Thanks for this! We will check it soon.

Hello!

I use the passive viewer from time to time, and I'm wondering if you can achieve the same with something like this (untested pseudo-code):

_IS_RUNNING = True
def _key_callback(key: int) -> None:
  global _IS_RUNNING
  if key == 32:  # Space bar
    _IS_RUNNING = not _IS_RUNNING
 
handle = mujoco.viewer.launch_passive(
      model,
      data,
      key_callback=_key_callback,
  )
while handle.is_running():
   with handle.lock()
      if _IS_RUNNING:
        mujoco.mj_step(model, data)
      else:
        # Forward to reflect direct position perturb
        mujoco.mj_forward(model, data)
      handle.sync()
   time.sleep(0.001)

The advantage of this is that you can also create much more complex behavior.

Sure, I also used to toggle run/pause like this. But this PR comes with further benefits:

1. UI shows "PAUSE" banner properly
2. Interactive manipulation of qpos and free pose

This feature does not contradict with "more complex behavior", since the run flag can be toggled from both UI side and program side. It just exposes a mechanism to pause in a more "standard" way than a _IS_RUNNING flag.

There are several other simulator features disabled in passive mode, like step forward/backward, and reload. These features suit well with the key_callback mechanism if needed.

I am curious about what a user should do to visualize a trajectory of saved states without running mj_step. Is the user expected to set handle.run = False, and not press spacebar? EDIT: Sorry to be completely honest the viewer code still confuses me sometimes.

RE(2), this is currently possible by calling mjv_applyPerturbForce and mjv_applyPerturbPose.

Seems like a pretty useful feature - been looking to pause and modify joint qpos.