Learning-based gain-scheduling of trajectory tracking controllers for agricultural mobile manipulators under off-road conditions

Currently, mobile robotic applications in the agricultural field are being called for safety and accurate handling of farm crops. During the crop yield, manoeuvring tasks on changing and heterogeneous terrain surfaces lead autonomous vehicles to loose motion precision due to slipping or sliding phenomena, thus requiring an adaptable motion strategy to overcome a deteriorated control performance. In this scenario, this work proposes a gain-scheduling technique based on three non-supervised learning algorithms. In particular, clustering and self-tuning strategies are combined to obtain the best control parameters of trajectory tracking controllers. The proposed approaches are real-time implemented on two motion controllers and tested on an omnidirectional holonomic mobile manipulator -KUKA youBot. Results from trials with different reference trajectories and navigation terrains showed that the control performance could be enhanced, reaching around a 30.1% of tracking error reduction and 54.6% of total cost when using the proposed clustering approaches. The latter may impact on the energy resources of the mobile manipulator throughout harvesting tasks.