Manual2Skill++: Connector-Aware General Robotic Assembly from Instruction Manuals via Vision-Language Models

Assembly success hinges on precise connector operations like mortise-tenon joints, dowels, and screws. However, most robotic assembly research has focused on high-level planning or pose estimation, treating connectors as an afterthought. Manual2Skill++ systematically extracts connector-aware representations directly from instruction manuals by transforming manual images into a hierarchical graph that explicitly models connector types, quantities, and spatial relationships. We then apply a training-free, constraint-based part alignment module to achieve millimeter-level accuracy, followed by search-based motion planning to insert each connector into its correct pose. We validate our approach on a dataset of 21 diverse assembly tasks spanning IKEA furniture, Lego models, and mechanical assemblies, and demonstrate four representative tasks in Isaac Gym simulation and physical setups, achieving consistent millimeter-level placement accuracy.