8.5 Perception

Based on sensorimotor models, the present work showed how visual perception might emerge:

• The location and orientation of an object within the reach of an arm could be perceived by simulating an arm posture suitable to grasp the object. This simulation was demonstrated with the robot arm (chapter 6). The abstract RNN associated an arm posture given an image representation of the object. The arm posture represented the real location and orientation accurately (mean distance between gripper tip and object: about 7 mm, mean difference in orientation: about 4^o).

• The geometry of the surrounding world, the real distance to obstacles, could be perceived by `mentally' simulating a movement toward the obstacles. This simulation was carried out in the mobile robot (chapter 7). The robot predicted the number of 2 sec intervals required to reach an obstacle in front.

• Symmetry within the world could also be perceived by simulating a movement. This kind of perception was demonstrated on the example of a circle (chapter 7). A simulated turn in the center of the circle predicts an invariant visual input, and this invariance can be observed only at the center. This way, humans might detect symmetry, despite having a distorted sensory representation in the visual cortex.

These experiments show that perception of space and shape can be explained by the sensorimotor approach. Instead of using a pure sensory representation, all is reduced to the simulation of motor commands and their sensory consequences.

In the mobile-robot study, depth perception was based on computing the time-to-contact; in the robot-arm study, depth perception was based on associating an arm posture. This difference implies that for unlike tasks different spatial representations exist, for example, a far field for locomotion and a near field for grasping (Mallot et al., 1992).