Question related to how it has been trained the architecture. Trained from video of human doing a task with hands or moving a robot by jogging?

[kappa95]

Is it not clear to me the purpose of the camera in the work. Are they used for monitoring the robot state? For the training side, how the neural network learned the task? is it trained using only a video of the operator doing the task? or teleoperation of the robot while camera observing the scene? Besides, how are translated those frames to training data?