Prevalent approaches for endowing robots with autonomous navigation capabilities require the estimation of a system state representation based on sensor noisy information. This system state usually depicts a set of dynamic variables such as the position, velocity and orientation required for the robot to achieve a task. In robotics, and in many other contexts, research efforts on state estimation converged towards the popular Bayes filter. The primary reason for the success of Bayes filtering is its simplicity, from the mathematical tools required by the recursive filtering equations, to the light and intuitive system representation provided by the underlying Hidden Markov Model. Recursive filtering also provides the most common and reliable method for real-time state estimation thanks to its computational efficiency. To keep low computational complexity, but also because real physical systems are not perfectly understood, and hence never faithfully represented by a model, Bayes filters usually rely on a minimum system state representation. Any unmodeled or unknown aspect of the system is then encompassed within additional noise terms. On the other hand, autonomous navigation requires robustness and adaptation capabilities regarding changing environments. This creates the need for introducing contextual awareness within the filtering process. In this thesis, we specifically focus on enhancing state estimation models for dealing with context-dependent sensor performance alterations. The issue is then to establish a practical balance between computational complexity and realistic modelling of the system through the introduction of contextual information. We investigate on achieving this balance by extending the classical Bayes filter in order to compensate for the optimistic assumptions made by modeling the system through time-homogeneous distributions, while still benefiting from the recursive filtering computational efficiency. Based on raw data provided by a set of sensors and any relevant information, we start by introducing a new context variable, while never trying to characterize a concrete context typology. Within the Bayesian framework, machine learning techniques are then used in order to automatically define a context-dependent time-heterogeneous observation distribution by introducing two additional models: a model providing observation noise predictions and a model providing observation selection rules.The investigation also concerns the impact of the training method we choose. In the context of Bayesian filtering, the model we exploit is usually trained in the generative manner. Thus, optimal parameters are those that allow the model to explain at best the data observed in the training set. On the other hand, discriminative training can implicitly help in compensating for mismodeled aspects of the system, by optimizing the model parameters with respect to the ultimate system performance, the estimate accuracy. Going deeper in the discussion, we also analyse how the training method changes the meaning of the model, and how we can properly exploit this property. Throughout the manuscript, results obtained with simulated and representative real data are presented and analysed.