The possibility of developing increasingly sophisticated robots, and the availability of cloud-connected resources, have boosted the interest in the study of real world applications of service robotics. However, in order to operate under natural or less structured conditions, and given the information processing bottleneck and the reactivity required for a secure execution of the task, it is desirable that the agent can exploit more efficiently the local information available, so that being more autonomous, and relying less on remote computation. This study explores a strategy for obtaining reliable approach tasks. It considers the anticipation of perception, by taking into account the statistical regularities and the information redundancies induced in the sensory-motor coupling. From an initial perception of the object assisted by remote computation, contextual features are defined for capturing bodily sensations emerging in the task. The observations based on proprioceptive and visual data are fused in a Bayesian Network, which is in charge of assessing the saliency during the object approach, thus constituing a local discriminative processing of the object. The strategy proposed reduces dependency on context-free models of behavior, while providing an estimate on the degree of confidence in the progress of the task.