LRPE: A Lightweight Robot Pose Estimation Network

In fenceless Human-Robot Collaboration (HRC), safety relies on precise robot pose perception. However, training robust deep learning models for this task typically demands expensive real-world annotations. To address this data bottleneck without incurring high computational costs, we propose a lightweight robot pose estimation network. Unlike existing methods that rely on extensive domain randomization, our approach introduces a novel Feature Filtering Module (FFM). This module explicitly filters out domain-specific style to isolate domain-invariant geometry using instance normalization and orthogonality constraints. Trained solely on synthetic data, our method achieves 77.69% AP \(_{50:95}\) on the challenging Panda-Orb real-world benchmark, significantly outperforming the state-of-the-art YOLOv12-pose baseline. Notably, our approach reduces computational cost by approximately 35% and parameter count by 25%, offering a superior accuracy-efficiency trade-off for real-time deployment.