Excavator Pose Estimation from Synthetic Point Clouds: A Zero-Shot Generalization Approach

Excavators are a frequent cause of safety accidents due to high degrees of freedom. To mitigate such risks, 3D point cloud-based excavator pose estimation has emerged as a crucial technique. However, this task faces two significant issues: (1) a lack of data collected under real-world construction site conditions, and (2) the labor-intensive nature of manually annotating excavators’ 3D poses. In recent years, data synthesis has proven to be an effective way, but the distribution gap between synthetic and real-world data poses a challenge to the cross-domain generalization performance. To enhance the model’s generalization performance on real-world data, synthetic data augmentation is proposed to expand the diversity and improve the fidelity of the point clouds. Experimental results demonstrate that the proposed method achieves an average pose estimation error of 0.4 m on unseen real-world data, with only a slight decline in performance compared to synthetic data.