Clarification on ALOHA Cartesian control #187
Description
I'm trying to figure out how to use the ALOHA 2 model here with the original ACT code from Tony Zhao.
In the original ACT code, the trajectory is first executed with cartesian control while recording the joint positions. Then the joint positions are replayed with joint control, and the state and action are recorded.
How would this be accomplished with the ALOHA model shared here? I was confused by the following comment in filtered_cartesian_actuators.xml:
The dynamical parameters in this file are not the result of system identification.
They were chosen for MJPC efficiency. The resulting qfrc is still limited by every joint's actuator_frcrange,
as specified in the main model.
Am I correctly interpreting that when using these cartesian actuators the joints will still be limited by the actuatorfrcrange? This does not seem to follow from the Mujoco docs. My understanding of these cartesian actuators is that they model exerting a force at the end effector (as if you were backdriving the robot by hand), which would not be limited by the joint motor force limits. Which interpretation is correct?
If the cartesian control does actually model the dynamics estimated by the least-squares optimization, then it would be unnecessary to do the "replay" for ACT.