Surprisingly, a thorough experimental evaluation of properties related to the accuracy of visual feedback control has never been performed to our knowledge. The experimental results in most papers have been limited to a few runs, serving only to validate that a particular algorithm indeed works in practice. The closest to a general evaluation we have found is work by Wijesoma et al. [Wijesoma et al., 1993] that showed, for a 2 DOF implementation, the advantage of visual feedback over open loop control when model errors are large. Chen et al. [Chen et al., 1994] show that two heavy industrial robot arms can perform a peg-in-hole parts mating task with a clearance of 0.7mm between the peg and the hole. The aim of this section is to provide a quantitative evaluation of visual servoing for both robot arms and hands, with respect to positioning repeatability and convergence properties.
We present a thorough experimental evaluation of visual servoing on several real manipulators (PUMA 761 and 762 6 DOF robot arms, and Utah/MIT 16 DOF dextrous hand). Most of these results are not limited to only a specific algorithm, but apply to broader classes of algorithms, see Chapter 8. Each result is based on numerous repetitions of randomly generated positioning tasks. The positioning accuracy of visual feedback is evaluated, and compared to traditional joint feedback positioning in a set of repeatability experiments in Section 5.6 and 5.10. Controller convergence is related to the accuracy of the controller parameters. The convergence of adaptive and non-adaptive controllers are experimentally compared. in Section 5.7. Section 5.8 presents results on how positioning accuracy is dependent on the number and quality of visual feature measurements. Additional experimental results can be found in the technical report [Jägersand and Nelson, 1994].
Figure 5.4: Experimental setup for controller experiments. Above: Initial
configuration of the robot as seen in the left and right cameras. A
visual goal near the floor is given to the algorithm. Below: Goal
configuration achieved by the controller. The image space trajectory
of the tracked features is overlaid.
The following experiments use a common setup. Two cameras are positioned to view the desired portion of the robot workspace. We have found that the exact placement is not particularly important, and that we have a sufficiently well conditioned problem as long as the angular separation of the cameras is greater than about 30 degrees. On the robot end effector we have mounted a 30 by 40 cm T-shaped foam piece with three small light bulbs, one at each end. A fourth light bulb is connected to a rod extending 40 cm above, and 40 cm out of the plane of the T. We use the same general feature trackers that we use later in the visual specification experiments. The light bulbs simply make the tracking more reliable and accurate.
The main purpose of these experiments is to study kinematic accuracy. To make sure the trackers keep up with the features, and that the test rig does not vibrate very much, we run the robot at low speed. At the start of an experiment series a first Jacobian approximation is obtained by executing a sequence of test moves near the middle of the (visually defined) workspace.
A typical experiment is performed as follows: The robot is moved to a pre-defined pose specified by joint angles. Readings are taken from the trackers to define a visual goal. This ensures that the visual goal is actually attainable, something which is typically not true of a set of randomly selected feature values. After registering the goal, the robot is deflected to a random position in the selected workspace. During the deflection all control algorithm input is shut out, disabling model updating. At the end of the deflection additional noise is injected into the Jacobian model, and other adjustments are made to the control algorithm depending on the purpose of the experiment.
The controller is then turned back on, and relevant measuring processes are started, acquiring, for example, timing and visual trajectory information of the sort plotted in Fig. 5.4. The controller is turned off when the visual goal is achieved within a preset accuracy, or when it is no longer moving any closer to the goal. Endpoint accuracy measurements are taken at this point. Then the whole procedure is repeated a number of times (between 50 and 200 times), for different trajectories that have been selected with random sampling in the manipulator workspace, until the results are statistically significant.