Thanks for releasing the code! It's really helpful.
I have one confusion about the demo augmentation process.

In the inner loop of fill_replay(), line 256 of launch_utils.py, you call _add_keypoints_to_replay() every demo_augmentation_every_n steps. For example, if i = 10 and if the episode_keypoints is [33, 43, 55], then in _add_keypoints_to_replay() the following transitions will be added:

obs of frame 10, target gripper pose from frame 33,
obs of frame 33, target gripper pose from frame 43,
obs of frame 43, target gripper pose from frame 55,

The latter 2 are not dependent on i in the loop. Therefore, for i = 20, the latter 2 transitions will be added to the replay again. Is this what's happening? Or is my understanding not correct?

Thanks!