identify the vehicle, then mark it

(see ﬁgure 10). This leads to a re-planning process,

that should answer the questions, which UAV can take

over the new tasks and will there still be any UAV that

A GENERIC COGNITIVE SYSTEM ARCHITECTURE APPLIED TO THE UAV FLIGHT GUIDANCE DOMAIN

297

Figure 10: Manned-unmanned teaming use case.

is able to work as communications relay in the given

time frame. In case the temporal constraints cannot

be satisﬁed, it may be an option to re-plan the whole

mission, e.g. slow down the manned helicopters so

that the constraints become loosened.

4.2 Future Work

To solve such a problem without too much human in-

teraction, the intelligent agent needs to predict the dy-

namic development of the situation. Furthermore, the

agent needs to keep track of time constraints, consum-

able and renewable resources. As already mentioned,

this will require more elaborate planning algorithms,

potentially operating directly on the working memory

graph. Interesting improvements could come from

features of recent AI planners, e.g. strong and soft

constraints and preferences, that were introducedwith

PDDL3 (Gerevini and Long, 2005).

As furthermore mentioned in section 2.5, another

current research activity focuses on the control policy

between reactive and deliberative layer. Finding this

right balance of ﬂexibility is a major question in the

agent community - other hybrid agent architectures

have already introduced sophisticated (but sometimes

for our purposes too complex) concepts.

REFERENCES

Billings, C. E. (1997). Aviation Automation - the Search

for a Human-Centered Approach. Lawrence Erlbaum

Associates, Mahwah, NJ.

Endsley, M. R. and Garland, D. J. (2000). Situation Aware-

ness Analysis and Measurement. Lawrence Erlbaum

Associates, Mahwah, NJ.

Ferguson, I. A. (1992). TouringMachines: An Architecture

for Dynamic, Rational, Mobile Agents. PhD thesis,

University of Cambridge. Clare Hall.

Fikes, R. and Nilsson, N. (1971). Strips: A new approach

to the application of theorem proving to problem solv-

ing. In Artiﬁcial Intelligence, volume 2, pages 198–

208, Heidelberg, Germany. Springer.

Freed, M., Bonasso, P., Ingham, M., Kortenkamp, D., Pell,

B., and Penix, J. (2005). Trusted autonomy for space-

ﬂight systems. In AIAA First Space Exploration Con-

ference, Orlando, FL.

Gerevini, A. and Long, D. (2005). Plan constraints and pref-

erences in PDDL3. Technical report, Dipartimento di

Elettronica per l’Automazione, Universita degli Studi

di Brescia.

Laird, J. E., Newell, A., and Rosenbloom, P. S. (1987).

SOAR: An architecture for general intelligence. In

Artiﬁcial Intelligence, volume 33, pages 1–64, Hei-

delberg, Germany. Springer.

Matzner, A., Minas, M., and Schulte, A. (2008). Efﬁcient

graph matching with application to cognitive automa-

tion. In Applications of Graph Transformations with

Industrial Relevance, pages 297–312, Berlin, Ger-

many. Springer.

Meitinger, C. and Schulte, A. (2009). Human-uav co-

operation based on artiﬁcial cognition. In Engineer-

ing Psychology and Cognitive Ergonomics, pages 91–

100, Heidelberg, Germany. Springer.

Onken, R. and Schulte, A. (2009). System-ergonomic

Design of Cognitive Automation in Work Systems.

Springer, Heidelberg, Germany.

Rasmussen, J. (1983). Skills, rules and knowledge, signals,

signs and symbols, and other distinctions in human

performance models. In IEEE Transactions on Sys-

tems, Man, and Cybernetics, volume SMC-13, pages

257–266, Heidelberg, Germany. Springer.

Rauschert, A., Meitinger, C., and Schulte, A. (2008). Ex-

perimentally discovered operator assistance needs in

the guidance of cognitive and cooperative uavs. In

Proceedings of HUMOUS conference, Brest, France.

Springer.

Schulte, A., Meitinger, C., and Onken, R. (2008). Human

factors in the guidance of uninhabited vehicles: Oxy-

moron or tautology? the potential of cognitive and

co-operative automation. In International Journal on

Cognition Technology & Work, Heidelberg, Germany.

Springer.

Uhrmann, J., Strenzke, R., Rauschert, A., C.Meitinger, and

Schulte, A. (2009). Manned-unmanned-teaming: Ar-

tiﬁcial cognition applied to multiple uav guidance. In

NATO RTO SCI Symposium on Intelligent Uninhab-

ited Vehicle Guidance Systems, Neubiberg, Germany.

ICAART 2010 - 2nd International Conference on Agents and Artificial Intelligence

298