Behavior-based Perception for Autonomous Robot Navigation

For navigation, many animals are known to use their visual system in combination with a process called dead reckoning, in which the animal knows its position through egomotion, for learning landmarks in an unknown environment. In robotics, previous research has been done using such an approach. The goal was to learn salient perceptual features in the environment using an unsupervised neural networ, and learn their relative locations using odometry on the robot. The resulting map showed that the information was there, but it was too approximate. The main problem was that a general sense of direction was missing and landmark representations were ambiguous. The present study aims at using biologically inspired behaviors to yield a better encoding of the perceptual landmarks. Behaviors of this type used by animals are for example the head-scanning of rats for the detection of configurations of landmarks. These could be modeled on the robot by moving the camera in three directions, detecting three landmarks in one perceptive movement. The hypothesis is that this triple-view approach results in coupling landmarks more strictly to the environment and to each other since (1) the triple-landmark view is a less ambiguous percept than a single-landmark view, and (2) the triple-view is unique for a given direction. Furthermore, the learning of perceptual landmarks is done on the basis of an adaptive unsupervised neural network. New items are added to the set of landmarks incrementally when needed, where the number of nodes will depend on the visual complexity of the environment. A neural-network approach for incremental learning landmarks is proposed and tested in the first phase of the study. In the second phase, experiments were done to test the main hypothesis.