The main task of a voice-enabled service robot is to engage people (users) in dialogue providing an efficient ac- cess to the services it is designed for within a reasonable (limited) time. In managing service dialogues, extracting the user goal (intention) for requesting a particular service at each dialogue state is the key issue. In service ro- bots deployment conditions speech recognition limitations with noisy audio input and uncooperative users may jeopardize user goal identification. In order to reduce the risk of communication failures due to incorrect user goals, we introduce sequential dialogue repair techniques motivated by the theory of grounding in conversation, and exploiting the inherent multimodality in the perceptual system of a service robot. The error handling method is based on Bayesian networks fusing audio and non-audio based modalities during user goal identification and serving as input to graphical models known as decision networks. Decision networks allow the definition of dia- logue repair sequences as actions, and provide an explicit strategy for selecting actions. The paper demonstrates how the above graphical models can be used for designing and implementing of repair strategies for avoiding communication failures in spoken dialogues with mobile tour-guide robots in mass exhibition conditions. The benefits of the proposed repair strategies are tested through experiments with the dialogue system of RoboX, a tour-guide robot successfully deployed at the Swiss National Exhibition (Expo.02).