Skip to main content

Full text of "mit :: ai :: aim :: AIM-330"

See other formats

Artificial Intell Igence Memo No. 330 December 17, 1974 

LOGO Memo No, 17 



Howard Austin 

This research has as its basic premise the belief that physical and mental 
skills are highly simitar* enough so In fact that computation paradigms 
such as the ones used in Artificial Intelligence research about predomi- 
nantly mental skills can be usefully extended to include physical skills, 
This tiiesis is pursued experimentally by the categorization of "juggling 
bugs" via -detailed video Observations, A descriptive language for juggling 
movements is developed and a taxonomy of bugs is presented. The remainder 
of the paper is concerned with an empirical determination of the character- 
istics of an ultimate theory of juggling movements. The data presented is 
relevant to the computational issues of control structure, naming, addressing 
and sub procedu ri 7a t i on . 

This report describes research done at the Artificial Intelligence Laboratory 
Of the Massachusetts Institute of Technology.- Support for the Laboratory s 
education research is orovided in part by the National Science Foundation under 
Srant EC-407Q8X, 

Table of Contents 

1) Introduction I 

a) Thesis: Physical and Mental Skills Are Very Similar In Nature 

b) Thesis: h CoiflpuLdcrur.dl Psratiiyn ran Bo -j'a". id I r ^.rolltiJ 

2) Procedural Paradigm Explained 2 

a) Terminology 

b) Taxonomy definition &f a theory 

3) The Experimental Setup 3 

<0 Role of the Teach&r 
• b) Ise of Video 

4) The Role of Model for Physical Skills 4 

a) Cascade vs. Shouerg: Whidh Model Do You Have 

b) Implications for a Theory 

5) Performance Data 

a) The EXCHANGE Stop 

b) Performance Graph 

6} Host.Cofflflfl&n Bugs 

4) FORWARD RUG and Fix 
b) EARLY -LATE BUG and Fix 

7) Tftj^o^on^ 1 of Juggling Bugs 

a) Importance of A Good Descriptive Language 

b) Bugs and fixes 

a) Human Constraints 

b) Ballistic Constraints 




8) Timing Analysis ?fi 

9) Basic Cycle ,a 

a) Overall Movement Pattern 

b) Reaction Theory 

1 ) Anticipation Theory £g 

a} Frame by Frame Timing Analysis 

b} Latency Considerations 

c) The Cycle From tfie Control Point of View 

1 1 ) Bugs Come in Bunches 34 

a) Observation 
■ b) Implications for Specifying Time Information 

12) The BreakthrouqN Phenomenon 35 

a) Observation 

b} ImpT 1 cations for Waning/ Addressing 
c) Graphs for Si , S2, and 55 

13) Its Hard to 61 ve yourself Advice aq 

4) Observation 

b) Theoretical Implications 

M) Constructing Subroutine^ 42 

a) Tne Main Issue 

bj Empirical Data from tile Juggling Case 

15 J Summary and Conejos ions 45 

16) Apj>endix : Protocol for SI 47 

17) References 55 


I would like to gratefully acknowledge the help given by Dorothy 
Hummel in the preparation of this paper, without h&r patient assistance 
during many long hours of video analysis this experiment might never 
have been finished. 3 would also like to thank Karen Prendergast for 
her editorial connnents and assistance with the majority of the illus- 
trations, Draig Latham for his heln or tne remainder and Gene HorvRz 
for his patience in typing the original! as well as count le^ revisions. 


This paper .„!««! tha p^ical skill of j u5s ,i„ 9 frM1 the „„,„ af 

vta. of th* pnc^l parad tgm f the na*!, davelcpin., field of Artifice, 


The study of the mechanisms of physical or motor 
activity has been a classical theme of a wfde 
variety of scientific disciplines Including 
psychology and neurophysiology. The approaches 
employed In the investigation of sensorimotor 
phenomena have ranged from the v^y loc&T i.e. 
Single neuron studies to studies of global skfll 
acquisition strategies. Important issues this 
research has helped to clarify Include the nature 
of reflex and the precise extent to *h,ch sensory 
Input controls motor activity. 

Despite U ast amounts of sensorimotor data hwew, we are stni 
largely in the dark on important general Issues, such as the difference 
between "physical" and "Intellectual" activity as well as on specific 
questions such as how new physical skills are constructed from existing 
ones. The present research attempts to sharpen these issues by means of 
extreme Ty detailed empirical observations and by the development ■ 
of theoretical models. A major conclusion of the paper is that physical 
skills are considerably more "intellectual" in nature than hithertofore 

The Procedural Paradinir 

TMs SSCtfon is a brief aside about the Artificial Intelligence 

Philosophy which underlies the experimental approach. A mare exten- 
sive discussion can be found in Minsky h Papert and Uinoarad M-3). 

The most important aspect of any theory or paradigm is the degree to 
which it explains, or has the potential to explain, observed phenomena. 
Of course different paradigms use different criteria for testing to see 
whether or not a given theory "explains" a particular set of observations. 

The procedural paradigm of Artificial Intelligence 
(A, I.) holds that knowledge Is stored internally 1n 
a form which can be modeled accurately by means of 
computer procedures . Hence the relevant questions 
for a given theory have to do with the control 
structure associated with the procedures, the 
mechanisms by which that control structure retriev es 
and activates procedures, the bujjs (i.e. mistakes) 
encountered during activations and the process by 
which those mistakes Are debugged (corrected). 

Since recent work in A.I. provides a wide variety of control and acti- 
vation strategies from which to choose, the most cruciaT questions for judging 
the merits of a specific theory, for exampTe a theory of juggling, are 
■ what kinds of bugs occur when a person learns to juggle and how are these 
hugs fixed. In particular ,if it 1s possible to exhibit a taxonomy of 

"juggling bugs" and to describe precisely how these bugs are removed then 
you have given a complete theory of juggling,, i.e. 


Section 7 gives such a taxonomy for 3 ball cascade juggling, obtained 
by extensive analysis of video protocols of adult subjects learning to 
juggle. As we shall see later the precise description of a process by which 
observed bugs cart be corrected is considerably more difficult to obtain, 
Before proceeding further however the details of the experimental setup 
are given. 

The Experi mental Setup 

This sections explains two unusual aspects of the experinsntal s1 tua^ 
tion. The first aspect 1s the extensive use of instruction, i.e. a teacher, 
during the learning period. The second 1s the use of video equipment to 
record the experimental trials. 

The use of a teacher was very nearly mandatory given the decision to 
study the skill of juggling > Juggling was felt to be an excellent choice 
for detailed analysis since it appears to be a very complicated skill but 
In fact requires, little more than the ability' to toss and catch an object 
and the Ability to visually track moving objects. These prerequisite 
skills are usually highly developed 1n normal adults. Hence the learning 
required for juggling focuses on the recombination of existing well- 
developed movements rather than the format f on of entirely new ones. 


Unfortunately, informal experiments show that very few people team even 
the most simple kinds of juggling without the benefit of instruction. 
Although it is perhaps reasonable to conjecture that most people could ■ 
eventually Team to juggle on their own qjven sufficient motivation, the 
amount of tine required *ould most lifeely rule out short term learning 
experiments as well as unduly cmnpTfrate the data collection problem. 
These problems were avoided by the active use of a teacher during the 
learning trials , 

The use of the. video equipment to record data was likewise essentially 
forced. It is virtually impossible to accurately record experimental trials 
which involve the study of complicated human movement without some sort of 
vfsual recording device such as a film or video technology. Video tape is 
more desirable than film since it requires no development time and can he 
erased and reused . 


The following paragraphs explain in detail the precise nature of the 
experimental task. In addition, the first important empirical observation 
fs presented. 

The subjects were randomly selected undergraduate, graduate and faculty 
meters of the *U.T. community ranging in age from 18 to 50. Twenty " 
subjects were involved of which approximately half were video taped. The 
rest of the experiments were recorded by a coronation of third party 
observers and/or audio recordings. No significant difference was found in 

the two groups. 

The experiment proceeded as follows: after a brfef Juggling demon- 
strati on i the subject was asked to demonstrate, in slow motion, his or 
her model of juggling, 

JuggHng can be defined as keeping more objects 
1n the air than you have hands. For example keeping 
I objects in the air with one hand Tike this 

Figure 1 

is juggling since only one hand Is involved hut I 

people tossing 4 balls back and forth simultaneously 

are not juggling since a total of 4 hands are being 

Cascade juggling, perhaps the most common and certainly the most 
simple form of 3-hall Juggling is illustrated by the following sequence 
of sketches.: 

Figure £ 


It fs Interesting to note that almost all of the subjects tested did_ 
not describe the "cascade" model of juggling, Rather, they almost invariably 
described a form of what is tonrnonlj known as "showers" juggling in which 
one hand does the same throw over and over while the other hand does all of 
the catching. 




r igure 3 
The difference in the two techniques can be s.aid [at the risk &f over- 
simplification) to lie in the fact that during SHOWERS a given hand either 
tosses or catches, but never both* while in CASCADE a given hand alternately 
tosses and catches. In SHOWERS the catching hand either physically passes 
each ball to the tossing hand or tosses it horizon tally to the tossing 
hand, Whereas in CASCADE the tosses are all vertical no matter which hand 
they come from.* 

This observation, interesting in its own right* is important experi- 
mentally because the initial ' demonstration each subject was given was a 

*for an excellent description and further details, see CARLO (4) 

dernonstratfon of CASCADE juggling which fa considerably easier to learn 
than showers. The subjects persisted in the SHOWERS model even after the 
.CASCADE demonstration and, left to their own devices b would have undoubtedly 
failed even more miserably than subjects left on tneir own with tile 
CASCADE task. The Important observation her is that the formation of 
juggling models, be it CASCADE or SHOWERS or whatever, is an entirely 
mental process - It lias nothing whatsoever to do with muscles, neural 
pathos' in the usual sense or sensory feedback. Hence the first step 
tn the experiment provides us with our first clue as to the relationship 
between "physical" and "mental" acts, You cannot perform a physical 
act (at least a skilled one) until you have tn accurate mental model or 
representation, of that act! 

■ ..-■■ r- f o nr = n;g Driyi 

This section redefines the experimental task and gives performance 
data for the "typical juggler" in terns of that definition. 

After correcting "model errors " the subjects were led through a 
series of exercises in which they learned to control first one toss, then 
tyro, then three and so onu CASCADE juggling has the nice feature that 
both a SINGLE TOSS, 
figure 4 SINGLE TOSS 

arid a DOUBLE TOSS, 
Figure 5 


are very easy to do. Almost everyone does them reasonably we n on the 
first try. The TRIPLE TOSS is considerably pore difficult and hence de- 
serves further analysis, 

The essential feature of the TRIPLE -TOSS Is that it requires, after 
an Initial TOSS wft*i the right hand*, a rapid TOSS -CATCH sequence with 
first the Teft and then the right hand followed by a concluding left- 
hand CATCH, The TfiJFtE-TDSS Is illustrated in figure 6, 

♦Virtually all jurjgl frig algorithms are handwise reusable. 

Figure 6 














TOSS by 
right hand 

by left band 

by right hand 

Concluding CATCH 
by left nand. 

The real difficulty lies, in the rapid TOSS-CATCH sequences, for good 
reason as we shall s« later* For now we shall be content to observe that 
if we relabel these sequences, calling them EXCHANGES, then the description 

of CASCADE juggling reduces to; 


to irtUialfze 







an indefinite series 


of exchanges 





and & rnnrliirf-lnn /-=.+ - 

Given Oils description* It Is easy to see why the ability to do 3 

tosses is art important step in the learning experience. At that point 
the subject performs the hardest step, the EXCHANGE, once with each 





Heft hand) 
{right hand) 

THe rest of the teaming process consists of fine-tuning the basic 
EXCHANGE sequence. The experimental data shows that the avenge person 
tikes approximately 7 minutes to succesfully complete 3 tosses. Figure 
? gives performance data for a hypothetical average juggler in terns of 
" .number of tosses achieved after a given m*,ber of minutes pr*cti< 
















10 15 

TIKE {minutes J 




,^ peridental 
training period* Fraa this perspective it makes sense to a&y, 
for example, that the average juggler can do 2.5 tosses after 
5 minutes. It should be noted that several r:-f -he subjects 
Tilt ^efsre the X) ainute* elapsed, This was due to (a) frus- 
tration (4 people) or (b) the achievement of 10 or so tosses 
{5 people). Nevertheless, it is felt that the graph is repre- 
sentative of average performance. 

Connon Bugs and Their Fixes 

Host people exhibit a fairly standard sequence of bugs. The foiling 
section discusses In detail two of the most caron bugs and their 
recoumended ffaes, 


If we suspend now the narration of the experiment and return for a 
moment to the procedural paradigm, we see that the EXCHANGE task is Impor- 
tant for another set of reasons, It is here, at the EXCHANGE sequence, 
that the first serious bugs begin to show up fn the learning process. 
These bugs usually relate in one way or another to either subtle or gross 
timing errors. Thus it is fair to say, at the risk of precision, that 
the most common kinds of bugs in learning to juggle are timing 

The most frequently occurring t fining hug usually manifests Itself as 
a variation on the following theme. 


Figure 8 

Normal right hand 
TOSS followed by 
FORWARD left hand 


The first ball is tossed correctly. It h-as the proper height (slightly 
higher than the forbad) and width {slightly less than shoulder wide)/ The 
second ball h however, is tossed almost directly forward with little or no 
vertical velocity. This bug appears to be due to the conroon tendency to 
swing the forearm from the elbow when tossing an abject. The resulting 
trajectory is then critically dependent on the precise Instant the 
object is released by the fingers. An early release, perhaps caused by 
concern about the incoming ball, causes the second ball to go forward 
rather than up* A late release causes the ball to be thrown back towards 
the juggler. 



Figure 9 

Of course the correct solution is either to release the outgoing ball 
at approximately the vertical tangent point of the swing or to move the arm 
in a vertical direction only so that it doesn't matter when the ball is 
released. The 2 "files' 1 are accomplished respectively by having the subject 
juggle while standing near (within arms reach) and facing a wall or by 

teaching the subject a special largely vertical throwing motion known as 




Figure 10 

The FORWARD TOSS bug sometimes occurs on the second toss but usually 
appears on the third toss or at other points of difficulty. 


It should be noted at this point that the FORWARD bug is relatively 
ease to fix in comparison to another member of the timing class, the 
EARLY-LATE bug. The EARLY-LATE bug occurs a& the name suggests, when one 
or more of the TOSSES is early or late with respect to the evenly spaced.' 
rhythm of smooth juggling. Unevenly spaced TOSSES necessitate especially 
rapid for slow) EXCHANGES one step later which are usually teyond the reach 
of beginning jugglers. 


Figure 11 

In this example an early second toss has necessitated a rushed right 
hand EXCHANGE at the Same time a left hand CATCH fs required. 

The fixes for the EARLY-LATE hug all involve helping the 
subject to identify a good timing sequence for TOSSES, One way of doing 
this is a game called OROPCATCtf. In DROPCATch" the subject is asked to 
execute <a modified double toss in which the first ball is tossed as usual 
but the second ball is dropped or thrown onto the ground at the last 
possible instan t before the Incoming ball arrives. This of course focuses 
attention on the timing involved In holding on as long as possible. With- 
this newly discovered "timing point" fresh in mind [which is very similar 
to the correct time for EXCHANGE Initiation) it is frequently possible to 
achieve a good TOSS rhythm on the next trial. Perhaps a better method is 
to prevent the Occurrence of the EARLY-LATC hug as much as possible by 
Strongly emphasizing a proper rhythm during the demonstrations. 

A Taxonomy of Juggling Bujj S 

This section presents the main body of empirical mutts produced by 
the experiment i namely a taxonomy of juggling bugs and their associated 
f i . 

The preceding timing bugs, FQRWARD-TQSS and EARLY-LATE as well as the 
URONG-MENTAL-MaD£L bug mentioned earlier have been analyzed in detail 
(along with their associated fixes} since they are representative of the 
whole set of bugs encountered during the experiment. This set is summarized 
in the taxonomy of "juggling bugs" given in figure 12, An important aspect 
of the taxonomy has to do with notion of a good descriptive language. In 
this case and in the protocol in the appendix, the names in the taxonomy 
attempt to be both a name and a description of the physical event. This 
device has two fold purpose. It greatly simplifies the difficult task of 
real-time analysis of complicated movements. It also facilitates communi- 
cation about those movements, Fnr more information about the Insert ant 
but difficult problem of developing descriptive languages for human move- 
ment See Birdwhistle (5 ). 

Figure 12A 










POPPTWG (when it occurs 
on the first toss as Is most 
carmen). Later occurrences 
are usually caused by a gross 
error on proceeding cycles and 
hence disappear u+ien the 
prior error is fixed 


Figure 12B 



The next three bugs usually 
occur either as the result of 
a barf model of what a toss 
should look Tike or as the 
consequence of an earlier 
mistake somewhere in the 
proceeding cycles. In the 
first case the fix is a careful 
demonstration of a good toss 
along with verbal guidelines 
like "head high" or "shoulder 

In the latter case they usually 
disappear along with the 
prior mistake. (The VERTIC A!_ 
hug has a separate name because 
ft is so disruptive when it 
occurs that it deserves special 
attention) . 


Figure ]££ 

a) Demonstrate the dffference 
between underthrowing and 
aver throning. 

b) Pretend there's a "Urge 
pipe beside and "inside" the 
trajectory of the In Mining 
ball Into which you must 
throw the outgoing ball. 


This bug most frequently 
occurs on the 4th boss and 
may perhaps be due to the 
pedagogical sequence used 
which emphasises "getting to 
3" Or to an overly cautious 
"dynamic lo oka head" ability 
which predicts disaster on 
the upcoming toss. In any 
event the fix is to force the 
occurrence of the inhibited 
toss by either changing the 
toss cue tfa instructions 
like "don't throw until you 
hear me shout HDLJ " (this is 
a variant of DROPCATCH) or 
by removing the fear of 
failure by instructions like 
"don't worry about catching it 
for now, just throw it," 


Figure"! 2G 





The catching errors shown 
here almost always stem 
from poorly executed 
previous toss or sequence 
of tosses. Hence the 
fix in most cases involved 
correcting the offending 


Figure 1ZE 



(severe hip rotation but no step) 

(Usually found only after 
serious BACK or timing bugs} 


(involved one or more steps, lunges, etc. i.e., a last ditch attempt) 


figure 12F 














I* 1 1 

\ < 

» / 

Although all of the bugs 
exhibited so far are„ at 
some level b Accurately 

described as timing bugs, 
some are more directly 
concerned with the sequencing 
of major events, such as* , rather than details, 
within a given event such 
as a release within a given 
toss. Hence they deserve 
special attention and names. 
The fixes in all cases in- 
volve establishing a good 
major event sequence by 
games like DFLQP CATCH, coun- 
ting out loud, demonstrations, 

TIMING BLJES {front, ) 

Figure 12G 




I 1 
J \ 
1 1 



In the SLAP bug the outgoing 
TOSS is so late that there 
Is ho tfme left to throw. 
In a desperate attempt to 
both threw and catch at the 
same instant the hand 
compromises both tasks by 
"slapping" [while still 
holding unto) the supposed 
outgoing ball into the 
incoming ball. The fix 
for SLAPPING is to forbid 
the subject to do it {since 
it 1s a voluntary act). 
The ball must be either 
caught or allowed to drop 
untouched. This helps to 
focus attention, on the proper 
EXCHANGE initiation point. 

The Outlines of A Computational Theory of Juggling 

The preceeding sections of this paper were concerned with the 
presentation of empirical data about the skin of juggling approached fro-* 
a computational point of view. From that point erf view the experimental 
observations with respect to the role of model in physical skills* the most 
coranon and most severe bugs and associated fixes, as well as the notion of 
a complete taxonomy of bugs all have to do with what computational experts 
would call software aspects of juggling. By definition however, the compu- 
tational paradigm consists of a set of software instructions [by which the 
programmer fulfills his model of the problem, creates bugs, etc.} and a 
set of hardware devices by which those instructions are interpreted and 
executed. For people of course the hardware devices by which movement 
Instructions are executed are muscles, neural networks, the motor cortex. 
and so on. 

It is not possible at present to give a complete theory of this hard- 
ware device for physical "commands 1 .' It is difficult enough to give detailed 
theories about very small components of what might he called the "Physical 
computer" part of the human tody, The rest of this paper is concerned with 
establishing empirical constraints; on the range of theories which night be 
offered for the various components of such a computer ,■ The issues considered 
have to do with timing constraints, potential control structures k naming 
and the construction of subroutines. 

Timing Analysis 

It should be clear, a^rtitltmaTlj, speaking that jljqq ii nq 1 S a ™i 
tire process and further, that the strain* loosed by the physics 
involved ar* fairly severe, The folding discussion has as fts goal a 
Clair foliation of the timing constraints faced ty an accomplished juggler 
. as he performs his pontine. The constraints derived here coined with 
nhat is know: of human response times should be of considerable value 1 n- 
designfng a control structure for a theory of Juggling movements. 


Human Constraints 

It is apparently true that any willed skeletal movement takes approxi- 
mately ZOO n5 . to initiate. That is to say, when your head tells any part 
of your iody to move, Including the eyes, on the order of 200 ms elapse 
before the movement actually begins. Although there is some variation 
depending on the task, the individual, and to some extent the amount of 
practice, the 2Q0 m elapse latency 1s essentially invariant, particularly 
for large movements of the kind involved In juggling. Likewise the maximum 
acceleration for the hand is approximately 2400 in./sec 2 . 

These parameters can he verified empirically by the 
following two experiments: 

Try to catch a ruler dropped (by someone else} between 
your almost closed fingers. The z^ro mark should he 
even with the top of the hand. The hand should be 
resting on some fixed object so it can't "chase* the 

ruler. The fingers „ill usually grasp the ruler some- 
where between the 7 and S inch marks which corresponds 
roughly to a 2W m S delay. This explains why ft is 
nearly fusible to a catch A 6" dollar bill in the 
popular version of the experiment. 
The second experiment involves trying to swat a r«Ter 
dropped past some flat surface m e a taljle , Thg hand 
1s positioned 3" from the edge of the table and moves, 
when the ruler is dropped >in a direction which is 
horizontal to the table. The ruler is held so that 
the end is even with the tap of the table, tffth a 
little practice you can just nfck the end of the 
ruler as it goes by. Sfnce the ruler falls 12* tn 
■-250 ms N the hand moves 3" in SO ms, assuming a 
SOD m latency. This corresponds to an acceleration 
of 2400 fn./sec. . These examples are due to Lacy (6). 

Ballistic Constraint*! 

Given that the hand operates erltk a 200 n* latency and can accelarate 
at no nore than Zm u. n ^\ how do ^ ^^ ^^ ^ 

the Ohes noosed by the la*s of physics? The first step Is to determine 
the velocity of the hall in normal catching regions. For a person 6 feet ' . 
tall, a toss typically rises to a maximum height of 6 1/2 ftet ajlfl ^ USMlly 
caught at slightly more than shoulder level. This situation can b e modeled 
from the timing point of view as a ball dropped frjn a height of 6 1/2 feet 
and caught at a height of 5 1/2 feet. Thus the ball reaches a velocity of 
% fn./sec. and hence requires 250 ms to travel the 12 inches from the top " 

of its arc to the level at which ft usually. caught. If for some reason 
tne distance traveled becomes as much as 24 inches the velocity of the 
ball increases to 136 in, /sec. so that only 350 ms are require* to cover 
the longer distance. 

Basic CycTe 

It should be noted that the hand which is about to catch the incoming 
call (f,e. i r our model the dropped call) is currently occupied with tossing 
the outgoing ball and hence at the conclusion of that toss is very close to 
where it should be in order to catch the incoming ball, in fact if atten- 
tion is focused on say the jugglers left hand the following pattern of i»*e- 
mencs can be observed ♦ 



Figure 13. The Basic 

1 ft, hence 250 ms 



For each hand the basic cycle consists of 

-j TOSS 

and the shift part of that cycle has to cover only s or 4 inches for most 
tosses. By the second ruler ex peri merit it seems reasonable to believe the 
hand can cover 5UC h distances withfn the available 250 ms and hence that 
a kind of control structure one might call a "react iott" theory could 
support juggling rrovenents Satisfactorily, Such a theory would claim, for 
example, that at the end of the TOSS movement the eyes perceived the position 
of the incoming ball and "reacted" by issuing a cojimand to move the hand 
under the ball. Likewise the actual contact of the hand and the ball causes 
another reaction in the form of a GRASP movent. The central t*nent of 
a reaction theory of movement control 1s that the commands whfch initiate 
movements are issued in response to rather than in anticipation of external 

An Anticipation Theory 

Unfortunately the preceeding analysis does not include latency times 
for the various movements and as yet says nothing about the control of eye 
movements. These considerations make it obvious that a more complicated 
"anticipation" theory will be required for any control structure the overall 
theory might select. For example ,figureti presents a frame by frame analysis 
of a brief (a toss} juggling sequence. This analysis shows that the most 
crucial part of the EXCHANGE, the SHIFT movement, lasts on the 


Hgure 14 

Basic Cycle 

(330ms) SHI 


\ (165ms) 

IE 1 

(Hall travels 1 ft fn 25C nsl 

TOSS f36GHis) 

5HIFT movement initiated here early 
in the TOSS movement 


average 330- ms after ft begins. Of course there is a theoretical question 
as to whether or not the 330 ms consists of a 20Q ms waft and a 130 nts move 
or consists entirely of a 330 ms move. The empirical evidence from the video 
tape is clear on that point however. There 1s no observable pause it the 
end of the TOSS movement and certainly none of ZOO ms duration. Rather 
the whole TOSS, RELEASE, SHIFT, CONTACT, HOME sequence appears to be One 
smooth movement* 

LE?T HAim 


(11 j Upsweep 
ftel&aae — 

[7) Position for Cat eh. 

'3-ontac t , , w- 

(5} &raep| 


{4S} Hone fc R B *t 

right ir^^fn 

Downs weep (9) 

(11) Upsweep 


(1£) PoBi-tion far Catah 

(4) iiraflj>| 
(+1) Home 1 Hant 

llEWvaep ( 1 1 ) 


Home £ Rest (4o) 


(10) Upsws e p 


(5) Position fur Catch 
on tac ■ 

(5 j CiraspJ 

Horn* (25) 
(3SJ & 

Rest (14} 

(11} UpS-rffiOP 



(1 > J j Position fir J.^i.nL 

V nnt«lC t — n. - 

flHs6 (S3) 

(42.} 4 

R*»t (9) 

Upkeep (12) 


Posl^l^n for Catch (1u) 




] j rno p { 6 j 

Kocie £ Rest (4?) 

Upowaep (10} 

' ■ Release 

^SSlti&n Tor Catch (10} 

— I „ ,,_ — Contact 

1 |ara B p (5) 

Hone i Seat (4E) 



Upaweep (10) 


Position for Catch (9) 

i - Coaiac t 
Gr*iap (6) 

Home ft Rest (42) 


Fig, 15. Tiding Analysis of £asis_ Cascade 
Juggling C.ycfe (For an accojnplished Juggler) 
Times, are given in terms of video frames and 
are indicated in parentheses for- convenience 
One frame - 1/30 sec ■ 33ms + So for example 
the 1st UPSWEEP by the right hand tools 
It frames, I.e. 363 ms. 

The only way this could have occurred from a control point of view fs for the 
SHIFT iiovauent to have been Initiated so™ 160 ms Into the TOSS movement; the 
GRASP movement to have Initiated 130 ms into the SHIFT movement, etc, (see 
figure rd&h 

To return to the theoretical discussion T the Issue becomes forced when' 
eye movements are- considered. At least one eye fixation is required to deter- 
mine the trajectory of the ball in flight. Yet eye movements also have a 
200 ms latency So seeing the ball and causing the arm to catch it Involves 
at least a 400 ms delay assuming an instantaneous trajectory calculation 
and zero time for the arm movement. But the model allows only 250 ms for 
the ball to fall the 12 inches Involved in a normal catch. Doubling this 
distance adds only 100 ms or so to the time available. Thus it is clear 
that the SWIFT movement is anticipatory in nature in the sense that It 
must have been ."planned," 1,e. , initiated, during a prior movement, notice 
that the preceding discussion does not consider the possibility of quick- 
reaction low-teveT feedback loops wnich obviously could affect the details 
of the analysis, Nevertheless the primary conclusion still stands. Any 
theory of Juggling movements must contain a control structure which is ■ 
capable of supporting parallel, real-time, events of an '■anticipatory re- 
sponse" nature. 



Toss J*,hal»d 


"fcss bftj.ns 

(360 ms? 

(330 ma) 

OWoa Odcws M«nt Begins 
Grasp Begins 

(IBS ms) 


IH . . 

J*ti. h 

•■■•.. j 


**0- ■ 

IBS ■ 

dOCJ m*J 

HI > • 

36*0- r 

I 1 00 - - 

i](KJ, . 

(?00 nigl 

Shifl m-^et 
C2C0 ms) ' 

Grasp Inured 
h+jmn n haied 

(300 n-^1 


■ ;• ■ - 

>34— -7 

jugs Come In Bunches 

The neat e*peri rental observation also has to do with timing information 
but in a dffferent me. The observation is that hugs apparently come in 
bunches, at least that Is tHK for the juggling variety anyway. Its probably 
safe tD say that above the DOUBLE T0S5 level most of the subjects were Initially 
grappling «i th two or more simultaneous bugs. One fluent combination- was 
mu. LATE/FORWARD. The protocol in Appendix I cattalrs many other eumplH, • 

This observation is significant for both theoretical and pedagogical 
reasons. One of the most important differences between mentaT and physical 
events is the fact that alT physical events reduce ultimately to So™ pattern 
of muscle start and stop times. Although 1t could likewise be said that 
all mental events reduce ultimately to something like current flow, in some 
deep sense it seems to be true that physical events are considerable more ' I \ 
ti me dependent than mental events. We have seen for example the whole 
taxonony „f bugs which tan be generated from subtle shifts in the timing 
of events in the basic juggling cycle. It is perhaps reasonable to conjec- 
ture then that simultaneous b-jgs are due to a combination of timing mis-speci- 
fications in the learning process and the narrow range of permissible 
specifications afforded by physical constraints, The Important theoretic*! 
implication Is that any theory of movement must have a mechanism, perhaps 
a whole language, which is capable of describing and modifying timing events 
at the millisecond level. 

Pedagogical ly the observation 1s important because 1f, as conjectured. 

dueling f 5 a rHisora^c represents™ o' physical ski I Is <n general . t.hen 
the indicated experimental direction is the creation of learning environ- 
ments in which the same bugs occur, but one at a time. For juggling, this 
might take the form of a machine which allows EXCHANGE practice in slow 
motion, or as a separate entity, EXCHANGE bugs could -then be dealt wfth 
one at a time independently of the rest of the process. 

The Breakthrough Phenomenon^ 

The single most surprising discovery made during the experiment uas 
that for most individuals progress, as measured by either the number of the 
quality of successful tosses, was not at all continuous but P ra the reappeared 
to consist of a series of breakthroughs or plateaus. These breakthroughs were 
not initially stable, frequently being attainable for only one or 
two trials, but became more frequent and stable over time. The really 
astonishing piece of information is that the magnitude f measured in tosses) 
of the breakthrough was usually 2 - 4 times that of the previous level J 
Jumps from 3 tosses to ID or 11 tosses were common. Se* for example the 
sixth minute of £1 or the eleventh minute of 52 in figure 17. Almost every 
subject in the experiment exhibited a breakthrough which virtually doubled 
the previous best effort. 

One explanation of this observation is that several bugs vanished or 
were temporarily overcome simultaneously thereby allowing a dramatic break- 
through to occur,' The fact that the breakthrough is not initially stable 
but improves with time suggests that some form of trial and error or hill- 

dinting search is going or. a the lower levels. of timing specification. 
This argues against a strictly syirbalie naming/addressing mechanism in the 
overall theory. 








l l l l l l 

(\f T~ 

O 01 

f i i i i 


















r**^ ru t- 


I I i I i i i i i f 

«» lf\ ^J- i*\ CM t- O cr. (fi C- 



■ .-■■ 










It 1 5 Hard Tp Stye Yourself Advice 

Hera's an observation which is true only sometimes and 1s intriguing 
both when Tt does and when it doesn't apply. The question is why is jt 
hard to give yourself advice when learning new physical skills. . Tne most 
extreme e^mple of this situation is the disastrous ."blind trials- approach 
herein the subject feverishly tries the task over wd over without even 
taking the trouble to describe his mistakes, let alone advise himself on 
how to fix them. 

In many cases though, thoughtful analysis has 
produced a reasonably accurate description of the 
mistake but It's still difficult to translate that 
information into corrective action, For example, 
most people -can sucessfully teTl themselves to 
"toss higher' 1 . (to gain mire time) but apparently 
it's quite difficult for many people to telT 
themselves "toss later" or "toss right before 

It's also f (iterating to note that frequently the best advice you can 
give yourself appears on the surface to have nothing to do with the task 
at hand but in reality corrects exactly that portion of the movement which 
was previously erroneously specified. For eaamole in tennis the advice 
"Serve as though you were trying to throw your racquet away But nold on " 
frequently helps the novice master the service movement more quickly. In 
juggling the advice "Practice white facing a wall" is frequently useful for 
correcting the FORWARD and other bugs, 

Tbe theoretic* 1 implications of this observation have to do with the 
mechanisms by which existing movement specifications can be brought to bear 
on the task of learning new movements. The fact that existing knowledge is 
relevant to similar task should he obvious and has been demonstrated expert 
mentally by tasks like learning to write with the non-preferred hand. The 
apparent ability to advise yourself on certain tasks but not on otters 
raises questions about the mechanism by which advice is accomplished and 
the degree to which it is a symbolic process. Thus the overall theory of 
physical movement must explain why the symbolic advice "throw this way" 
(in the serve example) is useful and can be acted upon, but the equally 
symbolic advice "toss earlier" (in juggling} is only marginally useful. 

The Construction of Subroutines 

A constant theme which runs throughout the exoeri mental data has to 
do with the question of how "chunks lh of movements are grouped into Targer 
chunks. Computationally this is the Issue of suh-procedurization or 
subroutine construction. It appears to be perhaps the most central Issue 
in designing a theory of physical movement as well as one of the mast 

Part of the difficulty lies in ascertaining the real primitives 
of movement specification and the decree to which the process 1s 
symbolic. At least some of the problems posed by this gao in our knowledge 
can be avoided if we- assume that for advanced skills [lite Juggling) the 
relevant aspects of the component movements are predominantly symbolic and 
hence can be treated like ordinary subroutines. Of course this assumption 
which may be entirely unwarranted, merely postpones the issue of what 
happens inside a given subroutine. However it allows us tD begin the task 
of describing the overall learning process from a symbolic point of view. 

For example T 1n juggling t the basic learning process apparently centers 
around combining the first part of an existing TOSS subroutine 




with the body of an existing CATCH subroutine 


and Inserting MOVE'S as appropriate to produce the EXCHANGE procedure ■ 



where TOSS* is TOSS minus the FOLLOW-THROUGH line 


Notfce that the final version EXCHANGE cannot be a simple linear 
combination of TOSS, MOVE and CATCH. As previously discussed, the latency 
times and the real world constraints rule this out. It is certainly true 
however that, as in Sussman's thesis (see reference 7) the linear approxt- ' 
(nation subroutine 1s a good initial mafe^ f r ow which to start, i.e + there Is a 
good chance that the debugging process Kill converge to the desired final 
structure. The linear modsl idea also brings to mind the discussion of . 
cascade guggling models in section 4 r Sussman and Goldstein (8 ) also 
,-roYide clues as to how' that convergence process might go 1n different 
problem domains. The degree to which those rethods apply to the domain 

of movement depends a great deal on the extent to which moveitent primitives 
are symbolic. Issues such as annotation, invocation and patching however, 
appear to be good starting points for the discussion of whether or not the 
Goldstetn/Sussman theories can be applied to physical procedures. In any 
event it appears to he true at least it the more advanced levels, that 
physical movements can successfully be treated as subroutines from a 
computational point of view. 


Suumary and Concision 

In summary, *e have seen in "cons f derail e detail the extent to which 
the computation] paradigm can be applied to the skill of juggling, a skill 
which is hopefully representative of most forms of higher-leveT skill. The 
di session initially focused on the external or performance aspects of 
juggling acquisition . A descriptive language for juggling movements* 
formulated in bug/fix terminology was developed to the point of exhibiting 
an extensive taxonomy of cascade juggling bugs. 

The latter part of the paper was concerned with establishing the gross 
outlines of a theory of the internal mechanisms by which movement commands 
are interpreted end executed. Empirical evidence which was found useful 
in this regard included an observation about the importance of having a 
model of the physical task at hand, a study of human ballistic constraints, 
the unfortunate fact that hugs come in bunches, the very surprising break- 
through phenomena and the Intriguingly difficult task of advising yourself 
during physical skill acquisition. 

This evidence shows that while physical skills have an obvious symbolic 
component, the precise extent to which the overall theory should be symbo- 
lically oriented is not clear. Evidence which supports the symbolic point 
of view includes the data on models and the successful instances of advice. 
Evidence which argues against 8 completely symbolic theory includes 
the unsuccessful advice cases as well as the "trial -and-errar-1 ike" 
parameter search which is assumed to underlie the breakthrough phenomenon. 

The issue of control structure is sccnewhat more dear-cut. Human and 
physical constraints make ft obvious that a theory of movement must Include 
a control regime which is capable of supporting real-time, parallel events 
of an antic Ipatbry nature. There must also be a mechanism for describing 
and modifying event times at the millisecond! level. The present experiment 
yields little information about feedback mechanisms. 

The theoretical outline can eludes with the observation that the central 
issue appears to revolve around the problem of how existing movement patterns 
(subroutines) are combined to fonti new movements. This statement is confirmed 
by a subroutine analysis of the Juggling learning task* The procedural 
modification theories of Sussman and Goldstein appear to be partially 
relevant to physical skill acquisition from that point of view. Hence the 
issues of invocation, annotation, and self- debugging appear to be good 
topics for further research. The non-symbolic nature of movement primitives 
appears to be a central issue in such endeavors. Despite the existence of 
these problems however, it is felt that the primary conclusions as to the 
similarities of physical and mental activity and the relevance of the 
computational paradigm are thoroughly justi fieri, 


Protocol for SI* 

1] H-Show me your model of juggling. Make the bags move as though you were 

2) Si-Hell let's see. You thro* this one (f,e. the left one) in the air. 
When it reaches its high point you then begin (the transfer operation?) 

3) E-Just make them move through the motions., 

4) St-This one goes up [i.e. the left one) and this one goes over fi.e the 
nght one J . 

Tosses L-+R, passes R-*L r 

5) E-?ou have the model of juggling that most people have, namely this 
{demonstrates 'shower" juggling}. You can_ jugg le that way but it's 
much harder,. I'm going to teach you a new kind of juggling Let's 
get some terms. This 1s a toss f demonstrates a toss), This is a pass 
(demonstrates a nass). There's no passing in the juggling I'm about 
to do. It s all tosses (demonstrates cascade juggling), 

6) Sl-Un huh. 

7) E-I'd like for you to do this. One in each hand TOSS ft L h TOSS L R 
How notice I don't want simultaneous tosses (demonstrates) and I don't 
want al most simultaneous (demonstrates!. I want a distinct cadence 
toss, toss. 

8) unintelligible 

9) 51-L^R, 

R-*L».|»ss, dropped, 

10) E-You iPfl&ted 

11) SI -Yeah. 

-12) L-*R "hands lip" 
R-*L wide, forward. 


13) L+U fingernail 

R-+ L wide* forward. 

14) E-Toss the second one more like that (demonstration) 

Sl-I'm used to tossing with this hand (left) but not this one (i.e. right). 

E-Toss it in a nice parabolic arc. 

*In this protocol 51 is the subject, E is the experimenter t L-*fl is a toss from 
the left hand to the right handXmeans a satisfactory toss and terms like "hand- 
si Ip", "wide," "fprward" the names of specific bugs from the taxonomy , pgs.17-2?, 

15} U + L J (i/meam satisfactory TQBS), 
l&j R+L^ 

1?) E-R1ght H now do the exchange, 

1BJ R+L slightly forward, handslip 

L->R back, "p^ch catch" (i.e. thumb and ffr^t two fingers). 

19) L+Rv 

R^L sHghtly wide, slightly forward, 

20) E-O.K., quit. That took exactly 2 1/2 nrfnutee Khu a. *l_ -. 

£1) L-Rv/ 
R-*L Tow. 

22) L-*R k> 


23) L R back, dropped, 
F L very forward. 


24) L^fi ^ 

R-* L forward 1 

Sl-I have trouhle tossing with tMs (right) hand. 
£5) L-R *-" 

E-Please talk out loud. 

26). L^R W 

R->L fingernail 

27) E-L»k S good to ™. NM what I'd like you to do f S *,«, you'r* at tho 

stage when the second one 1 * coming down, toss the third one (demonstrates) 
it goes like this (slow demon-strati on), 

£8) L^fi slightly forward 

R**L very forward, dropped 

L-^R wide and forward, caught by stepping forward. 

E -Almost, 

29) L-*R ^ 

R^L very forward* very wide, stepped forward, dropped 
L^R caught directly over shoulder due to previous step. 

E-Conment ta cameraman, 

30) L^R *•* 
R-*»L vertical 

L^R very forward, lunge cateh 

E-Your second toss went straight jp and straight down. 
SI -Yeah 

31) L-*R *-" 

R-*L slightly back, dropped 
L-*k slightly forward 

E-Ah you had it but you dropped the second one. 


SI -It's hard for me (to concentrate on catching?). 

32} L -► R *■*" 

R-*L slightly hack, dropped 
L^fi very forward* lungecatch 

E-You're getting a very nice cadence there, femember that. cadence. 

33) L-sR^- 
. R-*L slightly back, chestcatch 
L-*R fingernail 


E-O.K., quit. That took you a total of 3 minutes, 35 seconds. 

34} Normally we have people stop here because there's about 3 important 
points in juggling; 3. TO and 100. But you got there so quickly we'd 
like to have you go on. Let's see how quickly you get to 10 

E-Expla1ns how to keep going, the apex rule for tlmfno th F n ^t **,=* 
and about stopping. • tl[ "Tng uie next toss, 

£-R1ght, keep talking out Toud. 

35) L^L high, vertical, caught with left hand (unintell fgible) 

36) t_^R very high, back [unintelligible) 

37) L-*fl high back 

R^L forward, fingernail 

Sl-It's Tfke throwing tennis serves, 

38) L-*ft high, slfghtly back, dropped 
K ™ L 

39) L^R^ 

R^L late, caught Kith R hand, 

40) L» ft vertical, chestcatch. 

- . 

41) E-.Notice thai it ccnei and g«s, All learning aoes Tike that f w *™ 
thin, gees wrong then it „ back for ."oSpW IS^SL *" 

SI -I know how that is + 

43) L^R flngern.riled 
R^*L late 
L-+ R back, wide , dropped 

■ " * L llll fslapdrJJ?)'^ 1 ^ ° Ut ° f hl " S hand wh1le «* chlT| 9 the th1 ^ 

E-O.K. , you almost got Hrj of the fourth toss. 

43) L^ft back, hands Up 
R*L late, very forward 

44) 1) L^fi V 

Z\ R^L very forward, lungecatdi 

3) L^R late 

4) R^L late, back, leanback catch 
| J L-^R very forward, lungecatch 
of ft-*L very forward, lungecatch 


7) L-»R fnot visible) 

a) R^L (partly visible) slightly back 

9) L^R (not visible) 

10) R*L (partly visible) lungecatch bounced loose , chestcatch 

11) L-'R very back, stepback- catch. 


45] E-That took a to^il af 6 minutes IS seconds, O.K., let's see What fine 
timing looks like. Comments on juggling as an unstable (in the 
mathematical sense) process (advice about "laying" a toss up beside the 
incoming one and a demonstration) f 

45) L-*fi *■** 

fr*L late, vertical* caught with R 

47) L^R^ 

R-*L late, vertical, caught with R 

A little more flip, you don't want to wait so long. 

4a) L^R-^ 

R-*L forward, lungecatch 

E-You have too much follow-through 1n your left hand there. 

46) L-*R slightly back, fingernail 

R^L forward, lungecatch 

E-Right T much more like that. 

50) L->R-^ 
■ R-*L forward t lungecatch 

51) L**R 
R^L late, forward V lungecatch. 

52) E- Ad vice ref. "popping "theory ( dentins trati on. 

53) Si-Practices popping (six times) 

E-O.K. Hit it. lfou've al ready done IK 

54) L-*RV 

R-*L forward, lungecatch 

55) L-^R ** 

R-K slightly wide. 

56) L^R^ 

R ■* L wi dfi , 

57) L^Rv^ 

R-*L slightly forward, sHghtly wide 
{unintel ligible) . 

58) L"*R' 



R->L w--" 
■ E-Q.K.. Go ahead. You've dene 12 already. 
50) L-^R vertical, taught with L. 
61] L-*fi vertical caught with L 


62) L*R low, hands lip 

R^L later vertical caught with R h 

Sl-unintel 1 f -g i h 1 e . 

63) L^R late 
R"*L late 

L-^R slapthrowt both dropped 

E- Don't slap it. Toss tt higher. 

64] L->l/ 

R*L hack, handslip 

L-*R vertical, caught tuith L 

51 -(Getting in trouble because r'm not getting them out far enough?? } 

65] L~*R fingernail 
L+R very lew, very forward, lunge catch 

E- fou; keep getting lower* 

66) 1) L-*R 

1} R^L late, forward 

3) L-*R forward, lungecatch 

4) R + L hack 

5} L^R almost vertical, slightly low 

6) R*L late 

7) L+R 

S] R^L late, che&tcatch, dropped 
9) L-*R two handed catch 

S1-0.K., [ H m not getting the tosses (unintelligible). 

E-That's O.K., you're getting good, Keep doing that and you'll clean 
it up yourself. 

67} 1) L-R^ 

2) R-H. forward, lunge catch 

1) L^R slightly vortical, bands lip 
4} R-*> 1 ate* two handed catch 

E-Most people don't use the height enough to cheat and get more time 
yp there, 

66) 1] L-^R fingernail 

2) ft^L forward, lungecatch 

3) L-»R forward, lungecatch 

4) R-*L very wide, leancatch 

5) L-*R not visible 

6) R*>L not visible 

7) L-*R not Visible 

69) L R vertical. 

slightly forward and left 

wide, rotates, left to catch it 
vertical, rotates even further left 
not visible 

L-*R not visible, apparently slightly vertical 

R-H forward 

L->R vertical, forward, lungecatch, dropped 

R^ L forward ,. lungecatch 

E-Looks good to me. 

SI -Still moving around too much. 

71) 1) L^Rv^ 

2) R-H forward 

3) L-*R forward. leaneatch 

4) R^L forward, lunge catch 

5) L-M? forward, lean catch 

6) R-*L forward, handslip 

7) L"*R forward, lunge catch 
. 8) R-^L very back, dropped. 

72) E-O.K., you do in fact have some 1nd1 cation of a migration hug. The 
way jou fix a migration bug is face a wall (demonstration, more Advice 
ref. migration bug) 

73) (now ag-ainst wall) 
L + R wide* back 

74) L^RV' 
R-?L passes! 


75) L->S wide, fingernail 
R^ L hack, handslip 

L - * R late, low, hits the wall 

76) 1) L-»R fingernail 
2) R>Lv 

3} L-* R fingernail 

4) R^L wide, leancatch, fingernail 

5) L^R forward 

6} R-*L forward, leancatch 

7} l_-> R back* chestcatch, dropped 

8) R->L wide, fingernail 

9) l»R slightly vertical 

E-The one that got you was the ore that came back. 
SI -Yep. 

77) L^R forward 
R->L wide 

78) L->R vertical T caught with L 

E-They tend to go straight up and down, 

79} L^R hands lip 

R^ L vertical, caught with R. 


R^L vertical, fingernail t chestcatch. le^niack 
L^R forward, hands! ip 
R^L leancatch 

81) L+R fingernail 
L-*R forward 
R-H hands Up 

L-*R vertical, forward, almost hit the wall 
R*L wide, forward, leancatch. 

52} L^R--" 
R^L *-"" 

L^-R forwards vertical 

R-*L slightly wide, handsli? 

L->R early forward, caught against wall 

R-*L hands! 1 p. 

83) E^O.K., quit wasting film on him. That took a total of 11 minutes 
SI -I was just getting the hang of it + 


.1) Wlrcograd, T>, Understanding natural Language . Academic Press (1972). 

2) H1 fishy, IL , Ed ,, . Semantic Information Processing . M.J ,17 Press (196a] . 

■ , 5 

3) MTfisky $ Pap&rt t Artificial Intel licence , Condon Lectures, Oregon State 
System of Higher Erfu cation , Eu ge n e , Oregon (1974} + 

4) Carlo, The Jugglfng t .Random House (1974 J. 

5) Sfrdwhistle. - Kinesics and Context , U. Fenn h Press. 

o") Lacy,: R.» " Juggling" (unpubl ishaed paper), rLI.T. A.I* Laboratory 
[Dct>.1971). ■ .j_, . ■ i . 

7) SuasHian., G.J., k Coirajtaticna l Model c* Sri ill /inquisition. rtI-TR-257 , 

M , .1 ♦TV A , I , . Laboratory [Sept-, l974).-~± - L--. \-\ ■ — 

(J) Goldstein, I .P. v Understanding Single Picture Programs, AI-TR-294, M.I-T- 
*♦!/ Laboratory/ Sept. 1974 J," .-■-■'r^. ,.-