106 TEACHERs COLLEGE COLUmbIA UNIvERsITy
Paul the Robot as a Naive Drawer
continuously publish new components reusable for
other projects. Paul is currently using YARP as a
framework. The choice of YARP was following the
participation in the BCBT20096 summer school,
where YARP was introduced.
For Paul’s successors the system is being ported
to ROS as it is now a more complete system, seems
to be becoming a standard in robotic research
around the world and also provide a large range of
reusable components. An overview of Paul’s current
framework is presented in Figure 3.
Paul’s sketching cycle
Overview- Scan the environment by moving the eye
until a face is detected. Use a line feature
detector. - Focus the eye onto the sitter.
- Limit the region of interest (ROI) to a close-
up framing. - Convert the ROI to gray levels.
- Make an image pyramid with n levels.
- Draw salient lines with increasing precision.
- Perform the shading behavior.
- Execute the signing script.
- Human operator mechanically and boringly
detaches the paper, gives it to the sitter and
puts a new sheet of paper while Paul cools
down and waits for the next sitter.
Discussion
Since June 2011 Paul has sketched more than
450 individuals and has been exhibited in various
locations (London (a few times), Istanbul, New
York, Camden (Maine)). The response from the
audience has been very positive, especially when
considering the responses from artists and other art
professionals. It is surprising that such a relatively
simple entity can produce so constantly interesting
and often unexpected drawings.
We can propose a range of hypotheses that could
explain why Paul draws relatively well.
An artistic evaluation. Patrick Tresset, Paul’s
main developer, is an artist who practiced drawing
extensively and as such has used his insights to craft
the program. Patrick has also evaluated Paul’s out-
put as if it was his own, and has used this feedback
to adapt the software until the output became satis-
factory. As such we could assume that Patrick has
taken charge of a part of the drawing process. The
drawing should be seen as the result of a collabora-
tion between a human artist and a robot.
A naive drawer. Paul is a naive drawer. Paul has
no memory, no concept of what a face is. As such
his drawings are not negatively influenced by the
knowledge of a subject, what Van Sommers calls the
“conceptual bias” (Van Sommers, 1984).
The depiction of salient features. When Paul
draws salient lines they are extracted from the
response of Gabor filters which are accepted as
good simulations of simple cells in the early visual
cortex. In computational models of visual attention
such as Itti’s (Itti & Koch, 2001), Gabor filters are
used to build one of the saliency maps. Areas that
display high orientation disparities are the salient
regions. In effect Paul puts an emphasis on regions
that would be expected to be perceived as salient.
The influence of decisions based on visual
feedback. Although there is very little use of feed-
back data to control the otherwise random shading
process, it seems that this is sufficient to produce
patterns that are perceived as not due to chance,
and as such perhaps richer in emotive content for a
human observer.
Paul’s physicality. Like a drawing produced by
a human, Paul’s drawings are the results of move-
ments: as such they are the record of a process. We
can hypothesize that this adds to the richness of the
experience when an observer is reading the sketch.
Furthermore due to the configuration of Paul’s arm
the type of errors it makes might be perceived as
rather natural by a human observer.
These factors are at this stage hypotheses that
will be investigated in the near future. It is even pos-
sible that the combination of these characteristics
facilitates the production of satisfactory drawings.Beyond Paul
The successors to Paul will in the near future
evolve in two directions. On the one hand some
will evolve towards being more theatrical and be
presented in an art context. Their the first evolution
will be to shorten the time it takes to do a drawing
from 20-30min to 10-15min and tell the sitter that
they have to sit and remain still for the duration
of the drawing session. This will require that Paul
keeps the sitter interested, for example by pretend-
ing to look at the sitter more often, and perhaps by
doing some entertaining actions or gestures. On the
other hand we are continuing work on the research
version that will provide a more accurate model of
the sketching activity. As such the system will make