Visual landmark learning Giovanni Bianco Alexander Zelinsky Miriam Lehrer Computer Science Serv. Dept. of Syst. Eng. - RSISE Dept. of Neurobiology University of Verona Australian National University University of Zurich Via S. Francesco 8 Winterthurerstr. 190 I-37129 Verona - Italy Canberra, ACT 0200 - Australia CH-8057 Zurich - Switzerland bianco@chiostro.univr.it Alex.Zelinsky@anu.edu.au Miriam@zool.unizh.ch Abstract Biology often offers valuable example of systems both for learning and for controlling motion. Work in ro- botics has often been inspired by these findings in di- verse ways. Though, the fundamental aspects that in- volve visual landmark learning and motion control mech- anisms have almost exclusively been approached heuris- tically rather than examining the underlying princi- ples. In this paper we introduce theoretical tools that might explain how the visual learning works and why the motion is attracted by the pre-learnt goal position. Basically, the theoretical tools emerge from the nav- igation vector field produced by the visual behaviors. Both the learning process and the navigation scheme influence the motion field. We apply classical mathe- matical and dynamic control to analyze the efficiency of our method. 1 Introduction Animals, including insects, are proficient in navi- gating and, in general, several biological schemas for solving navigational tasks seem to be promising for robotics applications. Basically biological systems in- volve two fundamental mechanisms: visual learning and visual motion control. There are, of course, di- verse ways of implementing such mechanisms among animals. Nevertheless, insects provide basic methods that are extremely valuable for robotics. To this ex- tent, the use of landmarks for piloting and for goal identification is fundamental to orientation, although other mechanisms, such as the optomotor response, path integration and skylight navigation are clearly in- volved [19]. However, whereas the optomotor response and the path integration mechanism require no more than an appropriately hard-wired neural network, the use of the skylight compass and of landmarks requires, in addition, a learning process that must take place in every new situation. On the other hand, referring to motion planning, it is widely agreed that landmark-guided navigation in insects is based on storing the image of the land- mark(s) as some kind of a snapshot ora template [1, 4]. On its next journey, the insect navigator strives to ac- complish a match between the currently viewed image of the landmark(s) or of the goal and the snapshot stored in memory. Many examples of the above men- tioned abilities transposed to real robots can be found in the review [18]. From these case studies the imple- mentation of biological mechanisms appear to follow a pragmatic approach where the final behavior is the main objective. The learning and control aspects are of utmost importance but a thorough study on the the- oretical principles has only recently been performed [13, 3] Basically, the core of the theory is represented by the navigation vector field, whose study provide two main results: • the visual potential function generating the field represents the driving principle to perform vi- sual guidance. When proven to be a Lyapunov compliant function, we can state the navigation system exhibits convergence to the goal. • The conservativeness of the navigation vector field deals with the concept of repeatability of the trials and provides key information to per- form landmark learning. In this paper we address, in particular, principles involved in landmark learning. Details about the navi- gation system and the nature of the potential function can be found in references [2] and [3] respectively.