Minimum-time row transition control of a vision-guided agricultural robot

This paper presents a vision-based, subspace optimal controller aiming to improve the row transition performance of an agricultural robot in a strawberry field. The contribution of this paper is twofold. First, only RGB cameras, instead of complicated sensor suites, are used for cross-bed navigation and row alignment. Second, a real-time adaptive dynamic programming-based algorithm is designed for an optimal row transition. The conditions for row alignment are derived in an augmented pixel coordinate frame. Based on these conditions, a simple motion rule is utilized to reduce the search space dimension so that the proposed algorithm can be implemented in real-time. Additionally, the inverse-dynamics policy of the algorithm is updated using vision feedback at each control step to adapt to uncertainties. The proposed controller is tested in both simulations and field experiments. In a simulation comparison, the minimum-time solution achieved using the proposed algorithm is 44.7 s, which is very close to that of a benchmark algorithm (44.4 s). However, the CPU time required by the proposed algorithm is only 4.3% of time needed by the benchmark algorithm. Twenty field experiments using the presented design were all successful in row transition, with a mean final alignment error of 0.5 cm.

Automated summary

This paper introduces a vision-based controller for improving the row transition performance of an agricultural robot in a strawberry field. The controller utilizes only RGB cameras for navigation and row alignment, and features a real-time adaptive dynamic programming algorithm for optimal row transition. The proposed controller shows promising results in simulations and field experiments, achieving efficient row transition with a minimal alignment error.