Transcript AAMAS - Desktop Warfarex

Jeremy Straub
Department of Computer Science
University of North Dakota



Introduction
Background
Applications Requiring Humans
◦ Prospective Applications
◦ Requirement Source

Autonomous Control
Goal-Based Autonomy
Human Collaboration
Integrated System
Pathway to Implementation

◦ Technical
◦ Social, Legal & Ethical Considerations
Conclusions & Future Works





Partial autonomous control approaches are poised to
provide benefit to the warfighter
The approach combines:

A collaborative control approach is discussed, it combines:

Goals

◦ Scalability of autonomous control
◦ Specialized skills and abilities that humans are either particularly
well suited to or which control software hasn’t been created for.
◦ goal-based autonomy
◦ human assistance
◦ teleoperation capabilities
◦ to maximize system efficiency through the use of autonomous
control wherever possible
◦ maximize task completion and accuracy through human support
or control, as required

Robotic sensing and weapons platform
control technologies fall into two categories;
however, the boundaries of these are blurring
◦ The First presumes complete robot system
autonomy in achieving goals and completing tasks
assigned by controllers.
◦ The second presumes that human control (or at
least supervision) is required

The control of multiple robots by humans,
however, presents a particular challenge due
to the necessity to concurrently assess and
command these robots.




The prospective applications for unmanned craft
(teleoperated or autonomous) are numerous.
Applications include:
◦
◦
◦
◦
◦
◦
intelligence
surveillance and reconnaissance missions
munition location and disablement missions
military attack/defense operations
rescue missions
search and rescue operations
Robots are able to facilitate activities in remote areas,
harsh environments and constrained spaces
Teleoperation facilitates attaining these benefits
while still allowing the missions to benefit from
human knowledge and judgment.



From a moral perspective, Sauer and Schornig
site Kahn in arguing that “a soldier’s right to kill
his or her opponents depends on the condition of
mutual risk”.
Human judgment is also required for ethical
decision making (though Sauer and Schornig
suggest that in the longer-term, artificial
intelligence systems may be act in a “more
‘humane’ fashion” than humans.
Technically-dictated need:
◦ identification of targets in non-warfare applications
◦ system goal setting and/or tasking
◦ valuable intuition-based guidance




Autonomous control, when a suitable solution, is highly desirable.
Completely autonomous systems do not require human operators (and
the associated expense).
They can be controlled by local software (either onboard or in-region)
preventing the need maintain expensive long-distance communications
channels and reducing the potential for the system to fail or be
compromised due to communications channel failure or compromise.
Autonomous control can take several forms:
◦
◦
◦
◦

Script
Script with error handling capabilities
Adaptive planning
goal-based autonomy
With goal-based autonomy, the AI is given the programmatic tools
needed to determine what actions are required to complete a set of
goals. Controllers supply initial goals and refine them as necessary.
The system creates and refines an operations plan, based upon the
supplied (and updated, if applicable) goals.





With goal-based autonomy the controller
sends a high level goal in a goal definition
language that is analyzed and decomposed
by the AI.
The AI, considering current operating
conditions, situational knowledge and
other factors develops an optimized plan
for the performance of tasks required to
meet these decomposed goals.
Dependencies for these tasks are identified
and an optimized schedule for their
performance is created.
Management by exception can be used to
validate whether each prospective task has
been suitably successful.
If an exception is found to exist, the
system attempts to determine if it can be
fixed by the system (without requiring
human intervention).
Controller supplies
goal or goals
Decompose goal
into composite tasks
Schedule tasks
Plan next task
Perform next task
No
Can be fixed
autonomously
Exception
condition exists?
Yes
No
Fix exception
Tasks complete
Notify controller
No
End
Controller supplies
goal or goals






Decompose goal
into composite tasks
Many tasks can be performed successfully by an AI without
human intervention.
Some (e.g., object recognition in a chaotic environment)are
beyond the current capabilities of software systems.
Schedule tasks
Others or may be possible, but infeasible with the hardware
capabilities available onboard the craft or in the operating
region.
Prompt controller
for required changes
Problem
No problem
In these cases, human involvement can expedite task
completion and increase accuracy and system performance.
Plan next task
The previous system is now presented augmented with
human involvement
Solicit required
human input
Enhancements:
◦
◦
Each task is evaluated to determine whether human input is
required. If input is required, it is sought from human
controllers

the system attempts to resolve the exception autonomously

if this fails, human input is sought; this allows more complex
task exceptions to be effectively resolved
Yes
Task requires
human input
No
validation process that occurs after scheduling; clarification
from controllers is sought to resolve validation failure, if
applicable
addition of task types that require human intervention

Validate tasks /
schedule
Perform next task
No
Can be fixed
autonomously
Exception
condition exists?
Yes
No
Fix exception
Tasks complete
Solicit required
human input
Notify controller
End
No


it is necessary to consider how multiple instances
of the collaborative control process work in
conjunction with each other.
Several approaches have been utilized in other
work:
◦ security-monitoring-style approach where the controller
has multiple screens (or multiple windows) and attempts
to watch all of the craft concurrently, scanning for
irregularities
◦ approaches where the controller switches between the
craft regularly, checking for issues resolving them with
commands and moving to the next screen.
◦ autonomously identify and prioritize situations which
require the controller’s input and present these to him
or her

Work is required on a variety of fronts to facilitate the use
of teleoperated and autonomous craft in more battle
scenarios
Technical challenges range from specific focus areas

to more general concerns

Keeping humans in-the-loop:

◦ such as the development of technology that rivals human
performance in target identification
◦ such as ensuring that human values are properly implemented in
robotic decision making systems.
◦ emotional effects of remote warfighting on human participants
◦ enhancing training (or providing autonomous decision making
support, etc.) to facilitate better decision making about craft and
human operator capabilities
◦ understanding human perception and how to improve human
situational awareness when commanding a remotely controlled
craft

Sauer and Schomig suggest that a plethora of social, ethical and
legal considerations must be evaluated when contemplating
remotely controlled vehicles and their further advancement.
◦ The effect on the conduct of war must be considered:
 reduced the threshold to engage in combat
◦ The impact of teleoperation on the rules of war and what is considered
acceptable must be considered:
 attacks upon civilian areas may be provoked in response to unmanned craft
conflict.
◦ Lack of risk to the operators of unmanned vehicles (who are far removed
from the war) removes the right, born from “mutual risk” to harm or kill
enemy combatants
 the impact that these actions (if perceived as unfair by craft operators) will
have on operators’ mental state.
◦ Utilization of autonomous control technology may create a spiral driving
further autonomy.
 each side would realize the comparative competitive advantage and that each
upgrade would trigger a virtual need by the other side to match
 the prospective conclusion to this cycle may be wars initiated and fought
before humans are even fully aware of them.



This paper has provided an overview of
ongoing work related to the initial design of
an integrated system for controlling
Future work will involve the completion of the
implementation of this system and testing to
quantify the qualitative benefits from
controller involvement discussed herein.
Work will also be undertaken to quantify the
relative performance of the AI-driven
controller software, as compared to the
alternate approaches traditionally used.