The augmentation of robotic controllers with proprioceptive inputs in Evolutionary Robotics have been shown to drastically improve controller performance on complex tasks. Previous research used a neural network to produce approximate location information as proprioceptive inputs to a robot controller which do not have other access to location information. This software-only technique produced controllers that were significantly more effective. However, it was found that for the approach to be used for an extended period of time, it becomes necessary to periodically reset and correct the approximate location information. Resetting has that drawback that it may not be possible in all situations and may require hardware intervention. This paper presents an alternative approach to correcting location information which makes use of a task-specific rectification neural network. Two real-world experiments are conducted to evaluate the approach and to identify limitations. Experimental results show clear performance benefits of using the rectification network.