Star Lightweight Convolution and NDT-RRT: An Integrated Path Planning Method for Walnut Harvesting Robots.

To address issues such as slow response speed and low detection accuracy in fallen walnut picking robots in complex orchard environments, this paper proposes a detection and path planning method that integrates star-shaped lightweight convolution with NDT-RRT. The method includes the improved lightweight detection model YOLO-FW and an efficient path planning algorithm NDT-RRT. YOLO-FW enhances feature extraction by integrating star-shaped convolution (Star Blocks) and the C3K2 module in the backbone network, while the introduction of a multi-level scale pyramid structure (CA_HSFPN) in the neck network improves multi-scale feature fusion. Additionally, the loss function is replaced with the PIoU loss, which incorporates the concept of Inner-IoU, thus improving regression accuracy while maintaining the model's lightweight nature. The NDT-RRT path planning algorithm builds upon the RRT algorithm by employing node rejection strategies, dynamic step-size adjustment, and target-bias sampling, which reduces planning time while maintaining path quality. Experiments show that, compared to the baseline model, the YOLO-FW model achieves precision, recall, and mAP@0.5 of 90.6%, 90.4%, and 95.7%, respectively, with a volume of only 3.62 MB and a 30.65% reduction in the number of parameters. The NDT-RRT algorithm reduces search time by 87.71% under conditions of relatively optimal paths. Furthermore, a detection and planning system was developed based on the PySide6 framework on an NVIDIA Jetson Xavier NX embedded device. On-site testing demonstrated that the system exhibits good robustness, high precision, and real-time performance in real orchard environments, providing an effective technological reference for the intelligent operation of fallen walnut picking robots.