[TECH Weekly] MindAI Bets on the 'Robot Brain' ... Shifts Its Core to Physical AI

MindAI's predecessor was Minds Lab. Minds Lab grew out of a technology investment spin-off research lab company from the Electronics and Telecommunications Research Institute (ETRI) in 2014, and in its early days it ran big data businesses using Korean natural language processing and unstructured text analysis. The company is now led by CEO Yoo Tae-jun, and after its KOSDAQ listing in 2021, it expanded its business scope across AI software.

The company has AI engines such as STT (Speech-to-Text), TTS (Text-to-Speech), and vision, and has used them to supply enterprise AI platforms, AI call centers, chatbots, generative AI, and AI agents. In 2023, it also changed its name from Minds Lab to MindAI.

Building on that foundation, MindAI has recently shifted its business focus from generative AI and AI agents to physical AI. It is combining voice, vision, and on-device AI with robots and industrial equipment, while expanding its applications into defense, robotics, logistics, public safety, and regional manufacturing AX.

MindAI is pushing to strengthen the link between existing AI technologies and autonomous intelligence for the physical world, and is targeting the defense and robotics sectors as it shifts from AI software to the robot brain, the focus of this article. The company has the language foundation model MAAL, the real-time voice conversation model SUDA, the vision-language model BODA, and the vision-language-action (VLA) model WoRV, and is placing emphasis on enabling perception, judgment, and action in robots and industrial equipment through MAAL, SUDA, BODA, and WoRV.

That trend extends to its main products, the JINDO BOT four-legged robot and the MAIED on-device AI integration platform. JINDO BOT is being expanded for use in patrol, safety management, education, and research, while MAIED supports the operation of AI models in robots and devices.

MindAI is expanding its business areas by using JINDO BOT and MAIED. The expansion is moving into robot platforms, on-device AI, and data training systems.

MindAI is focusing on a physical AI business that links its own AI models to judgment systems in robots and industrial equipment. The goal goes beyond repetitive automation of fixed movements and aims to realize autonomous intelligence that recognizes surrounding conditions and chooses the next action based on the purpose of the task.

Jindo Bot is the platform that implements this vision in an actual robot. Jindo Bot is designed so that cameras and sensors can understand the surrounding environment, and it is configured to link AI judgments to movement and task execution. MindAI is pushing to combine on-device AI, robot management systems (RMS), precision motion control, and actuator technologies with Samjeong Automation, and is upgrading Jindo Bot for education and research use.

On this basis, MindAI is widening its business expansion into defense and industrial robotics. Representative examples include the unmanned K10 ammunition carrier developed with Hanwha Aerospace, unmanned loaders with Doosan Bobcat, and the development of a patrol robot in Thailand. The K10 project is being pushed toward unmanned driving and ammunition resupply, while the Doosan Bobcat project is being advanced toward autonomous loader movement and operation.

In public safety, MindAI is participating in the Cheonan city safety network demonstration and expansion project, taking on a task worth about KRW 1.6 billion. The project involves applying the HAECHI-02 public safety model in the JINDO BOT series, and the target areas are safety-vulnerable zones such as riversides, parks, walking trails, and old downtown areas. HAECHI-02 applies on-device AI and autonomous intelligence based on a vision-language-action model, with the goal of detecting risks and carrying out necessary actions.

MindAI is co-developing a patrol service for autonomous mobile robots with Toolbotics, and SUDA is being embedded in it. The patrol service is built on a combination of STT, TTS, vision AI, and a vision-language model, and operates while also holding conversations with people. Through this, it identifies fires, intrusions, and facility anomalies, and the target use cases are factories, industrial complexes, public buildings, smart cities, and military border support.

In this context, MindAI sees the internal AI brain as the target market rather than complete robot products. This AI brain handles perception, reasoning, and action, and is a common application target across different robots and industrial equipment. If applied across common platforms, it is expected to expand the customer base into defense, manufacturing, and logistics.

The target for physical AI-based technology development is the World Model. A World Model is a technology that learns the structure of a robot's real-world space, learns physical relationships, and predicts changes that occur after an action.

MindAI, LG Electronics, and the Ministry of Science and ICT are jointly carrying out the 'Physical AI Leading Technology Development Project.' The project period is the next 2 years.

While generative AI focuses on producing language and images, the World Model focuses on predicting the changes that robot actions will cause in physical space.

Over the next 2 years, a total of KRW 34 billion in government budget will be invested to build autonomous intelligence and data training systems based on a World Model for use in industrial settings. The goal of this project is to improve robot operation success rates by more than 20 percentage points.

It is carrying out a second-stage AI data factory construction project with Kongju National University. The project is aimed at collecting and processing data accumulated in real environments for robot and AI systems, and the target data are vision, voice, and behavior. The collected and processed data will be used to retrain models and improve performance. It will also verify learning models in virtual environments, and the data secured during on-site verification will be fed back into retraining.

In logistics, it is participating in the Ministry of Science and ICT and the National Information Society Agency (NIA)'s '2026 AI Application Product Rapid Commercialization Support Project.' Through this, it will develop an unstructured cargo recognition, classification, and movement system. MindAI is responsible for the World Action Model and on-device AI technologies in this project. It also plans to combine domestic AI models, robots, and AI semiconductors (NPU) to carry out demonstrations and operations in actual logistics sites.

Regional manufacturing AX has been set as a new business touchpoint. MindAI is participating in South Chungcheong Province's 'Regional-Led AI Great Transformation Project.' The project is jointly promoted by South Chungcheong Province, Cheonan, and Asan, and supports AI transformation in the region's key manufacturing industries, including semiconductors, displays, and mobility. The project is worth a total of KRW 29.8 billion, and aims to drive AI transformation at 90 manufacturing companies and train 410 practical AX professionals.

MindAI is taking part in building an education system for current employees, job seekers, and career changers. The curriculum ranges from data utilization to AI agents and physical AI. It is running a physical AI data factory specialist training program with Kongju National University, and based on that program, it is planning to nurture personnel who can operate and apply AI in manufacturing sites.

CEO Yoo also serves as chairman of the Korea Physical AI Association. Members of the association include MindAI, Crowdworks, Hyundai Movex, and Morai, and the participating fields are AI, robotics, and infrastructure companies. The association is pursuing university-company joint research, technology verification (PoC), and talent development as key tasks, aiming to expand the physical AI ecosystem.

In line with this trend, the securities industry is paying attention to MindAI's business restructuring. Hana Securities published a report in May and analyzed that MindAI is shifting its business focus from large language models (LLM) to physical intelligence, including robot foundation models (RFM), starting in 2025.

As its existing agent AI revenue declines, MindAI is expanding its business by cultivating physical AI as a new pillar. Across its businesses, contraction in legacy operations and expansion in new ones are happening simultaneously. Agent AI revenue fell from KRW 7.5 billion in 2024 to KRW 6.8 billion last year, and reached KRW 1.3 billion in the first half of this year.

By contrast, physical AI is growing in scale as new revenue is being generated. Physical AI revenue rose from under KRW 100 million in 2024 to KRW 2.6 billion last year, and reached KRW 1.3 billion in the first half of this year. The company continues to invest in R&D and infrastructure for its physical AI business.

However, even though physical AI revenue has increased rapidly, it has not yet offset the decline in its existing businesses. As a result, second-quarter revenue this year fell 56.5% year over year to about KRW 100 million, while operating loss widened from KRW 1.6 billion to KRW 3 billion.

Ultimately, the point at which new businesses shift into full-scale production revenue is expected to determine the overall earnings trend.

Total orders rose quickly on a quarterly basis, from about KRW 100 million in the first quarter of this year to KRW 2.4 billion in the second quarter, and most of the increase came from physical AI projects. The order expansion was driven by Hanwha Aerospace, Doosan Bobcat, and the Thai patrol robot project.

However, major projects still have a high proportion of technology verification and demonstration. The Hanwha Aerospace, Doosan Bobcat, and Thai patrol robot projects have not yet entered mass production. Government projects in the World Model, city safety network, and logistics fields are also still in the technology development and on-site verification stages, and the challenge remains to connect them to actual product supply and long-term contracts.

There are variables in the cost structure of physical AI. Robots, sensors, and edge devices used differ by site, and additional data acquisition may require model training and verification. Accordingly, as the customer base expands, the number of development personnel may increase, and costs may also rise.

A structure in which core technologies are repeatedly applied to multiple customers is presented as a condition for improving profitability. MAIED, MAAL, SUDA, BODA, WoRV, and the World Model are technologies that can be applied in common to different robots and industrial equipment. If the same technology is repeatedly applied to multiple customers and devices, the likelihood of recovering one-time development costs across multiple projects expands, and room for recurring revenue also increases.

On the other hand, each customer may separately require data collection, model development, and system integration. In that case, even if the business expands, it becomes an SI-type structure in which costs rise in proportion to staffing. Accordingly, other conditions for improving profitability are said to include conversion to mass production after PoC and the ability to supply software and models repeatedly across multiple devices. The gist of the statement that repeated applicability of core technologies and the ability to supply across multiple devices are key conditions for improving profitability also aligns with this comparison standard.

MindAI started with big data and enterprise AI software, then moved through generative AI and agent AI. Its current business shift is toward physical AI centered on defense and robotics, and its technology scope has expanded to World Models, data infrastructure, on-device AI, and robot control. In this trend, the importance of commercialization scale will grow more than the number of demonstrations.

Projects are underway with Hanwha Aerospace and Doosan Bobcat, and if these projects lead to actual mass production and follow-on contracts, and if the same AI technology is repeatedly applied to other robots and industrial equipment, they could establish a new growth engine for physical AI. On the other hand, the structure centered on PoC and state-funded projects may continue for a prolonged period, and if that structure does continue, additional time will be needed for revenue growth and profitability improvement.

Source: TECHWORLD · Kim Seung-gi
Original: https://www.epnc.co.kr/news/articleView.html?idxno=406573

References

• https://www.epnc.co.kr/news/articleView.html?idxno=406573