How to read the "act" state in a physical robot actuator

[sanjainarainperspectalabs]

This issue was posted on the MuJoCo mailing list but it was advised to (a) ensure that sensors provide the information needed and (b) post follow-ons here.

We are trying to use MJPC to control a crane with three actuators. These actuators have to be intvelocity. But that seems to mean we have to provide the act part of the MuJoCo state to MJPC. If we don't or we estimate it, MJPC trajectories diverge considerably. We have looked at the documentation of our crane motors and there is no obvious thing that we can read. Have also contacted the manufacturer.

Meanwhile, is there a workaround? Is it possible to just work with qpos and qvel part of the state even with intvelocity actuators? Any links to papers where MJPC has been interfaced with physical robots would be much appreciated.

Thanks.

[yuvaltassa]

Setting act = qpos is a not-horrible estimate, or at least better than 0.
Did you try that?

Basically you need an estimator. @thowell added some estimators to this repo, perhaps he can point you in the right direction.