Motivation

Our dexterous telemanipulation system uses the CyberGlove®, from Immersion Corp., as an input device. When performing dexterous telemanipulation, it is imperative to calibrate the CyberGlove® to each individual user to ensure repeatable and accurate results. Therefore, we need to devise a scheme to calibrate the user-glove system in a way which is suitable for our telemanipulation task. For our current set-up, we use only the index finger and thumb, to mimic the pinching grasp which the robot is capable of.

Parameter Identification

Three issues are central to calibration of a user wearing the CyberGlove®: the fit of the glove, the properties of the sensors, and the kinematic model of the human hand used to extrapolate the finger positions. The dataglove measures finger joint angles using a proprietary analog bend sensor. Therefore we need to know the conversion factor ("gain") from the analog sensor values to radian angle measurements, as well as a constant offset term. This results in two parameters for each sensor, according to the equation:

In addition, the human thumb has complex kinematics which are not accurately measured by the CyberGlove® in its default configuration. The thumb base joint (trapeziometacarpal joint or TMJ) is a 2-degree-of-freedom rotational joint: its two axes of rotation are neither intersecting nor perpendicular. The two sensors on the CyberGlove® associated with that joint do indeed measure two degrees of freedom, but these rotational axes too are neither perpendicular nor intersecting. Rather than develop a complicated relationship between the sensors and the actual kinematics of the thumb, we chose to assume that the two TMJ angles were each linear functions of the two sensors:

We can extend this cross-coupling to include other joint-sensor relationships as well. The second thumb joint (metacarpophalangeal joint MPJ) actually has two degrees of freedom: in addition to the obvious flexion, it twists and yaws as well (they are inter-related). These additional degrees of freedom are not easily controllable by the human, but are mostly responsive to forces exerted on the thumb.

Our model assumes Hookean pin joints at each bone intersection: the bone lengths themselves are constant. These bone lengths are additional parameters to be calibrated. Finally, although most rotational axes in the fingers are nearly parallel, the thumb MPJ axis is skew to the two axes of the TMJ joint, by some unknown angle. Due to the twisting motion of the MPJ, this 'metacarpal twist', or MCtwist, as we call it, may be a function of some of the other thumb angles. We can employ the cross-coupled gain idea to the MCtwist parameter as well.

In summary, we are now calibrating for the following 30 parameters:

• One offset for each sensor (there are 8)
• One gain for each sensor (again, 8)
• Several cross-gains for cross-coupled sensors, including the effect of sensors on the MCtwist (we used at least 5 of these)
• the length of each bone (there are 8 of these as well)
• the base MCtwist offset angle (only 1 of these)

Calibration Theory

Rohling and Hollerbach, in an earlier paper, describe a method of calibrating a kinematic model of a human hand, concerned with the offset angles only (i.e. no bone lengths or gains). We extend that model to include sensor gains, bone lengths, and cross-coupling factors.

Without going into the gory details, the calibration method proceeds as follows: several 'poses' of the hand are captured, while the user is wearing the glove. The exact positions of the finger tips are recorded, as well as the uncalibrated joint angles. Using knowledge of the kinematics of the hand, we can come up with an iterative solution which minimizes the error between a calculated fingertip position and the actual, measured fingertip position. After several iterations, and provided that the set of 'poses' is sufficiently large, we can obtain actual values for the unknown parameters which allow us to accurately extrapolate the 3-dimensional position of the user's fingertips, given only the sensor readings from the CyberGlove®.

Calibration Implementation

Rather than use an external measurement device (such as a stereoscopic vision system) to capture the actual positions of the user's fingertips, we used a different, rather sneaky approach. If we instruct the user to touch his or her fingertips together while gathering the poses, we can approximate a closed kinematic chain. With the fingertips held together, and the fingers moving around, we can gather approximately 20 seconds of data, and calibrate from this data. Granted, it's not entirely accurate: we assume that the fingertips are in exactly the same position, when in reality they move relative to each other, due to rolling contact and soft tissue deformation. It is, however, sufficiently accurate to provide us with accuracy on the order of ±4mm. Compare this to the uncalibrated accuracy, which is in general worse than ± 30mm, which is quite unacceptable for fine telemanipulation tasks. From our experiments, we have determined that this calibration accuracy is sufficient for the tasks which we wish to perform.

The calibration results are demonstrated in this MOVIE!!. Watch the correspondence between the user's fingers and the computer model, both before and after calibration.

Additional Material

The calibration method is covered in the following publication:

W.B. Griffin, R.P. Findley, M.L. Turner, and M.R. Cutkosky, "Calibration and Mapping of a Human Hand for Dexterous Telemanipulation", Presented at ASME IMECE 2000 Symposium on Haptic Interfaces for Virtual Environments and Teleoperator Systems

More indepth coverage can be found in Michael Turner's PhD Thesis in Chapter 4 and Appendix A and B:

Michael L. Turner, Programming Dexterous Manipulation by Demonstration, 2001 (.pdf 1.4 MB)