Doosan Robotics Selected to Lead Two South Korean Physical AI Projects with Nearly $74M R&D Budget

The South Korean government selected Doosan Robotics to lead two physical AI projects, with a combined R&D budget of approximately 98.9 billion won (about $74 million), of which about 68.1 billion won (about $51 million) comes from government funding. The two projects are jointly managed by South Korea's Ministry of Trade, Industry and Energy and the Korea Institute for Advancement of Technology, with Doosan Robotics serving as the lead institution for both projects.

In terms of funding structure, government funding accounts for about 69% of the combined budget, and the specific contribution structure of the remaining portion is not disclosed in the report. The two project durations are 54 months and 45 months respectively, both medium- to long-term plans, with unclear short-term deliverables.

It should be noted that the above information all comes from a single reporting source, The AI Insider, with no first-party announcement links from the South Korean government or Doosan Robotics provided. This means the fact that the projects were approved is itself relatively credible, but regarding the specific technical commitments and commercialization timelines around the projects, currently there is only evidence at the level of company self-statements.

The first project is titled "Development of Next-Generation Collaborative Robot Technology Based on On-Device AI Semiconductors," with the goal of enabling robots to complete perception, decision-making, and control without relying on external computers or cloud services. According to Doosan, the system will use South Korean domestically developed AI system-on-chips installed directly into collaborative robots; the chip integrates a neural processing unit for AI computation as well as real-time communication and robot control functions, allowing the robot to process camera, tactile, force/torque sensor, and voice inputs locally.

At the product level, Doosan stated it plans to develop three collaborative robots with payloads of 5 kg, 10 kg, and 20 kg respectively, and plans to launch an industrial humanoid robot with dual-arm operation and omnidirectional mobility capabilities. These systems will be demonstrated continuously for 72 hours at at least two customer sites, and Doosan plans to apply for international safety certification and aims to begin commercializing them as standard products in 2031.

The first project runs for 54 months, with participants including Aidin Robotics, Mobilint, the Korea Electronics Technology Institute, the Korea Institute of Industrial Technology, and SAIGE, as well as industry-academia cooperation institutions from Korea University, Dongguk University, Sejong University, and Yonsei University.

The key uncertainty here is: the on-device AI system-on-chip is domestically developed in South Korea, and its actual performance and mass-production feasibility are not given data in the report; the industrial humanoid robot's dual-arm operation and omnidirectional mobility capabilities are currently at the planning stage, with no prototype or demonstration evidence; and there are no intermediate milestones verifying the 2031 commercialization target before it.

The second project is titled "Development of Intelligent Welding Solution Based on Collaborative Robots and AI," and will use collaborative robots, AI, and digital twin technology to automate more difficult welding work. According to Doosan, the vision system will enable the robot to identify welding targets and automatically adjust position, and the system will also use imitation learning to learn from skilled welders and generate welding paths and weaving motions.

Doosan has set a set of quantifiable targets for this project: robot-generated welding paths with over 95% similarity to skilled worker paths, path error not exceeding plus or minus 3 millimeters, and plans to automate multi-layer multi-pass welding of five or more layers and reduce processing time by over 50%. These figures constitute the most specific commitments of the entire project, but they are precisely the part lacking third-party verification.

Testing will be conducted at the Doosan Enerbility nuclear power center, with initial work covering spot welding, preheating, and welding in nuclear equipment manufacturing; after obtaining relevant nuclear quality certification, Doosan plans to expand the welding technology to shipbuilding, plant engineering, and defense applications. The welding project runs for 45 months, with DeepX, SAIGE, and the Changwon National University industry-academia cooperation group participating as research institutions, and Doosan Enerbility participating as the end user.

Certification is the most critical constraint here: the time required to obtain nuclear quality certification and the probability of passing are not specified in the report, and expansion to shipbuilding, plant engineering, and defense is entirely premised on that certification. In other words, the imaginative space for the welding project beyond the nuclear power scenario currently depends on an approval step with no timeline.

Doosan stated that it ultimately plans to integrate the technologies of the two projects into a single physical AI system for collaborative robots, combining AI chips, software, and robot hardware. This integration vision is central to understanding the relationship between the two projects: the on-device AI project provides the chip and local perception and decision-making capabilities, while the welding project provides a specific, highly difficult industrial scenario and data.

However, from an evidence standpoint, the feasibility of ultimately integrating the two projects' technologies into a unified physical AI system has not received any independent assessment. The two projects have different durations (54 months and 45 months), different participants, and different testing scenarios, and the integration timeline and interface standards are not mentioned in the report.

A more realistic issue is that whether the 95% welding path similarity and plus or minus 3 millimeter error targets are achievable lacks support from completed test results; key parameters such as the computing power, power consumption, and reliability of the on-device AI chip are also not disclosed in the report. With these basic data missing, the integration vision can currently only be regarded as a corporate-level planning statement.

Looking at the above information in layers, the confirmed facts include: Doosan was selected to lead two projects, the combined budget is approximately 98.9 billion won, government funding is approximately 68.1 billion won, the two projects are jointly managed by South Korea's Ministry of Trade, Industry and Energy and the Korea Institute for Advancement of Technology, the first project is 54 months and its participant list, the second project is 45 months and its participant list, and the official names of the two projects.

Content belonging to Doosan's self-statements (source_claim) includes: the specific architecture and local processing capabilities of the on-device AI system-on-chip, the payload specifications of the three collaborative robots and the industrial humanoid robot plan, the 72-hour customer site demonstration and 2031 commercialization target, the vision recognition and imitation learning mechanisms of the welding project, quantified targets such as 95% similarity and plus or minus 3 millimeter error, the nuclear power center testing and plans to expand to shipbuilding/plant engineering/defense, and the vision of ultimately integrating the two projects' technologies.

What belongs to inference is: the above information all comes from a single reporting source, The AI Insider, with no first-party announcement links from the South Korean government or Doosan Robotics provided. This inference does not negate the fact that the projects were approved, but it alerts readers: all performance metrics and commercialization timelines currently have only the layer of evidence of company self-set targets, and the report provides no third-party verification or completed test data.