MULTIRESOLUTIONAL SCHEMATA FOR UNSUPERVISED LEARNING OF AUTONOMOUS ROBOTS FOR 3-D SPACE OPERATION

This paper describes a novel approach to the development of a learning control system for autonomous space robot (ASR) that presents the ASR as a ''baby''-that is, a system with no a priori knowledge of the wor ld in which it operates, but with behavior acquisition techniques that allow it to build this knowledge from the experiences of actions with in a particular environment (we will call it an Astro-baby). The learn ing techniques are rooted in the recursive algorithm for inductive gen eration of nested schemata molded from processes of early cognitive de velopment in humans. The algorithm extracts data from the environment, and by means of correlation and abduction, it creates schemata that a re used for control. This system is robust enough to deal with a const antly changing environment because such changes provoke the creation o f new schemata by generalizing from experiences, while still maintaini ng minimal computational complexity, thanks to the system's multiresol ution nature. Experimenting with ASR is especially interesting because the rules of input control do not coincide with human intuitions. Act ually, we want to see that the simulated device can learn unexpected s chemata from its own experience. Although the traditional approach to autonomous navigation involves off-line path planning with a known wor ld map (such as the potential fields algorithm), in most of the real t asks the environment is not well known because of ever-changing condit ions such as absence of gravity and because of sophisticated, hard-to- predict obstacles like components of the space station. Astro-baby gat hers data from its sensors and then, by using a schema-discovery syste m, it extracts concepts, forms schemata, and creates a quantitative/co nceptual semantic network. When the Astro-baby is first dropped into s pace, it does not have any experiences and its sensors and actuators a re sets that do not have any distinction among elements. Then, by tria l and error, the ASR learns the function of its actuators and sensors and how to activate them to achieve the goal given by its creator or t he sub-goals that it finds. In our simulation, the initial goal is to minimize the distance to a beacon. The results of simulation are posit ive: Astro-baby displays the ability to learn a number of maneuvers.