Recent years have seen the development of wind energy conversion systems that can exploit the superior wind resource that exists at altitudes above current wind turbine technology. One class of these systems incorporates a flying wing tethered to the ground which drives a winch at ground level. The wings often resemble sports kites, being composed of a combination of fabric and stiffening elements. Such wings are subject to load dependent deformation which makes them particularly difficult to model and control. Here we apply the techniques of evolutionary robotics i.e. evolution of neural network controllers using genetic algorithms, to the task of controlling a steerable kite. We introduce a multibody kite simulation that is used in an evolutionary process in which the kite is subject to deformation. We demonstrate how discrete time recurrent neural networks that are evolved to maximise line tension fly the kite in repeated looping trajectories similar to those seen using other methods. We show that these controllers are robust to limited environmental variation but show poor generalisation and occasional failure even after extended evolution. We show that continuous time recurrent neural networks (CTRNNs) can be evolved that are capable of flying appropriate repeated trajectories even when the length of the flying lines are changing. We also show that CTRNNs can be evolved that stabilise kites with a wide range of physical attributes at a given position in the sky, and systematically add noise to the simulated task in order to maximise the transferability of the behaviour to a real world system. We demonstrate how the difficulty of the task must be increased during the evolutionary process to deal with this extreme variability in small increments. We describe the development of a real world testing platform on which the evolved neurocontrollers can be tested.