Yangsheng Xu, Yongsheng Ou. Control of Single Wheel Robots

Подождите немного. Документ загружается.

1445

rther

pics

Learning-based

Con

trol

0 10 20 30 40 50 60 70

100

120

140

β [deg]

time(sec)

lean angle of Gyrover

0 10 20 30 40 50 60 70

−60

−40

−20

[deg]

time(sec)

tilt angle of flywheel

Fig. 5.18. Ve

rtical balancing by

CNN mo

del,

trail

#3.

ave

rage lean angle

βDOFDOF

flywheel

CNN control#19

7260677497..26

0 . 6774

CNN control #2 95 60 0403995..60

0 . 4039

Human control 87 31 0737287..31

0 . 7372

Table 5.12. Performance measures for tiltup motion.

control, while the human control is shown in Figure 5.22. The performance of

these motions are summarized in Table 5.12.

Since a large portion of the ﬂywheel’s motion is contributed to tiltup the

robot, the overall degree of freedom of the ﬂywheel in tiltup motion is much

lower than that of lateral stabilization. For the CNN model control in Figure

5.20 and 5.21, the robot is lying on the ground initially, with β

≈ 150 , after

, after

a few seconds, the model tiltup the robot and brings the robot back to the

upright position.

5.2

Imp

lemen

tationo

earning

Con

trol

145

0 10 20 30 40 50 60 70

100

120

140

β [deg]

time(sec)

lean angle of Gyrover

0 10 20 30 40 50 60 70

−30

−20

−10

[deg]

time(sec)

tilt angle of flywheel

Fig. 5.19. Ve

rtical balancing by

man op

erator.

Combined Motion

We observed thatthe CNN modelsfor lateral balancing andtiltup motionare

subjected to some initial condition, the

problem canb

esolved by

combining

the two motions to form asingle motion.

Consider thecase that therobot is in thefall position, that is, with β ≈

150. In Figure 5.20 and5

.21,although the CNN

tiltup model is

able to

keep

the robottostayaround at 90

foracertain moment, the robot will fall back

to the ground eventuallyb

ecause the

ﬂywheel has reachedanill-condition

( β

= ± 90

). Moreover, thetiltup model is unable to let the robottoconverge

to 90

sometimes, whichcausesalarge ﬂuctuation in thelean angle about90

Figure 5.24.

To dealwith this problem, we combine thetiltup motiontogether with

the lateralbalancedmotion,Figure 5.23. Sincethe CNNmodel is unable to

keep the robotatthe verticalposition,afterthe robothas tiltup, we ask the

model to balance the robotat90

The experimental result forthe whole tiltupand stabilization process after

thecombination is showninTable 5.13 andFigure 5.25. Initially,the robotis

in afall position, by executingthe tiltupcontrol of theCNN model, the robot

is recovered to thevertical position. Afterwards, the lateralstabilization is

controlled by another model whichspeciﬁcally trained forkeepingthe robot

into the verticalposition.Fromthe results, thecombinedmotion can keep

1465

rther

pics

Learning-based

Con

trol

0 5 10 15

100

120

140

160

180

β [deg]

time(sec)

lean angle of Gyrover

0 5 10 15

−20

100

[deg]

time(sec)

tilt angle of flywheel

Fig. 5.20. Tiltup motion by

CNN mo

del,

trail

#1.

average lean angle

β DOF

flywheel

CNN control#1 88. 40

0 . 8998

Table 5.13. Performance measures for combined motion.

5.2.3Discussions

In this chapter, the CNN modelsfor lateral balancing andtiltup motionare

being veriﬁedbyexperimentalimplementations. By combining thetwo mo-

tionsintoasingle motion, the robot is able to recoverfromthe fall position,

and thentoremain stable at thevertical position after tiltup.Therefore,we

have completed the low-level behavior module within thebehavior-based ar-

chitectureshown in Figure 6.4. With this module completed,weare goingto

develop asemi-autonomous controlfor Gyroverinthe nextchapter.

the robotatthe verticalposition well after tiltup from the ground foramuch

more longer periodoftime.

5.2

Imp

lemen

tationo

earning

Con

trol

147

0 5 10 15 20 25

100

150

200

β [deg]

time(sec)

lean angle of Gyrover

0 5 10 15 20 25

−20

100

[deg]

time(sec)

tilt angle of flywheel

Fig. 5.21. Tiltup motion by

CNN mo

del,

trail

#2.

0 1 2 3 4 5 6 7 8 9

100

120

140

160

180

β [deg]

time(sec)

lean angle of Gyrover

0 1 2 3 4 5 6 7 8 9

−40

−20

100

[deg]

time(sec)

tilt angle of flywheel

Fig. 5.22. Tiltup motion by human operator.

1485

rther

pics

Learning-based

Con

trol

Verticle position

Lateral stabilizationTilting-upInitial status

REAR VIEW OF GYROVER

flywheel

spinning axis of flywheel

motion of

flywheel

motion of robot

Fig. 5.23. Combined

motion.

0 1 2 3 4 5 6 7 8 9 10

100

150

200

β [deg]

time(sec)

lean angle of Gyrover

0 1 2 3 4 5 6 7 8 9 10

−100

−50

100

[deg]

time(sec)

tilt angle of flywheel

Fig. 5.24. Fluctuation in the lean angle made by the tiltup model.

5.2

Imp

lemen

tationo

earning

Con

trol

149

0 10 20 30 40 50 60 70 80 90 100

100

120

140

160

180

β [deg]

time(sec)

lean angle of Gyrover

0 10 20 30 40 50 60 70 80 90 100

−50

100

[deg]

time(sec)

tilt angle of flywheel

Fig. 5.25. Tiltup and ve

rtical balanced motion by

CNN mo

dels.

Shared Control

Basedonthe successful implementationsofthe lateralbalancing andtiltup

motion, in this chapter, we aregoingtodevelopashared control framework

forG

yrover. In

fact,any

situationofa

system usings

haredc

ontrol will

in-

volvehumaninteractions. Under shared control, thehuman operator actsas

upervisor foro

verall con

trol, while ther

obot itself can handle some

cal

motionswhichinturn assistthe human. In order to distribute thecontrol

tasks systematically,wedevelop an expression to makesuchadecision. Ex-

perimental resultswill be given in order to verify ouridea.

6.1 Control Diagram

In fact, shared

controlhappens in

man

ailyexamples, esp

ecially for human-

animalinteractions. First of all, let’s consider the horse riding case [103], it

is afairly good example of asemi-autonomous system, or morespeciﬁcally,a

shared control system.

he horse whichisb

eingr

idden by

uman,itisusually able to take

care of all low-level tasks suchascoordination of legmotion,stability, local

obstacle av

oidance andprovide enough po

werand sp

eedfor diﬀerenta

ctions.

On the other hand, the riderprovides global planning,interacts with the horse

to arriveatdiﬀerentlocationsand achieve various goals.Atthe same time,

therider can override anyhorse behavior by pulling thereins or hitting on

the horse’s body if necessary.Throughout thejourney,the riderrelies on the

horse motoric abilitiesand the horse’s behaviorsbecome more intelligentby

receiving the rider’s commands. Theinteraction between the two individuals

happens in anaturaland simpleway.

Another example we want to illustrate is to ask aroboticarm to handle

acup of tea[120]. Thewhole task canbedecomposed into two subtasks:

(i) to handlethe cup of tea safelywithout pouring the tea (localbalancing),

and(ii) to reachthe desiredlocation (global navigation). In ateleoperated

environment, it maybediﬃcult for ahuman operator to perform both tasks

Y. Xu and Y. Ou: Control of Single Wheel Robots, STAR 20, pp. 151–173, 2005.

1526

Shared

Con

trol

simultaneously,oritcould be mentally taxing.However, if an autonomous

module is introducedfor thelocal stabilityofthe cup, the operator in the

control loop is only responsible for the navigation task, whichgreatlyreduces

theb

urden

ators.

reov

lear

that

the

rformance

the

system

uld

tter

table

than

ein

trolled

ingle

tit

man/mach

ine).

Gyroverisacomplex system not only in terms of the diﬃculties in deriving

its mathematicalmodel, butalso in terms of its controlbyhumanoperator.

The robot can be controlled manuallythrougharadio transmitter with two

independentjoysticks, one of them is assigned to control the drivemotor,

while the other oneisassigned to control the tilt motor. Similar to abicycle,

Gyroverisasingle trackvehicle whichisinherently unstable in its lateraldi-

rection.

Therefore,

diﬀeren

romc

ollinga

quasi-static

mobile

rob

human operator notonly handlesthe global navigation forthe robot, but also

needs to payattentiontogovern the lean angle of therobot simultaneously.

Moreover, thehighly couplingeﬀect between the wheel and the internal ﬂy-

wheel also complicates the controlofGyrove

r. To

this end,for suc

complex

system, instead of creatingfullyautonomous control, it is much more practical

to developacontrol methodwhichcan “share” theworkload of thehuman

operator.

Recently,shared controlhas been widely applied into many robotics man-

machine systems, from health care [10,103, 26,

2, 97, 20]

to telerobo

tics

[121, 64, 39, 120, 29].For rehabilitation applications, atypical example is

robotic wheelchairs. Although the wheelchair itself canprovide alevel of au-

tonomyfor theusers, it is still desirable that

the

user can augmentthe con-

trol by the on-board joystickinsome specialsituations (e.g.docking, passing

through adoorway). Atelerobotic system usually consists of ahumanoper-

ator ands

everal autonomous controllers.

uman op

erator usually interacts

with the system in diﬀerentways. One of theimportantissues is to develop

an eﬃcientm

ethodt

ombine ah

umanand mac

hine intelligences so thatthe

telerobotic system can perform taskswhichcannotbedonebyeither ahu-

man or autonomous controller alone[39]. In these shared controlsystems, the

autonomous modules exist in the system to assist the human operator during

navigation,inorder to relievethe stressofthe operatorsinacomplex system.

Usually,t

he hu

man op

erator is

responsible for some

high-level

con

trol (e.g.

global navigation), whilethe machine performs low-levelcontrol (e.g. local

obstacleavoidance).

In fact,the two behaviorswehavementioned in the previous chapters, (i)

Lateral balancing and (ii) Tiltup motion, aredesigned to tackle the robot’s

instabilityprobleminthe lateral direction. Sincewehavesuccessfully modeled

andimplemented the two behaviorsbyamachine learningapproachand ver-

iﬁed them in experiments, the next step is to incorporatethesemotionswith

human controlinorder to develop asharedcontrol framework forGyrover.

We preferusing ashared controlscheme rather than afullyautonomous one

because of thefollowing reasons:

6.2

mes1

• Sophisticateddynamic system. As mentioned before, it is diﬃcultfor

us to obtainacomplete mathematical modeltogovernthe motionsof

Gyrover, due to itscomplicated dynamicand nonholonomicnature. This

mak

coun

ter

man

iﬃculties

dev

elo

pinga

ful

autonomous

system

for

his

stage.

• Hardware limitations. Due to the special physical structureofGy-

rover, the currentprototype of Gyroverweare using stilldoesnot have

anynavigation devices equippedon-board(e.g. vision), whichmakes it

impossible forthe robottonavigate itself.

• Importance of human operators. Practically, some complicated tasks,

whichmay be trivial forhumans, areoftennot performedwell by robots.

Therefore, ahumanoperatorisessentialtoexist in the controlloopin

order to monitor andoperate the executive system.

• Time and cost. Building afullyautonomous system which provides safe

and robust performance would be time consumingand costly,interms of

computations and resources. In contrast,itisfar morepractical andmuch

eaper

develop as

emi-autonomous system.

• Accuracy vs Reliablilty. Machines are excellentinperforming repetitive

tasks quickly andaccurately buttheirabilitiestoadapt to changes in the

environmenti

he other hand, hu

mans are usually reliable, with

tremendous perceptive abilityand good decisionmaking in unpredictable

situations, but their accuracy is relatively

lowe

hanm

achines. Shared

control canlet them compensate each weaknesswhichwould result in

tter control.

• Te

leoperations.

Gyrove

an be

operatedb

umansthrougha

radio

transmitter,whicha

llows hu

mans to participate in

the

control

therobot.

Themain diﬃcultyindeveloping asharedcontrol forGyroverisdue to

the access to

the

tilt motor. Since the

lean angle of

therobot

con

trolled

the tilt

motor, not only the

autonomous module will access the

tilt motor

to achieves

tabilityi

he laterald

irection, the human operator also needst

accesst

he tilt motorduring na

vigation.A

particular time

instan

these

commandsmay

con

tradict eac

ther.T

herefore,i

big issuet

et the

system decidewhichcommandisgoingtobeexecuted,and at thesame time,

managethe contaminated commandsinareasonable way.Tothis end, we have

developed an expression formaking this decision, whichwill be discussedin

the later part of this chapter. With better sharingbetween the machine and

human operator,the performance of thesystem canbeenhanced, andthe

rangeoftasksthatcan be performedbythe system canalso increase.

6.2Schemes

In fact, there aremanyaspects of “sharing”insharedcontrol, whichvaryfrom

application to application.Basically,semi-autonomous control canbecatego-

rized into serialand parallel types[120]. In theserial type,manualcontrol and

1546

Shared

Con

trol

autonomous control cannot be executedsimultaneously,only oneofthem will

be selected at atime. In parallel type,bothmanualand autonomous control

can be executed simultaneously.

the

llo

wing

sections,

wil

rieﬂy

discusst

hreeo

rating

shared

con

ol,

namely:(

Switc

istributed

(3)

Comb

ined

de.

6.2.1Switch Mode

In switchmode, manual andautonomous controlare switched in serial, as

sho

Figur

.1.

The

condition

triggert

switc

end

pplica-

tions, for example, if an operator acts as asupervisor of thecontrol system,

the human control will only be activatedwheneverthe system reaches an “ill

condition”. No matter whichcontrol module is switched, the robotwill be

fullycontrolled by theselected one. If high cooperation between the machine

and operator is required, we must have afunction(Π )whichcan “smartly”

switch between the two control modules.

Autonomous

module

Human

operator

ROBOT

Fig. 6.1. Switchmode.

6.2.2Distributed Mode

Figure 6.2 illustratesthe architectureofdistributed control. Diﬀerentfrom

switchmode, both manual andthe autonomous controlcan be executed in

parallel in thismode. The controlofvarious actuators (

)inthe entire system

will be distributedtoeither of the two modules.

Therefore, the two entities can exist in the system peacefullywithout

disturbing each other.However, thisalso shows the weaknessofthis mode