autonomy: from deep space to inner space - ieee entity … · · 2012-01-0912 principal...

Funded by the David and Lucile Packard Foundation

[email protected]

Principal Researcher for Autonomy

Monterey Bay Aquarium Research Institute (MBARI)

Moss Landing, Californiahttp://www.mbari.org/autonomy/

Autonomy: From Deep Space to Inner Space

Friday, December 10, 2010

Autonomy: Outer to Inner Space Kanna Rajan, MBARI 2010

Outline

Introduction

Issues in Ocean Observation

MBARI - brief intro

Scientific Motivation

Background - Autonomy & Legacy & Lessons

Partitioned Interleaved Planning and Execution (and state

Estimation)

Field Results

Next Steps - CANON


Collaborators

Frederic Py: Autonomy

Jnaneshwar Das: Adaptive Sampling/USC

Thom Maughan: Project Mgr and Sampling

John Ryan: Biological Oceanography

Julio Harvey: Molecular Ecology

Robert Vrijenhoek: Ecology

Gaurav Sukhatme: Robotics and Adaptive Sampling/USC

Maria Fox: Model Learning/Strathclyde

Angel Olaya: Adaptive Sampling/Madrid

Sergio Jiménez Celorrio: Learning/Madrid



Most of the previous century could be called a “century of undersampling.”

Walter MunkSecretary of the Navy Research Chair in Oceanography, Scripps Institute of Oceanography,

UCSD

Testimony to The U.S. Commission On Ocean Policy, 18 April 2002

On Ocean Observation



In preparing for battle I have always found that plans are useless, but planning is indispensable.

&

Failing to plan is planning to fail.

General David Dwight G. Eisenhower, Supreme Commander Allied Forces, WWII and

34th President of the United States

On Planning



“We should…send machines and instruments out to sea, not people…”

David PackardMBARI Founder

Co-Founder Hewlett-Packard &

United States Under-Secretary of Defense, Reagan Administration

On Automation



“We should…send machines and instruments out to sea, not people…”


Co-Founder Hewlett-Packard &

United States Under-Secretary of Defense, Reagan Administration

On Automation

“““WWWWWWWWWWWWWWeeeeeeeeeeeeee ssssssshould…send mmmmmmmmaaaaaaaaaachines and inssssssssssssssttttttttttttttrrrrrrrrrrrrrruuuuuments out to sssssssseeeeeeeeeeeeeaaaaaaaaaaaaa, not people…………”


Co-Founder Hewlett-Packard &&&&&&&&&&&&&&&

United Stateeeeeeeeeeeeeessssssssssssss UUUUUUUUUUUUUUnddddddddddddddddeeeeeeeeeeeeeerrrrrrrrrrrrr-SSSSSSSSSSSSeeeeeeeeeeeeecccccccccccrrrrrrrrrrrrreeeeeeeeeeeetary of Defense, RRRRRRRRRRRRRRRReeeeeeeeeeeeeagggggannnnnnnnnnnnnn AAAAAAAAAAAAddddddddddddddddmmmmmmmmmmmmmiiiiiiiiiiiinnnnnnnnnniiiiiiiiiisssssssssssttttttttttrrrrrrrrrration



Key Problems in Ocean Observing

• Tools and techniques are inadequate to understand dynamics of coastal ocean processes

�“We’ve been doing Oceanography the way Darwin did more than a 100 yrs ago. We need new tools and techniques to better characterize our environment, especially the oceans” Marcia McNutt, ex-MBARI President/CEO, now Dir. US Geological Survey/Science Advisor Sec. of Interior

• Often the phenomenon to be observed, cannot be sampled directly

� Use proxy variables (e.g Chl. Fluorescence, backscatter, temperature, salinity)

• Obtaining power & comms. in the water-column is difficult/non-existent

• Synthetic ocean models are poor predictors of change

• Persistence presence necessary to understand spatio-temporal variation

• Sub-sampling the large ocean is expensive and unsustainable

� Ship and labor costs are going up

� Large expeditions with multiple ships/crews are logistically difficult

� Oceanographers prefer “Terra firma”



Why is Ocean Sciences interesting....now?

• Multi-disciplinary in nature–Physicists, Chemists, Mechanical, Civil, Electrical Engineering, Biological,

Physical, Chemical Oceanographers, Environmental Engineering, Computer Science, Economists, Numerical Analysts…

• At a cusp:–Realization that the Oceans are regulators for Global Climate processes

• Synoptic views, not point data can tell us how ocean processes actually work

–Realization of possible anthropogenic influences on our environment and the impact to the oceans

–Advances in sensors, platforms, robotics, control, AI are substantial over ~40 years

–Large science is slowly coming to the fore in Oceanography• E.g NSF funded $350 M Ocean Observatory Initiative (OOI)

–Global/Coastal/Regional Cabled Observatories proposed



Does a domain matter?

T

Space• Power is not such a big deal (payload

scaling is the issue)

• Observability is by and large not such a big deal

• Communication is generally not a problem given observability

• Reachability is definitely an issue

• Launch costs are disproportionally high

• But space has substantial funding

• embedded in the public’s imagination

The Oceans• Power is a big deal; you have to carry

it with you

• Observability is fundamentally lacking

• Communication is a huge problem given lack of observability

• Reachability is not as much an issue given a support vessel

• Launch costs are disproportionally high

• Marine science/engineering has nowhere near as much resources



MBARI:Why is it unique?

• Privately funded

–By the David and Lucille Packard Foundation

–Dependence on Congress is minimal for for its operation

–By its charter < 25% of its internal budget for external (NSF, NASA, ONR) monies

• Long term view of Ocean Science and Engineering is strongly encouraged

• Strong applied technology focus for inter-disciplinary science

• Science, Engineering and Operations are in-house

• Scientists and engineers (mostly) free from undue management & budgetary interference

–MBARI’s influence in the Ocean sciences in the US is disproportionate to its size

• A strong desire to make an impact in Oceanographic sciences by sharing and collaborating with external entities



MBARI: What niche does it occupy?

OceanographyFor the most part

Pasteur’s Quadrant



~200 Staff• Science• Engineering• Operations• Outreach 12 Principal Scientists (Biology, Chemistry, Ecology,

Genetics, Sensors, Platforms, Autonomy)

Deep-dive ROVs, Test TankAUVs & shipsMoorings & buoys

Annual Budget ~$40M

Funded by the David & Lucille Packard Foundation

nggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggns






Annual Budget ~$40M


ngggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggns



The MBARI Fleet

R/V Pt. Lobos: Day boat with 1400m rated ROV

R/V Western Flyer: Deep Ocean with 3600m rated ROV R/V Zephyr: Day boat for AUV operations





Annual Budget ~$40M



The MBARI Fleet





Annual Budget ~$40M



The MBARI Fleet





Annual Budget ~$40M



The MBARI Fleet





Annual Budget ~$40M


Kanna Rajan 2009

Cou

rtes

y C

hris

Sch

olin

, MBA

RI

And then the unique Platforms…:ESP


Kanna Rajan 2009

Cou

rtes

y C

hris

Sch

olin

, MBA

RI

And then the unique Platforms…:ESP



Intermediate Nepheloid Layers (INLs)• Fluid sheets of suspended particulatessss. Originate from the sea floor

through diverse fluid dynamics [McPhee-Shaw 2004].

• Large Horizontal Scales (Kms)

• Small Vertical Scales (meters)

• Patchy

Focus of Interest: Event Response for Coast Ocean Processes



Intermediate Nepheloid Layers (INLs)• Fluid sheets of suspended particulatessss. Originate from the sea floor

through diverse fluid dynamics [McPhee-Shaw 2004].

• Large Horizontal Scales (Kms)

• Small Vertical Scales (meters)

• Patchy




Courtesy John Ryan, MBARI


Algal Blooms• Regular in episodic

appearance in coastal waters everywhere

• Some are harmful; toxins (Demoic acid) generated which cause significant impact to coastal economy and health� Econ. impact ~$75 M ’87-00*� Beach closures� Large mammal deaths� Human health impacts

* D. Anderson, P. Hoagland, Y. Kaoru, and A. White, “Economic impacts from harmful algal blooms

(HABs) in the United States,” tech. rep., Woods Hole Oceanographic Institution Technical Report: WHOI 2000-11, 2000.



Technological Foci

• Adaptation of a mobile robotic asset in response to dynamic signals in the coastal ocean

• Large scale (interruptible) water-column surveys to target specific features of scientific interest

• To track advected patches of water with specific properties



Technological Foci

• Adaptation of a mobile robotic asset in response to dynamic signals in the coastal ocean

• Large scale (interruptible) water-column surveys to target specific features of scientific interest

• To track advected patches of water with specific properties

Sense

Plan Act

World



Distinction between Autonomy & Automation

• Autonomous Systems:

–Sense-Deliberate-Act cycle is intrinsic

–Machine-based decision making that results in behavior that is emblematic of human behavior

• Automation:

–Deliberation is not intrinsic

–Machine-based execution of activities which alleviates human work



Distinction between Autonomy & Automation

• Autonomous Systems:

–Sense-Deliberate-Act cycle is intrinsic

–Machine-based decision making that results in behavior that is emblematic of human behavior

• Automation:

–Deliberation is not intrinsic

–Machine-based execution of activities which alleviates human work

…..is notautonomous



Key Ingredients of an Autonomous System

• FDIR (Fault Diagnosis Isolation and Recovery)

• Deliberative Planning

• Plan Execution

DeliberativePlanning

Fault Diagnosis

CONTROLLED SYSTEM

EXECUTIVE

ThrustGoals

Attitude Point(b)

Engine Thrust (b, 200) Off

Delta_V(direction=b, magnitude=200)

Power

Warm Up

meets met_by

contained_by

contained_by

equals

Sensing Real World information



• Communication is not just a bottleneck; its damn difficult in the water-column!

• Powering instruments/platforms is a serious problem

• Costs effectiveness

– Costs associated with ship based science is increasingly prohibitive • Given funding profiles, autonomy is a strong leverage to squeeze more science out of existing

ocean observing systems

– Need for looking at the next generation of sampling and observation methods

– Platforms are more robust and increasingly capable

– Longer durations mission will need:• Goal-based commanding

• Opportunistic re-planning for adaptability

• Ability to reason about resources

• Onboard fault diagnosis, isolation and recovery (FDIR)

• Reduce the cognitive burden of mission operators

• And to do so efficiently (even at 3am in PIs time-zone)

Why is Autonomy important for the Ocean Sciences?



• Communication is not just a bottleneck; its damn difficult in the water-column!

• Powering instruments/platforms is a serious problem

• Costs effectiveness

– Costs associated with ship based science is increasingly prohibitive • Given funding profiles, autonomy is a strong leverage to squeeze more science out of existing

ocean observing systems

– Need for looking at the next generation of sampling and observation methods

– Platforms are more robust and increasingly capable

– Longer durations mission will need:• Goal-based commanding

• Opportunistic re-planning for adaptability

• Ability to reason about resources

• Onboard fault diagnosis, isolation and recovery (FDIR)

• Reduce the cognitive burden of mission operators

• And to do so efficiently (even at 3am in PIs time-zone)

Why is Autonomy important for the Ocean Sciences?

pabbbllleee

FFDDDDDDDDDDDDDDDDDDIIIIR)

rss

-zzzoooooooooooooonnnnnne)

The Sampling Conundrum1. Finite Time2. Finite Energy

3. Finite Sampling Resources

4. Uncertainty of occurrence of phenomenon

5. Uncertainty of location, size, shape and strength of feature signal



ROVs AUVs Gliders

Platforms: Ocean going Robots



Intellectual/Operational Legacy of our work

Real-Time Execution

FlightH/WFault

MonitorsPlanning Experts (incl.

Navigation)

GroundSystem

RAX Manager

Mode ID and Reconfig

MissionManager

SmartExecutive

Planner/Scheduler

Remote Agent

ScienceActivities

ScienceSupport Tasks

Flight rules, science

constraints, and resource

constraints

EngineeringActivities

Science OperationsWorking Group

EngineeringTeam

Planner Activity Plan

SequenceGeneration

MAPGEN:Mixed-Initiative Planner

Mar

s Exp

lora

tion

Rov

ers (

MER

)

MAPGEN: Mixed-Initiative Activity Planning

– Jan 14th 2004 - ongoing

– Principal Investigator

Remote Agent Experiment (DS1)

– Flew onboard Deep Space 1, May

17-21, 1999

– 65 Million miles from Earth, during

Ballistic Cruise

– One of six principals

– NASA’s 1999 Software of the Year Rem

ote A

gent

on

Dee

p Sp

ace

One

(DS1

)



Lessons Learned from Space Missions

• How (Lessons Learned from DS1 & MER)– Customer training and buy-in is not a luxury

– Requirements change and change often: learn to live with it

– Test what you fly and fly what you test

– Spiral mode of development often works well for model-based approaches to s/w engineering

– All or nothing approach = nothing

• Incremental use of automated approaches is necessary

• Solving an existing problem using a dumb but automated method; ++

• Migrate to more abstract/higher levels of autonomy but slowly.

– Usability issues should not be ignored

– Mission-critical software requires good systems engineering

• Connecting the pieces in the software puzzle

• Software not developed in isolation to its operating environment

– Work-practice ethnography is very useful

• How do people use tools and processes to get the ‘job done’?

DomainModel

ngggSearchEngine

SearchControl

AutonomyLevel

0 100

507525

Mars Exploration Rovers/Science Meeting



Architectural relationships in the AUV domain

Mis

sion

Sop

his

tica

tion

TimeEarlyAUVs

TodaysAUVs

FutureAUVs

Todays Future

SequentialActivities

No in-situ adaptation - requires complex

programming



Architectural relationships in the AUV domainM

issi

on S

oph

isti

cati

on

TimeEarlyAUVs

TodaysAUVs

FutureAUVs

Todays Future



programming

Easy to build but no in-situadaptation - not scalable for

complex missions (a la Brooks’ Subsumption Architectures)

ReactiveArchitectures

lSequentialActivitiesSequential



Architectural relationships in the AUV domain

Mis

sion

Sop

his

tica

tion

TimeEarlyAUVs

TodaysAUVs

FutureAUVs

Todays Future



programming

DeliberativeApproaches

Goal-oriented, capable of re-planning and

adaptation - requires model infrastructure





Ap

ctive



Architectural relationships in the AUV domainM

issi

on S

oph

isti

cati

on

TimeEarlyAUVs

TodaysAUVs

FutureAUVs

Todays Future



programming

DeliberativeApproaches

Goal-oriented, capable of re-planning and

adaptation - requires model infrastructure





Ap

ctive

Distributed Control

High-level multi-level controladaptive sampling

in lossy comm. environments

berative



•Abstraction plays a significant role

• all computation on the robot is not equal

• all effort to generate a solution is not equal

• All entities can be seen as Sense/Plan/Act (SPA) loops

• Functional and temporal scope of computation can and should be exploited

• Partitioning computation should be a necessary and important driver for an agent architecture

• Deliberation and execution are intertwined :

• decision of one actor can be impacted by observations of another

Our Approach for Autonomy

Vehicle

Arm operatorPilot

Scientist



Asset control (commands)

Navigation (where and how to go)N(

Science operation ( payload)

Mission (science objectives, and operational constraints)

Tem

poral Scop

e

go)o)goo)

Functional Scope









Our Approach for Autonomy



T-REX: Teleo-Reactive EXecutive









Our Approach for AutonomySense

Plan Act

World



Partitioning Computation

• Functional scope along the lines of what

• Temporal scope along the lines of when

• What does partitioning buy us?

• “Divide and Conquer” in problem solving/Scalability

• Different search strategies can be encapsulated

• Incremental model development

• Allows targeted commanding at any level of abstraction

• For embedded real-world systems, robustness in failure

• Identical interfaces making software development easier

• Reactor contents are agnostic to the T-REX agent

T-REX

Agent

Iridium interface

Vehicle Control interface

Mission Manager

Skipper

NavigatorScience

Legend :Interface reactor

Planning reactor

Goal flow

Observation flow

Vehicle Functional LayerVehicle Functional LayyyyyyerV

Functional ScopeT

emp

oral Scope

look-ahead�=22hr

latency �=5mins

look-ahead �=5hrlatency �=2mins

look-ahead �=10seclatency �=1sec

look-ahead �=1hrlatency �=10sec



Partitioning Computation

• Functional scope along the lines of what

• Temporal scope along the lines of when

• What does partitioning buy us?

• “Divide and Conquer” in problem solving/Scalability

• Different search strategies can be encapsulated

• Incremental model development

• Allows targeted commanding at any level of abstraction

• For embedded real-world systems, robustness in failure

• Identical interfaces making software development easier

• Reactor contents are agnostic to the T-REX agent

Sense

Plan Act

World

T-REX

Agent

Iridium interface

Vehicle Control interface

Mission Manager

Skipper

NavigatorScience

Legend :Interface reactor

Planning reactor

Goal flow

Observation flow

Vehicle Functional LayerVehicle Functional LayyyyyyerV

Functional Scope

Tem

poral Scop

e

look-ahead�=22hr

latency �=5mins

look-ahead �=5hrlatency �=2mins

look-ahead �=10seclatency �=1sec

look-ahead �=1hrlatency �=10sec



• constraints

• restrictions over the tokens behavior

• temporal or parametric

• Each reactor is composed of timelines:• Internal : represent the state of the world as viewed

by this reactor

• External : a view of a state variable Internal to another reactor

• for each External timeline there’s one and only one reactor that defines it as Internal

• timelines :• A sequence of tokens that describe the evolution of a

state variable

• tokens :• atomic entity describing a predicate that holds over a

temporal scope

• start, duration and end can be described as intervals (e.g. start=[0, 10])

Definitions


Mission Manager

Volume_Survey (x1,y1,xn,yn 10, 40) Goals

Path

Navigation

Command

Idle

Transect(Yo-Yo, Min=10, Max=40, x1,y1,xn,yn)


• constraints

• restrictions over the tokens behavior

• temporal or parametric

• Each reactor is composed of timelines:• Internal : represent the state of the world as viewed

by this reactor

• External : a view of a state variable Internal to another reactor

• for each External timeline there’s one and only one reactor that defines it as Internal

• timelines :• A sequence of tokens that describe the evolution of a

state variable

• tokens :• atomic entity describing a predicate that holds over a

temporal scope

• start, duration and end can be described as intervals (e.g. start=[0, 10])

Definitions


Volume_Survey (x1,y1,xn,yn 10, 40) Goals

Path

Navigation

Command

Idle



• Tokens represent a flexible temporal extant

• Temporal flexibility allows plans to deal with

• “fail operational” modes

• uncertainty in plan execution

• use of Allen Algebra for all computation

• Plans are composed of logical assertions of tokens connected by explicit constraints on timelines

Definitions



Inactive

Volume_Survey (x1,y1,xn,yn 10, 40)

Move_To(Recover) Move_To(Deploy)

Goals

Path

Navigation

Command

At_Surface

At(Start)

Idle

Transect(Straight, Depth=10)


At_Surface

Transect(Straight, Depth=10)

Descend(10)

WayPoint(x1,y1)) Descend(40)

Ascend(10)

Ascend(10)

WayPoint(xn,yn))

Ascend(0)

contained_by

contained_by

met_bymet_by meets

contained_by[0,0][1,100]

Go(x1,y1) Go(xn,yn) Go(x2,y2,Yo-Yo)







Definitions



Inactive Surface Localization

meets meets

meet_by

50

Δd = [40, 50]

time

10 15 25 contains

Obtain GPS fix







Definitions

Teleo-Reactive EXecutive(T-REX)

Functional Layer

eautive

EUROPAPlanner

OPAer

NDDLDomain Description

Language


Reactormodel

At(base)Move

10 120 [121 +inf]

Navigation

IdleMoveTo(x,y)

10 120 [121 +inf]

Commands

At(x,y)At(x',y')At(x",y")

118 119 120 [121 +inf]

Position


• Synchronization is the operation during which

a reactor identifies its current state

• merging observations with expectations

• computes the current value of its Internal

timelines

• The Internal state depends on External timelines

• Cannot synchronize unless the reactors it depends on are synchronized

Sensing : Synchronization


Reactormodel

At(base)Move

10 120 [121 +inf]

Navigation

IdleMoveTo(x,y)

10 120 [121 +inf]

Commands


118 119 120 [121 +inf]

Position


Mission Manager

Navigator

Science

Vehicle Control Interface

Skipper

Syn

ch o

rder





timelines





Reactormodel

At(base)Move

10 120 [121 +inf]

Navigation

IdleMoveTo(x,y)

10 120 [121 +inf]

Commands


118 119 120 [121 +inf]

Position


Mission Manager

Navigator

Science

Vehicle Control Interface

Skipper

Syn

ch o

rder





timelines



• Role of the TREX agent:

• Order reactor synchronization based on their dependencies

• Propagate Internal state updates to the corresponding External timelines




Reactormodel

At(base)

120 [121 131]

Navigation

Idle

120 [121 131]

Commands

At(x,y)At(x',y')

119 120 121

Position

MoveooooooooMoMMMMMMMMMMMMMMM At(Loc1)

1 31131131]]]1 1

Move

[211 294] [294 +inf]

MoveTo(x,y)vevvvvvvvvoooooooooMoMMMMMMMMMMMMMMM

131]1

MoveTo(xloc, yloc) Idle

[211 294]

At(xloc,yloc)

[211 294]

...

• Deliberation is purely internal to the reactor

• find how to alter External states in order to be

in a desired Internal state

• 2 possible triggers for deliberation :

• a new goal on its Internal state

• External observations contradict reactor’s plan leading to plan repair/failure

• Each reactor defines two important parameters :

• latency � : maximum time required for producing a plan

• look-ahead � : how far ahead the reactor needs to look for deliberation

Plan/Deliberate



Reactormodel

At(base)

120 [121 131]

Navigation

Idle

120 [121 131]

Commands

At(x,y)At(x',y')

119 120 121

Position

MoveoooooooooMoMMMMMMMMMMMMMMM At(Loc1)

1 131]1 1

Move

[211 294] [294 +inf]

MoveTo(x,y)vevevvvvvvvvvvooooooooMoMMMMMMMMMMMMMMM Idle

[211 294]

At(xloc,yloc)

[211 294]

...

� �

MoveTo(xloc, yloc)MoveTo(xloc, yloc)MoveTo(xloc, yloc)

• In order to execute its plan the reactor

should be able to request External timelines

to change according to its plan

• This dispatching can be done when a planned

state overlaps the planning window of the owner of this External state variable

• planning window = [�+�exec, �+�exec+�]

• �exec is the execution latency of a reactor including :

• reactor’s deliberation latency (�)

• and the maximum �exec of its External states

• A newly dispatched token becomes a goal of the owner of this External state (i.e. the

one declaring it as Internal)

Act : Dispatching



Reactormodel

At(base)

120 [121 131]

Navigation

Idle

120 [121 131]

Commands

At(x,y)At(x',y')

119 120 121

Position

MoveoooooooooMoMMMMMMMMMMMMMMM At(Loc1)

1 131]1 1

Move

[211 294] [294 +inf]

MoveTo(x,y)vevevvvvvvvvvvooooooooMoMMMMMMMMMMMMMMM Idle

[211 294]

At(xloc,yloc)

[211 294]

...

� �

MoveTo(xloc, yloc)

MoveTo(xloc, yloc)

• In order to execute its plan the reactor

should be able to request External timelines

to change according to its plan

• This dispatching can be done when a planned

state overlaps the planning window of the owner of this External state variable

• planning window = [�+�exec, �+�exec+�]

• �exec is the execution latency of a reactor including :

• reactor’s deliberation latency (�)

• and the maximum �exec of its External states

• A newly dispatched token becomes a goal of the owner of this External state (i.e. the

one declaring it as Internal)

Act : Dispatching



survey(west, north)

pos(x)

Idle

Mission

Manager

Navigator

Exec.

ScienceOp

Path

Path

check-in

Command

Position

Command

Position

� = 1 s

� = 0 s

� = 5 s

� = 1 s

� = 22 hrs

� = 60 s

Illustration

• 3 reactors

• Mission Manager : select and order scientific goals to do in the mission

• Navigator : Waypoint based navigation control including operational constraints (surface every ~30mins for localization, ...)

• Executive : dispatch commands and collect data from the vehicle



survey(west, north)

pos(x)

Idle

Mission

Manager

Navigator

Exec.

ScienceOp

Path

Path

check-in

Command

Position

Command

Position

� = 1 s

� = 0 s

� = 5 s

� = 1 s

� = 22 hrs

� = 60 s

Illustration

• 3 reactors






survey(west, north)

pos(x)

Idle

Mission

Manager

Navigator

Exec.

ScienceOp

Path

Path

check-in

Command

Position

Command

Position pos(x)

Idle

At(X)

surf.

At(X)

Wait

� = 1 s

� = 0 s

� = 5 s

� = 1 s

� = 22 hrs

� = 60 s

Illustration

• 3 reactors






survey(west, north)

pos(x)

Idle

Mission

Manager

Navigator

Exec.

ScienceOp

Path

Path

check-in

Command

Position

Command

Position pos(x)

Idle

At(X)

surf.

At(X)

Wait

Go(west, north, Yo-Yo, ...)

Done

� = 1 s

� = 0 s

� = 5 s

� = 1 s

� = 22 hrs

� = 60 s

Illustration

• 3 reactors






survey(west, north)

pos(x)

Idle

Mission

Manager

Navigator

Exec.

ScienceOp

Path

Path

check-in

Command

Position

Command

Position pos(x)

Idle

At(X)

surf.

At(X)

Wait


Done

At(north)At(east)

� = 1 s

� = 0 s

� = 5 s

� = 1 s

� = 22 hrs

� = 60 s

Illustration

• 3 reactors






survey(west, north)

pos(x)

Idle

Mission

Manager

Navigator

Exec.

ScienceOp

Path

Path

check-in

Command

Position

Command

Position pos(x)

Idle

At(X)

surf.

At(X)

Wait


Done

At(north)At(east)

Going

Go(X, east, ...)

� = 1 s

� = 0 s

� = 5 s

� = 1 s

� = 22 hrs

� = 60 s

Illustration

• 3 reactors






survey(west, north)

pos(x)

Idle

Mission

Manager

Navigator

Exec.

ScienceOp

Path

Path

check-in

Command

Position

Command

Position pos(x)

Idle

At(X)

surf.

At(X)

Wait


Done

At(north)At(east)

Going

Go(X, east, ...)

x'

� = 1 s

� = 0 s

� = 5 s

� = 1 s

� = 22 hrs

� = 60 s

Illustration

• 3 reactors






survey(west, north)

pos(x)

Idle

Mission

Manager

Navigator

Exec.

ScienceOp

Path

Path

check-in

Command

Position

Command

Position pos(x)

Idle

At(X)

surf.

At(X)

Wait


Done

At(north)At(east)

Going

Go(X, east, ...)

x'

x'

� = 1 s

� = 0 s

� = 5 s

� = 1 s

� = 22 hrs

� = 60 s

Illustration

• 3 reactors






survey(west, north)

pos(x)

Idle

Mission

Manager

Navigator

Exec.

ScienceOp

Path

Path

check-in

Command

Position

Command

Position pos(x)

Idle

At(X)

surf.

At(X)

Wait


Done

At(north)At(east)

Going

Go(X, east, ...)

x'

x'

� = 1 s

� = 0 s

� = 5 s

� = 1 s

� = 22 hrs

� = 60 s

Illustration

• 3 reactors






survey(west, north)

pos(x)

Idle

Mission

Manager

Navigator

Exec.

ScienceOp

Path

Path

check-in

Command

Position

Command

Position pos(x)

Idle

At(X)

surf.

At(X)

Wait


Done

At(north)At(east)

Going

Go(X, east, ...)

x'

x'

plan horizon

Go(X, east, ...)

� = 1 s

� = 0 s

� = 5 s

� = 1 s

� = 22 hrs

� = 60 s

Illustration

• 3 reactors






survey(west, north)

pos(x)

Idle

Mission

Manager

Navigator

Exec.

ScienceOp

Path

Path

check-in

Command

Position

Command

Position pos(x)

Idle

At(X)

surf.

At(X)

Wait


Done

At(north)At(east)

Going

Go(X, east, ...)

x'

x'

plan horizon

Go(X, east, ...)

descend waypoint

nav

� = 1 s

� = 0 s

� = 5 s

� = 1 s

� = 22 hrs

� = 60 s

Illustration

• 3 reactors






survey(west, north)

pos(x)

Idle

Mission

Manager

Navigator

Exec.

ScienceOp

Path

Path

check-in

Command

Position

Command

Position pos(x)

Idle

At(X)

surf.

At(X)

Wait


Done

At(north)At(east)

Going

Go(X, east, ...)

x'

x'

plan horizon

Go(X, east, ...)

descend waypoint

nav

x''

x''

H

descend� = 1 s

� = 0 s

� = 5 s

� = 1 s

� = 22 hrs

� = 60 s

Illustration

• 3 reactors






survey(west, north)

pos(x)

Idle

Mission

Manager

Navigator

Exec.

ScienceOp

Path

Path

check-in

Command

Position

Command

Position pos(x)

Idle

At(X)

surf.

At(X)

Wait


Done

At(north)At(east)

Going

Go(X, east, ...)

x'

x'

plan horizon

Go(X, east, ...)

descend waypoint

nav

x''

x''

H

descend

survey(west, north)

pos(x)

Idle

Mission

Manager

Navigator

Exec.

ScienceOp

Path

Path

check-in

Command

Position

Command

Position pos(x)

Idle

At(X)

surf.

At(X)

Wait


Done

At(north)At(east)

Going

Go(X, east, ...)

x'

x'

Go(X, east, ...)

descend waypoint

nav

x''

x''

descendddededededededede

ddededededededede

� = 1 s

� = 0 s

� = 5 s

� = 1 s

� = 22 hrs

� = 60 s

Illustration

• 3 reactors






survey(west, north)

pos(x)

Idle

Mission

Manager

Navigator

Exec.

ScienceOp

Path

Path

check-in

Command

Position

Command

Position pos(x)

Idle

At(X)

surf.

At(X)

Wait


Done

At(north)At(east)

Going

Go(X, east, ...)

x'

x'

Go(X, east, ...)

descend waypoint

nav

x''

x''

descendddedededededededede

ddededededededede

� = 1 s

� = 0 s

� = 5 s

� = 1 s

� = 22 hrs

� = 60 s

Illustration

• 3 reactors





“informal”description

Dataacquisition

ised

ctorsers in

ds of

rs

rk

Neighbourhood

Landscapby c

Lands

Sort cellhighest counneighbourhcell, where

inversely prRepeat unt

Clustering


Nepheloid layer :suspended material

logs clusters

Nepheloidlayerdetected !

OnOOnnOnO bobobobob araraaa dddddclclasasassisisiiii ififififififififiififififififfffff rerrrrrererrrerrererereeeeeee

Feature(s)detection & sampling

• Need a generic way to identify features from sensor data in real-time

• Idea : learn a model from previous missions

• Technique : Use clustering to extract information from raw-data

Integrating HMM’s with Planning


Outside InsideTrigger

Sampler

c1 c2 cn...

.99 .98 .98

.01

.01

.01

.02

.26 0.02

.13 .05.01

.3.1 .05

Observations

States

StateTransition

Probabilities

ObservationProbabilities

.05


Integrating HMM’s with Planning: Next Steps

Objective: To provide the context of the sequence of observations to enable state estimation


Outside InsideTrigger

Sampler

c1 c2 cn...

.99 .98 .98

.01

.01

.01

.02

.26 0.02

.13 .05.01

.3.1 .05

Observations

States

StateTransition

Probabilities

ObservationProbabilities

.05

Human Labeling

SOM Learning

Label Propagation

HMM Learning

Clusters

Phase I: Processing Raw Sensor Data

I. Partial Labeling II. Semi-supervised clustering

III. Cluster Extraction

....

C1

C2

C3

Cn

Outside

Boundary

Inside

Centroid

Raw Sensor Data

Phase II: Modeling Sensor Dynamics

I. Partial Labeling II. Semi-supervised clustering

III. Cluster Extraction

DataD

Phase II: Modeling Sensor Dynamics

V. Probabilistic Estimation

Raw Sensor aRaR w Sensor Saa

IV. Label Propagation

.....

O B I C

C1 C2 C3 Cn

HMM

SOM Clusters


Integrating HMM’s with Planning: Next Steps

Objective: To provide the context of the sequence of observations to enable state estimation

F. Py, S. Celorrio, K. Rajan in review



GimbaledTail-cone

CTD

Vehicle Computers

Batteries

Gulper, USBL Transponder

GPS

MBARI’s CTD* AUV

Speed 4knots Length: 4.3 m (typical), Diameter: 0.53m Endurance ~22hrs

Depth rating 4500m/typical 1000m Missions: Upper water-column, time-series, engineering testing

Frequency of use: 4 days/week

CPU: 1 300 Mhz PC-104 Functional Layer/QNX, 367 Mhz 1 EPIC EPX-GX500/Fedora RH7

Launch/Recovery: R/V Zephyr

Cost: ~ $1.5 M

HydroScat/HS2, LOPC, LIZT

* Conductivity, Temperature, Depth



Mission

Script

Creation

Main VehicleComputer

(MVC)

Data

Analysis

MissionScript

Data log

Real-time

Interaction

Tracking &

Navigation

?

What was done before

T-REX’s Embodiment



Main VehicleComputer

(MVC)

Data

Analysis

AdaptiveMission

Controller

Data log

Real-time

Interaction

Tracking &

Navigation

?

Adaptor

Adaptor

MissionScript

AdaptorAdaptor

What we can do now




Gulper : water sampling device. Needs AUV adaptability Autonomy

computer stack

Basic AUV control stack



p(INL)

S1

S2

S3

S4

S5


Results:Volume Surveys

INL mapping and sampling survey around M0

• Ability to detect in-situ INLs (Intermediate Nepheloid Layers)

• Reactively takes water sample when needed

• Change the resolution of the survey

• Mission duration up to 6 hours

• Average CPU load ~20% including execution control, planning and state estimation

Location: Off M0 Monterey Bay Nov 2009


p(INL)

S1

S2

S3

S4

S5


Middle/High res. (500~250m)










p(INL)

S1

S2

S3

S4

S5



Low res. (~750m)










p(INL)

S1

S2

S3

S4

S5



Low res. (~750m)

samples taken nearby the strongest INL signature











Adaptive Sampling: The Role of Sample Utilityp(IN

L)

S1

S2

S3

S4

S5

• Problem: To determine when and where samples are to be acquired for maximum information gain, while respecting spatial and temporal sampling constraints.

• Proposed solution: Use utility functions as a means to inform plan-time decisions

• Weakly Informed: tracking history - past samples

• Mission Aware: plan analysis - time/traverses remaining

• Feature Aware: a priori knowledge of feature

Purely Reactive



Adaptive Sampling: The Role of Sample Utility

p(INL)

S1

S2

S3

S4

S5






Anticipatory (Distance)



Adaptive Sampling: The Role of Sample Utilityp(IN

L)

S1

S2

S3

S4

S5






Anticipatory (Samples)



Adaptive Sampling: The Role of Sample Utility

p(INL)

S1

S2

S3

S4

S5






• Given certain space and a set K of points in this space, the Voronoi region for any ki � K contains all the points of the space that are closer to ki than to any other kj.

• Good indicator of spatial distribution of samples

• But spatial distribution is not all that matters:

• large areas could correspond to small probability areas...

• Solution: integral of the probability of all the points contained in a certain Voronoi region

4069.5

4070

4070.5

4071

4071.5

4072

4072.5

592.5 593 593.5 594 594.5 595 595.5 596

2008.315.02

Coefficient of variation: 1.3



Front identified

Data sent to shore continuously during transect

Results: Following Fronts

Location: Off M0 Monterey BayJuly 2009



PR2 control

European Space Agency rover testbed

• T-REX is a general purpose Open Source framework (in Google Code)

– Anything that can be described as Sense/Plan/Act can be integrated as a reactor

• T-REX is being used to coordinate multiple planners for advanced service robot control (http://www.willowgarage.com):

• T-REX will be the core controller for ESA rover testbed:– IP-CNR (Italy): APSI Planner

– LAAS (France) : GenoM functionnal layer

– Verimag (France) : BIP compositional Verification

Results: Other TREX Instantiations



Next Steps: CANON- A Controlled, Agile and Novel Observing Network




• Mixed-Initiative control from shore/ship

• Loosely coupled multiple heterogenous vehicles

• Each vehicle has onboard plan synthesis capability

• Limited information exchange• Human as a separate ‘agent’

brings substantial cognitive capability

• Synoptic views are generated by the fusion of disparate map data, combined with onboard autonomy



Decision Support System (DSS)

forCANON



Decision Support System (DSS)

forCANON


Autonomy: Outer to Inner Space Kanna Rajan, MBARI 2010Friday, December 10, 2010

595.5 596 596.5 597 597.5 598 598.5 599 599.5 600

4082

4082.5

4083

4083.5

4084

4084.5

4085

4085.5

4086

4086.5

Easting (km)

Nor

thin

g (k

m)

DrifterDorado 181Dorado 180


t0t1

t2t3

t4

r =Sdrifter

Sauv cos(pitchyoyo)

r =Sdrifter

1.5 cos([20, 30])=

Sdrifter

[1.30, 1.41]

d

r*d

P

IX

initial path

drifter displacementprojected path

−→IP = k.

−−−−−→Sdrifter


Ch

l. F

luo

resc

ence

Oct

201

0


595.5 596 596.5 597 597.5 598 598.5 599 599.5 600

4082

4082.5

4083

4083.5

4084

4084.5

4085

4085.5

4086

4086.5

Easting (km)

Nor

thin

g (k

m)



t0t1

t2t3

t4

t0

t1

t2

t3

t4

r =Sdrifter

Sauv cos(pitchyoyo)

r =Sdrifter

1.5 cos([20, 30])=

Sdrifter

[1.30, 1.41]

dr*d

P

IX

initial path


−→IP = k.



Ch

l. F

luo

resc

ence

Oct

201

0


595.5 596 596.5 597 597.5 598 598.5 599 599.5 600

4082

4082.5

4083

4083.5

4084

4084.5

4085

4085.5

4086

4086.5

Easting (km)

Nor

thin

g (k

m)



t0t1

t2t3

t4

t0

t1

t2

t3

t4

r =Sdrifter

Sauv cos(pitchyoyo)

r =Sdrifter

1.5 cos([20, 30])=

Sdrifter

[1.30, 1.41]

d

r*d

P

IX

initial path


−→IP = k.



Ch

l. F

luo

resc

ence

Oct

201

0


595.5 596 596.5 597 597.5 598 598.5 599 599.5 600

4082

4082.5

4083

4083.5

4084

4084.5

4085

4085.5

4086

4086.5

Easting (km)

Nor

thin

g (k

m)



t0

t1

t2

t3

t4

r =Sdrifter

Sauv cos(pitchyoyo)

r =Sdrifter

1.5 cos([20, 30])=

Sdrifter

[1.30, 1.41]

dr*d

P

IX

initial path


−→IP = k.



Ch

l. F

luo

resc

ence

Oct

201

0


Autonomy: Outer to Inner Space Kanna Rajan, MBARI 2010Friday, December 10, 2010

Autonomy: Outer to Inner Space Kanna Rajan, MBARI 2010Autonomy: Outer oooooooooooooooooooooooooooooooooooooooooooooooooto to InInIn InnInnInInnnInInIn InInInnnnnInInnInInInnnnInnInInIn In InInInIn In In InnInnInInInIn InInIn InInInInInIIIIIIIIIIIIIIIIIII ererererernenenennnnnnnnnnnnnnnnnnnnnnnnnnn SpSpppSpSpSppSpSpppppppSpSppppSppSpppSppppppppSpppppSpSpppppppppppppSppSpppppSpSpSpSppSppppSp SpSpSpSppp SpSpppppppppSpppSpSpp SpSSSSSSSSSSSSSSSSSSSSSS SSSSSSSSSSSS aceaceceeaceaceaceeaceeeeaceaceacacacacacacacacacaccacaccacccaaaaaaaaaaaaaaaa KaKaKaKaKaKaKaKaKaKaKaKaKaKaaaaaaKK nnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnn aaaaaaaaaaaaaaaaaa aaaaaRaaRaRaRaRaRaRaaRaRaRaRaaRaRaRaaaRaRaaaaRaRaRaRaRaaRaRaRaaaaaaRaRaaaaRaRaRaRaaaaRaaaaaaRaaRaaRaRRaRRRRRRRRRRRRRRRRRRRR jjajjajajajajajajaaaajajajajajajjajaaajajajajajajajajajajajajaajajajjajajajajajajajajajaajajajjaajajajajajjajajajajjaajjaajajajajjjjajajajajjjajajjajajajajjjjjjjjjjjjjjjjjjjjjj nn,n,n,nn,n,n,,,n,n,n,n,n,n,n,nnn,nn,n,nnnnnnnnnnnnnnnnnn BBMBBBBBBBMBMBBBMBBMBBBMBBMBBMBMBBBMBBMBMBBBBBBBBBBBBBBBBBBMBMBMBMBMBMBMBMBBMBMBBBBMBMBMBMBMBMMMMMMMMMMMM ARARARARARARARARARRARRRARARARRARRRRARRARRARRRRARARARARRRARARRRRARRARRRRARARARARARRARARRRARARARARRARRARRARRRRRARRRARRRRRRRRRRARARRARARRARARRRRARARARARARARRAARARARARARARAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAA IIIIIIIIIIIIIIIIIIIIIIIIIIIIIIIIIIIIIIIIIIIIIIIIIIIIII 22222 22222 2 22222222222222222222222222222222222 11110101010101101011111101101011101110101101011101110111010100000000000000000000 000000000000000000000000000000000000000000000000000000000Friday, December 10, 2010


Composite of tracks of all CANON assetsfor sampling bloomsOct 9th - 22nd 2010



• The complex coastal ocean requires the ability to sample precisely

–Ocean models are imprecise and require substantial hand-holding

–Embodied robots with statistical models coupled to control add scientific value

• We have demonstrated a formal framework for partitioning a complex control problem into multiple Sense-Plan-Act control loops

• Strong execution semantics

• Each reactor becomes the executive of the reactors depending on its state

• Coupled state estimation allows modeling dynamic features in the coastal ocean

• Future directions:

• shore/ship side DSS

• experiment further system scalability/flexibility

• diverse solvers for deliberation

• multi-vehicle shore side control under limited (lossy) communications

Concluding Remarks



References

• J.Das, F. Py, , T. Maughan, J Ryan , K. Rajan & G. Sukhatme,, “Simultaneous Tracking and Sampling of Dynamic Oceanographic Features with Autonomous Underwater Vehicles and Lagrangian Drifters”, Accepted, Intnl. Symp. on Experimental Robotics (ISER), N. Delhi, India, Dec 2010.

• S. Jiménez, F. Py & K. Rajan, “Learning Identification Models for In-situ Sampling of Ocean features” Working notes of the RSS'10 Workshop on Active Learning for Robotics. Robotics Systems Sciences, Spain. 2010

• J. Das, K. Rajan, S. Frolov, J. Ryan, F. Py, D. Caron & G. Sukhatme, “Towards Marine Bloom Trajectory Prediction for AUV Mission Planning” ICRA, May 2010, Anchorage

• Ryan, J.P., Johnson, S., Sherman, A., Rajan, K., Py, F., Paduan, J., Vrijenhoek, R.,“Intermediate nepheloid layers as conduits of larval transport” Limnology & Oceanography: Methods, 2010.

• Py. F, K. Rajan, C. McGann “A Systematic Agent Framework for Situated Autonomous Systems”, AAMAS10, Toronto

• Rajan, K, Py, F, McGann, C., Ryan, J., O’Reilly, T., Maughan, T., Roman, B., “Onboard Adaptive Control of AUVs using Automated Planning and Execution”, Unmanned Underwater Systems Technology conference (UUST), New Hampshire, August 2009.

• McGann, C., Py, F., Rajan, K., Thomas, H. Henthorn, R., McEwen, R. “A Deliberative Architecture for AUV Control”, ICRA, Pasadena, 2008.

• McGann, C., Py, F., Rajan, K., Ryan, J. Henthorn, R. “Adaptive Control for Autonomous Underwater Vehicles”, AAAI, Chicago, 2008.

• McGann, C., Py, F., Rajan, K., Thomas, H. Henthorn, R., McEwen, R. “Preliminary Results for Model-Based Adaptive Control of an Autonomous Underwater Vehicle”, ISER, Athens, 2008.

• McGann, C., Py, F., Rajan, K., Thomas, H. Henthorn, R., McEwen, R. “Automated Decision Making For a New Class of AUV Science” Proc. American Society of Limnology and Oceanography, ASLO/Ocean Sciences, Orlando, 2008.

• Fox M., Long, D., Py, F., Rajan, K., Ryan, J. “In-situ Analysis for Intelligent Control” Proc. IEEE/OES OCEANS Conference, Aberdeen, Scotland, 2007.

• Bellingham, J, K. Rajan “Robotics in Remote and Hostile Environments” Science cover article, v318 Nov 2007.

• Web Site: http://www.mbari.org/autonomy/

g fdoooooooooooooooodooooooooooooooooooooo, 202020200020002000200202020200200200202020200020200000020000020000000020200000000202222222222222222222222222222222222222222222222222222222222 08000000000000000000000000000000000000000000000000000000 .

c. IEEIEEEEEEIEEEEEEEEEEIEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEIEEEEEEEEEEEIEEEEEEEEEEEEEEEEEEEEIIIIIIIIIIIIIIIIIIIIIIIIIIIIIII EE/OES OCEANS Conferenc

clllle,e,,,,,,,,,,,,,,,,,,,,,,,,,,,,e,,,,,e,,,,,,,eeeeeeeeeeeeeeeeeeeeeeeee vvvvvvvvvvvvvvvvvvvvvvvvvvvvv 3131131311111311131113131311311131131313333333333333333333333333333333 8 8 8 8 8888888888888888888888888888888 NoNoNoNoNoNoooooNooNoNoNoNoNoNoNoNooooooNooooooNoooooNoooooooooNNNNNNNNNNNNNNNNNNNNNNNNNNNNNNNNNNNNN v v v vvvvvvvvvvvvvvvvvvvv 2020202002002002000222222222222222222222222 00700000000000 .



Backup Material



0

Goto(W2) Surface(W2) Goto(W3) Terminate(W3)

0 120 220 260 480 500

Program P

Traditional Approaches for Dispatching Timed Programs

00000 1222

Goto(W1)



0


0 120 220 260 480 500

Program P


00000 1222

Goto(W1)

handleStart(“Goto(W1)”)


handleStart(“Surface(W2)”)


handleStart(“Terminate(W3)”)



0


0 120 220 260 480 500

Program P


00000 1222

Goto(W1)

• Does not model uncertainty in action and environmental response

• Is brittle in execution• Is not responsive to environmental

feedback








0


[0, 5] [110, 120] [220,225] [260,290] [480, 520] [500, 600]

Program P

A less Traditional Approach of Dispatching Timed Programs

00[0 5]00 [11000

Goto(W1)



0


[0, 5] [110, 120] [220,225] [260,290] [480, 520] [500, 600]

Program P


000[0 55]00 [11000

Goto(W1)








0


[0, 5] [110, 120] [220,225] [260,290] [480, 520] [500, 600]

Program P


00[0 5]00 [11000

Goto(W1)

• Flexible start/end times can deal with some measure of uncertainty

• Can be robust to fail-operational modes of recoveryhandleStart(“Goto(W2)”)







What is Constraint-Based Planning?• CBP systematizes the use of

– relationships (real or implied)

– provides a sound mathematical framework for representing axiomatic equations

– encapsulates a systematic approach to evolving state

• State-Variable based approaches further decompose the planning problem

– timelines describe state evolution for pre-specified sub-systems

– by merging the representations of time with state-variables we can use systematic representations to depict state evolution realistically while still being discrete

• The essential elements of timeline based CBP are:

– state variables: describe evolution of state or a single thread of execution in a concurrent system

– tokens: describes a procedure which instantiates and maintains state

– timepoints: instances of time when a significant change in state is likely to occur

– constraints: explicit representation of relationships between entities within and across timelines


50

time 10 45 60

Waypoint_Yo-Yo

State Variable Token

TimepointsAutonomy: Outer to Inner Space Kanna Rajan, MBARI 2010














Waypoint

Waypoint

meets meets

meet_by

50

time 10 45 60

Waypoint

Δd = [40, 50]

Waypoint_Yo-Yo (?threshold)
















X =[10 20] Y=[30 100]

[30 38]

X=[10 20] Y=[30 100]

38

-30

STN

Distance Graph

CONVERSION: Y-X � [30 38] � Y-X <= 38 ^ X-Y <= -30

Origin={0}

20 -10 -30 100

Upper Bound on Path Length: 20 + 38 -30 = 28

�Fast Incremental Propagation�Detect Inconsistency with a negative cycle�Backtrack free search if there is a feasible solution

Foundational Representation: Simple Temporal Networks



X =[10 20] Y=[30 100]

[30 38]

X=[10 20] Y=[30 100]

38

-30

STN

Distance Graph

CONVERSION: Y-X � [30 38] � Y-X <= 38 ^ X-Y <= -30

Origin={0}

20 -10 -30 100

Upper Bound on Path Length: 20 + 38 -30 = 28

�Fast Incremental Propagation�Detect Inconsistency with a negative cycle�Backtrack free search if there is a feasible solution

Foundational Representation: Simple Temporal Networks



How does one build Timelines?



Waypoint(Depth, Heading)

Wait_At_Surface

Waypoint_Yo-Yo(Min_Depth, Max_Depth, Heading)

[0, 1800]

meets

[0,∞]

[0,∞]

I





Wait_At_Surface


[0, 1800]

meets

[0,∞]

[0,∞]

I





Wait_At_Surface


[0, 1800]

meets

[0,∞]

[0,∞]

I


Wait_At_Surface


[0, 1800]

meets

[0,∞]

[0,∞]

II





Wait_At_Surface


[0, 1800]

meets

[0,∞]

[0,∞]

I


Wait_At_Surface


[0, 1800]

meets

[0,∞]

[0,∞]

II


Wait_At_Surface

Waypoint_Yo-Yo(Min_Depth,

Max_Depth, Heading)

[0, 1800]

meets

[0,∞][0,∞]

III





Wait_At_Surface


[0, 1800]

meets

[0,∞]

[0,∞]

I


Wait_At_Surface


[0, 1800]

meets

[0,∞]

[0,∞]

II


Wait_At_Surface

Waypoint_Yo-Yo(Min_Depth,

Max_Depth, Heading)

[0, 1800]

meets

[0,∞][0,∞]

III



meets meets

meet_by time


Wait_At_Surface

meet_by

IV



Plan Formulation Example (NASA’s Deep Space 1)

Max_ThrustIdle IdleTimer

Accumulation

SEP Action

SEP_Segment

Th_Seg

Standby

Th_Sega Th_Seg Th_SegIdle_Seg Idle_Seg

Thr_Boundary

Th_SegTh_Seg

Op_NAV_Window

No_Accum

contained_by

Final_Accum

ends




Accumulation

SEP Action

SEP_Segment

Th_Seg

Standby


Thr_Boundary

Th_SegTh_Seg

Accumulate

meets

Op_NAV_Window

No_Accum

contained_by

Final_Accum

ends





Accumulation

SEP Action

SEP_Segment

Thrust_Segment

Standby

Idle_Seg Idle_Seg

Thr_Boundary Accumulate

meets

Final_Accum

ends

Op_NAV_Window

No_Accum

contained_by

Starting_up





Accumulation

SEP Action

SEP_Segment

Thrust_Segment

Standby

Idle_Seg Idle_Seg


meets

Final_Accum

ends

contained_by

Op_NAV_Window

No_Accum

contained_by

Starting_up





Accumulation

SEP Action

SEP_Segment

Thrust_Segment

Standby

Idle_Seg Idle_Seg


meets

Final_Accum

ends

contained_by

Thrusting

equal

Op_NAV_Window

No_Accum

contained_by

Starting_up





Accumulation

SEP Action

SEP_Segment

Thrust_Segment

Standby

Idle_Seg Idle_Seg


meets

Final_Accum

ends

contained_by

after

Starting_Up

Thrusting

equal

contained_by

met_by

Pointing

contained_by Op_NAV_Window

No_Accum

contained_by




Max_ThrustIdle Idle

Poke

Timer

Attitude

Accumulation

SEP Action

SEP_Segment

Th_Seg

contained_by

equals equalsmeets

meets

contained_by

Start_Up Start_UpShut_Down Shut_Down

Thr_Boundary

Thrust ThrustThrustThrustStandby Standby Standby


Accum_NO_Thr Accum_ThrAccum_Thr Accum_ThrThr_Boundary

contained_by

CP(Ips_Tvc) CP(Ips_Tvc) CP(Ips_Tvc)

contained_by

Th_SegTh_Seg




What about modeling?



class Front extends AgentTimeline { predicate None {}

predicate Info { float temperature; DEPTH depth; DEPTH mapDepth; eq(duration, 1); }

predicate Begin {}

predicate Search { float temperature; DEPTH depth; NORTHING xTo; EASTING yTo; bool frontDetected; NORTHING xFront; EASTING yFront; } predicate Map { NORTHING xFront; EASTING yFront; float gulpSeparation; } }


Front::Map { float distanceToFrom, distanceToTo; bool closeToFrom, closeToTo; float x1, y1, x2, y2;

contained_by(FrontTracker.Track trk);

calcDistance(distanceToFrom, trk.xFrom, trk.yFrom, xFront, yFron testLEQ(closeToFrom, distanceToFrom, trk.moveIncrement); calcDistance(distanceToTo, trk.xTo, trk.yTo, xFront, yFront); testLEQ(closeToTo, distanceToTo, trk.moveIncrement);

if( closeToFrom==false ) { float ratioFrom, dxToFrom, dyToFrom, dx2, dy2; addEq(trk.xFrom, dxToFrom, xFront); addEq(trk.yFrom, dyToFrom, yFront); addEq(x2, dx2, xFront); addEq(y2, dy2, yFront); mulEq(trk.moveIncrement, ratioFrom, distanceToFrom); mulEq(dx2, ratioFrom, dxToFrom); mulEq(dy2, ratioFrom, dyToFrom); } if( closeToFrom==true ) { eq(x2, trk.xFrom); eq(y2, trk.yFrom); }..............



class Front extends AgentTimeline { predicate None {}

predicate Info { float temperature; DEPTH depth; DEPTH mapDepth; eq(duration, 1); }

predicate Begin {}

predicate Search { float temperature; DEPTH depth; NORTHING xTo; EASTING yTo; bool frontDetected; NORTHING xFront; EASTING yFront; } predicate Map { NORTHING xFront; EASTING yFront; float gulpSeparation; } }

Allen Algebra*


Front::Map { float distanceToFrom, distanceToTo; bool closeToFrom, closeToTo; float x1, y1, x2, y2;

contained_by(FrontTracker.Track trk);

calcDistance(distanceToFrom, trk.xFrom, trk.yFrom, xFront, yFron testLEQ(closeToFrom, distanceToFrom, trk.moveIncrement); calcDistance(distanceToTo, trk.xTo, trk.yTo, xFront, yFront); testLEQ(closeToTo, distanceToTo, trk.moveIncrement);

if( closeToFrom==false ) { float ratioFrom, dxToFrom, dyToFrom, dx2, dy2; addEq(trk.xFrom, dxToFrom, xFront); addEq(trk.yFrom, dyToFrom, yFront); addEq(x2, dx2, xFront); addEq(y2, dy2, yFront); mulEq(trk.moveIncrement, ratioFrom, distanceToFrom); mulEq(dx2, ratioFrom, dxToFrom); mulEq(dy2, ratioFrom, dyToFrom); } if( closeToFrom==true ) { eq(x2, trk.xFrom); eq(y2, trk.yFrom); }..............

*J. Allen, “Towards a General Theory of Action and Time,” Artificial Intelligence, vol. 23(2), p. 123154, 1984



So how does one execute a (temporal) plan?



So how does one execute a (temporal) plan?MaI I

Po

Ti

At

Acc

SE

SEPT

cont

e emm

cont

St StSh Sh

Thr

T TTTSt St St

T T ThId Id

Accu AccAcc AccThr

cont

CP( CP( CP(cont

TT

Tem

po

ral P

lan




Execution Engine

Plan(s)

Commands

MaI I

Po

Ti

At

Acc

SE

SEPT

cont

e emm

cont

St StSh Sh

Thr

T TTTSt St St

T T ThId Id

Accu AccAcc AccThr

cont

CP( CP( CP(cont

TT

Tem

po

ral P

lan




Execution Engine

Plan(s)

Environmental State/ObservationsCommands

State Update

MaI I

Po

Ti

At

Acc

SE

SEPT

cont

e emm

cont

St StSh Sh

Thr

T TTTSt St St

T T ThId Id

Accu AccAcc AccThr

cont

CP( CP( CP(cont

TT

Tem

po

ral P

lan




Execution Engine

Plan(s)


State Update

With

one cav

eat

• Use a single temporal database to enforce planning constraints & for dispatching • Solves dispatachability issues

• Retains causal structure for “reasoning” by the executor

• Execution engine accesses the same plan causal structure and uses same inference mechanism as Planner to the database

MaI I

Po

Ti

At

Acc

SE

SEPT

cont

e emm

cont

St StSh Sh

Thr

T TTTSt St St

T T ThId Id

Accu AccAcc AccThr

cont

CP( CP( CP(cont

TT

Tem

po

ral P

lan




Execution Engine

Plan(s)


State Update

With

one cav

eat




Teleo-Reactive EXecutive(T-REX)

Functional Layer

MaI I

Po

Ti

At

Acc

SE

SEPT

cont

e emm

cont

St StSh Sh

Thr

T TTTSt St St

T T ThId Id

Accu AccAcc AccThr

cont

CP( CP( CP(cont

TT

Tem

po

ral P

lan


NDDLDDEUROPA



Execution Engine

Plan(s)


State Update

With

one cav

eat




ROTeleo-Reactive

EXecutive(T-REX)

Functional Layer

MaI I

Po

Ti

At

Acc

SE

SEPT

cont

e emm

cont

St StSh Sh

Thr

T TTTSt St St

T T ThId Id

Accu AccAcc AccThr

cont

CP( CP( CP(cont

TT

Tem

po

ral P

lan


autonomy: from deep space to inner space - ieee entity … · · 2012-01-0912 principal...

Documents