approaches based on lightweight, fully convolutional networks (FCNs) [16]. The latter showed superior performance and better generalization capability. To effectively generalize to new conditions (different fields, weather, and so on), we exploited geometric patterns resulting from the fact that several crops were sown in rows. Within a field of row crops, the plants share a similar lattice distance along the row, whereas weeds appear randomly. In contrast to visual cues, this geometric signal is much less affected by changes in visual GPS Onboard Computer DJI M100 Onboard GPU Computer VI Sensor (a) (b) Laser Scanner (Velodyne) Laser Scanner (Velodyne) Laser Scanner (FX8) High-Resolution RGB Camera and/or VI Sensor Lever Arm Lever Arm PPP-GPS + RTK-GPS Arms Lever Arm Encoders Weed-Intervention Module Wheel Encoder (c) Figure 2. The two main robots used in the experiments and demonstrations: (a) a DJI Matrice 100 UAV multirotor performing an autonomous flight over a sugar beet field, (b) the UAV with the installed sensors, and (c) the Bosch BoniRob farming UGV. PPP: Precise Point Positioning. 32 * IEEE ROBOTICS & AUTOMATION MAGAZINE * SEPTEMBER 2021