Multi-sensor perception have an important role in robotics and autonomous systems, as inputs for critical functions such as obstacle detection, localization, etc. This Multi-sensor perception begins to appear in critical applications, such as drones and ADAS (Advanced Driver Assistance Systems). However such complex systems are diffcult to validate entirely. In this paper we study these systems under an alternative dependability method: fault tolerance. We propose an approach to tolerate faults in multi-sensor data fusion based on the more classical method of duplication-comparison, and offering detection and recovery services. We detail an example implementation using Kalman filters data fusion for mobile robot localization. We demonstrate its effectiveness in this case study using real data and fault injection