Non-linear Model Predictive Control for Multi-task GPS-free Autonomous Navigation in Vineyards

TL;DR

The paper addresses GPS-denied autonomous navigation for a rover in row-based vineyards by developing a position-agnostic NMPC controller that uses a single RGB-D camera to drive along intra-row space while avoiding obstacles and approaching targets. The method couples 2D point-cloud processing with a tailored NMPC objective that includes a lane-centering term $C_{lane}$, an alignment term $C_{align}$, a quadratic input-change term $r$, and a terminal term $m(x)$ under a constrained unicycle model with quaternion orientation, enforcing obstacle constraints $- (x_1 - o^i_1)^2 - (x_2 - o^i_2)^2 + R^2 \\le 0$. Key contributions include a robust, cost-effective RGB-D based navigation pipeline, a 2D border estimation with safety margin $R$, and comprehensive validation in both simulated and real vineyards across straight, curved, and pergola configurations. Experiments demonstrate speeds approaching the limits $v_{x,max}$ with centimeter-scale trajectory errors, supporting practical applicability to precision agriculture without GPS or expensive RTK systems. The approach reduces sensor cost and expands GPS-denied capabilities, offering a viable path toward scalable autonomous farming deployments.

Abstract

Autonomous navigation is the foundation of agricultural robots. This paper focuses on developing an advanced autonomous navigation system for a rover operating within row-based crops. A position-agnostic system is proposed to address the challenging situation when standard localization methods, like GPS, fail due to unfavorable weather or obstructed signals. This breakthrough is especially vital in densely vegetated regions, including areas covered by thick tree canopies or pergola vineyards. This work proposed a novel system that leverages a single RGB-D camera and a Non-linear Model Predictive Control strategy to navigate through entire rows, adapting to various crop spacing. The presented solution demonstrates versatility in handling diverse crop densities, environmental factors, and multiple navigation tasks to support agricultural activities at an extremely cost-effective implementation. Experimental validation in simulated and real vineyards underscores the system's robustness and competitiveness in both standard row traversal and target objects approach.

Figures (7)

• Figure 1: Vineyards used for testing the proposed navigation system in Agliè, Turin, Italy.
• Figure 2: Computation Graph of the ROS 2 overall application system. A Behavior Tree manages and coordinates the NMPC controller for row traversal, target object approach, and recovery behaviors for robust multi-task navigation.
• Figure 3: The black points represent the input PCD, filtered and flattened on a 2D map (obstacles), the green area is the free space in front of the rover, while the two straight lanes represent the lane borders. Finally, the dotted green line is the expected trajectory as computed by the NMPC controller.
• Figure 4: Aerial view of vineyards in Gazebo used for testing in simulation.
• Figure 5: Satellite view of the vineyard. In red the trajectory followed by the Husky rover during a test session.