I want to do e.g. 10 "controlled" steps per every policy action output, that is, I compute 10 motor targets for the next 10 physic steps for every action that the policy sends.
I put my opt_timestep=0.002 and n_frames=1.
based on the MJX example, I tried to do it by modifing the pipeline_step. The original:
I want to do e.g. 10 "controlled" steps per every policy action output, that is, I compute 10 motor targets for the next 10 physic steps for every action that the policy sends. I put my opt_timestep=0.002 and n_frames=1.
based on the MJX example, I tried to do it by modifing the pipeline_step. The original:
and mine:
The leading axis of motor_targets_times is 10. This seems to work, however it is introducing a notable speed decrease about 1.5x~2x.
Don't know why this is happening. Is this expected or am I doing something wrong, and is there a better way to do this?