It seems ok to use contact forces from pybullet. I did not do a quantitative analysis (which would be by the way difficult to perform) but the behavior seems ok. Here a screenshot of the arm in touch with a planar surface (not visible).
The wrench is measured cumulating the normal forces at the contact point between the collision meshes. And not measuring forces at the joints which introduce gravity biases.
It seems ok to use contact forces from pybullet. I did not do a quantitative analysis (which would be by the way difficult to perform) but the behavior seems ok. Here a screenshot of the arm in touch with a planar surface (not visible).
The wrench is measured cumulating the normal forces at the contact point between the collision meshes. And not measuring forces at the joints which introduce gravity biases.
The work is here
FYI: @marco-tognon-robotics @JenJenChung