In Section 5.5 we have presented an extensive evaluation of visual servoing, based on numerous real positioning experiments, giving the following four main results:
Our results show that visual feedback control yields significant repeatability improvement in an imprecise PUMA robot arm, while only marginal improvement in the Utah/MIT hand. This is explained by the different characteristics of the manipulators. The inaccuracies in positioning the PUMA arm stem mainly from backlash, which the visual feedback can correct and the model adaption is robust against. In the hand a combination of stiction and flexibility makes it much harder to control accurately. When trying to make small movements, the hand will initially not move at all, and the controller ramps up the signal. When the hand finally unsticks it grossly overshoots the intended goal. This makes it very hard to control with a feedback controller, and even harder to estimate the visual motor Jacobian. To estimate the Jacobian, large scale movements where the relative effects of this stiction is small need to be added to fine manipulation tasks.
One can argue that the improvement in repeatability found on the PUMA arms when using visual servoing is due to deficencies in the mechanical contruction. The joint angle feedback angle encoder is on the motor and is linked to the joint through the cogwheel transmission. There are certainly modern advanced robots that have overcome problems like this. However the argument that one can buy new expensive robots which do as well without visual feedback misses the main point. The main point is that using visual feedback relaxes the requirements on the robot mechanical construction and now allows us to use inexpensive or older robots in place of high prescision ones. In the future relaxed requirements on mechanical rigidity can allow new families of significantly lighter and faster robots to be built.
While non-adaptive visual servoing methods have been shown to converge in simple settings (e.g., stationary cameras, world coordinate robot control) [Conkie and Chongstitvatana, 1990, Hollinghurst and Cipolla, 1993], we have found that in more complex cases the adaptiveness is crucial. Eye-in-hand type manipulation for instance does not work well without online Jacobian estimation [Jägersand and Nelson, 1995]. The Utah/MIT hand control we presented, as well as control of large, complex object rotations we have shown earlier [Kutulakos and Jägersand, 1995], needed the trust region controller, and we could not make the 12 DOF non-rigid control in Section 5.9 work without the way points and the homotopy method. In addition the visual space way point generation is a natural way of doing trajectory planning in a hand-eye system [Jägersand and Nelson, 1995].
We believe that the results we show are of some general value in that the repeatability for a visual feedback method is insensitive to the exact controller parameters. So long as they converge, different controllers can be expected to converge with about the same repeatability, although the time and trajectory taken may vary, and whether they converge or not depends much on the controller. The convergence results in Section 5.7 should be similar among non-adaptive methods and also should be valuable for adaptive methods using other Jacobian estimation schemes (eg. [Hosoda and Asada, 1994]) or (partially) model based schemas (e.g. [Feddema and Lee, 1990]), so long as the estimate accuracy is similar.