NC AI Builds Autonomous Welding System for Hanwha Ocean

According to reporting by Seoul Economic Daily, Digital Today, Chosun, and Chosunbiz, NC AI secured a Hanwha Ocean project to develop vision-based welding-specific models and collaborative-robot autonomous welding models. The outlets describe a scope that includes a vision-recognition welding model and a collaborative-robot autonomous welding model intended for next-generation commercial and special-purpose vessels.

Digital Today and Chosun report that the effort aims to fuse AI vision recognition with precision robot control, moving beyond trajectory-following automation toward robots that recognize and judge weld areas and perform welding in real time. Digital Today identifies NC AI's industry-specialized vision-language model VAETKI Vision as the core engine, while Chosun's English coverage refers to the model as Baeki Vision and describes its expansion into a vision-language-action (VLA) system that converts visual and textual or voice instructions into robot control commands. An NC AI official, quoted by Chosun, characterized shipyard welding as a cost- and quality-critical task carried out under harsh conditions for vision systems, including intense arc light and sparks, real-time welding fumes, and camera-lens contamination in rough outdoor environments. Seoul Economic Daily and Digital Today add that the project will use real-time worksite data and Hanwha Ocean engineer feedback to develop shipbuilding-specialized recognition capable of extracting weld-line geometry and detecting defects amid noise and contamination.

combining a vision-language model with closed-loop robot control to form a VLA system aligns with research that integrates multimodal perception and action for physical autonomy. Companies attempting similar integrations commonly face three hurdles: robust perception under extreme lighting and particulates, low-latency mapping from perception to motion commands, and safe, certifiable collaboration between robots and human workers. Reliable weld-line extraction and defect detection in arc-heavy environments typically require domain-specific data augmentation, sensor fusion, and on-robot inference optimization to meet real-time constraints.

Observers should monitor field-trial results and metrics such as weld-defect rates, cycle-time gains, and system uptime in live production, along with the sensor and model choices used to harden perception against arc light and contamination, perception-to-actuator latency, and any safety or certification steps for deploying collaborative robots. Note the model-name discrepancy across outlets - Digital Today cites VAETKI Vision, Chosun cites Baeki Vision - which may reflect translation or branding differences; direct documentation from NC AI would clarify.