The stochastic control problem of partially observed systems is approximated by a sequence of partially observed impulse control problems with non-zero but vanishing impulse costs. For the latter problems, we give an explicit construction of the so-called &-optimal controls in a separated form.As a consequence of this approach, we prove that the infimum of the expected cost over the set of separated controls-i.e. which depend non anticipatively on the filtering process-is equal to the infimum of the expected cost over the set of admissible controls-i.e. which depend non anticipatively on the observation process.For the classical situation of linear systems, this method yields the existence of an optimal separated control.