In this paper, a probability-based balance monitoring concept for humanoid robots is proposed. Two algorithms are presented that allow us to distinguish between exceptional situations and normal operations. The first classification approach uses Gaussian-Mixture-Models (GMM) to describe the distribution of the robot's sensor data for typical situations such as stable walking or falling down. With the GMM it is possible to state the probability of the robot being in one of the known situations. The concept of the second algorithm is based on Hidden-Markov-Models (HMM). The objective is to detect and classify unstable situations by means of their typical sequences in the robot's sensor data. When appropriate reflex motions are linked to the critical situations, the robot can prevent most falls or is at least able to execute a controlled falling motion. The proposed algorithms are verified by simulations and experiments with our bipedal robot BARt-UH. © SAGE Publications 2009 Los Angeles, London.