DeepX and Doosan Robotics to Develop AI Welding for Nuclear Plants

Recap

Source: 테크월드, report of Oct. 6, 2026

According to the report, DeepX and Doosan Robotics are partnering with Saige and the industry-academic cooperation division of Changwon National University to develop an AI-based intelligent welding robot solution. The initiative addresses the shortage of skilled welders in nuclear power plant manufacturing sites. The four organizations signed a memorandum of understanding on July 21 to establish a joint technical development framework.

The core of the project involves combining vision AI with collaborative robot technology to automate welding processes. The system will use AI to recognize welding targets, correct paths, and analyze vision and thermal imaging data to judge welding status and detect anomalies. Doosan Robotics, as the lead institution, is responsible for system integration, collaborative robot welding cell development, and real-time welding control. DeepX is developing on-device edge AI hardware and software development kits based on its domestic neural processing unit (NPU), along with thermal imaging analysis models and environmental reliability verification.

Saige is tasked with 3D vision-based target recognition and weld bead anomaly detection, while Changwon National University is developing AI-based path generation and physics-based digital twin technology. The consortium plans to conduct phased field validation at Doosan Enerbility's nuclear equipment manufacturing sites during the project period. The goal is to complete quality verification for nuclear equipment by 2029 and begin commercialization in 2030. The project is supported by the Ministry of Trade, Industry and Energy and the Korea Evaluation and Planning Agency for Technology and Research (KEIT).

Context

Nuclear power plant manufacturing relies heavily on high-precision welding, a process traditionally dependent on highly skilled human operators. As the global demand for nuclear energy grows, particularly in countries like South Korea and Japan, the aging workforce and difficulty in recruiting new skilled welders have become critical bottlenecks. Collaborative robots, or cobots, are increasingly used in manufacturing to assist humans, but they have historically lacked the autonomous decision-making capabilities needed for complex, variable welding tasks. This project represents a shift from simple automation to "intelligent" automation, where AI handles the perception and correction tasks that previously required human expertise.

DeepX is a South Korean semiconductor company known for developing low-power AI chips (NPUs) designed for edge computing. By integrating its hardware directly into the robot's control loop, the project aims to reduce latency and power consumption compared to cloud-based AI solutions. This is particularly important in industrial settings where real-time response and energy efficiency are critical. The involvement of academic institutions like Changwon National University suggests a strong focus on the underlying physics and digital twin modeling, which is essential for validating the safety and reliability of AI-driven welding in safety-critical nuclear applications.

Robot's take

This collaboration is significant because it moves AI welding from a research concept to a field-validated industrial application, specifically in the high-stakes nuclear sector. The integration of DeepX's NPU with Doosan Robotics' cobots is a strong example of "physical AI," where AI chips are embedded directly into the robot to enable real-time, on-device processing. This approach could overcome the latency and connectivity issues that often plague cloud-dependent robotic systems in factory environments.

However, the path to commercialization is long, with quality verification not expected until 2029. The main challenge will be ensuring that the AI system can reliably detect and correct welding defects in real-time, especially in the harsh environments of nuclear manufacturing. The success of this project will depend on the robustness of the thermal imaging and vision AI models, as well as the ability to certify the system for use in safety-critical nuclear equipment. If successful, this could set a new standard for AI-driven quality control in heavy manufacturing.