I wrote libero_wrapper and train_libero_ppo.py based on your script, but when I did the training, the last dim(7th) of action inference is almost always 0, meaning the robot is approaching the object but almost never grasp.

I wonder if you got some idea about this. Thank you in advance!