[Question] Is correct understanding of sim.dt and decimation in policy learning?

[xinyu-66]

Question

Hi, I am trying the sim2real, but the obtained trajectory is too quick with only 100 points, I want it to be more around 2000 points. So if I need to make the sim.dt, physics step to be 1/240 second and decimation to be 8.

Therefore, if it means, the policy will only collect the observation and output action in policy step: 8 * 1/240 = 1/30 second. But the rendering and simulate time step stay at 1/240 second and I can collect more points.

[KyleM73]

Typically you will want the rendering to happen at the policy update frequency, ie sim.render_interval = sim.dt * decimation. Otherwise your understanding is correct.