8.1 Data collection and preprocessing

Sensorimotor models were acquired in three setups: a kinematic arm model (section 4.5), a real robot arm equipped with a camera (chapter 6), and mobile robot also equipped with a camera (chapter 7). The raw data contained the following:

• For the kinematic arm model, the motor part are six joint angles, and the sensory part are the three corresponding coordinates of the end-effector and a collision variable.

• For the real robot arm, the motor part are the six joint angles of an arm posture suitable to grasp an object, and the sensory part is an image of the object.

• For the mobile robot, the motor part are the wheel velocities, and the sensory part are two successive images.

Training data were collected during random exploration; that is, the robot chose random motor commands and observed the sensory consequences of its actions. This exploration is the actual unsupervised part of the learning. Whether the following interpretation of this data is called unsupervised is more a matter of definition (see, for example, Meinicke (2000)).

Different from the kinematic arm model, the studies with the real robots revealed that it was necessary to preprocess the collected data. The dimensionality of the images was too high. The following three processing strategies proved to be useful:

• Keep positions (here, of an object or of the mobile robot) that are close together in Cartesian space also close together within the distribution of the preprocessed sensory data.
• Leave a residual redundancy in the preprocessed sensorimotor data. Population codes, as observed in nature, were a good example (chapter 6).

• Separate points in the final sensorimotor space that do not relate to each other, such as redundant arm posture for the same sensory input.

A processing that extracts lines, rectangles, or the like from an image was avoided. Including such more complex sensory representations may lead to the same conceptual problems as discussed for the symbolic approach (section 1.4.1). These representations depend on the designer's choice and may distract from the real difficulties of a behavioral task (Möller, 1999; Brooks, 1986b).