In a GPS-hindered indoor environment, precise positioning is a challenge for mobile robots that perform accurate operations. For instance, during ridge-raising operations in plant factories, robots require positioning accuracy upto 5 cm, which is a pending issue in industry. To address this problem, we propose a ultraprecise simultaneous localization and mapping (SLAM) positioning method, namely, preciseSLAM, which tightly couples LiDAR, inertial measurement unit (IMU), and ultrasonic sensors to achieve ultraprecision, robustness, and real-time positioning performance. First, the preciseSLAM framework is established by fusing multimodal data from LiDAR, IMU, and ultrasonic sensors. Second, to effectively reduce naive SLAM positioning drift after long-term operation, a confidence zone optimization method is proposed for ultrasonic global positioning information. Finally, a factor graph optimization algorithm is developed to tightly couple four preciseSLAM factors, i.e., LiDAR odometry, IMU odometry, loop closure, and the specific ultrasonic global positioning factor. The proposed method effectively addresses the challenges in plant factory environments and can provide reliable and accurate positioning for intelligent robots. Experimental results demonstrate that preciseSLAM achieves a remarkable positioning accuracy of 5.4 cm in the indoor environment inside a plant factory. Compared with that of existing methods, preciseSLAM exhibits ultraprecision, superior robustness, and real-time performance. preciseSLAM provides a general-purpose positioning and navigation solution for robots in large-scale plant factories, which is among the very few methods that can deal with precise indoor positioning.

Efficient drivable region segmentation is a critical for greenhouse robot navigation. State-of-the-art deep learning based road segmentation methods rely largely on labeled datasets to deal with the complexity of unstructured facility agriculture environment. However, the scarcity of annotated datasets limits the model performance. To break the bottleneck, this paper proposes a semi-supervised domain adaptive learning method for unstructured road semantic segmentation. Firstly, we establish a training framework for segmentation models through the transfer learning approach from a synthetic road dataset to an unstructured road dataset. Secondly, we determine the optimal pre-training strategy for solving the greenhouse road segmentation problem. Finally, for the long-tailed distribution of image data in the process of drivable area segmentation, we optimize the loss function to obtain an effective segmentation model for greenhouse robot navigation. For unstructured facility farming scenarios, we created an unstructured road dataset with annotation. Experiments show that, with a small number of labeled data, the road mIoU reaches 98.6%, which is about 10% greater than the existing unstructured road segmentation models to deal with ambiguous boundaries, complex obstacles, and shadow interference. It shows that the proposed method is feasible to leverage the successful existing city self-driving models and datasets to enrich and improve the road segmentation under agricultural scenarios.

Agricultural multi-robot systems in narrow and dynamic environments require global coordination, semantic event interpretation, and responsive trajectory execution. This study presents an end–edge–cloud fast–slow semantic planning framework. The cloud maintains a farm topology and generates fleet-level dispatch policies; the edge hosts an asynchronous agentic vision–language planner and a fast trajectory planner; and the robot performs sensing, LiDAR odometry, low-level control, and execution. The fast planner reuses the latest valid semantic condition until an event-triggered update becomes available. The fast branch is pretrained on nuScenes and adapted using the training and validation subsets of a 3780-sample agricultural dataset comprising synchronized front- and rear-view images, robot states, motion histories, and future trajectories, with an independent 630-sample test set reserved for final evaluation. On an edge-side RTX 4080 SUPER, the complete planner achieves an average L₂ error of 0.67 m, a fast-step latency of 96.3 ms, and a throughput of 10.4 Hz. In the four-robot topology experiment, the framework achieves a 100.0% success rate under the representative single-blockage condition and maintains an 86.7% success rate under the dual-blockage condition. During an approximately 30 min operation at a nominal semantic update rate of 2 Hz, the cloud and robot communication round-trip times average 24.43 and 3.85 ms, respectively, with no robot deadline misses, while the mean trigger-to-updated-trajectory latency of the full event-driven pipeline is 2357.37 ms. These results demonstrate the feasibility of assigning global coordination to the cloud, semantic reasoning and trajectory inference to the edge, and sensing and execution to the robot.