[Question] Tuning Stiffness and Damping arm robot

[AntonioClaudiossf]

Question

Hi, I am trying to control the robot arm (Franka and UR10) using JointPositionActionCfg() and RelativeJointPositionActionCfg(), and I noticed some issues with ImplicitActuatorCfg(). Let me explain clearly: when the robot's joint setpoints are defined, the actuator is unable to reach or maintain the desired position.

To verify this, I performed two tests using both control modes by placing the robot in an initial configuration (I used usd model available in IsaacLab with the same configurations.):

For the control with JointPositionActionCfg(), I set the initial joint positions as the target. Therefore, the expected result is that the robot should not move.

For the control with RelativeJointPositionActionCfg(), I set all joint commands to zero so that there would be no increment in joint positions. Again, the expected result is that the robot should not move.

However, when I start the simulation, the robot simply collapses in both cases, meaning that the actuator is not capable of maintaining the robot’s position.

I show this behavior in the video:

Screencast.from.05-25-2025.09.59.06.PM.webm

When I disable gravity, I notice a better behavior from the actuator. However, I would like to keep gravity enabled in the simulation for my Articulation robot. Therefore, I would like to know how the stiffness and damping values were determined for the robots. I suspect these parameters might be the cause of the issue. Were they chosen arbitrarily, or was there a specific method used to define them?

Another point: I noticed that when I load the robot through the Isaac Sim GUI, it stays in position and does not collapse.

[Jdowdy05]

The consensus is that you aren't supposed to use ImplicitActuator, so if you can change to an ideal PD actuator, then you should do that. However, if you can't change from implicit actuator for whatever reason, what is your render interval set to? If you increase this, you will see that the robot starts to track. I found that if my physics dt is at 500 Hz and render interval is set to 1, then I get good tracking performance, though you can play with this yourself and see what works best for you!

[RandomOakForest]

Thanks for posting this. Also note that there are some fixes underway related to #1552 on the default values for stiffness and damping.

[AntonioClaudiossf]

Hi @Jdowdy05 and @RandomOakForest thanks for reply.

I tried to change the frequency to 500Hz, but that didn’t completely solve the problem (and this might cause some issues when transferring the simulation to the real robot). I also tried switching to idealPDActuator, but the robot still shakes a lot. The best result I got for controlling the robot to follow the target was using ImplicitActuator with high stiffness. I believe I’ll use these settings. That way, I can interpret the neural network output as the command velocity for each joint, clip it to the robot's limits, and multiply by env_dt to get my dq.

Apparently, it worked well for the learning process, and the robot is able to follow the target joint values.