atively easier instead. The task nonspecific nature of
a developmental program is a blessing. It relieves
human programmers from the daunting tasks of
programming task-specific visual recognition, speech
recognition, autonomous navigation, object manip-
ulation, etc, for unknown environments. The pro-
gramming task for a developmental algorithm con-
centrates on self-organization schemes, which are
more manageable by human programmers than the
above task-specific programming tasks for unknown
or partially unknown environments.
Designing and implementing a developmental pro-
gram are systematic, clearly understandable using
mathematical tools. Designing a perception program
and its representation in a task-specific way using a
traditional approach, however, is typically very com-
plex, ad hoc, and labor intensive. The resulting sys-
tem tends to be brittle. Design and implementa-
tion of a developmental program are of course not
easy. However, the new developmental approach is
significantly more tractable than the traditional ap-
proaches in programming a perception machine. Fur-
ther, it is applicable to uncontrolled real-world envi-
ronments, the only approach that is capable of doing
this.
Due to its cross-environment capability, the SAIL
robot has demonstrated vision-guided autonomous
navigation capability in both complex outdoor
and indoor environments. The Hierarchical Dis-
criminant Regression (HDR) engine played a cen-
tral role in this success. Although ALVINN at
CMU (Pomerleau, 1989) can in principle be applied
to indoor, however the local minima and loss of mem-
ory problem with artificial intelligence make it very
difficult to work in the complex indoor scenes.
The SAIL robot has successfully developed real-
time, integrated multimodal (vision, audition, touch,
keyboard and via wireless network) human-robot in-
teraction capability, to allow a human operator to
enter different degrees of intervention seamlessly. A
basic reason for achieving this extremely challeng-
ing capability is that the SAIL robot is developed
to associate over tens of thousands of multi-modal
contexts in real-time in a grounded fashion, which is
another central idea of AMD. Some behavior-based
robots such as Cog and Kismet at MIT do online in-
teractions with humans, but they are off-line hand
programmed. They cannot interact with humans
while learning.
The perception-based action chaining allows the
SAIL robot to develop complex perception-action se-
quences (or behaviors) from simple perception-action
sequences (behaviors) through real-time online hu-
man robot interactions, all are done in the same
continuous operational mode. This capability ap-
pears simpler than it really is. The robot must
infer about context in high-dimensional perception
vector space. It generates new internal representa-
tion and uses it for later context prediction, which
is central for scaling up in AMD. David Touresky’s
skinnerbot (Touretzky and Saksida, 1999) does ac-
tion chaining, but it does it through preprogrammed
symbols and thus the robot is not applicable to un-
known environments.
7. Conclusion
For a robot, every action is context dependent, i.e., it
is tightly dependent on the rich information available
in the sensory input and the state. The complexity
of the rules of such context dependence is beyond
human programming, which is one of the fundamen-
tal reasons that traditional ways have been proved
to be extremely difficult to develop robots running
in a typical human environment.
We introduced here a new kind of robots - develop-
mental robots that can develop their mental skills au-
tomatically through real-time interactions with the
environment. Motivated by human mental develop-
ment from infancy to adulthood, the proposed the-
oretical framework have been proved on the SAIL
robot in multiple tasks, from vision-guided naviga-
tion, grounded speech learning, to behavior scale-
up through action chaining, all learned and per-
formed online in real time. The main reason be-
hind this achievement is that the robot does not rely
on human to pre-define representation. The repre-
sentation of the system is automatically generated
through the interaction between the developmental
mechanism and the experience. We believe what we
have achieved is a starting point of the promising new
direction of robotics. While there are yet plenty of
practical questions waiting for us to answer, it opens
a wide range of opportunities for future research.
Acknowledgements
The work is supported in part by the National
Science Foundation under grant No. IIS 9815191,
DARPA ETO under contract No. DAAN02-98-C-
4025, DARPA ITO under grant No. DABT63-99-
1-0014, an MSU Strategic Partnership Grant, and
a research gift from Microsoft Research and Zyvex.
Many thanks to M. Badgero, C. Evens, J.D. Han,
W.S. Hwang, X. Huang, K.Y. Tham, S.Q. Zeng and
N. Zhang for their contributions to the SAIL and
DAV pro jects as referred.
References
Almassy, N., Edelman, G., and Sporns, O. (1998).
Behavioral constraints in the development of
neuronal properties: a cortical model embedded
in a real-word device. Cerebral Cortex, 8:346-
361.