Leveraging the robot's state improves deep learning models' performance in non-egocentric perception tasks.