Is there a way to change the actions only after 150ms?


assuming self.sim_params.dt= 0.05 is set to 0.05, I want the agent changes its action after 150ms. Basically after every 3 simulation iterations. Is there a way for this?

Hi @hosei2

You can use the controlFrequencyInv configuration parameter for it. This parameter simulates as many specified iterations as shown below:

For example, see AllegroHand task

1 Like

Thank you for the answer.
I am a little bit confused. So in this example of AllegroHand, the sim.dt is set to 0.01667 (which 60Hz). Every 0.01667 seconds, the controller is simulated twice, which makes the control frequency to 30Hz.

However, in my case, the sim.dt is set to 0.05s (which is 20Hz) and the control frequency should be set to 0.15s (~6.67Hz). So, ignore in the first two simulation steps (dont apply self.gym.simulate(self.sim) and only in the third simulation step apply it.

Wouldn’t this badly confuse the agent? Because the computed actions sometimes produce results and sometimes not…?

Hi @hosei2

According to tho the framework documentation (and the implemented code) controlFrequencyInv is defined as the control decimation, ie. how many simulator steps between RL actions. Defaults to 1.

In your case, setting it to 3 will apply new actions at 150ms (control frequency will be lower than a simulation frequency for physics propagating forward) for dt=0.05.

By the way, I think that (depending on the nature of the task) for that dt, there may be a loss in simulation accuracy… unless substeps be higher enough (See Isaac Gym documentation faqs.html in PATH_TO_ISAAC_GYM_PREVIEW/docs folder)

1 Like

According to the code from

# step physics and render each frame
          for i in range(self.control_freq_inv):
              if self.force_render:

In each of the simulation step (which is set to 60Hz), self.gym.simulate(self.sim) is applied twice, as the self.control_freq_inv is set to 2 (see the AllegroHand.yaml). This will make the control frequency to 120Hz, right and not 30Hz, as they have stated in AllegroHand.yaml??

Hi @hosei2

For AllegroHand it means executing the following sequence on each interaction of the agent with the environment

pre_physics_step (apply actions)
simulate (0.01667 sec)
simulate (0.01667 sec)

Then, the actions are applied every 2 * 0.01667 sec (29.9940012 Hz) simulation time

1 Like

This topic was automatically closed 14 days after the last reply. New replies are no longer allowed.