A Robot Visual Homing Model that Traverses Conjugate Gradient TD to a Variable λ TD and Uses Radial Basis Features

The term ‘homing’ refers to the ability of an agent – either animal or robot to find a known goal location. It is often used in the context of animal behaviour, for example when a bird or mammal returns ‘home’ after foraging for food, or when a bee returns to its hive. Visual homing, as the expression suggests, is the act of finding a home location using vision. Generally it is performed by comparing the image currently in view with ‘snapshot’ images of the home stored in the memory of the agent. A movement decision is then taken to try and match the current and snapshot images (Nehmzow 2000). A skill that plays a critical role in achieving robot autonomy is the ability to learn to operate in previously unknown environments (Arkin 1998; Murphy 2000; Nehmzow 2000). Furthermore, learning to home in unknown environments is a particularly desirable capability. If the process was automated and straightforward to apply, it could be used to enable a robot to reach any location in any environment, and potentially replace many existing computationally intensive homing and navigation algorithms. Numerous models have been proposed in the literature to allow mobile robots to navigate and home in a wide range of environments. Some focus on learning (Kaelbling, Littman et al. 1998; Nehmzow 2000; Asadpour and Siegwart 2004; Szenher 2005; Vardy and Moller 2005), whilst others focus on the successful application of a model or algorithm for a specific environment and ignore the learning problem (Simmons and Koenig 1995; Thrun 2000.; Tomatis, Nourbakhsh et al. 2001). Robotic often borrow conceptual mechanisms from animal homing and navigation strategies described in neuroscience or cognition literature (Anderson 1977; Cartwright and Collett 1987). Algorithms based on the snapshot model use various strategies for finding features within images and establishing correspondence between them in order to determine home direction (Cartwright and Collett 1987; Weber, Venkatesh et al. 1999; Vardy and Moller 2005). Block matching, for example, takes a block of pixels from the current view image and searches for the best matching block in stored images within a fixed search radius (Vardy and Oppacher 2005).