|
The robot is placed on a baby mat where different objects can be bitten, bashed, or just visually detected. The robot is equiped with the Intelligent Adaptive Curiosity engine. By trying different motor primitives, which it can modulate continuously, it progressively discovers that some objects are easier to interact with than others. As the experiment continues, the robot learns to master particular sensory-motor trajectories, its behavior changes qualitatively and gets more complex.
Here is a curve which summarizes one experiment, in which the robot performed a sequence of 15000 actions.
This curve shows that after a phase of random body babbling, its
behavior self-organizes in a first phase (P1) where it focuses on turning the head,
gazing at objects and discovering where they are. Then, in a second phase (P2) the robot begins to
focus on the simple biting action, but with no particular targets (it tries to
open and close its mouth even in front of objects that can not be bitten for
example). Then, in a third phase (P3) it shifts its focus of action to a more complicated kind of
motor program: it tries to move its legs, but again with no particular targets.
Then, in a fourth and fifth phase (P4 and P5) the robot begins to perform actions towards objects which have the
correct affordances: it now tries to bite the bitable object, then to bash the
bashable object. We see that there is a continuous increase in the
complexity of the robot’s activity, which self-organizes without being
pre-programmed for doing so.
|
