Why
Tendon-driven hands move the actuators out of the hand and into the forearm, then route force through cables. That buys compactness and low distal inertia, and it costs you: tendon length depends on every joint the cable crosses, friction and wrapping are hard to model, and the joints are coupled even when the mechanism looks simple.
I wanted a simulation model where I could change the routing, break the transmission, and see what happens — before committing to hardware.
What I built
- A parametric hand model in MuJoCo, with tendon actuators whose length is computed from joint coordinates rather than hard-coded.
- A joint-space outer loop that tracks reference angles, plus the low-level tendon force allocation underneath it.
- Scripts to sweep routing parameters (moment arms, pretension, friction) and log tracking error.
What I’m figuring out
- How much tracking error comes from friction and wrapping versus model error in the moment arms.
- Whether a simple Coulomb-plus-viscous friction model is enough for pick-and-place, or whether hysteresis has to be modeled explicitly.
- How to validate the simulated transmission against a real tendon-driven finger, not just against itself.
Status
In progress. Code and results will be linked here once they are presentable.