Fault tolerance for multi-sensor perception: application to smart vehicle localization.

Summary Perception is a fundamental input of robotic systems, in particular for localization, navigation and interaction with the environment. However, the data perceived by robotic systems are often complex and subject to significant inaccuracies. To remedy these problems, the multi-sensor approach uses either several sensors of the same type to exploit their redundancy, or sensors of different types to exploit their complementarity in order to reduce inaccuracies and uncertainties on sensors. The validation of this data fusion approach poses two major problems: first, the behavior of fusion algorithms is difficult to predict, which makes them difficult to verify by formal approaches. Moreover, the open environment of robotic systems creates a very large execution context, which makes testing difficult and expensive. The aim of this thesis is to propose an alternative to validation by implementing fault tolerance mechanisms: since it is difficult to eliminate all faults from the perception system, we will try to limit their impacts on its operation. We have studied the fault tolerance intrinsically allowed by data fusion by formally analyzing data fusion algorithms, and we have proposed detection and recovery mechanisms adapted to multi-sensor perception. We then implemented the proposed mechanisms for a vehicle localization application using Kalman filtering data fusion. We finally evaluated the proposed mechanisms using real data replay and fault injection technique, and demonstrated their effectiveness against hardware and software faults.