This paper presents a novel failure-tolerant architecture for future robotic spacecraft. It is based on the Time and Space Partitioning (TSP) principle as well as a combination of Artificial Intelligence (AI) and traditional concepts for system failure detection, isolation and recovery (FDIR). Contrary to classic payload that is separated from the platform, robotic devices attached onto a satellite become an integral part of the spacecraft itself. Hence, the robot needs to be integrated into the overall satellite FDIR concept in order to prevent fatal damage upon hardware or software failure. In addition, complex dexterous manipulators as required for on orbit servicing (OOS) tasks may reach unexpected failure states, where classic FDIR methods reach the edge of their capabilities with respect to successfully detecting and resolving them. Combining, and partly replacing traditional methods with flexible AI approaches aims to yield a control environment that features increased robustness, safety and reliability for space robots. The developed architecture is based on a modular on-board operational framework that features deterministic partition scheduling, an OS abstraction layer and a middleware for standardized inter-component and external communication. The supervisor (SUV) concept is utilized for exception and health management as well as deterministic system control and error management. In addition, a Kohonen self-organizing map (SOM) approach was implemented yielding a real-time robot sensor confidence analysis and failure detection. The SOM features non supervized training given a typical set of defined world states. By compiling a set of reviewable three-dimensional maps, alternative strategies in case of a failure can be found, increasing operational robustness. As demonstrator, a satellite simulator was set up featuring a client satellite that is to be captured by a servicing satellite with a 7-DoF dexterous manipulator. The avionics and robot control were integrated on an embedded, space-qualified Airbus e.Cube on-board computer. The experiments showed that the integration of SOM for robot failure detection positively complemented the capabilities of traditional FDIR methods.