To control assistive devices, people with severe paralysis must reorganize their residual body motions for carrying out new functional tasks. Body-machine interfaces are a new technological tool to facilitate this reorganization. In this study, we investigated motor learning on a group of unimpaired subjects that were trained, via a body-machine interface, to solve a reaching task and to play a computer game by controlling a cursor with coordinated motions of their upper body. While reaching involves continuous motions, the computer game was based on selecting a finite set of keys. Since there are multiple body configurations corresponding to each cursor position and each game command, it is possible that different movement sets or tasks lead to the formation of different representations of the map between body configurations and cursor coordinates. In contrast, we found that subjects tended toward a single stable map. We discuss how the stability and robustness properties of the learned map are consistent with the possibility to exploit redundancy for optimizing performance in different tasks.