Robust and fault-tolerant Neural Architectures are critical for the NASA's development of autonomous spacecraft and rovers. Such systems are of key importance to space missions in which traditional (human-guided) control is infeasible due to the light-time delay. Ideally, they would form the spacecraft or rover's brain and nervous system, capable of autonomous decision given the particular environment and circumstances. We propose a new paradigm for growing and evolving, rather than designing, neural networks architectures that could lead to the evolution of computational structures that could form the ``central nervous system'' of a fault-tolerant, self-regenerating, intelligent spacecraft control system. We have a model for the evolution of neural networks that encompasses both development and evolution in which information about a network's structure, connections, weights, and their behavior under stimuli is entirely encoded in a genetic string, whose expression (evaluation) gives rise to the network itself. Evolution is enforced by a Genetic Algorithm (GA) acting on pools of genetic strings, whose fitnesses are evaluated by actually growing the networks that correspond to the particular string. The artificial brains we can evolve have several advantage over their human-designed counterparts. They are fault-tolerant, and operate reliably in noisy and only partially-known environments. Further, they have the capacity to regenerate. We expect to have novel neural architectures that can be used on actual rovers (for navigation as well as robotic arm control) that can outperform their human-designed counterparts within ten years. The path towards achieving this goal was mapped out at a workshop held at JPL in September 2003, which brought together the leading teams in the US and in Europe now working on this approach.