Optimizing the Trajectory of Agricultural Robots in Greenhouse Climatic Sensing with Deep Reinforcement Learning

Having a complete climate mapping of a greenhouse at one's disposal is undoubtedly an invaluable task in agriculture. However, this ability usually comes at the price of additional costs and complexity. For this purpose, the concept of virtual sensors, where physical sensors are substituted with more or less complex models that have been trained with previously collected data, has been proposed in the literature. This data collection can be performed in different ways, for instance, through temporary sensors that are removed after a limited period or through an agricultural robot that collects measurements during its tasks. In this regard, choosing the best visiting sequence of the Points of Interest (PoIs) where to perform the measurements deserves particular attention. In this paper, we propose a method based on Deep Reinforcement Learning to determine the trajectory of visiting the various PoIs by adequately balancing the number of measurements and their utility. The proposed approach has been applied to a real-world case regarding a greenhouse in Verona (Italy) and compared with another state-of-the-art approach proposed in the literature.