The lack of robustness and of efficiency of current Embedded Perception and Decision-making systems is one of the major obstacles to a full deployment of self-driving cars. In this talk, it will be argued that three enabling technologies are required for improving the capabilities of such systems: (1) a framework for fusing multiple-sensor data in the presence of uncertainty and for interpreting in real-time the surrounding dynamic environment, (2) a method for predicting future environmental changes using perception history, contextual information and some prior knowledge, and (3) a decision-making approach having the capability to continuously evaluate the risk of future collisions and to provide on-line maneuvers recommendations for a safe navigation. New approaches developed at Inria for Embedded Multi-sensors Perception, Situation Awareness, Collision risk assessment and on-line Decision-making for safe navigation, will be presented. It will be shown that Bayesian approaches are mandatory for developing such technologies and for obtaining the required robustness in the presence of uncertainty and of complex dynamic situations. The talk is illustrated by experimental results obtained in real traffic situations, in the scope of several collaborative projects involving Toyota, Renault, CEA-LETI, or the French IRT (Technological Research Institute) Nanoelec.