We explored how human subjects learn to move objects through space using neuromuscular control signals having more degrees of freedom than needed to specify object location unambiguously. Subjects wore an instrumented glove that recorded finger motions. A linear transformation matrix projected joint angle signals (a high-dimensional control vector) onto cursor position on a video monitor (a low-dimensional workspace). We assessed how visual information influences learning and generalization of novel finger coordination patterns as subjects practiced using hand gestures to manipulate cursor location. Three groups of test subjects practiced moving a visible cursor between different sets of screen targets. The hand-to-screen transformation was designed such that the different sets of targets (which we considered implicit spatial cues) varied in how informative they were about the gestures to be learned. A separate control group practiced gesturing with explicit cues (pictures of desired gestures) without ongoing cursor feedback. Another control group received implicit spatial cueing and feedback only of final cursor position. We found that test subjects could learn to produce desired gestures as well as subjects provided explicit cues, although training efficacy decreased as the amount of task-relevant feedback decreased. While both control groups learned to associate screen targets with specific gestures, only subjects provided with online feedback of cursor motion learned to generalize in a manner consistent with the internal representation of an inverse hand-to-screen mapping. Thus, spatial learning and generalization appear to require dynamic feedback of object motion in response to control signal changes; static information does not suffice.