Resilient Hybrid AI Navigation for Mobile Robots in Hazardous Cyber-Physical Environments: A Review

In this work, we present a comprehensive review on how to tackle the challenging tasks of navigation of mobile robots in cyber-physical environments which are prone to failures. The main contribution of this paper is to present a system partitioning criterion, i.e. how to divide the control authority of the different subsystems in order to prevent a single failure from triggering a chain of system failures. To this end, we analyze the effects of three types of environmental degradation, i.e. environmental disturbances, structural instability and cyber-physical disturbances, on the different layers of a typical robot, i.e. perception, SLAM, planning and control. We then review all the existing methods for navigation, i.e. probabilistic methods for uncertainty estimation, learning methods and hybrid systems for various layers and provide a review of the ways to test them. We also provide a number of observations such as the high navigation performance of multi-layered hybrid learning systems which can reach up to 99% while single networks fail to reach even 48% and the high performance of embedded perception which reaches up to 93% when the onboard perception fails. We also present a modeling method for such systems by combining an adaptive learning representation with a model-based representation which can be used for recursive feasibility and safety verification. Finally, we present the open challenge of cross-layer verification for such complex systems.