A multi-robot platform, made of a hybrid parallel kinematic machine and 2 KUKA LWR arms, is dedicated to open skull neuro-surgical tasks. Sub-millimeter accuracy is clearly required for both the absolute tool tracking and for good performances in motion compensation when the head is set free to move. An analysis of the sources of inaccuracies, mostly derived from the calibration phase, illustrates that errors are insufficiently reduced by stand-alone calibrations of the single robots. A method for volumetric compensation of errors is reported. A compensation transform is, in fact, computed during an offline training phase for a set of discretized subregions of the constrained head workspace. At runtime, a compensation motion is applied to robots so as to reach the desired real targets on anatomical parts. The resulting end-to-end static accuracy is distributed with median 0.75 mm and below 1 mm for the 95% of tests, with a 1:36 reduction factor from the starting conditions. The accuracy is evaluated also in dynamic tests with mild oscillatory patterns.