jgvictores
commented
5 years ago A couple of things we've commented:
- It would be nice for OpenRAVE to expose the motor command, thus not expose the individual joints, rendering these internals not extremely relevant.
- I'd like to conserve our sign conventions. Considering hands a separate entity with the base at the bottom (even with different DH models for right/left hand), this would mean most of the Z axes (joint rotations) would point right, and close upon a positive rotation. Note that these are internals of the simulation, but it's good to have a clear criteria to always speak in a unified manner that applies to simulation and physical hand.
At https://github.com/roboticslab-uc3m/sign-language-papers/issues/7, we trained a series of NN models to express motor joint commands given actual angles between consecutive finger joints. Those models were given the opposite relationship, i.e. resulting finger joint angles given a motor command. The problem is: OpenRAVE's dextra models work with a different configuration regarding directions of rotation, i.e. finger joints move between
-7degrees (almost fully open) and-85degrees (fully closed, perhaps unattainable?) whereas said models are always given positive angles.Decide on either updating OpenRAVE models accordingly or applying a conversion on the angle inputs to the trained NNs (or: generate those models again). Perhaps initial pose and limits should be reviewed, too.
PS: mind the thumb!