zero dynamic actuators

[jrmch12]

Correct me if I am wrong, but I think actuators always lead to a qfrc_actuator. Therefore actuators are always dynamic.

I think a zero dynamic actuator, i.e. an actuator that directly override the joint qpos data, could be useful in scenarios where the control of such a joint is pretty simple, in comparison with the rest of actuators. I know you can have this behaviour simply by overwriten this qpos at each step, but it would be cool for the mujoco pipeline to consider it as an actuator. Then it is easier to connect our mujoco agent to a RL pipeline for example, using the python api for actuators.

My particular case, is a pixel to torque RL controller for a drone using Madrona-mjx. The four actuators for the motors control the dynamics of the drone, which is something RL agent have to learn. But for my camera, which have a hinge joint to tilt , I would like to control this tilt inmediatly (just kinematics) and do not add physics burden to the simulation, as in real life this control is super simple. Also to avoid tuning the Kp and Kv of a position actuator.

:)