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

Precision agriculture has made significant progress in recent years by utilizing technology to optimize crop production, enhance farming efficiency, and automate harvesting processes. Autonomous navigation is a critical component for ground rovers in the agricultural field. This thesis 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 line-of-sight. This breakthrough is especially vital in densely vegetated regions, including areas covered by thick tree canopies or pergola vineyards. The primary objective of the control system is to navigate through entire rows, effectively avoiding obstacles in its path. To ensure versatility across crop types with different row spacing, the rover is designed to operate within the entire inter-row area for crops with small row spacing or predefined lanes for crops with larger ones. The navigation system utilizes a vision-based approach, relying on an RGB-D camera for real-time video streaming analysis to detect and identify row spaces and obstacles. Then, a Non-linear Model Predictive Control (NMPC) strategy is used to compute trajectory and control sequence. The proposed navigation system is implemented in Python and runs into a ROS2 (Robot Operating System) dedicated subsystem. Moreover, the strategy proposed can also be employed in navigation with similar constraints, i.e., a long straight path between two "walls", such as in passages, galleries, etc. A distinctive feature of this system is its ability to recognize and approach objects of interest, such as fruit boxes. Upon identifying a target, the system adjusts its navigation to approach the target object and then resumes its row traversal until it reaches the end of the row. However, the primary scope of this work is the navigation system, so basic image segmentation techniques are employed as a demonstration to identify the targets and validate the approaching and recovery maneuvers. Extensive experimentation is conducted on realistic simulated and real vineyards to demonstrate the competitive advantages of the proposed solution. The controller has exhibited robustness in handling heterogeneity in crop density, height, and other environmental factors. Moreover, it successfully navigates through pergola vineyards and maintains functionality on rough terrains. This research contributes to the ongoing efforts to advance precision agriculture and autonomous navigation in row-based crop environments.