2
COSA
– A Cognitive System Architecture with
Centralized Ontology and Specific Algorithms
{stefan.brueggenwirth|axel.schulte}@unibw.de
Application: UAV Guidance in Manned-Unmanned Teaming
• UAV operator located in a
transport helicopter cockpit
• Guidance of multiple UAVs
• UAVs perform reconnaissance
of routes and landing sites
semi-autonomously
• real-time data delivered to
helicopter crew
• UAV operator assistent system
UAV Guidance
Solution: Cognitive Automation
• Human centered automation approach
• Inspired by Rasmussens theory on human performance
and Endsley‘s work on situation awareness
• Collaboration with a human operator on the basis of
shared work objectives
• Artifical Cognitive Units are
intelligent agents using the
COSA Framework
UAV Camera View Helicopter Displays
UAV Operator
Helicopter Pilot
Fig.: The Cognitive Process
Implementation: COSA2
Implementation: COSA
• Originally implemented using
Soar as inference engine and
CORBA as middleware
• Reimplemented using graph
transformation techniques
with an evolutionary search
plan optimizer
• Using reimplemented
inference engine
• External PDDL planner
• Linear optimizer CPlex
for plan adjustment and
scheduling
• Spread-based middleware
Fig.: Original COSA Architecture
Identification
Goal Determination
• Rule-based inference step
• Transforms low-level, subsymbolic
sensory information into matching
high-level concepts
• Object oriented, typed syntax
• Based on Eclipse
Plugin CML
(Cognitive Modeling
Language)
procedure-based concept-based
behaviour
behaviour
• Reactive layer with prestored
behavior decreases response time
• A matching pattern immediately
fires associated procedure
• Concept-based behavior is
suspended, but may project
expected outcome using
forward chaining to evaluate
benefit of reactive rule
skill-based
behaviour
Procedure Based Behavior
Planning
• Rule-based goal instantiation step
• Activates currently pursued goals
dependent upon matched concepts
• Support for hard goals (removed-atend, may-never-occur type)
• Support for soft goals (penalty on
occurence)
Fig.: Situation awareness according to Endsley
Fig.: CML model of a-priori knowledge
Concepts
identificationrelevant
cues
Identification
Motivational Contexts
matching
concepts
Goal
Determination
task-relevant
cues
Cue Models
Feature
Formation
sensations
Fig.: Interpretation of Rasmussen‘s model of human performance
• Projection of feasible tasks candidates
against identification and goal
determination rules
• Generates task agenda that satisfies
hard-goals while optimizing soft-goals
• Strips based notation with
preconditions and expected effects
• Uses PDDL 2.2 AI planner
Fig.: Task agenda showing causal structure during execution
Task Options
goals &
constraints
Planning
task
agenda
Projection of task candidates
Task Situations
Task
Determination
Fig.: COSA2 Architecture
Procedures
current task
(intent)
Task
Execution
action
instructions
Sensori-Motor Patterns
control-relevant cues
sensations
Action
Control
Work
Environment
effector
commands
Plan execution and
Monitoring
•Task-Agenda contains tupel
with task and triggering cue
• Using ILOG CPLEX for linear
optimization and scheduling
of parameters
•Deviation between actual
environment situation and
expected preconditions or
effects triggers replanning