A Biorobotics Approach to the Study of Insect Visual Homing Strategies

Biorobotics is a methodology where models of animal behavior and of the underlying neural mechanisms are implemented and tested on robots in the real world. This eliminates erroneous assumptions about the physical interaction between an agent — an animal or a robot — and its environment, and therefore provides a stronger proof for the validity of the model than computer simulations do. Moreover, insights gained in this modeling process can directly be transferred to technical applications. Thus, besides being a method of biological modeling, biorobotics is also a sub-discipline of bionics, an interdisciplinary direction of research which aims at exploiting the sophisticated biological solutions found by evolution for applications — in this case primarily in robotics. This work uses the biorobotics methodology to study models of insect visual navigation and to develop parsimonious methods of robot navigation. A class of local navigation models was investigated where a single goal position has to be reached (homing); such methods are employed by insects e.g. to find back to the nest after foraging excursions. Local methods are building blocks for methods of long-range navigation based on topological maps of the environment. All methods studied share the basic assumption that a target location is characterized by a signature derived from the panorama of its surroundings. When returning to the goal, the agent relates the stored signature to the signature derived from the image of the current location, and moves in a way that the two signatures become more similar to each other. Different models of this class were tested in robot experiments or partly in computer simulations. Besides the development of novel models of navigation, a substantial portion of the work was invested in the development of robot hardware including special sensors (panoramic cameras, color-contrast sensors), of software for the readout of the sensors and the control of the robots, and of an interactive simulator software for visual landmark navigation. In addition, methods of analog computation were employed which may see a renaissance in robotic applications. Starting point of the work was the snapshot model of bee navigation, where the signature is an image. This model was adapted to real-world sensory input and tested together with a model of the polarized-light compass of insects on a mobile robot. The high precision of homing achieved in these experiments provides support for the snapshot model as a model of insect navigation. A neural version of the snapshot model was developed to add plausibility to this model from the neurobiological perspective and to provide a starting point for physiological or anatomical investigations. For the real-world test of the average landmark vector model, a parameter-based model with a two-component vector as signature, a particularly strict way of biorobotic modeling was chosen: the model was implemented in analog hardware which shares basic computational principles with biological neurons, thus ensuring the neurobiological plausibility of the model. From the architecture of the analog implementation, insights into the neural apparatus that might be employed by insects for visual homing could be derived. Based on a software implementation it could be demonstrated that the average landmark vector model is applicable to robot navigation in indoor environments, however, a comparison with the alternative image-warping method revealed disadvantages of this and other models which are based on a separation between landmarks and background. How insects could employ the above-mentioned parsimonious methods for outdoor navigation was explored by constructing a visual detector with the same spectral characteristics as insect photoreceptors. Based on a collection of visual samples from natural objects and sky it was shown that a simple contrast mechanism between UV and green receptors would enable a robust separation of skyline as foreground and sky as background under varying conditions of illumination. This color contrast mechanism could be exploited for robot outdoor navigation. The suspicious behavior of desert ants in experiments with expanded landmark arrays was investigated in a series of computer simulations. While models of insect visual homing widely accept the so-called template hypothesis, according to which the target signature is an image, a newly developed model was shown to better reproduce the observations from these experiments. This contour model is an instance of the alternative parameter hypothesis, which assumes that only a set of parameters is extracted from the image and stored, in this case length and eccentricity of the landmark contours. Mechanisms of this class could be especially suitable in combination with the color contrast mechanism of skyline detection. As a building block of the topological methods of long-range navigation, the local methods studied in this work contribute to a change of paradigms in robot navigation. Based on biological insights, the classical, position-based approach of robot navigation is increasingly replaced by the topological approach. It is not attempted to solve the difficult problem of determining position and orientation in space from the sensor data, but the sensor data are immediately used for reaching a goal position without the intermediary step of position estimation. Acknowledgments I was greatly benefitting from working with Prof. Rolf Pfeifer, Artificial Intelligence Lab, and Prof. Rüdiger Wehner, Neurobiology Group, who provided an excellent environment for this interdisciplinary project. I would like to thank my colleagues for the stimulating and friendly atmosphere in the AI Lab. Many of them and many other people contributed to this work with help, advice, and encouragement, and I would especially like to thank Christian Gfeller, Verena Hafner, Bärbel Herrnberger, Koh Hosoda, Sonja und Markus Knaden, Jürgen Knobloch, Hiroshi Kobayashi, Axel Könies, Thomas Labhart, Dimitrios Lambrinos, Marinus Maris, Peter Paschke, Thorsten Roggendorf, and Sepp Ruchti, as well as Bernhard Schmid and colleagues from the workshop of the Physics department. In the UV/green project, I received valuable help with the software from Mark Spaeth (MIT) and Paul Ricchiazzi (UCSB). Financial support from the Swiss National Science Foundation and the “Kommission zur Förderung des akademischen Nachwuchses” of the University of Zurich is kindly acknowledged.