Figure 9: SAIL robot navigates autonomously using its autonomously developed visual perceptual behaviors. Four
movies are available at http://www.egr.msu.edu/mars/ to provide more results.
Figure 12: The SAIL robot learned the chained action after verbally instructed by human trainers.
First level LBE primed context
Figure 10: A hierarchical developmental learning archi-
tecture for action chaining.
interacting with human trainers. For more details,
the reader is referred to another paper in the pro-
ceeding of this workshop (Huang and Weng, 2002).
6. Comparison with others’ work
What is the most basic difference between a tradi-
tional learning algorithm and a developmental algo-
rithm? Autonomous development does require a ca-
pability of learning but it requires something more
fundamental. A developmental algorithm must be
(a) (b) (c) (d)
ɔ OS O¾}
(e) (f) (g)
Figure 11: Gripper tip trajectories of the SAIL robot.
(a)-(d) are basic actions, each of which starts from the
black dot. (e)-(g) are composite actions chaining some
or all of the basic ones.
able to learn tasks that its programmers do not
know or even cannot predict. This is because a de-
velopmental algorithm, once designed before robot
“birth,” must be able to learn new tasks and new
skills without requiring re-programming. The rep-
resentation of a traditional learning algorithm is de-
signed by humans for a given task but that for a de-
velopmental algorithm must be autonomously gener-
ated. As a working example, humans’ developmental
algorithm enables humans to learn new skills without
a need to change the design of their brain.
However, the motive of developmental robots is
not to make robot more difficult to program, but rel-