Based on psychophysical evidence about how learning of visuomotor transformation generalizes, it has been suggested that movements are planned on the basis of movement direction and magnitude, i.e., the vector connecting movement origin and targets. This notion is also known under the term “vectorial planning hypothesis”. Previous psychophysical studies, however, have included separate areas of the workspace for training movements and testing the learning. This study eliminates this confounding factor by investigating the transfer of learning from forward to backward movements in a center-out-and-back task, in which the workspace for both movements is completely identical. Visual feedback allowed for learning only during movements towards the target (forward movements) and not while moving back to the origin (backward movements). When subjects learned the visuomotor rotation in forward movements, initial directional errors in backward movements also decreased to some degree. This learning effect in backward movements occurred predominantly when backward movements featured the same movement directions as the ones trained in forward movements (i.e., when opposite targets were presented). This suggests that learning was transferred in a direction specific way, supporting the notion that movement direction is the most prominent parameter used for motor planning.