Abstract
Traditional ways of livestock herding and management are a complicated task that relies on human intervention (the herdsman) and training of the animals. The herdsman is expected to know the behaviour of his herd, where the animals are at any given moment, which animal usually breaks away from the herd, and, finally, he needs to move them without causing any panic and stampedes. In such a setup, scaling and monitoring become a very difficult task. An integrated system that uses drones and unmanned ground vehicles (UGV) is proposed. The system maintains geofencing boundaries for the livestock and also acts as a virtual herdsman. Neural networks are used for predictions of animal behaviour, Internet of Things (IoT) sensors for environmental telemetry, and large language models (LLMs) are for planning at higher levels. Natural language commands from the herdsman (farmer-in-the-loop) who cannot code are translated into executable robotic swarm paths. A natural language interface is used. Animals are equipped with an IoT device for data collection. Drones use convolutional neural networks (CNNs) for real-time instance segmentation performance for identifying the centroid of the herd. LLMs are the directors that receive data from the soil sensors and weather to make decisions on ideal grazing sites. Robots receive tactical subtasks. A transformer is used in the herd trajectory prediction. The model analyses the minimal force that should be exerted by the robots in order to direct the herd without causing pressure on the animals. The swarm of UGVs that are on the ground uses reinforcement learning (RL) to execute movement and to push the herd to a selected grazing area.