Previous studies have used transfer of learning over workspace locations as a means to determine whether subjects code information about dynamics in extrinsic or intrinsic coordinates. Transfer has been observed when the torque associated with joint displacement is similar between workspace locations - rather than when the mapping between hand displacement and force is preserved - which is consistent with muscle- or joint based encoding. In the present study, we address the generality of an intrinsic coding of dynamics and examine how generalization occurs when the pattern of torques varies over the workspace. In two initial experiments, we examined transfer of learning when the direction of a force field was fixed relative to an external frame of reference. While there were no beneficial effects of transfer following training at a single location (Experiment 1 and 2), excellent performance was observed at the center of the workspace following training at two lateral locations (Experiment 2). Experiment 3 and associated simulations assessed the characteristics of this generalization. In these studies, we examined the patterns of transfer observed following adaptation to force fields that were composed of two subfields that acted in opposite directions. The experimental and simulated data are consistent with the idea that information about dynamics is encoded in intrinsic coordinates. The nervous system generalizes dynamics learning by interpolating between sets of control signals, each locally adapted to different patterns of torques.