A robot herding strategy combining CMFF and MPC for shepherding

Agricultural robots are considered a key technological path to alleviate labor shortages, but existing applications are mostly focused on crop production, and research on herding scenarios is still relatively weak. To address the limited effectiveness of herd guidance and control in complex pastoral environments, as well as the lack of real-world validation for robotic herding, this paper proposes a robot herding strategy that integrates the collective motion flow field (CMFF) with model predictive control (MPC), enabling stable driving and guidance of non-cooperative herds. First, a coupled collective motion model of the heterogeneous multi-agent system is established based on the multi-machine collaborative herding platform, and a unified dynamical foundation is thereby provided for subsequent strategy design. Second, at the collective level, a guidance mechanism based on CMFF is formulated to characterize the global motion tendency of the herd and generate dynamic guidance cues. Next, at the robot level, a nonlinear MPC framework is introduced, in which the predicted herd evolution over the prediction horizon is incorporated into the optimization process. Constraint-satisfying optimal control actions are then planned to achieve anticipatory guidance of the herd. Simulation results demonstrate that at a herd size of 300, the success rate of the proposed strategy is 90 %, which is significantly higher than those of the three baseline methods. Compared with the V-formation push strategy, the robot travel distance is reduced by approximately 70 %, and the completion time steps are reduced by approximately 60 %. Meanwhile, the proposed strategy demonstrates strong adaptability to collective behaviors under different self-organization rules. Finally, real-world experiments further demonstrate the engineering feasibility. With appropriate parameter configuration, the real-world experiments of both scales achieve a 100 % success rate. Through the robustness analysis of the real-world experiments, it is confirmed that the proposed strategy can achieve stable and reliable herd guidance in complex pastoral environments, offering new insights and practical references for non-cooperative multi-agent guidance and robotic herding.