Comments by MarkCutkosky

Karlin:

Overall I found this paper hard to follow...

-I had some trouble with the equations: equation 9- it wasn't obvious to me how they arrived at this equation from "taking the derivative of eqn 2"- derivative with respect to what?; equations 12 and 13- again- it just seems like their descriptions of how the equations were derived aren't helpful- for example, Eq.12 = "substituting eqn 6 into 1...", but I think they mean "substituting eqn 7 into 1" but even then, the variable declarations are not exactly the same which makes it hard to see immediately how the equations were derived.

MRC: This is a standard stiffness transformation from joint space to Cartesian space. (It is in my thesis and papers by Kao and Cutkosky and by Salisbury and others.) Think of stiffness as rate of change of force with respect to displacement. So, let K = dF/dX or let K = dTau/dTheta. Generally, I did not have problems with their derivations but there are some known issues with the stiffness mapping stuff. It only works for infinitesimal motions.

I personally (in my humble opinion) think the Kao and Cutkosky derivation is better

http://www.humanneuro.physiol.umu.se/PDF-SCIENCE/1997_k_c_j_ieee_DIST.pdf

%ENDCOMMENT%

-lots of formatting inconsistencies and questionable placement of figures. Fonts in the figures don't match fonts in the text, figure numbers are inconsistent, spacing of captions needs to be fixed...

Yes.

-figure 5, of the experimental setup is not helpful...the drawing of the "person" is really unclear.

Yes.

-I was surprised at the lack of data/results- there's some time spent describing an experimental setup and trials, but then not too much on the real data. It took me awhile to look at the figure of the isoeffector space (figure 6) and realize what it was. Also, in that figure there's a circle that indicates motor noise, and I think it's a poor way of depicting the size of motor noise on that particular figure...it just seems so abstract. perhaps if it was a shaded region under the region "C", it would make more sense.

-at the end of the abstract, they state that "we use this control strategy to control our ACT hand and investigate the neural control strategies..." but unless I am missing something, I didn't see any implementation of a control strategy on their ACT.

-their results show a large space of various forces/configurations to produce a force/stiffness at the fingertip, and human test subjects only used a small portion of that space, but were the test subjects told to change their configurations? I guess that is expected simply from their setup. If the subject was only asked to press the end of a phantom with their index finger to generate a specific amount of force, there would be no reason for them to move their finger and use different muscles. In my head, I picture a subject keeping their index finger in the same geometric position throughout the experiment...I thought the purpose of the experiment was to validate the derivations presented, but I don't think the experiment really does that.

- I also have some problem with the test. It doesn't really show that their mapping works in any general way. They only looked at one dimension of force and stiffness -- for which there could be many possible mappings of the muscles and joint stiffness as the middle and distal joint.

In fact, I am surprised they did not say anything about the coupling between the middle and distal and metacarpophalangeal joints... Although there are 7 actuators and 4 DOF in 3 space there is generally coupling.

---

Li:

The paper described a control framework inspired by structure and control of human hand. For each joint on human hand, there are at least two muscles controlling the motion of the joint. Muscle forces oppose each other to generate a joint torque, however the stiffness adds up to produce the overall stiffness. The authors use this basic relationship to define a framework that can control the end effector's force and stiffness separately.

I think this paper has some good research presented in there, but it is not very well organized.

On the third page of this paper, the figure on the right side has a figure number "4.2". It seems the author just copied the figure from his/her thesis without changing the figure number.

Indeed...

On the second page, for the equation (1), it was described that the model is decaying exponential.

Is "decaying exponential" the right word to use here? It confused me at first, when I first read "decaying exponential", I thought about something like exp(-f). But in this case, it is not. The k is monotonic increasing with f, but it increase very fast when f is small and get slower and slower as f get bigger.

I agree, this is confusing to the controls community

The last sentence of the abstract said that they use the control strategy to control their anatomically correct testbed (ACT). However, they didn't mention how they use the control strategy on the testbed in this paper. And it seems there is no need to mention too much about their ACT in this paper, since this paper is mainly about the control framework and they didn't show how they use the framework to control the ACT.

Karlin noticed this too.

In the paper, they keep mention in their index finger model there are 3 joint, 4 degrees of freedom. But usually 3 joints provide 3 DOF. Is the fourth DOF come from the yaw of the first joint? They should explain more about this.

It is a little ambiguous...

On figure 5, the drawing of finger is very confusing. At first I thought the person is holding a box. The box has a fake finger mounted on it. And the person used this thing to apply force on the sensor. It is much better to show a real picture of the setup here.

Karlin noticed this too.

---

Jason:

Here are my comments on the IJRR paper by Afshar and Matsuoka, as requested. I should mention that I was holding it to a high standard given the reputation of the journal.

Generally, the idea was interesting though not completely novel. I like the marriage of the two fields (biomechanical modeling and neural control meets robot manipulation) The significance of the application as stated in the introduction is real, however I felt that the conclusions reached fell short of the stated goals. I would expect more significant (and generalizable) conclusions in a journal of this stature. I also felt that it was less polished that expected (there was a figure out of place (Figure 4.2?) and some of the variables used were not clearly defined (though discernable with a little effort).

I agree

Here are a few more specific comments.

I would have like to see more justification for the biomechanical models and parameters used as this is critical to the outcome. For instance equation (1). Also, did they model tendon compliance (I assume not). This may be significant for the long tendons in the hand. What about assuming constant muscle moment arms (page 4).?

Exactly what the 4 DOF and 3 joints used int he finger model was not obvious. I assume the base joint has 2 DOF base on the D-H table but I'm not 100% sure.

I like that they used a simpler 1 DOF model to explain the concept but it leaves little room for details.

The experimental results from 3 subjects show consistency in their iso-effector space. The results for the iso-force and stiffness are not shown. I would like to see that to know if the clustering also exists there, in which case the results would be less impressive. There are also no statistics shown or attempts to quantify the effects they are seeing only figure 6. The discussion section could do more to tie the results to the stated goals.

I agree

---

Summary

The authors describe a procedure to describe and specify the simultaneous control of forces and stiffness terms in a biologically inspired, redundantly actuated mechanism. The motivating example for the work is the ACT hand. The topic and the presentation are interesting. However, the paper requires some revisions before I would recommend it for publication:

1. The title may be misleading. I would suggest something like "Framework for the control of force and stiffness with neuromusculoskeltal redundancy." The broader problems of manipulation, including trajectories, dynamics, grasp stiffness, etc. are not dealt with. The analysis is instantaneous and static.

2. The manuscript draft is sloppy. There are mis-numbered figures (e.g. "Figure 4.2" on page 3) and improperly formatted captions (Fig. 2.). Figure 5. is so iconic and stylized that it gives the reader very little sense of what the setup and display actually were like. Figure 6 also requires more explanation.

3. The transformations in eq. (2) - (10) are a fairly straightforward specialization of cartesian stiffness control as applied to robot and human hands (e.g. [Kao 1997]). This approach has some known limitations (e.g. only infinitesimal motions about an equilibrium configuration, and issues of coordinate frame invariance as acknowledged by Kao et al.)

What is new is that treatment of redundant actuators and the consideration of the iso-stiffness space as a function of muscle properties.

4. The experiment does not entirely convince one that the authors' mapping process works as they claim. Only the vertical stiffness is considered -- and in this particular direction there are multiple possible ways in which muscle force and stiffness could produce the desired effect.

The more general case would involve looking at simultaneous force and stiffness specification in multiple directions. In this case, coupling terms would become important. For example, the authors do not talk about the coupling between the torque that people produce at the middle and proximal metacarpophalangeal joints. Can we really control the torque and stiffness at each of these joints independently or is there coupling? Depending on the finger configuration and the mapping from joint space to task space, it may or may not matter.

5. The results of the Validation Experiment are not described in sufficient detail. Figure 6 is really not enough to determine how well the results match. The authors say they "...performed a linear regression on the .... EMG..." -- How well did it fit? The "Motor noise" circle in Fig. 6 is shown to scale with the contour of M and C? What are the units along the horizontal and vertical dimensions of Fig. 6? How large are the M and C boundaries, actually?

Additional points:

Even if one considers only infinitesimal motions with respect to an initial configuration, there is an issue with stiffness control that one needs to be careful about: If bias forces or torques are high in the original configuration, even small motions with respect to the configuration may produce changes in forces (due to small changes in the Jacobians) that are comparable to those produced directly by the [K]*dx terms. Could this be the true in the case considered in the paper? (It depends on the finger configuration and on how hard people are pushing.)

The notation could be cleaned up. There are terms like script upper-case bold-face M^(f) in Figure 2 and a different font used in the text and caption.

The authors say just after eq. (5) that "and thus these matrices are constant with respect to theta..." Actually, the point is that R and J are held constant for a single configuration. It's not that they are independent of theta (or of time, or anything else).

The authors say the stiffness is a decaying exponential in equation (1). It might be clearer to say that it approaches a constant.

In the most general case, biological systems may be redundant with respect to some DOF and under-actuated with respect to others. The framework could be extended to address this case. It will transpire that there is controllability, via the null space, in some DOF and not in others.

---

Refs:

[Kao 1997] KAO I, CUTKOSKEY MR, JOHANSSON RS. "Robotic stiffness control and calibration as applied to human grasping tasks," IEEE Trans on Robotics and Automation 1997 13:557-566.