Experiments in Implicit Control
Katia P. Sycara
School of Computer Science
Carnegie Mellon University
Michael Lewis
School of Information Sciences
University of Pittsburgh
Outline
• Agent Roles & Functions
• Implicit Control
• Examples from our work
– Slippery Motion in VR
– Path planning for a military coordination task
– Robot control for urban search & rescue (in
progress)
Don Norman’s 7 stages of Action:
Agents can assist anywhere in this loop
Secretary
scheduler
eSnipe
Spam
filter
bidder
Problems in Human-Machine Interaction
Synchronous Commands
Difficult for long sequences or many
parameters ex: composing
complex queries, setting up
spreadsheet
Asynchronous Commands
Difficult for long sequences or
branching ex: programming
languages
Implicit Commands
Difficult due to ambiguity ex: plan
recognition
Implicit Control as a special case
• Direct correspondence between interaction
with automation & the domain
User directly performs some analog of task
• Restrictive, recognizable, & repeatable
automation subtasks
Automated subtask is unique & precise
• Means for unambiguous communication of
intent
Initiating conditions unique & distinctive
Design Pattern
• User expresses intent by acting upon
domain (imprecisely or failing to account
for full range of constraints)
• Agent infers intent and “elaborates” action
(do what I mean)
• Agent elaborated action is more consistent
with user’s intent than original action
Collision Handling in VR
Non-augmenting strategies:
Clunk- collide & stop
Ghost- collide & pass through
Implicit control (many possible)
Slip- collide & redirect
SLIP moves the actor along
surfaces rather than “into” them
The Baffles Maze
Implicit Control was faster
MokSAF & Deliberative Planning
Agents:
• have access to digital information in the infosphere
• cannot consider intangible objectives which are not part of
that digital infosphere
Humans:
• Understand Idiosyncratic and situation-specific factors
– local politics, non-quantified information, complex or vaguely
specified mission objectives
• Dynamically changing situations
– Information, obstacles, enemy actions
Problem:
• To share and combine human and agent information and
resources
Route Planning in MokSAF
Control
Autonomus
Cooperative
Path Length, Route Times, and Fuel Usage
were uniformly better for Aided Teams
Path Length
Route Times
Errors in Vehicle Choice session 2
13
12
11
Errors
10
9
8
7
6
Control
Autonomous
Cooperative
MokSAF & Implicit Control
Augmentation improved path planning
But
Implicit control of the Cooperative RPA
(elaborated user action rather than responding
to commands) improved overall Task
performance (path & vehicle selection)
Robots in Urban Search &
Rescue (USAR)
Earthquakes, fires, war, or terrorism can leave human victims
trapped in unstable structures hidden within rubble
Robotic Searchers are:
Expendable
Can reach otherwise inaccessible areas
Heightened sensory capabilities: FLIR, Acoustic, Ladar, chemical
Problems:
Expense
Locomotion over irregular terrain
Perceptual limitations
NIST’s Urban Search & Rescue Reference Tasks
(from Jacoff et al. 2003)
•
Yellow Region
– Simple to traverse, no agility requirements
– Planar (2-D) maze
– Isolates sensors with obstacles/targets
– Reconfigurable in real time to test mapping
•
Orange Region
– More difficult to traverse, variable floorings
– Spatial (3-D) maze, stairs, ramp, holes
– Similarly reconfigurable
•
Red Region
– Difficult to traverse, unstructured environment
– Simulated rubble piles, shifting floors
– Problematic junk (rebar, plastic bags, pipes…)
Orange & Yellow Arenas from Jacoff et al. 2003
Human Factors Challenges
• World through a straw (restricted FOV)
• Camera control for search/navigation (Hughes &
Lewis, HFES 2002)
• Survey knowledge (mapping environment) from
restricted FOV & impeded movement
• Visual “smearing” from close surfaces
• unfamiliar ground level perspective
• Difficult distance judgments from degraded 2D
image
• Difficult orientation judgments from visual cues in
disorderly environment
• Difficult locomotion due to out-of-view &
negative obstacles
Orange Arena Simulation
(January-February 2003)
• ProEngineer solid model converted to Unreal
format
• Digital photographs used to create textures to be
applied to the model
• Glass, mirrors, orange safety fencing, and other
“special effects” added
• Rubble, debris, and victim models added to
simulation
• Robot characteristics adapted from Karma vehicle
class
Simulation of Orange Arena
Orange Arena Platform: photo &
simulation
First generation interface
(runs with both Corky & simulation)
Robot interface demo
Implicit Control for teleoperation
• Hidden obstacle avoidance/safeguarding
• Camera control & attention direction
• Automated scanning/scene reconstruction
END
Corky in real life & simulation
Manual (naïve) path
Autonomous Agent with Constraints
Road
Soil
Rendezvous
Point
River
Forest
Commander’s route
Building Teammate’s route Freeway
Cooperative (hi-liter) Agent
Descargar

Document