Autonomous navigation is the key to achieving the full automation of agricultural research and production management (e.g., disease management and yield prediction) using agricultural robots. In this paper, we introduced a vision-based autonomous navigation framework for agriculture robots in trellised cropping systems such as vineyards. To achieve this, we proposed a novel learning-based method to estimate the path traversibility heatmap directly from an RGB-D image and subsequently convert the heatmap to a preferred traversal path. An automatic annotation pipeline was developed to form a training dataset by projecting RTK GPS paths collected during the first setup in a vineyard in corresponding RGB-D images as ground-truth path annotations, allowing a fast model training and fine-tuning without costly human annotation. The trained path detection model was used to develop a full navigation framework consisting of row tracking and row switching modules, enabling a robot to traverse within a crop row and transit between crop rows to cover an entire vineyard autonomously. Extensive field trials were conducted in three different vineyards to demonstrate that the developed path detection model and navigation framework provided a cost-effective, accurate, and robust autonomous navigation solution in the vineyard and could be generalized to unseen vineyards with stable performance.

翻译：自主导航是实现农业机器人全面自动化科研与生产管理（如病虫害管理、产量预测）的关键。本文提出了一种基于视觉的农业机器人自主导航框架，适用于葡萄园等篱架式种植系统。为此，我们提出了一种新颖的基于学习的方法，可直接从RGB-D图像估计路径通行性热力图，进而将热力图转换为优选通行路径。我们开发了一套自动标注流程，通过在葡萄园首次部署时采集实时动态差分全球定位系统（RTK GPS）路径数据，并将其投影到对应的RGB-D图像中作为真实路径标注，从而构建训练集，实现无需昂贵人工标注的快速模型训练与微调。将训练后的路径检测模型融入完整的导航框架（包括行跟踪与行切换模块），使机器人能自主在作物行内通行并在行间转换以覆盖整个葡萄园。在三个不同葡萄园中开展的广泛田间试验表明，所开发的路径检测模型与导航框架提供了一种经济、精准且鲁棒的葡萄园自主导航解决方案，且能稳定泛化至未知葡萄园。