Local visual homing navigation using gradient-descent learning of haar-like features

The autonomous mobile technology of mobile robots has been developed. Visual navigation is one of non-trivial problems and it has been tackled with biologically inspired models. Especially, ant navigation system inspires robot navigation. The visual cell structure of ants was modeled with Haar-like features. Those features can be obtained with computationally efficient process. In this paper, we handle visual homing navigation where an agent is supposed to return home after exploration in the environment. We apply a learning process based on gradient-descent algorithm to estimate the homing vector at an arbitrary position of a mobile agent. Our approach is simple but very effective to find the homing vector and its performance is better than the conventional algorithm. From our results, the Haar-like features in the snapshot images are sufficient to estimate the homing vector.