S${3}$aDPWo: Spatial-, Semantic-, and Shape-Aware Diffusion Policy Toward Autonomous Wound Repair

Abstract

Imitation learning (IL) offers a promising pathway for enabling surgical robots to perform autonomous wound repair. However, existing methods often neglect spatial semantics and wound-shape information, leading to poor generalization and low success rates. This paper presents the <bold>S</bold>patial-, <bold>S</bold>emantic-, and <bold>S</bold>hape-<bold>a</bold>ware <bold>D</bold>iffusion <bold>P</bold>olicy towards autonomous <bold>Wo</bold>und Repair (S<inline-formula><tex-math notation="LaTeX">$^{3}$</tex-math></inline-formula>aDPWo), a framework integrating two visual perception modules tailored for wound repair: the Spatial Semantic Perception Module (SSPM) and the Wound Shape Perception Module (WSPM). These modules supply the action predictor with semantically enriched point clouds and keypoint-based wound geometric descriptors, enabling S<inline-formula><tex-math notation="LaTeX">$^{3}$</tex-math></inline-formula>aDPWo to jointly perceive spatial-semantic and wound-shape information. Experimental results demonstrate the effectiveness of the proposed algorithms in wound segmentation and keypoint prediction, and further validate the overall framework on wound approximation—a key contact-rich sub-task of wound repair essential for facilitating subsequent suturing and promoting healing. Notably, S<inline-formula><tex-math notation="LaTeX">$^{3}$</tex-math></inline-formula>aDPWo achieves success rates of 90% and 80% on seen and unseen wound instances, respectively, while maintaining mean errors below 3 mm across inter-edge distance, edge-height difference, and edge consistency. This substantially outperforms SOTA IL baselines in both generalization and performance.