Abstract
INTRODUCTION: Automating shiitake mushroom picking is critical for modern agriculture, yet its biological traits hinder automation via target recognition, path planning, and precision challenges. Traditional manual picking is inefficient, labor-heavy, and unsuitable for large-scale production. In human- robot collaboration, computer vision - based human motion prediction enables efficient picking coordination, yet methods like LSTM and static graph networks struggle with robust spatiotemporal correlation capture and long-term stability in complex agricultural settings. METHODS: To address this, we propose the Chaos-Optimized Adaptive Spatiotemporal Graph Convolutional Network (CSO-ASTGCN). First, it integrates three core modules: the Adaptive Spatial Feature Graph Convolution Module (ASF-GCN) for dynamic joint correlation modeling (e.g., wrist-finger coupling during grasping). Second, the Dynamic Temporal Feature Graph Convolution Module (DT-GCN) captures multi-scale temporal dependencies. Third, Chaos Search Optimization (CSO) globally optimizes hyper parameters to avoid local optima common in traditional optimization methods. Additionally, a flexible control system fuses CSO-ASTGCN motion prediction with GRCNN grasp pose estimation to optimize grasping paths and operational forces. RESULTS: Experiments show our model reduces the Mean Per - Joint Position Error (MPJPE) by 15.2% on the CMU dataset and 12.7% on the 3DPW dataset compared to methods like STSGCN and Transformers. The human - robot collaborative system boosts picking efficiency by 31% and cuts mushroom damage by 26% relative to manual operations. DISCUSSION: These results validate CSO - ASTGCN's superiority in spatiotemporal modeling for fine - grained agricultural motions and its practical value in intelligent edible fungi harvesting.