I'd like to discuss a detail in the training process. Should the gripper undergo pre-training initially? Since the gripper and the arm correspond to different RL algorithms, I think train_bc is primarily designed for the arm. During the initial training phase, the gripper would open and close randomly, which could make human demonstrations very difficult, requiring constant repositioning of objects (like wooden blocks) for the gripper and re-grasping them with the 3D spacemouse.
I'd like to discuss a detail in the training process. Should the gripper undergo pre-training initially? Since the gripper and the arm correspond to different RL algorithms, I think train_bc is primarily designed for the arm. During the initial training phase, the gripper would open and close randomly, which could make human demonstrations very difficult, requiring constant repositioning of objects (like wooden blocks) for the gripper and re-grasping them with the 3D spacemouse.