Edge-AI Driven Distributed Visual Perception System: Real-Time Pose Perception for Robotic Assembly of Large-Scale Aerospace Components

Real-time, high-precision pose perception over large workspaces remains a major challenge in the assembly of large-scale aerospace components. Traditional fixtures function only as passive locating and clamping devices, lacking the integrated sensing capabilities that are critical for advancing toward intelligent and flexible assembly. To address this, we propose an edge-computing-based visual perception architecture with embedded AI functionality. A target detection algorithm based on lightweight neural networks was developed and deployed. It was distributed across multiple binocular edge sensing nodes and was capable of autonomously detecting key datum features such as datum holes and artificial reflective marking targets. Each node performs local inference, including target detection and 3D pose estimation, and communicates via a time-synchronized industrial internet for low-latency coordination. A unified calibration framework aligns multi-node coordinate systems, enabling cross-domain pose fusion and perception directly anchored to the fixture itself. The proposed system is validated in a robotic assembly scenario involving large aerospace parts. Results demonstrate superior deployment flexibility, sensing efficiency, and reduced network and computation load compared to centralized vision systems. This work introduces a scalable paradigm for real-time, intelligent perception in large-component robotic assembly.