Behavior-Based Early Language Development on a Humanoid Robot



Figure 7: Activation of Kismet’s motorskill internal variable (solid line) and the “orient” fixed concept (ragged
line). Motorskill = 0 corresponds to the activation of the orienting behavior.


The simplicity of the update rules is possible be-
cause of the finite and very precisely described set of
pre-existing releasers on Kismet. For a more scalable
system, if we wish to maintain natural-time respon-
siveness, we will want to consider more sophisticated
techniques.

5.2 Acquisition of novel concepts

ConceptMap is the single behavior which acts as a
very high-level manager of concepts. Described as
such it sounds like precisely the homunculus which
we are trying to avoid by creating this distributed
architecture of behaviors. However, the functional-
ity of
ConceptMap is very tightly tied in with that
of
HeardThis and should perhaps be implemented as
two extra states within that behavior. That would
perhaps be more acceptable in terms of the spirit of
the project. The current implementation has singled
out
ConceptMap mainly for the sake of readability.

The vocal labels of concepts serve also as the key in
the
ConceptMap and may be used to remove a concept
by finding the vocal label key. The map stores point-
ers to all
Concept behaviors, i.e., all non-grunting,
non-babbling protoverbal behaviors. The map is ini-
tialized to contain a small set of fixed concepts, which
correspond to Kismet’s existing behaviors.

Every time a new spoken input is heard, a new
Concept will be generated by ConceptMap, with the
heard phonemic string as its vocal label, unless in-
hibitory signals are received by
HeardThis via any of
its
N input connections. When a new protoverbal be-
havior is automatically generated from
ConceptMap,
its vector of
Receptors is initialized to include one for
every kind of releaser that is present in the system.
All receptor gains are set to a default value. This
exhaustive assignment will be remedied later as the
gains are updated individually within each behavior.

6. Preliminary Experiments

In (Varchavskaia, 2002), evaluation of ConceptMap
workings, the acquisition and labeling of concepts
was hindered by poor phoneme recognition and the
often unpredictable nature of the robot’s behaviors.
Here we have attempted to remove some of these
obstacles. We have created an initial small vocabu-
lary of 20 words, chosen arbitrarily as most often
used in interactions with the robot. This vocab-
ulary was transcribed phonemically and used with
ViaVoice exclusively for the purposes of more con-
sistent phoneme recognition. This initial vocabulary
was augmented with novel phonemic strings when-
ever novel concepts were added to the map. In this
way, the robot is biased to hear the “words” it knows
again, and we can eliminate much of the spurious
creation of new concepts.

6.1 Labeling an existing behavior

During a 15 minute interaction, the human teacher
attempted to teach Kismet labels for some of its
original behaviors, such as orienting, search, and ex-
pressing an emotion. The words used for this pur-
pose were not present in the initial ViaVoice vocab-
ulary. Figure 7 shows the correspondence between
the activation pattern of the “orient” concept and
the robot’s internal variable representing current mo-
torskill. The table below shows the progression of la-
bel updates within the “orient” concept; the teacher
wanted Kismet to associate the phrase “look here”
with the orienting behavior.

cycle:  last heard:    label:         C:    d:

200:    h uh k ih h    uh k iy       0.3   3.3

400:    h uh k iy      l uh k iy h 0.9   1.2

600:    l uw           h uh k ih     1.1   2.8

A cycle refers to one scan cycle of the vocal behav-
iors system, in which each behavior has exectuted for
one state, and takes an average of about half a sec-
ond. Figure 9 shows that appropriate labels were
associated with fixed concepts, e.g.,
/l uh k iy h/
with “orient”, /w ah n/ with “seek” (in this case,
the robot was taught the label “wanna” for the seek-
ing behavior).

6.2 Acquiring a perceptual category

During another 15 minute interaction, the human at-
tempted to teach Kismet the English words for the
colors green and yellow through the following sce-
nario. The teacher shows a toy of the corresponding
color and says the word when Kismet’s attention is



More intriguing information

1. CAPACITAÇÃO GERENCIAL DE AGRICULTORES FAMILIARES: UMA PROPOSTA METODOLÓGICA DE EXTENSÃO RURAL
2. EXECUTIVE SUMMARY
3. The name is absent
4. The name is absent
5. Draft of paper published in:
6. The Veblen-Gerschenkron Effect of FDI in Mezzogiorno and East Germany
7. The name is absent
8. Pupils’ attitudes towards art teaching in primary school: an evaluation tool
9. DEVELOPING COLLABORATION IN RURAL POLICY: LESSONS FROM A STATE RURAL DEVELOPMENT COUNCIL
10. Biologically inspired distributed machine cognition: a new formal approach to hyperparallel computation
11. The name is absent
12. A COMPARATIVE STUDY OF ALTERNATIVE ECONOMETRIC PACKAGES: AN APPLICATION TO ITALIAN DEPOSIT INTEREST RATES
13. The name is absent
14. The duration of fixed exchange rate regimes
15. Regulation of the Electricity Industry in Bolivia: Its Impact on Access to the Poor, Prices and Quality
16. Antidote Stocking at Hospitals in North Palestine
17. Human Resource Management Practices and Wage Dispersion in U.S. Establishments
18. The name is absent
19. Willingness-to-Pay for Energy Conservation and Free-Ridership on Subsidization – Evidence from Germany
20. The Response of Ethiopian Grain Markets to Liberalization