The law of path steering, as proposed by Accot and Zhai, describes a quantitative relationship between human temporal performance and the path's spatial characteristics. The steering law is formulated as a continuous goal crossing task, in which a large number of goals are crossed along the path. The steering law has been verified empirically for locomotion, in which a virtual driving task through straight and circular paths was performed. We revisit the path steering law for manipulation tasks in desktop virtual environments. We have conducted controlled experiments in which users operate a pen input device to steer a virtual ball through paths of varying length, width, curvature and orientation. Our results indicate that, although the steering law provides a good description of overall task time as a function of index of difficulty ID = L/W, where L and W are the path length and width, it does not account for other relevant factors. We specifically show that the influence of curvature can be modeled by a percentage increase in steering time, independent of index of difficulty. The path orientation relative to the viewing direction has a more complex effect on the steering law, which is moreover for instance asymmetric, i.e. it differs when moving to the left or right. A detailed analysis of our results indicates that a 3D steering movement can probably not be modeled as a sequence of individual goal crossing subtasks. Rather, we can postulate that the overall steering task is likely better described as a sequence of smaller movements that are closer to ballistic movements. One argument for this is that we established that the time for subtrials with continuous steering is related to ID by a power law, with an exponent in the range 0.5-0.6, rather than being equal to 1 as required by the steering law.