Robot dynamic identification techniques rely on the quality and completeness of the signals available as inputs, typically joint positions and motor currents or joint torques. These signals are often noisy and filtering operations are required before using them for identification. Moreover, some robot control units (e.g., in the KUKA KR5 Sixx) return the user only the absolute values of the motor currents (or of the torques), thus preventing a correct dynamic estimation. We present a method for the identification of the robot dynamic model when the motor torques/currents have unknown signs. The method consists in solving a sequence of constrained optimization problems, exploiting physical feasibility constraints. A tree of solutions is built, and the branches leading to unfeasible solutions are pruned. As output, the torque signs are estimated together with the resulting robot dynamic model.