DMF-TIA: a data-model fusion-driven triply integrated assembly framework oriented toward modular antenna structures
Yulin Zhang, Tuanjie Li, Ziang Li, Lixiang Ban, Xudong Liu, Shichao Yang, Yuming Ning, Yan Zhang, Xiangyuan LiPurpose
This paper aims to address the autonomous assembly of modular antenna units in ground-based unmanned manufacturing stations for future on-orbit deployment. Existing robotic assembly approaches suffer from separated perception-planning-execution pipelines, heavy reliance on manual calibration, and limited generalization. A unified scheme that tightly integrates perception, reasoning and skill execution is needed to enable flexible, human-free modular antenna assembly.
Design/methodology/approach
A data-model fusion-driven triply integrated assembly (DMF-TIA) framework is proposed. An incremental perception module uses coarse-to-fine HSV filtering and point-cloud matching with variable-step optical-center adjustment. An assembly reasoning module couples a relational graph convolutional network with a hidden semi-Markov model to infer feasible assemblies and decode duration-aware skill chains. A skill execution module employs adaptive oscillatory primitives with a joint-space trajectory mapper for inverse-kinematics-free, few-shot execution.
Findings
Real-world experiments on a 7-DOF manipulator achieve a 93.3% assembly success rate across multilevel tasks. Compared with the five baseline methods, DMF-TIA achieves the lowest execution time and target position error, corresponding to 85.4% and 34.3% of the worst baseline values, respectively. Module-level ablations confirm the complementary contributions of the three modules, while robustness tests characterize the operating envelope under illumination, occlusion and viewing-angle variations.
Originality/value
To the best of the authors’ knowledge, this work presents the first triply integrated perception–reasoning–learning framework for modular antenna assembly. The key novelty lies in the complementary coupling of graph neural networks and hidden semi-Markov models, simultaneously exploiting assembly topology and temporal skill sequencing. The adaptive oscillatory primitive–based skill module eliminates inverse kinematics dependence, enabling efficient few-shot learning and providing a practical backbone for unmanned manufacturing stations.