A Perception Skill for Herding with a 4-Legged Robot

While predators are important for the health of ecosystems, they can also pose challenges in certain situations, especially when they interact with human activities such as agriculture and livestock farming. In this way, effective predator control in extensive livestock farming could be achieved. This work proposes the design, development and integration of a robotic skill to distinguish between different species, grouping them according to whether they are potential predators or harmless species for a flock of sheep. This skill is integrated into a cognitive architecture for helping in scene understanding and integrated with the planning layer of a robotic sheepdog. Thus, if the perception system detects a potential predator, the action to be taken is to scare the predators and make the herd flee. Initial experimental results on images taken by a 4-legged robot achieve a Top-1 Accuracy of 0.9145 and a Parent Accuracy of 0.9576.