frankaemika/libfranka

Missing features related to singularity and zero endeffector wrench

Opened this issue · 2 comments

ndehio commented

I am controlling my panda (version 4.0.2) via the libfranka interface.
Evaluating the robot's state, I noticed that the (estimated) endeffector wrench (see methods K_F_ext_hat_K and O_F_ext_hat_K) sometimes contains only zeros: [0,0,0,0,0,0]. This effect appears for both our two pandas and both for applying velocity-control and torque-control.

I understand that the problematic wrench estimation relates to singularities: it is not possible to infer the full wrench from joint torque measurements if the jacobian becomes rank-deficient, due to the required pseudoinverse. According to my experience, libfranka uses a threshold of 0.08 for the smallest singular value when outputting the zero wrench.

Please add the following features:

  • mention/explain this impossible wrench-estimation problem in the documentation (in the first moment, I was kind of irritated)
  • add the six singular values to the robot state (apparently they are computed anyway)
  • add a boolean to the robot state that indicates if the robot is close to singularity or not
  • add a function that allows specifying the singularity threshold (i.e. 0.08 as default)

However, even in singularities it is possible to compute/estimate a "reduced" end-effector wrench (the wrench estimation is of course missing one Cartesian direction, related to the small singular value). Why don't you provide this one in the robot state? I don't see a need to zero the entire wrench...

For those who are curios, here are two exemplary joint configurations that result in zero end-effector wrench:
[1.03634,0.715631,0.222965,-0.620327,0.112605,2.93128,2.06283]
and
[0.0300161,0.483728,-0.195585,-0.612593,-0.0135165,1.30887,2.04615]

kdesorme commented

Hi @ndehio ,

i think i ran in the same problem when i get close to a certain configuration:

traj_test_without_admittance

I can reproduce this problem every time i execute a certain trajectory.
If you have any news/found any solution for this problem i would appreciate your inputs!

Cordially,

Kevin Desormeaux.

ndehio commented

You can easily compute the jacobian pseudoinverse via SVD or QRD yourself and derive the end-effector wrench given the joint torques. So, you don't need to rely on this strange libfranka behavior.
Maybe you want to have a look at my PhD dissertation, see equations 2.8, 4.7, 4.11 and appendix B
https://core.ac.uk/reader/196653490
Feel free to cite it if it helps ;-)