ABSTRACT
Robotics research devotes considerable attention to path finding. This is the problem of moving a robot from a starting position to a goal avoiding obstacles. Also, the robot path must be short and smooth. Traditionally, path finders are either model-based or sensor-based. While model-based systems address the path finding problem globally using a representation of the workspace, sensor-based systems consider it locally, and rely only on robot sensors to decide motion. Both methods have limitations, which are rather complementary. By integrating the two methods, their respective drawbacks can be mitigated. Thus, in [15] a model-based system (a planner working on an artificial potential field) and a sensor-based system (a Hierarchical Extended Kohonen Map) which cooperate to solve the path finding problem have been described. Along related lines, several authors [5], [6], [8], [12] have proposed to build automatically the sensor-based system as the result of a learning process, where a local planner plays the role of the teacher. In particular, [5], [8] employ a Self-Organizing Map (SOM) and [6] use a dynamical variant of SOM (DSOM) based on a Growing Neural Gas network [2]. In these works, the decision of using a SOM-like network seems to be justified by its data topology-conserving character which is supposed to favor in some way the learning of suitable < perception, action > pairs. None of these works provide experimental evidence for this reasonable, but not obvious, claim.
In this chapter we describe a SOM-like neural network which learns to associate actions to perceptions under the supervision of a planning system. By reporting this experiment the following contributions are made. First, the utility of using a hierarchical version of SOM instead of the basic SOM is investigated. Second, the effect of cooperative learning due to the interaction of neighboring neurons is explicitly measured. Third, the beneficial side-effect which can be obtained by transferring motion knowledge from the planner to the SOM is highlighted.
