About the action space of Feedback Policy

[gouyinghong]

Hi, I assume that the action space of the Feedback policy is defined relative to the gripper frame coordinate. During training, five frames are selected, with the last frame serving as the goal and the other frames as the current state for action prediction. However, why is the rel_action of current frame considered the ground truth? Isn’t rel_action supposed to represent the end-effector state of the current frame? I would greatly appreciate it if you could help clarify this.