Repair2Skill: A Vision–Language–Action Framework for Robotic Furniture Repair

Abstract

Repair tasks differ fundamentally from robotic assembly, requiring fault localization, reverse dependency reasoning, and safe interaction with structurally degraded objects. While recent vision-language models have shown promise for high-level task reasoning, applying them to repair scenarios demands careful integration with perception, dependency modeling, and execution constraints. This paper presents Repair2Skill, a vision–language–action framework for robotic furniture repair that focuses on repair planning and simulation-based execution validation. The system detects damaged furniture components using a lightweight SSDLite-MobileNetV3 detector trained on procedurally generated synthetic damage data, constructs repair-oriented dependency graphs, and generates structured repair plans using a constrained large language model. Repair plans are restricted to a fixed action vocabulary and strict JSON schema to ensure deterministic parsing and feasibility. Generated plans are validated and visualized in a PyBullet simulation environment using kinematic action primitives, enabling interpretable verification of action ordering and dependency resolution without contact-rich physical manipulation. Experimental results demonstrate reliable part-damage association, achieving 76.8% mAP on real images, 100% valid repair plan generation under strict constraints, and stable simulation based validation across diverse repair scenarios. These results suggest that lightweight, constrained vision-language pipelines can support autonomous repair planning in bounded domains and provide a foundation for future work on physically grounded robotic repair.