In this paper we consider the problem of controller synthesis for timed games under imperfect information. Novel to our approach is the requirements to strategies: they should be based on a finite collection of observations and must be stuttering invariant in the sense that repeated identical observations will not change the strategy. We provide a constructive transformation to equivalent finite games with perfect information, giving decidability as well as allowing for an efficient on-the-fly forward algorithm. We report on application of an initial experimental implementation.