“headless mode” has lower rewords than “non-headless mode”

The problem that “headless mode” rewords are lower than “non-headless mode” rewords has not been resolved. I think that both should have the same curve, but what about?
Furthermore, in the case of “headless mode”, it seems that only one episode is calculated.
This task is my own task.
The obvious problem with my own task is that it doesn’t collide when units have to collide with each other. I’m currently ignoring it and running the task.

numEnvs: 512
horizon_length: 32
minibatch_size: 8192

  • Blue
    non-headless mode
    max: 508
    command: python train.py task=TASK headless=False
  • Orange
    headless mode
    max: 489
    command: python train.py task=TASK headless=True

“CUDA 11.5” may be the cause.
“Preview2”, which was working normally before, is now in the same state.
I use “conda” in “Preview2” and “venv” in “Preview3”, so the common part is “CUDA 11.5”.
I have a backup of the SSD before the change to “Preview3”, so I’ll check it.

The cause was found!

I commented out “gym.refresh_rigid_body_state_tensor (sim)” in “compute_observations()” function.

    def compute_observations(self):
        self.gym.refresh_rigid_body_state_tensor(self.sim)  #THIS!
        # self.gym.refresh_force_sensor_tensor(self.sim)
        # self.gym.refresh_dof_force_tensor(self.sim)

Even if you comment this out, for some reason it will be in a normal state in “non-headless mode”, but it will be in an abnormal state in “headless mode”. In other words, unlike “non-headless mode”, the position and pose of the rigid body cannot be obtained accurately in “headless mode”. Please be very careful.

“CUDA 11.5” is irrelevant.

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