During training, the robot may lose sight of the target, resulting in an empty point cloud input to the network. Have you considered this situation?

[Shaw-Way]