Autonomous navigation in orchards presents unique challenges due to dynamic environmental conditions and farming tasks such as monitoring, weeding, spraying, harvesting, and transporting goods. In field operations, GNSS-based navigation systems often experience signal interference caused by dense tree canopies, making them unreliable for precise local environmental perception. As an alternative, using cameras supported by computer vision techniques has emerged as a promising solution for local navigation in orchards. Compared to traditional computer vision techniques, the deep learning approach offers greater performance, better adaptability to complex environments, and efficient integration into field robots that require real-time processing. This chapter explores deep learning-based instance segmentation to support autonomous navigation in orchards by taking advantage of objects and their shape information. Some recent instance segmentation models are highlighted as fundamental components for navigation paths estimated from the tree trunk’s location. This chapter also describes the vision-based navigation system development pipeline: data acquisition, annotation, model training and validation, prediction, and model deployment on edge devices. Moreover, the extraction of the navigation path for robot guidance to track the tree row is explained. This chapter provides a comprehensive overview and practical insights into deep learning-based instance segmentation for autonomous navigation robots in orchards.