Browsing by Person "Boysen, Jonas"

Now showing 1 - 1 of 1

AI-assisted tractor control for secondary tillage
(2025) Boysen, Jonas; Bökle, Sebastian; Stein, Anthony
Modern agricultural machinery requires skilled operators to optimally configure their complex machines, while autonomous machines without operators must already optimize their configuration themselves to achieve optimal performance. During secondary tillage multiple performance measures need to be monitored and maximized: Seedbed quality, area output and fuel consumption. The seedbed quality can be measured with the soil surface roughness coefficient which can be computed with 3D-cameras attached to the machine. For our work, such cameras are mounted in the front and back of a Claas Arion 660 tractor with an attached power harrow seeding combination. The soil-machine response model of our prior work is utilized to model the soil-machine interaction for the training of a reinforcement learning agent and the application of a decision-time planning agent to assist in controlling the working speed of the machine. The control agents are tested in real-world field trials and compared to good professional practice. The decision-time planning agent achieves comparable results to a gold-standard while reaching significantly higher performance in terms of area output (29.1%) and more efficient fuel consumption (8.4%) than a baseline while the reinforcement learning agent performed worse during the field trials. The seedbed quality and field emergence are not showing significant differences between the variants. Further analysis shows that model training and selection for the reinforcement agent could have led to performance loss and models that are performing better in simulation have been trained after the field trials. Furthermore, we analyze the models when tested under the field conditions in the field trials (out-of-distribution) that are different from the field conditions during training data collection. The out-of-distribution testing leads to a reduced performance in terms of rRMSE of the decision-time planning agent and to some extend reward of the reinforcement learning agent compared to in-distribution testing.