m.

v-_

NASA

Techmcal

Memorandum

86847

Discovery of the Kalman Filter as a Practical Tool for Aerospace and Industry
Leonard A. McGee and Stanley F. Schmidt
:MASA-_N-6bS_7) r, ISCOVERY OF IHE K&L_&M _ILTS_ AS A PBACTICA '_CCL EC_ At.ROSPACE INL_Ub:_Y {l_,a._A) _/4 1: HC A_]2/MI; A0l CSCL

AND 17G Ui_c_3_ G3/O_ 0_9u5

November

1985

,,;

_cc_sDe]'...._'

kEJ

NASA

Technical

Memorandum

86847

Discovery of the Kalman Filter as a Practical Tool for Aerospace and Industry
Leonard Stanlev A. McGee, Ames Research Center, Moffett Associates, Field, Inc., California Mountain View, California F. Schmidt, Analytical Mechanics

Nuvember

1985

Nal_onat Space

Aeronautics Administration

and

Ames Research Center

Moffetl F_eld Calrfornra 94035

DISCOVERY

OF

THE

KALMAN

FILTER

AS

A PRACTICAL

TOOL

FOR

AERCSPACE

AND

INDUSTR?

Leonard

A. McGee Ames

and

Stanley Center

F.

Sc,hmidt*

Research

SUMMARY

The lyzing bility sule. ers after at

Kal_n

filter

in

its

various of

forms

has

become problems.

fundamental The early the first 1960s Apollo led

tool

for

anaknown

and

solving was for

a broad at

class Ames

estimation and

publicly during space the

application This Ames its

made

NASA

Research

Center sequence discovery Th Kalm_'s The

in the of of of

feasicapresearch-

studies

cifcumlinear recounts Center into neces3ary application is often the to the to

navigation furtunate the early

control

paper

events the

which Kalman

Research

filter

shortly and tool Kalman refora

introduction that were

literature. transform

sclentific work

breakthroughs into a useful extended Kalman the to

mulations specific as it filter's

for

aerospace known, gained on of length. the in the use

are still

described. referred in the and

resulting as the co,rnunity, led

filter, As of the

is now

to simply

filter. problems

in popularity sn_li filter The work _paceborne

scientific

implementation formulation ter word

airborne numerical

computers difficulties formulation and the

a "square-root" with compu-

to overcome that led to

associated is also

this

new

discussed, which was

including conducted formuiations industry.

first 1972. of the

airborn Since Kalman

_ computer then the

implementation of rapidly

flight

test and the

applications have grown

extended

square-root aerospace

filter

throughout

INTRODUCTION

In Filtering bution the

1960, and the

Dr.

Kalman

published Problems" linear

his

now-famous I). by was That

paper, paper the

"A

New

Approach

to Linear contriof state

Prediction field of

(ref.

made

a significant

to

filtering a solution age

removing particularly Co_cnly

stationary to the suited to as and

requirements linear the dynamical

Weiner

filter problem. needs

and

presenting

sequential

solution

time-varying

filtering estimation the cles. the new

Kalman's of the had space

(ref.

2). in

known

the

Kalman

filter, cir-

formulation Although the

a major uses

effect of the

related filter the

academic were

engineering

first

Kalman of

in aerospace resulted fiiter in in

applications, its its rapid

relative

simplicity for utili_ation

and

versatility in many other

formulation The Kaiman

adaptation various

fields.

*Analtyical View, CA 94043.

Mechanics

Associates,

Inc.,

2483

Old

Middlefield

Way,

Mountain

forms class

is clearly of The _timatlo_ events here

established problems. led to

as

funda_ntal

tool

for

analyzing

and

solving

broad

that to _h_

the of

current

widespread

use The

of

the

Kalman

filter given in

are (ref. 3)

recounted are as of

best greater

our

recollections. ir to the sequence of many

descriptions that

expanded, well the as the filter paper

giving and its

insight

problem8 that

had

to be in the

overcome, discovery

extremely

fortunate

events

resulted

successful how the

implementation. need lunar for a filter even such as Kalman's that was arose from selected at a

The NA_SA's as a

describes on the

early

work

manned Also

mission, is why

before work the

mission

was

national

program. time have of the and

described near-ideal

Kalman's to

introduced

near-perfect problem Schmidt might (one

why

this

solution for Kalm_n. but the

midcourse meeting

navigation between Dr.

_ne

undiscovere_ authors) step

except and

a fortunate

present was by a

Kalman's mission that was

work limited in

large

forward, of

its

usefulness We

for

the

Apollo the events

certain

features

formulation. the

recount that the

resulted

the

scientific extremely

breakthrough useful tool

and that

reformulation knowm as

transformed

Kal_n's Kal_n

work into filter." reasons for from which the

the

is now

"extended

The systems results filter,

for Apollo

and

events

that

took C-SA for

place aircraft

during

the

development and the of are way the

of

navigation the A exper-

and

Lockheed led suitable the

programs

in which Kalman

those had

key

programs

to a square-root aircraft of the

Formulation

features of

applications, first known

described. flight an

discussion iment

then

follows the

development of

aircraft for

in which was

performance

a square-root

for_Jlation

coded

airborne

computer The which highly the

validated. concludes and world. acknowledge V. Harper, with the their John indebtedness Rodney efforts; leading to Joseph H. Perry, their without to the R. Carlson, and Gerald and of help George L. support, the Kalman P. who the with a discussion Kalman of the tremendous are being range applied of in problems today's to

paper

extended

square-root

filters

technical The authors Eleanor _terially

Callas, assisted original filter

D. McLean,

Smith,

analysis

circumlunar would have

feasibility been impossible.

studies

discovery

THE

FIRST

APPLICATION

OF

THE

KALMAN

FILTER:

APOLLO

MISSION

The

Need

for

a New

Filter

In the Flight division Center at

fall

of

1959,

Dr.

Harry Dr.

J.

Goett, and meet

then other with

Director members the

of of

NASA's Dr. Task

Goddard

Space

(GSFC), Ames

i_vited

Sehmidt (ARC) to

Goett's Group

former located at

Research

Center

Space

LL_

Langley Research Center to discuss the future mannedspacecraft

(ref. 4) for a history of the way Apollo at the program.) NASA Dr. Goett centers was to feasibility the manned Of Analysis outcome potential guidance large, studies under lunar mission. interest at Ames to research

program. eager _efine to get

(See

guidelines

for

major Branch of many areas for

the

present

authors its

were

the

areas

in which The

the

Dynamics

should of the for

concentrate Space the Task

research was the

efforts.

principal _ two navigation design and for

meetings of

Group The the

identification midcourse autopilot

research

Branch. and

first

was

the

circumlunar body

mission,

secor._

was

the

flexible The Dynamics few

liquid-fuel Branch out two

boosters. had very only eight analytically oriented The research most logical design

Analysis to carry

persons--too choice, problem. and work

complex seemed

research to be

programs. the for booster the

considering But after problem, area. to

personnel many and Branch

experience,

autopilot

discussions, there was

enthusiasm finally an than was a had

grew

midcourse for one the that

navigation Branch to a

guidance in that

unanimous been

decision before. to

Certainly,

it was

ambitious

undertaking, faced brought resident high as

presented

greater The develop meant digital The with

challenge primary the the software

personnel in the mission Branch

emphasis and for with the

rapidly to be

bear

on

the

need

tc

concepts

technology

for would memory to have

a completely have and

self-contained in a

system. reliable,

This on-board speed.

computer

considerable would have

relatively efficient and Moon return our that

computational The

midcourse

algorithms

be as to

possible. the back

system, from satisfy-

pilot/navigator into

inputs,

would

navigate the on

guide and to

spacecraft to Earth,

injection ing very

a circumlunar entry the to

trajectory, corridor

around

restrictive selected solution with lunar we

requirements and directed

Earth. to a conceptual a rather the for required the the of

Having design massive tasks necessary Branch for a

problem the

area

attention we were

problem, that was

it was utterly

clear

facing of area in our

effort (such as effort

a staff would

inexperienced Perhaps our if we of with choice

in many had fully problem nothing be

trajectory have

analysis). reconsidered we an made was to do were IBM

comprehended

research

effort. tools.

Basically, Ames had was machine

starting 704,

the

way

analytical computer. efforts, when had that we

software

which

would

simulation research later on, we

Fortunately, but were software preparing all of our

Fortran for the own the

available quite the

to us

early For

in our example,

limited. filter

software

Kalman

matrix We soon slowly

operations, found, that

to write the

matrix-handling for matrix

subroutines. operations to use would single have ran

however, our matrixa time.

double

indexing had

needed to be

so

handling circumlunar The

subroutines trajectory first step

rewritten

indexing. taken program Clarence assistance timely and aid the hours

Otherwise, of computer of and in the versus

with to

a Kalman build a

filter trajectory

was Moon

analysis Dr. most Sun

capable Gates and was

simulating others form time. at of By

a trajectory the based an Jet on

to the work

and

return. (JPL) area. the

Fortunately, gave us Probably the of

Propulsion their

Laboratory in this

invaluable

counseling

ephemeris

tape

containing

positions

the

Moon

mid-1960, we could calculate

investigating implementing

free-return

trajectories

to

the

Moon

and

were and

linear perturbation methods that small velocity changes to simulate was going well at pilot observations

looked promising for calculating midcourse correction.

tion;

Most of the research that is, the use of

that time with the exception of of external bodies to esti_te

navigathe by

vehicle state. We the crew operating measure tion was the how the

had assumed from the beginning that this would an inertially referenced optical sensor on the azimuth, such and subtended angles of Earth To and this by

De accomplished spacecraft to Moon. end, JPL The we and We ques-

elevation, to process

data

in an

efficient

manner.

had decided would also a of had

reviewed that also put

iterative not only burden

weighted, too complex on the

least-squares for IBM 704 we were

estimator using for

in use

it was

state-of-the-art

on-board

computers

but

a severe taking

simulation. and filtering

considered polynomial.

multiple this

optical appeared

tracking to be not a

measurements simple for enough the

them the

with

Although state could

procedure, level of

accuracy measurement

the resulting accuracy that

estimation was be achieved.

adequato

optical

Some and had

of

the

staff

had

been

working of

with that

Weiner theory

filter to the

theory problems

for of

several guidance

years and

made

successful

applications

navigation for beam-rider and homing missiles. Because problem had obvious similarities to missile navigation, Weiner the filter theory and the We theory could be applied. for an The lunar difficulties vehicle that

the lunar vehicle navigation we wondered whether the were the of had nonlinearity an irregular continuous of the of

problem of

requirement, measurements could not find

navigation, navigation would that permit would

series

discrete

(whereas

missile

assumed

measurements). Weiner restrict approach filter the was

approach

applications either severely

without

making

approximations

observation needed for not

system or destroy the computing the estimated our simulation

inherent accuracy. state from on-board or an

Obviously, a new measurements--a computational

way that would capability.

overburden

facilities

on-board

Discovery

of

the

Kalman

Filter

In retrospect should nized. and Dr. have The taken authors for doing,

it seems place and

almost that those

incredible Dr. Kalman's in the for fall

that work

the

next be

sequence so

of

events recogat of Ames, the of his

should

quickly Dr.

recall the several called It was

events Institute In the

following Advanced of

way. Study

Schmidt, had unaware

Kalnm%n, at

Research years. and during of of

(RIAS),

been

acquaintances work mutual now the we were

1960, Dr.
Dr. Schmidt visit

Kal_n,

arranged this the

to visit

to discuss Kalman

topics presented

interest. paper had

very

f_rtunate of the as

that

famous staff

to members thinking responsive

staff

Dynamics a way of the because facing

Analysis _,andling sequential they with

Branch. the

Because the fea-

been hit a

filter chord. were

theory

problem,

presentation tures relieve even of Dr. some

In particular, of interest we were

solution

Kalman's of the the

formulation computational was linear

could the IBM

certainly 704. Thus, thought

problems and our

though

theory

application

nonlinear,

Dr.

Schmidt

the approach might have somemerit for

bers to carefully was not Dr. problems were examine _n easy Ka_man's used also Dr. by very Schmidt at the task paper. and That

our

application

and

assigned

key

staff

mem-

a great of

deal the

of

difficulty new well

was

experienced space

in approach used his

understanding to control by Dr.

paper Dr.

because Kalman.

relatively as

state as to at the

The for

notation,

concepts, On

Kalman

difficult arranged this became we had needed filter potential posed meeting

practicing with the

engineers Dr. method Kalman of He

grasp.

next the

trip paper

to GSFC, further.

to meet that to Dr. Dr. with

RIIAS to discuss Dr. that could the Kalman's the be used we combi-

It was system concepts system

meeting clear been

applying realized studies Thus,

_heory linear nation were and IBM

to a nonlinear perturbation the Dr. linear Kalman's using of of

Schmidt. Kalman's the to _inear

using

in our

guidance

to produce

to apply theory

theory.

linear us a

perturbation

methods

already also 704

gave the

solution by at the

the and

nonlinear storage in the

navigation limitations breakthrough

problem of we the

to some computer.

problems the

speed

Clearly,

RIAS

resulted

needed, evaluate provide remained

and we immediately and validate the the an state open

made plans to produce a digital Kalman formulation. Whether the accuracies was of necessary considerable for

simulation program to Kalman filter would navigation

estimation question and

circumlunar

concern.

Modification

of

the

Formulation

The given us on-board the

trajectory

analysis

work

that

had

been

progressing

in the

meantime of the use to

had simulated accomplish

some insight into the problem of how to schedule the time optical measurements and course corrections we planned to and guidance function. with The the schedule measurements long of periods the were course called for

navigation taken

optical spaced The the iD

measurements time. The time intekey over The a

to be

in a short

sequence, by and

evenly of

sequences of the variables gration features wide

could for routine of the and

be separated sequences analysis. to solve were to

relatively the timing These

time. by

starting would be

measurement later

corrections

features of

permitted for the

numerical The (time) time.

used

equations its be ability stopped to us let

motion

trajectory. step arbitrary IBM 704

routine the

to vary and

the

integration at any the the an

range

ability was

restarted it brought

variable time down Dr.

step-size

critical at

because us

computation

to a value Kalman's

that

least

consider have

doing required

problem. on-board in time our crew to make the and solua

original

formulation

would

a continuous lunar mission,

sequence an

of optical the

measurements original

equally to

spaced had

throughout measurement The into

impractical schedule, was

scenario.

Therefore, formulation the

implement to be

course-correction tion to the problem

revised.

obtained and natural

by decomposing a discrete-time and back, and intuitively the

original

formulation

discrete-time which Dr. the provided Kalman's time,

update a much

portion more

optical

measurement way of

update expressing

portion

appealing

algorithm. it was could

Looking a major have

decomposition step

seems and

almost one in

trivial; which an

at

however, error

critical

forward

unrecognized

been

disastrous.

The Extended Kalman Filter The original studies using perturbation methods and the above mentioned decomposition of the filter were based on a linearization about a nominal (reference) trajectory, it soon becameapparent that a relinearization about the current estimated state might offer substantial advantages over the technique previously used.
We reasoned or of that true, the An "on the average," to the the estimated or state would be closer and to the the or linreferextreme the had actual, earity ence, state than reference, be this retained nominal, than to state with out thus

approximation The correctness accidental to remain

would of in error in an

better was

a nominal, in a rather run

state.

presum_ption conditions Earth, transfer position run

borne

incident. true the

input

a simulation the

caused trajectory

trajectory proper

orbit for a

around lunar the the the

but

estimated The

starting

conditions used to

trajectory. of the the space

simulated were were inten-

optical tionally nonlinear

measurements sparse

estimate part of

vehicle

in the

early At

because

trajectories caused

quite relabut

in that

regicn.

first,

Kalman-filter

estimator

some

tively large the estimated modification filter."

overshoots in the estimated state state soon successfully converged to the implementation has come to

owing to the nonlinear effects, close to the true state. This be known as the "extended Kalman

Development

and

Testing

of

the

Fi_ter

Simulation

The was matrix was new

design the used state

of

the

digital how the filter a

simulation Kalman time update and

program w_s remained velocity

proceeded to be

fairly

rapidly The

once This

it

cle?rly for

understood discrete

filter

implemented.

transition matrix into a

a problem, error in state the

however. at one

to be error

to transform at a later

position It

time

time.

is a key

element

Kalman-filter

formulation. Another the programs group to at Ames the specializing transition spaced then in software matrix in using and accepted data as a from the a task stored of developing trajecwould A to
b

compute were not

lunar

tory. be set of

The: data to

equally the

time, and

result at the

interpolation time. data be

required

retrieve

desired would

position

velocity from this on

a specified trajectory would over by

variational the fast

equations matrix. reduce of

be evaluated thought that time

produce tively

transition and would of

It was the

approach IBM 70_

relaof it the

computational solving the second-order task was

the

that As

straightforward an additional turned thought. By Fortran matrix one and was the and time the out,

approach set

simply

variational

equations

integrating

18 complicated, development

differential more

equations than

the

software

much

difficult

originally

the coding their

simulation extensively software to give

program could the be

implementing the completed

the

Kalman

filter on the

was weeks. the

coded

in

verified,

group

working

transition Everytime came

expected extremely for

in a few program be a

additional try. When it was

eager

Kalm_n-filter program to

passed

the

transition

matrix

operational,

becoming

increasingly producing expected. order

clear software The only

that that

work was

should taking along was

be much with

stopped longer was those

on to to for

this

approach than the

because was

it was

execute integrate the IBM by

originally 18 secondand for the terms. estimated the

remaining equations this

alternative

additlona_

differential

true, 704

reference, budget by

trajectories. Kalman-filter that the At selves. computing directly program programs the

;[though simulation

a setback be inverted

in the merely

time

program, could about build

compensating

savings

were

made

discovery

transition this point

matrix we were to

rearranging we

6 months a test In

behind program way, program

the

schedule the

had

set

for

ourfor

It was the by was we

decided

to debug the modular

modular

software could

transition

matrix.

this

elements few

be used The test

the put had

Kalman-filter together developed. was by in a

simulation few days little already This a

with

very was was

changes. from

from

coding

that time

borrowed required

other most of

Very then

debugging ranning new

because and

complex was It

software very

in other

programs the

because

the

problem from Moon). perigee. tion

straightforward. the used translunar to produce were

program

produced from to

transition approach

matrix to the to

injection

into

trajectory transition punched laws to

to periselene matrix on

(closest trans-Earth be read

is also The matrix for

injection the

elements with the

then

cards

into

simula-

program In the

use

guidance the

compute matrix

course for

corrections. a time-update of the

Kalman-filter matrix in was the

program,

transition by

covariance tial

always

available program that the

integrating from injection

the

18 second-order time next the of and the so to

differenfirst on. the The time cor-

equations

simulation then from

to the to the from by

optical product of the

measurement of all these optical latest

and

measurement transition matrix was

time matrix

matrices

gave

injection

measurement.

This

required

mideourse

rection algorithm. The variational reference or the estimated states. (inbound) matrix trajectory periselene by reinitializing to Earth

equations Later the at

could be solved about either the program was modified to do the return with a,lother stored transition

periselene

from The

reentry. was the designed nominal to (or integrate reference) nonlinear trajectory, differential and the

digital for the

simulation true

program

equations estimated tions of accepted state. midcourse drive the

trajectory,

trajectory. The simulated the true state plus additive

optical biases

measurements and noise.

were calculated The Kalman-filter

as funcalgorithm

these measurements and provided incremental changes in the estimated The difference between the estimated state and nominal state was used correction estimated and perigee algorithm. position at the the to end This the of 6V algorithm gave the at _V vector at required the end

in the to of the

nominal the

position

periselene To the to

outbound course

inbound was

trajectory. added were the to

simulate

the

mid-

correction, Random simulate The

computed and system matrix

vector

estimated the added also

trajectory AV true in a

directly. :tot to

magnitude control covariance to the

pointing errors, of the of

errors and

added

computed to the

result state

was was

aJectory. manner

estimated the

increased errors. the IBM

corresponding When the

statistics was

simulated ready The for

control a test run

system run on

simulation over an

program of

finally time.

704,

it

took

a little

hour

execution

first

showed

disappointing

results
Kalman

in the state estimation scheme. Naturally,

filter did On not work? men, run an error was the second the everything program optical yield This we as we the had Our worked had we

we were
in the

shocked. way we

What were

if

the a

found fine{ been had

calling

subroutine. By extended results the early

1961,

si_lation and the on-board could

used

extensively OUr with

co very the

validate knowledge

the of

Kalman indicated of

filter that motion

guidance

laws

developed. combined for the

encouraging

measurements accuracy breakthrough for the

equations

adequate was the

circumlunar we had set out and result Kalman to guidance At for filter but compucould

navigation/guidance achieve for that the at the circumlunar it was

problem. when that mission clear

that Dynamics

beginning

had

selected area

the midcourse a potentially

navigation Analysis significant the extended

study

Branch.

time

achieved studies

on-board wculd with

navigation accuracies

systems.

indicated of weighted for

that

give

comparable

to

those

least-square computer extended

estimators, memory Kalman and filter at this

a tremendous speed. and

reduction We had not,

in requirements however,

on-board that the

tational be time. As the

verified with the

mechanized

operated

properly

flight

computers

available

it

turned

out, for

we

had

been

fortunate Kalman

in the

way

we

had

chosen

to mechanize to

computations

the

extended

filter.

The

order

chosen

in which

multiply the three matrices in the measurement update rather than K(HP) was the most numerically sbable of have we selected. done As a result, we had no apparent we

equation, that is, the two possibilities round-off found that

(KH)P we could If

computer have large much

difficulties. double-precision in computer for time, space-

had

things

differently, were necessary have _de

however, with the a

might

matrix which, craft

operations in turn,

resultant appear

increase less

would

filter

attractive

applicaticns.

Dissemination

of

the

Simulation

Results

As we recall, two of the first persons results were Dr. John V. Breakwell and Dr. Missiles ciently explain time. At about results. tute been was of Dr. this Battin same was and Dr. time at was t_ in early 1961, and Space Company with the (LMSC). results Both to

(outside Charlotte

of Ames) Striebel, and Dr.

that we told of our both of Lockheed Striebel their were part to at the suffi-

Dr.

Break-well further

impressed the

begin

research

on

equivalences

to other

trajectory

determination

methods

in use

Dr.

Schmidt Laboratory

told of

Dr. the

Battin

of

our Instihad an Procemade

Instrumentation in studies along lines work. titled In the

Massaehusets Dr. Dr.

Technology with

engaged in work

of

the

Apollo to 1961,

mission. those Dr. of Battin

Battin but

independently unfamiliar

engaged

similar

Kalman published

Kalman's Report 5).

In September "A Statistical introduction

Instrumentation dure the for Space

Laboratory Flight" statement:

Optimizing to his report,

Navigation Dr. Battin

(ref.

following

The

formulation

of

an the

optimum current to The

linear best produce author

estimator estimate a is still

as

recursion with

operation newly was

in which

is combined better to

acquired

information by Kalman.

estimate

presented

indebted

Dr. Stanley F. Schmidt for directing his attentions to Kalman's excellent work. In fact, the original application Kalman's theory his associates. of this paper to space navigation The work described was done without any was made by Schmidt and in the following sections detailed knowledge of

of

Schmidt's

activities

....

Dr.

Battin

continued system. square-root could in

to

investigate Potter handle

using (ref. random

the

filter

fo: with was

application Dr. Battin for out as as the

to at

the MiT,

Apollo devised system. for

navigation the His first

Later, filter not

6),

working which

implementation,

used

Apollo useful the of

formulation

forcing

functions, such

it was that part of of Some

implementation system. root Still

small-word-size Potter's that could

on-board implementation handle later. to spread the summer

computers, was forcing

Apollo squarethese

later,

included

other

implementations will of our

random

functions.

implementations The and word

be discussed work began During come

rapidly and

throughout fall we of had 1961,

scientific, many

academic, from of at all

engineering of two the

circles. country (refs. had

visitors summer NASA

parts 1961 (refs. an

to discuss 8), which (AAS)

the were

work year,

done.

In the of the two in San first

papers

7 and

shortened were

versions presented was This studies Because for with many

reports

9 and

10) published Astronautical of our work of and were

in the Society on

following

Francisco formal

American

conference. and guidance

introduction mission in the informal both before Apollo

navigation

for of

the the

circumlunar extreme interest of

a group program

scientists the held after

and the

engineers. our work held conference

potential

that

program,

extensive

discussions and

representatives

government

industry. Dr. Schmidt technical and his staff of researchers at Ames breakthroughs which led to this first can major

be

During credited

the early period, with the following of the Kalman

application I. problems. 2. current problem. 3. rate at

filter. Kalman's original theory could be adapted to nonlinear

Demonstration

that

Development best estimate

of the extended Kalman of the state to reduce

filter, which linearized about the effects of nonlinearities

the in the

Decomposition

and

reformulation

of

Kalman's so that

original

algorithm could

into b_

sepa-

time-update arbitrary

measurement-update time interval.

portions

measurements

prcce_3_d

any

9
%

4.
(I)-(3) spacecraft 5. tion

Demonstration, by meansof
above, of the Kalman and of filter's navigation Dissemination for guidance the

comprehensive potential

digital

simulation, to a

using

items on-board

application

nonlinear,

problem. of in the the simulation Apollo work to the M!T Instrumentaand control

results inclusion

Laboratory

possible

on-board

guidance

system. 6. segment papers. It should experienced Later, with be noted that at this point owing to in time to (1961) no problems had been Dissemination of the of inforn_tion and aerospace on the Ames Kalman filter work to a large and formal

_cientific

co,unities

through

presentations

in the

digital

simulations began on

truncation substantial effort

or modeling problems and

errors. of instability would De

however, Kalman

many

researchers problems

report

filtering,

which

considerable

research

expended.

FURTHER

FILTER

IMPROVEMENTS

Support Following siles filter and Space Dr. Schmidt's the

for

the from ARC for at

Apollo ARC in

Mission late 1961 to the Join Lockheed Mis-

departure staff to at NASA

Company, of

continued the ARC Apollo was

to apply program. directed

extended

Kalman

to problems From mid-1962

interest

to mid-1964, the same

research with

to and

three

general

areas trajec-

uslng essentially tories: I. design 2. most of 3. addition that the the To study

software

the

same

Earth-Moon

Moon-Earth

the on the

effect optimal effect computers effect

of

modeling

errors of

of

secondary vehicle word

importance trajectory.

and

of

off-

conditions To the To study

estimation of the of

a space

relatively time.

short

length

accommodated

by

airborne study the

that

on midcourse data. was first

guidance It was noticed. filter

of

using the

ground

radar of

tracking this chosen sensitive or a priori noticed work

in

to on-board "divergence" studies

observation problem with effects the of

during

early

phases the

Apparently, was not

problem

for

initial

extended computer the

Kalman

particularly parameters, was

to nonlinearities, statistics. and occurred

round-off, computer

unmodeled round-off

As mentioned when this at ARC the to was this

earlier, of

problem

first was only

sequence the

multiplying that to the Dr.

three computer Schmidt and

matrices round-off by Mr. the

together was the

changed. problittle we were to

Inittall}, lem. was using Our done

led at

assumption because When

discovery

comunicated _ime

Gerald

Smith, formulation first

but

problem some

filter the

seemed

quite

stable.

undertaken

time

later,

attempts

10

......

fix

the

computer

rour s-_3f wlth

problem nonnegative tried, by

involved

working

with After

the

covarianee P

matrix in First, the by

to make 1ormal forcing elements

it symmetric way, P and some to be

eigenvalues. four of which are the to

computing here. lower

fixes

were

mentioned upper or

symmetric on!y

selecting diagonal terms

either elements

off-dia{onal matrix. same as 3eccr_d, 5he ma_niz_aes of fix the assoin

using the

those

form

a symmetric Third, the

by averaglng second, were And then

off-diagonal the the a

to force

symmetry.

computing than by one, adding

off-diagonal coefficients positive

correlation would number and be to all

coefficients, printed of and the the diagonal

if any terms This an

greater fourth,

prograz, J 5toppeC.

s,_ll

eovariance ciated the

matri:[,

P, after

measurement errors with

time-update noise,

operations. thus _liowing

numerical

truncation matrix it, on after the

process

increase

covariance As we

numerical third and

operations. fourth was fixes being after causing would small be worked best but both very during at ARC the and time el2e-

recall was

that where. ments so

work

going

another was by

problem

uncovered, a series the of

The had

new that

problem processed

most the

apparent filter,

accurate P by

measureto become _he state would The

been

covariance

matrix ignored

"small"

additional happened, as true the

measurements only of The very

essentially to the

filter. would diverge use of

When

this

corrections and here quite

the

estimated state

be computed from the

result

a measurement, basic problem proved

estimated to

state. in the were

is due

modeling against

errors. this

"pseudonoise" researchers For 8, only also

time-update encountering

effective

problem. of K. Lee, by The after

Other round-off. dated July

computational Inc. interoffice

difficulties memorandum with P of more by

because R. C.

example,

a Honeywell, that the the by

1964, the update

recommended diagonal be plus that

the

syn_netry

problem

be the

solved matrix.

computing mei_orandum

upper problem using

triangular of P the

elements becoming but

recommended

negative-definite symmetric

a measurement expression

overcome

equivalent

Pt+1 where

= [I - Kt+IH]Mt,

I[I - Kt+IH] T + Kt+IRK

+I

'

Mt+1

: #p

T + Q

, -I

Kt+ I : Mt+IHT(HMtIHT

+ R)

This

equation

for

Pt+1

is

frequently

referred

to as

Joseph's

formulation Kt+ ! matrix is after

(ref. 11). opti_l, as a state The (refs. I. the

It reduces shown. It owing of

to Kalman's is also the arbitrary

measurement equation general equation for gain. in support of

when the gain the covariance

change results

to an the

filter the

research following:

the

Apollo

program

12-14) How

made to

clear

include process,

astrodynamic and bias how to

constant compute the

uncertainties performance

and of

bias-type

errors

in to

estimation or

a system

subjected

unrecognized

ignored

errors.

11

2.

Thai for the particular

cL._o_lunar

trajectories

being

in,/e_ti&Jted,

the

simulat[uon tions could the covariance

computer word mantissa of 2i bits was adequate_ but that some computabe carried out with the lesser precision. It also verified that when matrix P is a too-optimistic are data of the given too representation little weight. to use on-board of radar measurem data does nts not for of estimated state

errors, 3. estimating

external That the

measurements radar

ground

are

generally

superior but that

trajectory

spacecraft,

save significant midcourse is not greatly enhanced. primary backup Apollo system navigation on the

correction fuel, and control of the These results supported the ultimate conducted from It was the an ground, important using and

trajectory decisions radar result.

endpoint to have the data with a

ground timely

spacecraft.

Application Meanwhile, his first at the Lockheed to ARC. sta_e, process The using

of

t_e

Filter and

to

the

Agena Company, data was

Program in 1961, the Dr. Schmidt Kalman had

Missiles actual of

Space

oppcrt_nity at

measurement the effort tracking

with to

extended the

filter of the

developed Agena

purpose

validate from the

performance sta-

upper

Earth-based

data

downrange

tions and telemetry data from the vehicle. A general-purpose postflight program was developed which combined tracking data from several stationj model of the equations of biases for the motion and for the vehicle. errors, Agena as The error as state

analysis with the

included of a

tracking-measurement propulsion The data model procedure coast

location of the

well stage. to use

coefficients

thrust for

upper was and

developed phases

the

operation position

the and

tracking ccvarianee matrix of

during

to estimate ana the with

velocity

errors at the initiation thep handled by starting (from the coast-phase model. At the matrix da_..),

termination of the thrust phase. The thrust program with the initial state and covariance the error during state the expanded phase to include was

phase was matrix of normal

coefficients in the

thrust

Tracking termination used used as to

data of

thrust the the

processed from state the

manner. tracking covariance

the

thrust, of the

state

estimate

postburn with its % {

d. Ca was

a measurement characterize the

six-component

vector,

accuracy. techniques to the applications technology

of

This development the Kalman filter:

added

following

I. residual measurement

A bad-data magnitude update

rejection to its

technique

was

developed as

which

compared

the

measurement

standard If

deviation the resisual

computed magnitude

f_'o_ the exceeded value of n

Kalman-filter n times used was the 3,

algorithm.

standard deviation, corresponding to a 2. The Kalman

the measurement 3-3igma residual filter was used

was rejected. The ,_%gnitude test. as a data compression from the

algorithm of

to

form

an

equivalent measurement taken during the coast

and eovariance phases of the

matrix vehicle.

multitude

measure_nents

12

3. developed smoothing, mators 4.

An

iterative but simply at that

approach effects an time. was of

using

backward procedure

integration from to the the

and

forward _nis

filtering was not

was data-

to remove

nonlinearities

estimate.

equivalent

weighted,

least-squares

esti-

in use The

Kalman

filter models.

used

to estimate

parameters

in the

measurement

and

equation-of-motion

Development

of

the

Kalman-Schmidt

Filter

After 1962, program was GSFC. (ref. for solve Dr.

moving Sch_midt GSFC.

to Philco-WDL began During as work on

(now the

Ford

Aerospace of

and

CommLnications

Corp.)

in

development the the

a general-purpose Kalman-Schmidt of of Dr. to F. by 0.

error-analysis filter Von (ref. Bun at credit The rot 15)

for

this a

effort, of

so-called encouragement had been

developed, Earlier, 16) as

largely the the

result Kalman

extended

filter

referred

some

authors

Kalman-Schmidt technology filter as error

filter, which

probably

because

a desire Kalman of

to give filter. (but

the

application &_lected errors

resulted to here, As a result,

in the includes

extended the of

"Kalman-Schmidt" for}

referred states.

effects

does

a means which as

optimally errors

compensating in the in the filter

for modeling equations purpose are

is provided, This

when filter

it

is known was

model one 17).

significant.

provided

option

general-

error-analysis to the

program

delivered

to GSFC filter,

(ref.

In addition the following I. of the 2. ments

Kalman-Schmidt to recursive

the

developments

at

Philco-WDL

added

techniques

filtering for and

technology: saving machine of the time by taking account

General sparsity

computational of the

techniques matrix for

trarsition formulation data.

symmetry

covariance sensor

matrix. measure-

Mathematical

comb

ning

on-board

inertial

with

ground-tracking

3.
effects

Ad hoe technique _'ur adding of numerical errors.

pseudorandom

forcing

functions

to

minimize

the

DEVELOPMENT

OF

THE

SQUARE-ROOT

FORMULATION

The By When was 1966, the advantages the

C-SA of

Aircraft the

Navigation Kalman the

System filter were widely the C-5A recognized. Kalman navigation as a consultant and data This from was filter

extended for The

Lockheed specified was won

became for by the

prime

contractor system. which, The the

C-5A

aircraft, for Dr. the

navigation Corp.,

contract hired

system for the

Northrop

in turn, filter state had

Schmidt

Kalman-filter navigational opportunity

development. aids to produce of the

combined estimate been

inertial of the

data

various the

aircraft. to participate

first

either

authors

afforded

in the

13

development for a NASA knowledge, aircraft.

of a real-time system on program as early as 1961. was the first real-time

boa, u an aircraft, The C-SA system, of a

even though they had hoped to the best of the authors' filter on board an

application

Kalman

During of selecting with the working within and by this

the a

development limited word of work.

of error

the

C-5A and

navigation adding error the to

system, whole be

the

real-world to the system burden of

problems outputs, work model test

appropriate

states,

estimates

size,

making had

computational Many put that in to a use

time-frames

small

computers data time

confronted. 19) were metnods expense and the

the

numerical

compensation

techr !ques Also,

(refs.

18 and

demanding measurement were per-

developmental for John saving

compression at a small

averaging fected (ref. root erly method by 20). filter with was

computational while

accuracy at for

Weinberg

working definitely could

with

Schmidt out

Lukesh need

Northrop a general and squarepPopis, propaa

This

development that

pointed random

formulation the fixed-point wherein

include of the error

forcing

functions

wor_

arithmetic the filter's

available covariance

on-board would

computers; be computed

that and

needed

gated in square-root to maintain filter for this system, ad hoe (refs.

form and stability. standard

would therefore require However, because such extended was used Kalman for

less computational a formulation was with the

precision not available tech_lique caused by

the

algorithm,

epsilon

and other round-off The C-5A

procedures, 18 and 20). complexity had

controlling

numerical

problems

additional

of making

an

extended Dr.

Kalman Schmidt

filter of for an the

work need

with for a

the

navigation

system

thoroughly

convinced forcing

square-root tions. Apollo

formulation aware he of

including Potter's made

random square-root several

functions to find

airborne formulation efficient

applicafor the square-root

Being mission,

measurement attempts

update

personally

filter method that would allow use of random forcing algorithm. In early 1968, he successfully developed looked promising for application on a small,

functions a method

in the time-update (refs. 21 and 22) that computer. of the (The square-

fixed-point sum_ary of

on-board the

reader should consult reference root filter development during

23 for a good this time.)

history

Flight

Test

of

RAINPAL

with

the

First

Airborne

Square-Boot

Filter

In program

1969, to

Mr.

L.

A. McGee Shuttle

(one

of

the

present work by

authors) testing

proposed the Kalman

flight-test filter in an

support

the

development

on-board configuration device manufactured by indicated that would ity. Range the that the

to validate its performance with a highly accurate ranging Cubic Corporation of San Diego, California. Studies had navigational independent results these At were accuracy assessment to be was of of such its b, a system was so great

potential and the

a highly be The test

accurate if site

navigational the White system

performance communMissile on Sands

essential,

accepted the

scientific

meeting Mexico. was

requirements time, be the the

Army's

(WSMR) test

in New range

the to and

cinetheodolite accurate an

in operation nation for

WSMR

believed position

most of

system

in the

determining

the

actual

velocity

aircraft

during

approach,

14

ORiGt._LOF landing, gross XDS-920 the and takeoff with of rollout. weight of The flight-test lb. The and on a POOR

F;X'E QUALITY

aircraft on-board 32K

was

to

be

a Convair would on be a

CV-340

with

44,000

computer Based

ruggedized studies limited of

a 24-bit-word computer word

length length

memory.

previous and the

effect

Kalman-filcer need the

performance

amount culties project include In for the

of memory with the an

available, as Kaln_n filter of

well as the as possible,

to avoid decision of the

as many numerical dLffiwas made to incidde in this Kalman filter _hat _ould

evaluation forcing McGee,

a square-root

formulation

random 1970,

functions. G. L. Smith, to and others at and on be ARC issued (in a request flight) fixed-point ir order called to a for proposal

development filter with with such

of

software forcing

implement functions as would

test

square-root DS-920 develop was compua comto be a

Kalman ter plete

random other and

a 24-bit, required later

along

software system.

airborne

navigation

This

system, system.

RAINPAL,

precision The by

approach competition

landing the

navigation software

for

development Inc. (gMA),

work based had

on on

the the

RAINPAL square-root been

system

was

won

Analytical of

Mechanics handling who

Associates, forcing leading measurement to

formulation by

capable Dr. Potter's for the

random be a for in

functions the AMA update the used

which and

recently

developed used

Schmidt,

would

team.

The

original

formulation

algorithm time update

a modified forcing the by at

Gram-Schmidt functions. non-square the

algorithm The Gram-

order

include was

random to

Schmidt

triangularization covariance Mr. Paul

algorithm matrix to

reduce required

augmented filter

square-root formulation. {ref. the 23),

a square a doctoral

form

square-root University

Kaminski, to Dr. for

candidate another and time, made

Stanford

demor_strated could be Gram-Schmidt implemented,

Schmidt the time

that update At

triangularization save computational little it made to was

algorithm operations difference the Apollo faster

_y over which

Householder modified

used

algorithm. so the

this was

algorithm Householder the root RAINPAL filter The

was

decision Potter's

proceed used of an the

using in the the

algorithm. system is

Although believed to

algorithm first noise, ARC but,

system, square-

be

the

application on-board before

complete

technology, system

including was

process at

aircraft. flight tests were and that

RAINPAL

initiated

started, it became a joint the Army's Instrumentation of flight of testing was the Space new and to

program with NASA's Directorate at WSMR a square-root new concepts and for (SSV); promise

Manned Spacecraft (WSMR-ID). Ames' Kalman which filter might be with

Center (MSC) interest was inertial for was

validating investigate Vehicle offering

aiding; navigato at

MSC's tion

interest

suitable

Shuttle

the a

WSMR-ID future

interest

a desire system

investigate WSMR.

concepts

instrumentation

Overall, ure have caused been by

the

square-root

Kalman

filter round-off, of

was or the

under many

considerable other

scrutiny. causes technology.

Failwould _he

divergence, blow to

computer the

potential

a severe

proponents

square-root-filter

square-root the software Kalman dard filter

filter, newever, performed development, in fact, it above all the suspicion would when

flawlessly soon became software have

once it had been debugged earl_in possible to place the square-_ot occurred. because With of the its stan-

problems been

formulation,

filter

always

suspect

15

ORIGINAL

F;.:'_

I-3

OF POOR QUALITY propensity to diverge or develop negative eigenvalues, peculiar transients in the navigation estimate. It is rare in the development of a system which could cause very

such as RAINPAL

that the opportunity

for early validation is possible. The RAINPAL validation (refs. 24-26) was provided by the theodolite trackir_ system at WSMR, which is nationally recognized for its accuracy. The flight-test results showed that with only three range measurements versus seven theodolites, on the final approach to touchdown the RAINPAL system gave position and velocity estimates that were smoother and had accuracies of the same order as those of the WSMR system. Obviously, the RAINPAL system could be operated as an independent reference system against which other systems could be tested or evaluated. Since the fllght test of the RAINPAL designed several other square-root-filter (ref. 27), including systems on helicopters. system was used system in January 1972, Dr. Schmidt applications for aircraPt systems were flight has and

for ARC, which

tested on .... aircraft TOL

The same basic square-root approach developed for the RAINPAL in these filters. The early versions of these filters used an

experimental microwave landing system (MLS) ca!led MODILS as a primary landing aid (refs. 28-31). The later versions used a prototype MLS and were flight tested on a helicopter (refs. 32 and 33).

Other Airborne

Applications

of a Square-Root

Filter

During the time the RAINPAL system was being readied other researchers were also developing square-root Kalman

for flight test at WSMR, filters for airborne

applications. One of the best examples of this work was done by Intermetrics, Inc. for the Completely Integrated Reference Instrumentation System (CIRIS) unOer development at Hollo_an Air Force Base. This system was very :imilar to the RAINPAL system in that it employed inertial aiding from _n inertial navigation system also _mployed precision measurements to transponders many miles apart on the and

ground. The precision ranging system was a Cubic CR-100 system, which produced both range and range rate from the aircraft to the ground transponder. This was a later and much improved version of the Cubic system used by the RAINPAL system, which provided only precision range from the aircraft to the ground transponder system. Details of this work were published in 1973 by Widnall (ref. 34), who described simulation results for a new square-root algorithm devised by Carlson (ref. 35). Carlson's algorithm maintained the covariance square-root matrix in triangular form, but, more importantly, reduced the computation time from that of the Householder/ Potter method. This was a significant step forward, of states, Carlson's method would approach the speed Kalman-filter formulation. the most serious complaints because for a moderate number of the standard extended one of

This improvement in speed substantially overcame leveled against the square-root formulation.

16

The In 1975 Bierman

Square-Root (ref. 36)

Filter introduced

Successfully his U-D

Tested

in Space of the Kalman filter

factorization method standard

which appears it uses only (It should be

to be the most efficient s%uare-root slightly more computer time than the noted matrix there JPL were that the U-D are factorizatiol, required for of used in

to be developed extended Kalman a square-root matrix. This (ref.

'-date: _ -er. method With work, 37) on whose the

is not

actually the

in that this filter Mariner the a

_hree gorithm out at

prcduets is no on

to define square-root and the

covarance

requirement the work mission

operations.) in 1969 further for by

carried

draws

Dyer for

McReynolds navigational This success concept provided using the

algorithms

successfully space

computations

10 Venus-Mercury of except there the

1973.

established operaextended

reliability Thus, filter, future

sequential the to

square-root

filter

real-time standard

tions. Kalman in all

for seem

intuitive be ample

familiarity reasons for

the

square-root

formulation

applications. types the of square-root of root the the of filters standard the error have been developed. filter matrix. is defined has The The as first may it

Two be basically involves of the

different as leads the a to

regarded

factorization square of root

extended

Kalman

algorithm; second the to

covariance which

square

information Each of

matrix, these

inverse recom-

error

covariance certain

matrix.

algorithms

attributes

mend

it for

applications.

CONCLUDING

REMARKS

It Kalman tal

is clear filter

that by

the the

relatively authors in some

simple the

and

straightforward 1960's can be

sequential, adequate as a

extended fundamen-

used tool

early

analytical to

for

solving a Kalman

estimation without

problems. experience familiarity. factors,

Certainly, would be

persons tempted to take or roundany Some been solutions proper large of mismodela practior

attempting this route

construct of

filter

because this

the

intuitive may be

statistical dependent models

However, such as

success computer

failure off all

with

approach

on many and

errors, of which as often of ad

inadequate may we trigger yet have

statistical the

nonlinearities instability of these Which problem filter still of and

in the or

problem,

filter's

potential to some

inaccuracy. have pragmatic the on a of not

rather devised, to use amount machine,

pragmatic

effective seen on in the

solutions foregoing

problems the

history. the the may

depends hoc

correctly or

identifying "fix." may If be a

applying run to the

stabilization but an

is being solution remain

extra

precision

arithmetic stability

reliable

problem

numerical

instability, of

problems

because is often

ing. In the case cal alternative. It appears problems are

airborne

application,

extra

precision

that

those the

experienced ad hoc

in applying

Kalman techniques

filters and

to the

real-world standard filter The

abandoning in favor of

stabilization that are

formulation

algorithms

numerically

better

conditioned.

17

% %

square-root filter, by its than the standard extended Square-root superior reluctance ran slower algorithms as

nature, is inherently more Kalman-filter formulation. have gained for acceptance example stems

stable

and

better

conditioned

rather

slowly

despite

their Some of the

performance of

reported, users

i_: (refs. from the and

24,25,32-34). fact that the

potential the

probably Kalman used

early

algorithms

than too

standard and

formulation, too much

the

faotorization as well as

techniques computer
I

appeared time.

complicated

computer

storage

As mentioned Recently, devoted subject mination, ship to a special papers

earlier, issue on The

the of

Kalman IEEE that such

filter were

has as

been on

used

in a as

variety Control possible

of (ref. in orbit

fields. 38) their deterwas

the

Transations

Automatic ranging as radar and and

applications papers of cover

wide

matter.

diverse

subjects

spacecraft tracking,

prediction natural

cattle ray flow

populations spectroscopy rates, and on-line control were

in France, in oilestimation failure systems.

navigation, mea-

motion, of in

gamma

gas-well

exploration, of plant

surement variables

instantaneous industrial and were of the at

prediction

unmeasurable

processes, station and

detection In many

in nuclear cases, the Indeed, that that put

instrumentation, these papers the

power

solutions the have will years broad

in

implemented filter when the to

operationally unlikely

successful. suggests and be will

applieation scratched likely to come. This place at recording personal undoubtedly failures publication 1960's. the For of be

seemingly to

problems this filter

we we

only

surface

it comes

to possible which

applications,

amazed

applications

in the

history _RC, of those

of

Kalman

filtering and is elmost the

has totally

naturally dependent in

dwelt on this of

on the

the

events

that of there

took and are

LMSC,

Philco-WDL, events by

Analytical

Mechanics

Associates, work, some to of

Inc, which

The

recollections

contacts some, memory, of

made should the

authors.

Omissions to our

be

attributed

ignorance of which time, has the

pertinent the

events,

inevitable the in

limitations filter

and

mushrooming since a the early of are

papers

about

Kalman further and

taken

place counsel

those

interested cited

details, of the

authors

perusal

various to

references suggest.

herein

many

other

papers

they

in turn

certain

18

REFERENCES

I ,

Kalman, Trans. Kailath, Inf.

R. E: ASME, T.: Theory,

A new Set.

approach D, of J.

to

Linear Eng.,

filtering vol. of 82,

and Mar.

prediction 1960, pp.

problems. 35-45. IEEE Tran.

Basic

A view vol.

three

decades no. 2,

linear 1974.

filtering

theory.

I T-20,

Mar.

Schmidt, space p. 4

S.

F.:

The

Kalman AIAA

filter: J.

its

recognition Control, vol.

and 4,

development no.

for

aero1981,

applictions.

Guidance

I, Jan.-Feb.

Brooks,

C.

G.;

Grimwood,

J. M.; U.S.

and

Swerson,L. Printing

S.:

Chariots Stock

for

_pollo.

Th___ee

NASA History Series, 00768-0, 1979.

Government

Office,

Number

033-000-

Battin, 1961.

R.

H.:

A statistical R-341,

optimizing

navigation Inst.

procedure

for

space Mass., Sept.

flight.

Report

Massachusetts

of Tech.,

Cambridge,

Battin, pp.

R.

H.:

Astronautical

Guidance.

New

York:

McGraw

Hill,

1964,

303-340.

Smith,

G.

L.; 61-92,

and

Schmidt,

S.

F.:

The

application on-board Society

of

statistical Aug. 1961.

filter

theory

to optimal Paper

trajectory American

determination Astronautical

a circumlunar Meeting,

vehicle.

McLean, Paper

J.

D.; to

and an

Schmidt,

S.

F.:

Optimal system

filtering for the Meeting,

and

linear

prediction mission.

applied

on-board American

navigation Astronautical

circumlunar Aug.

61-93,

Society

1961.

i __ .

Smith,

G.

L.:

Schmidt, to the mission.

S.

F.; NASA

and TR

McGee, R-135,

L.

A.: of

Application and

of

statistical on-board a

filter

theory

optimal

estimation

position

velocity

circumlunar 10.

1962.

McLean,

J.

D.;

Schmidt, applied TN

S.

F.;

and

McGee,

L.

A.:

Optimal system for

filtering the

and

linear

prediction mission. 11.

to a mideourse D-1208, 1962.

navigation

circumlunar

NASA

Joseph, Course

P.

D.: Notes,

Space

control

systems--attitude, Extension Course,

rendezvous, University of

and

docking. at Los

Engineering

California

Angeles, 12.

1964.

Smith, G. L.: Secondary errors of space vehicle trajectories.

and

off-design conditions NASA TN D-2129, 1964.

in optimal

estimation

19

13.

McGee,

L.

A.:

Effect

of

reduced

computer

precision and

on

a midcourse prediction.

navigation NASA

and guidance system TN D-3382, 1966.

using

optimal

filtering

linear

14.

Smith,

G.

L.;

and

Harper,

E. data.

V.:

Midcourse TN

guidance 1964.

using

radar

tracking

and

on-board

observation

NASA

D-2238,

15.

Users

Manual

for Defense

the

Mark

II Error

Propagation Alto,

Program. Calif., Feb.

WDL-TR2758, 1966.

Philco

Western

Laboratories,

Palo

16.

Bellantoni, Schmidt

J.

F.;

and

Dodge, Paper

K.

W.:

A square-root New York, Jan.

formulation 1967.

of

the

Kalman-

filter.

AIAA

67-90,

17.

Schmidt,

S.

F.:

Application

of

state vol.

space 3, C. as

methods T.

to navigation ed. Tech. New Report

problems," York: No. Academic 4, July

Advances Press, 1964.)

in Control 1966.

Systems,

Leondes, WDL

(Originally

available

Philco

18.

Schmidt,

S.

F.:

Estimation Analytical

of

state

with

acceptable

accuracy Mountain

constraints. View, Calif.,

Report 67-4, Jan. 1967.

Mechanics

Associates,

Inc.,

19.

Schmidt, lems. Calif.,

S.

F.:

Compensation 67-16, 1967.

for

modeling Mechanics

errors

in orbit

determination Inc., Mountain

probView,

Report Nov.

Analytical

Associates,

20.

Schmidt,

S.

F.;

Weinberg, C-5 Mich.,

J.

D.;

and

Lukesh,

J.

S.:

Case IEEE

study Case

of

Kalman Semi-

filtering nar, 21. Ann

in the Arbor,

aircraft June

navigation 1968. (Also

systems. AGARDograph

Studies 1970.)

139,

Feb.

Schmidt,

S.

F.:

Final

report

for

mission

analysis

guidance Inc.,

study. View, Calif.,

Report 68-23, Jan. 1969. 22.

Analytical

Mechanics

Associates,

Mountain

Schmidt,

S.

F.:

Computational of Kalman

techniques

in Kalman

filtering. Feb. 1970.

Theory

and

Applications

Filtering,

AGARDograph

139,

23.

Kaminski, vol.

P. AC-16,

F.;

Bryson,

A. of

E.; Dec.

and

Schmidt, techniques.

S.

F.: IEEE

Discrete Trans.

square _uto.

root

filtering:

A survey pp.

current

Control,

727-736,

1971.

24.

Schmidt, and

S.

F.;

Smith, B.: Joint

G.

L.;

Hegarty, navigation Control

D.

M.; for

Carson, approach

T. and

M.;

Merrick,

R.

B.;

Conrad, 14-3,

Precision Automatic

landing

operations.

Paper

Conference,

1972.

25.

McGee,

L.

A.; S.

Smith, F.;

G. and

L.;

Hegarty, B.: landing

D. M.; Flight

Carson, results

T. from NASA

M.; a TN

Herri:k, st_idy of D-7302,

R. aided

B.; iner-

Schmidt, tial

Conrad, to

navigation

applied

operations.

1973.

20

26.

Schmidt, for

S. aided

F.;

BJork_n,

W.

S.;

and

Conrad,

B.: NASA

Ne,

mechanization 1973.

equations

inertial

navigation

systems.

CR-2352,

27.

Schmidt, AGARD

S.

F.:

Experiences Series No.

in 82,

the

development

of

aided

INS

for

aircraft.

Lecture

1976.

Schmidt, Schmidt, flight 30.

S. S

F.: F.;

A Kalman and Mann,

filter F. I.:

for

the

STOLAND

system. Kalman

NASA

CR-137668, the

!975.

A three-axis May

filter

for

STOLAND

test

system.

NASA

CR-137939,

1976.

Schmidt, of

S.

F.;

and

Mohr, NASA

R.

L.:

_avigat_on 1979.

systems

for

approach

and

landing

VTOL

aircraft,

CR-152335,

31.

Schmidt,

S.

F.;

Flanagan,

P.

F.;

and

Sorenson,

J.

A.:

Development area

and landing

flight opera-

tests of a Kalman filter for tions. NASA CR-3015, 1978.

navigation

during

terminal

32.

Foster,

J.

D.;

McGee, with Foster, landing vol.

L.

A.;

and

Dugan,

D.

C.:

Helical NASA

automatic

approaches

of

helicopters

microwave J. D.;

landing and

systems. G:

TP-2109, helical

1982. rotorcraft Control,

33.

McGee,

L.

A.; and

Xenakis,

Automatic system.

descent and

using 7, no.

a microwave 4, July-Aug.

landing 198_.

J. Guidance,

Dynamics,

3u,. Widnall,

W.

S.: and

Optimal precision

filtering ranging

and data.

smoothing AIAA

simulation 73-872,

results 1973.

for

CIPIS

inertial

Paper

35.

Carlson, J.,

N. vol.

A.:

Faster 9,

triangular Sept. 1973.

formulation

of

the

square-root

filter.

AIAA

11, no.

36.

Bierman,

G.

L.:

Factorization in Science and

methods

for

discrete vol. 128.

sequential New York:

estimation. Academic

Mathematics Press, 1977.

Engineering,

37.

Dyer,

P.;

and

McReynolds, J. 1969.

S.:

Extension Theory

of

square-root

filtering vol. 3,

to no.

include 6,

process pp.

noise.

Optimization

Applications,

444-459,

38.

IEEE

Trans.

Auto. of

Control, Kalman

vol.

AC-28,

no.

3,

Mar.

1983.

Special

issue

on

applications

filtering.

21 j-

Report

_.

2.

Go_mvn_t

Acc_sion

No.

3.

RIcililnt'l

Clillol

No.

NASA
4 T,tie _

TM-86847
_it;e

l
5. Report _te

DISCOVERY TOOL FOR

7 Aurar(s)

OF THE KALMAN AEROSPACE AND

FILTER AS INDUSTRY

A PRACTICAL

No_ember
6. Piriorminl

1985
Orlliniilitiorl Glde

fl.

lle,r'fixrilirlt

Orlllnization

Report

No.

Leonard
_f_mi_

A.
Crgln,_ti_

McGee
Name

and
e_ _

Stanley

F.

Schmidt
10.

85424
Work Unit No.

Ames Kesearcb Center Moffett Field, CA 94035

11.

Contrict

or

Gr_nt

No.

13, 12 S_nsocin 9 Agency _me and Aodresi

Type

of

Report

md

Period

Covered

Technical and Space Administration 14 S_inl _ 532-06-11

Memorand Code

National Aeronautics Washington, DC 20546

15 _D_e_ntary Not_

Point

of Contact: Leonard McGee and Research Center, Hoffett Field,

S. CA

F. Schmidt, MS 210-9, Ames 94035 (415-694-5443 or FTS

464-5443)

"_

A_tract

The

Kalman

filter

in

its

various

forms

has

become

fundamental

tool

for analyzing and solving a broad class of estimation problems. The first publicly known application was made at NASA Ames Research Center in the early 1960s during feasibility studies for circumlinear navigation and control of the Apollo space capsule. This paper recounts the fortunate sequence of events which led the researchers at Ames Research Center to the early discovery of the Kalman filter shoztly after its introduction into the literature. The scientific breakthroughs and reformulations that were necessary to transform Kalman's work into a useful tool for a specific aerospace application are described. The resulting extended Kalman filter, as Jt is now known, is often still referred to simply as the Kalman filter. As the filter's use gained in pepu!arity in the scientific community, the problems of implementation on small spaceborne and airborne computers led to a "square-root" formulation of the filter to overcome numerical difficulties associated with computer word length. The work that led to this new formulation is also discussed, including the first airborne computer implementation and flight test which was conducted in 1972. Since then the applications of the extended and square-root formulations of the Kalman filter have grown rapidly throughout the aerospace industry.

17

Key

W_

(SuijlJeSt_

by

Auth.(s))

18.

Oiitributi=

S_tement

Kalman filter State-estimation Navigation Square-root filter

Unlimited

Subject

category

- 04

19

_=ity

Omif.

(of thtirel)OCi)

I _"

Slllxi_

Clllit

(of

lhil

iiJl|

Unclassified
"For site

Unclassified
Technical Informltlol_ Service, Sprinifld,

i,,.,,,o.o,,>,, 24
Virllnil 22161

1,2 A02 _,"

by the Nliio_ll

i!