This paper presents a framework for learning autonomous visual navigation behavior from demonstration examples by integrating 3D range data and an omnidirectional camera. The approach uses programming by demonstration to learn the demonstrated trajectories as a mapping between visual features computed on the omnidirectional image and corresponding robot motion. Exhaustive tests are performed to identify discriminant features to mimic teacher demonstrations. The relationship between perception and action is learned from demonstrations using locally weighted regression and artificial neural networks. Experimental results on a mobile robot show that the acquired visual behavior is robust and can generalize and optimize performance in environments not presented during training. The decreasing cost and improved performance of computer vision systems make them attractive as primary sensors for mobile robot navigation. However, designing robust reactive visual robot behaviors remains a challenge due to the lack of depth information in 2D images and the variability of indoor environments. One feasible approach is to reconstruct the geometry of the local environment from single or multiple views. The main contribution of this paper is the presentation of a new framework for visual robot navigation of indoor environments. The approach intuitively generates a policy from mimicking sonar-based navigation behavior and transfers it onto a more complex vision-based behavior. The complete robot navigation framework is a three-stage approach towards acquiring autonomous visual behavior through imitation of other robotic behaviors, explicit demonstration of a particular skill by a human teleoperating the robot, and finally learning with a critic based on the robot's experiences. Robotic training examples are either generated from proximity sensor-based behaviors or human teleoperation. Learning from robot demonstrations allows the transfer of existing behaviors onto a novel sensor modality without the need for explicit reprogramming. However, in this type of learning, the competence of the visual behavior does not exceed its proximity sensor-based counterpart. The key challenge is the generalization of perceptions and actions in demonstrated scenarios onto novel unseen situations. The diversity of indoor environments and the ambiguity and complexity of visual perceptions severely complicate the identification of appropriate measures of similarity. The paper proposes a three-stage scheme for matching the current situation with demonstrated scenarios. The first level is concerned with the classification of the environmental context into a limited set of prototypical classes. In the case of omnidirectional vision, the classification relies on the distribution and shape of segmented regions.This paper presents a framework for learning autonomous visual navigation behavior from demonstration examples by integrating 3D range data and an omnidirectional camera. The approach uses programming by demonstration to learn the demonstrated trajectories as a mapping between visual features computed on the omnidirectional image and corresponding robot motion. Exhaustive tests are performed to identify discriminant features to mimic teacher demonstrations. The relationship between perception and action is learned from demonstrations using locally weighted regression and artificial neural networks. Experimental results on a mobile robot show that the acquired visual behavior is robust and can generalize and optimize performance in environments not presented during training. The decreasing cost and improved performance of computer vision systems make them attractive as primary sensors for mobile robot navigation. However, designing robust reactive visual robot behaviors remains a challenge due to the lack of depth information in 2D images and the variability of indoor environments. One feasible approach is to reconstruct the geometry of the local environment from single or multiple views. The main contribution of this paper is the presentation of a new framework for visual robot navigation of indoor environments. The approach intuitively generates a policy from mimicking sonar-based navigation behavior and transfers it onto a more complex vision-based behavior. The complete robot navigation framework is a three-stage approach towards acquiring autonomous visual behavior through imitation of other robotic behaviors, explicit demonstration of a particular skill by a human teleoperating the robot, and finally learning with a critic based on the robot's experiences. Robotic training examples are either generated from proximity sensor-based behaviors or human teleoperation. Learning from robot demonstrations allows the transfer of existing behaviors onto a novel sensor modality without the need for explicit reprogramming. However, in this type of learning, the competence of the visual behavior does not exceed its proximity sensor-based counterpart. The key challenge is the generalization of perceptions and actions in demonstrated scenarios onto novel unseen situations. The diversity of indoor environments and the ambiguity and complexity of visual perceptions severely complicate the identification of appropriate measures of similarity. The paper proposes a three-stage scheme for matching the current situation with demonstrated scenarios. The first level is concerned with the classification of the environmental context into a limited set of prototypical classes. In the case of omnidirectional vision, the classification relies on the distribution and shape of segmented regions.