Korean Robots Are Selling Bodies, but Their Brains and Wrists Still Depend on Foreign Tech
What makes robotics so confusing these days isn’t the flashy demos, but where the money actually sticks. More important than news that a collaborative robot sold well is who captured the value from the precision reducers, control software, vision recognition, and on-site SI inside it. The reality of Korea’s robotics industry is that it has a presence in platform assembly and applications, but the thick profit layers in core components and AI are still thin.
That’s why, when looking at listed Korean companies, you have to strip away the optical illusion. Doosan Robotics, Rainbow Robotics, and Neuromeka have built names in cobots; Robostar and T-Robotics are holding their ground in SCARA, delta robots, and process automation; and LG Electronics, Woowa Brothers, and Bear Robotics are expanding the field in service and delivery robots. But the really difficult question is this: do those achievements lead to technological leadership, or do they stop at good hardware packaging and sales execution?
The real money is made more in what comes before and after the robot than in the robot body itself
The robotics value chain is less straightforward than it looks. It starts with core components, layers on AI/software, bundles that into a robot platform, plugs it into application areas, and only at the end is real money actually collected through integration, SI, and operations. That last part is what the industry often misses. A robot is not a product that ends at shipment; revenue is only truly complete when it is deployed on-site and actually running.
Korean companies in particular are relatively strong in the robot platform and process automation/solutions segments, which is why integration and operational capabilities ultimately determine the quality of earnings. That is, in the end, exactly what T-Robotics and Raontech demonstrate. If you only sell the hardware, you get dragged into price competition. If you tie together the entire process, customers cannot switch easily. That difference is quite significant.
| Stage | Description | Sub-segments |
|---|---|---|
| Core components | Core robot components (actuators, motors, reducers, sensors, batteries) | Actuators, precision motors/BLDC, reducers/harmonic drives, sensors (vision, force, inertial), batteries/power supply, grippers/end effectors, etc. |
| AI/Software | Robot AI/SW (physical AI, VLA models, simulation, OS) | Foundation models/VLA, Robot OS/middleware, Vision AI/SLAM, simulation/digital twins, motion planning/control, teleoperation/remote control, etc. |
| Robot platform | Finished robot platforms (humanoid, industrial, service, mobility) | Humanoids, quadruped robots/quadrupedal locomotion, collaborative robots (cobots), industrial manipulators, AMR/AGV (autonomous mobility), service robots, etc. |
| Application areas | Robot application domains (manufacturing, logistics, medical, service, defense, space) | Manufacturing automation, logistics/delivery automation, medical/healthcare, service/hospitality, construction/job sites, agriculture/food, etc. |
| Integration/SI/Operations | Robot system integration and operational services (including RaaS) | Robot SI/integration, RaaS (Robotics-as-a-Service), operations/maintenance, data/fleet management, certification/safety, education/training |
Cobots have landed, but precision reducers and robot AI are still a long way off
The most dangerous thing when talking about localization is this vague, self-congratulatory idea that everything is going well. It isn’t. Korea has built a fairly distinct presence in collaborative robots, SCARA and delta robots, process automation and robot solutions, and service and delivery robots. That lineup includes Doosan Robotics, Rainbow Robotics, Neuromeka, Robostar, T-Robotics, LG Electronics, Woowa Brothers, and Bear Robotics.
The problem is that the areas with large technology gaps are far too obvious. Precision reducers, reinforcement learning-based control, manipulation AI, and VLA/robot foundation models are all gap level 4. AI vision and deep learning recognition are also gap level 3. In other words, Korea can build and sell the body, but the wrists and the brain are still weak.
The real bottlenecks in Korea’s robot industry are precision reducers and manipulation AI/VLA. If those two are weak, robots ultimately remain stuck as repetitive task machines. Even if cobots are strong, momentum fades when the industry tries to move into harder manipulation and greater autonomy. If reinforcement learning-based control is weak, robots are less flexible in changing environments. If grasping and manipulation AI is weak, applications that truly replace human hands remain limited. “Limited” sounds polite, but in reality it means the top end of the market is blocked.
On the flip side, the opportunity is also clear. Korea is strong at quickly bundling field-ready products such as cobots and process automation. The structure is not bad: Doosan Robotics, Rainbow Robotics, and Neuromeka push the platforms, while T-Robotics and Raontech connect them into solutions. But to be blunt, that means strength in the areas that sell today, not that Korea has secured the next hegemonic technology. You can feel it too, can’t you? The market’s attention is shifting away from cobots themselves and toward how well robots can see, grasp, and generalize their behavior.
Table below: Domestic vs. global TRL (technology readiness level 1–9) by technology, and key companies
| Technology | Domestic | Global | Assessment | Key companies | ||||||||||||||||||||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Drivetrain & core components | ||||||||||||||||||||||||||||||||
| Precision reducers (harmonic/RV) | 5 | 9 | Weak · Gap 4 | SBB Tech, SPG, Samick THK | ||||||||||||||||||||||||||||
| A reducer is the core “joint” component of a robot, converting the fast rotation of a motor into slower but much stronger motion. Like a car transmission, it reduces rotational speed while increasing torque. A harmonic reducer works by elastically deforming a thin cup-shaped gear into an oval so it meshes with the inner gear, allowing a high reduction ratio of around 100:1 in a single stage with virtually no backlash. An RV reducer uses eccentric rotation and is designed to withstand heavier loads, which is why it is used in large industrial robots. Because the teeth must mesh with microscopic precision and almost no vibration, machining and heat-treatment accuracy must be extraordinarily high. Two Japanese companies account for roughly 70% of the global market. In 2025, demand for reducers surged amid the race to mass-produce humanoids such as Tesla Optimus and Figure, pushing localization into the spotlight as a supply chain security issue. | ||||||||||||||||||||||||||||||||
| Servo motors & drives | 6 | 8 | Catching up · Gap 2 | Higen Motor, SPG | ||||||||||||||||||||||||||||
| A servo motor is a “smart motor” that uses sensor feedback to precisely match commanded position, speed, and force. A rotational angle sensor called an encoder continuously measures the actual position and corrects any error against the target in real time. Unlike a standard motor that simply spins, a servo motor can control stopping position down to 0.001 degrees, allowing robots to repeat highly precise motions. The drive is the control circuit that supplies current to the motor with great precision, adjusting current magnitude and direction thousands to tens of thousands of times per second to generate torque. In humanoids, quasi-direct drive (QDD), which uses motor force more directly without a reducer, is gaining attention, making motor power density increasingly important. Korea has achieved a degree of self-reliance in small and mid-sized industrial motors, but still depends heavily on imports in high-output, high-precision segments. | ||||||||||||||||||||||||||||||||
| Integrated actuators (joint modules) | 6 | 8 | Catching up · Gap 2 | Rainbow Robotics, Neuromeka | ||||||||||||||||||||||||||||
| These are “complete robot joints” that integrate the motor, reducer, sensors, and controller into one unit. Instead of buying parts separately and assembling them, a single module can form a joint. With built-in torque sensors and control boards, they can sense how hard they are pushing and smoothly adjust force on their own. That torque control is the basis for collaborative and humanoid motion that can stop safely even when colliding with people or objects. The key challenge is delivering high force and precision while keeping the module compact and lightweight, which makes motor power density and thermal design critical. Standardized modules are essential for quickly designing and replacing humanoids that require dozens of joints. In 2025, Rainbow Robotics, whose largest shareholder became Samsung Electronics, has been leading domestic technological self-reliance by developing humanoids such as the RB-Y1 with its own actuators. | ||||||||||||||||||||||||||||||||
| Precision bearings & mechanical structures | 6 | 8 | Catching up · Gap 2 | SBB Tech, Samick THK | ||||||||||||||||||||||||||||
| Bearings support rotating shafts with small balls or rollers to reduce friction and wobble, while the mechanical structure forms the robot’s frame and linkage system. For a robot to lift heavy objects without shaking and stop accurately, bearing precision (clearance error) and stiffness (resistance to bending) are extremely important. In particular, reducers contain thin, large-diameter cross-roller bearings that must withstand forces from all directions in a single component, making them difficult to manufacture. Material purity and heat-treatment technology that increases surface hardness determine lifespan and precision. As high-load cobots and logistics robots proliferate, demand for domestically made precision bearings is rising as well. | ||||||||||||||||||||||||||||||||
| Sensors & perception | ||||||||||||||||||||||||||||||||
| Machine vision (3D vision) | 6 | 8 | Catching up · Gap 2 | Yujin Robot, CMES | ||||||||||||||||||||||||||||
| This is the robot’s “eye,” analyzing camera images to identify an object’s position, shape, and type. 3D vision also measures distance (depth), using methods such as stereo vision, which calculates distance from the angle difference between two cameras, or structured light, which projects striped light and reads how it bends on a surface to reconstruct a 3D shape. Those 3D coordinates are what make bin picking possible, allowing a robot to select and accurately grasp a single item from a jumbled pile of parts. More recently, combining vision with deep learning has significantly improved recognition accuracy even for difficult targets such as shiny or transparent objects. Korea has commercialized 3D vision for logistics and manufacturing, but foreign players such as Cognex and Keyence remain strong in high-precision inspection. | ||||||||||||||||||||||||||||||||
| LiDAR & SLAM | 6 | 8 | Catching up · Gap 2 | Yujin Robot, SOS Lab | ||||||||||||||||||||||||||||
| LiDAR measures the distance to objects by emitting laser pulses and timing their reflections, then combines tens of thousands of points into a 3D point-cloud map of the surroundings. SLAM (simultaneous localization and mapping) is the technology that lets a robot build a map by stitching together scenes seen through LiDAR and cameras while moving, and at the same time estimate its own position on that map. It is a bit like a person closing their eyes, feeling their way around a room, memorizing its layout, and still knowing where they are. The key is correcting accumulated sensor error through loop closing, which recognizes places the robot has seen before. This allows robots to find their way in unfamiliar spaces without pre-installed rails or maps, making it the brain of autonomous mobile robots (AMRs). With the AMR market booming in 2025, demand is rising for Korean LiDAR and SLAM companies. | ||||||||||||||||||||||||||||||||
| Force & torque sensors | 6 | 8 | Catching up · Gap 2 | Aidin Robotics, Robotoous | ||||||||||||||||||||||||||||
| These tactile sensors are mounted on a robot’s wrist or joints to detect “how hard” and “in which direction” contact occurs across six axes: up/down, left/right, front/back forces, plus rotational force around three axes. The principle is to attach strain gauges, which change electrical resistance when metal stretches or compresses slightly under force, and then infer the magnitude and direction of contact force from that electrical change. Just as a person gently holds an egg without crushing it, a robot must sense and adjust force in real time for parts assembly, polishing, and safe collaboration with humans. These sensors also go into humanoid fingertips and soles, making them critical for balance and delicate grasping. The market had long depended on foreign suppliers such as ATI and Schunk, but companies like Aidin Robotics are rapidly advancing localization with technologies such as capacitive sensing. | ||||||||||||||||||||||||||||||||
| AI vision & deep learning recognition | 6 | 9 | Caution · Gap 3 | CMES, Yujin Robot | ||||||||||||||||||||||||||||
| This is the technology in which AI neural networks trained on vast numbers of images determine what an object is and where and how it is placed in a camera image. Unlike older rule-based approaches that required manually defining every condition, these systems learn features automatically from data, allowing them to recognize irregularly shaped items or even previously unseen objects with human-like flexibility. More recently, synthetic data generated in virtual simulation and large-scale pretrained models have been used to rapidly improve recognition accuracy even in industrial settings with limited real-world data. Nvidia’s simulation and GPU ecosystem has effectively become the standard training infrastructure. Korea is competitive in application solutions, but dependence on global players for foundation models and semiconductors remains a weakness. | ||||||||||||||||||||||||||||||||
| Control & AI | ||||||||||||||||||||||||||||||||
| Robot control software | 6 | 9 | Caution · Gap 3 | Neuromeka, Rainbow Robotics | ||||||||||||||||||||||||||||
| This is the robot’s “operating system” and brain, calculating and commanding how much each joint should move and when. To send the end of an arm to a desired position, the system must solve in reverse how many degrees each joint should rotate—inverse kinematics—and do so hundreds to thousands of times per second in real time to create smooth, accurate trajectories. At the same time, it must handle path planning to avoid collisions, speed control to reduce vibration, and safety stops. Global research and industry are converging around ROS (Robot Operating System), the open standard, while Fanuc and KUKA maintain share with their proprietary closed controllers. Korean companies such as Neuromeka and Rainbow Robotics are pursuing strategies that reduce dependence on foreign controllers while strengthening ROS compatibility. | ||||||||||||||||||||||||||||||||
| Reinforcement learning-based control | 5 | 9 | Weak · Gap 4 | Rainbow Robotics, Naver Labs | ||||||||||||||||||||||||||||
| This is an AI training method in which a robot learns optimal behavior on its own through endless trial and error, receiving rewards for doing well and penalties for doing poorly. Without being taught every step by humans, the robot can fall and get back up millions or even hundreds of millions of times in virtual simulation, automatically mastering complex actions such as walking and balancing. The key challenge is narrowing the sim-to-real gap—transferring behaviors learned in simulation to real robots with different friction and weight characteristics. Nvidia’s large-scale parallel simulators have accelerated this training by orders of magnitude and become the standard infrastructure. Tesla Optimus and Figure also base their locomotion and manipulation on this approach, while Korea remains in catch-up mode in terms of simulation and data scale. | ||||||||||||||||||||||||||||||||
| Manipulation AI (grasping & handling) | 5 | 9 | Weak · Gap 4 | CMES, Naver Labs | ||||||||||||||||||||||||||||
| This is the technology by which AI decides and executes how a robot arm should grasp and move objects of different shapes and materials. From rigid boxes to floppy bags to slippery bottles, the variety is effectively endless, and outcomes change depending on how the object is grasped. That is why “what to grasp, where to grasp it, and how” is considered one of the hardest problems in robotics. Recently, the field has advanced toward learning both visual scenes and tactile input from force sensors together, allowing robots to infer grasp points and force even for previously unseen objects. The dominant trend is combining imitation learning, in which robots learn from human demonstration videos, with reinforcement learning. The U.S., led by players such as Google and Physical Intelligence, is ahead in data and models, while Korea is catching up in logistics and industrial applications. | ||||||||||||||||||||||||||||||||
| VLA & robot foundation models | 4 | 8 | Weak · Gap 4 | Naver Labs, LG Electronics | ||||||||||||||||||||||||||||
| VLA (Vision-Language-Action) robot foundation models are large AI models trained jointly on language, vision, and action. Just as ChatGPT understands text and produces answers, a VLA model takes camera images and human instructions as input and directly outputs how the robot’s joints should move. The key is that vision, language, and action are not handled separately but connected within a single model, so a simple instruction like “put that cup in the sink” can flow seamlessly from seeing to understanding to acting. The goal is generality: one model that can generalize across multiple robots and new tasks without needing to be reprogrammed for each job. In 2025, the U.S. is running away with the field through data and models such as Google’s RT family and Physical Intelligence’s π0, while Korea remains in the early catch-up stage centered on Naver Labs and LG. | ||||||||||||||||||||||||||||||||
| Industrial & collaborative robots | ||||||||||||||||||||||||||||||||
| Collaborative robots (cobots) | 8 | 8 | Solid · Gap 0 | Doosan Robotics, Rainbow Robotics, Neuromeka | ||||||||||||||||||||||||||||
| These robots are designed to work alongside people without safety fences. By detecting collisions within 0.1 seconds through force sensors or current changes in each joint and stopping immediately, they can avoid injuring people even on contact. Unlike conventional industrial robots, which are heavy, dangerous, and require fencing, cobots are small and light enough to sit on a desk beside a worker. They can also be programmed easily by non-specialists through direct teaching, where the user physically guides the arm by hand. That has allowed them to spread quickly into small-batch, high-mix manufacturing and even services such as cafés and fried chicken preparation. Korea is a true powerhouse here, with Doosan Robotics, Rainbow Robotics, and Neuromeka all ranking among the global leaders. | ||||||||||||||||||||||||||||||||
| Industrial articulated robots | 6 | 9 | Caution · Gap 3 | Hyundai Robotics, Robostar | ||||||||||||||||||||||||||||
| These are large six-axis robots used for welding, painting, and transport in places such as automotive plants. Their six joints move freely like a human shoulder, elbow, and wrist, allowing the end effector to reach any posture. That makes them central to mass production, where heavy parts must be handled quickly and accurately over and over again. Because the technical difficulty of reducers and controllers that determine precision is high, the global big four—Fanuc, ABB, Yaskawa, and KUKA—control more than half the market. In Korea, Hyundai Robotics is in catch-up mode, supported by demand from its automotive group. More recently, these robots have been evolving beyond fixed motions by combining with AI vision and reinforcement learning to adapt to changing situations. | ||||||||||||||||||||||||||||||||
| SCARA & delta robots | 7 |
The FANUC-and-ABB era is over; now Tesla and Figure AI are shaking up the boardIf you track the leading companies by era, you can clearly see where the center of gravity in the industry has shifted. From the 1990s to 2008, the global leaders were FANUC and ABB, while in Korea it was Hyundai Heavy Industries. From 2009 to 2016, Universal Robots emerged and cemented collaborative robots as a category of their own, while Rainbow Robotics was taking shape in Korea. From 2017 to 2019, Doosan Robotics and Rainbow Robotics became the standard-bearers of Korea’s cobot industry. Then, moving through 2020–2024 into 2025 and today, the mood changes completely. Tesla and Figure AI rise to the top globally, while FANUC and ABB still hold their ground, and Unitree has built a real presence as well. In Korea, the key names are Rainbow Robotics, Doosan Robotics, and Hyundai Boston Dynamics.
Power has shifted from the precision repetition of industrial robots to general-purpose intelligent robots, and Korea is still in the leading pack of followers, not in the lead. The FANUC and ABB era was defined by the depth of factory automation. The rise of Universal Robots was a victory of usability. But now that Tesla and Figure AI are disrupting the top tier, the story is different. It is no longer enough to simply build great hardware; large-scale data, robot AI, and software integration capabilities have moved to the center of corporate value. Korea’s current position is quite interesting. Rainbow Robotics has gained symbolic weight through its connection with Samsung, and Doosan Robotics has built a clearly recognizable brand in cobots. Hyundai Boston Dynamics also carries major name value. But to say this lineup is already equivalent to the Tesla- or Figure AI-style narrative would be overstating it. Korea is competitive in manufacturing-site robotics today, but in the next round—led by robot foundation models and manipulation AI—it still has a lot left to prove. Frankly, a lot. The next battleground isn’t how many robots get sold, but who controls learning and operationsThe key thing to watch from here is not simple shipment volume. We need to see how far platform players like Rainbow Robotics, Doosan Robotics, and Neuromeka can actually move up the stack into AI vision and deep-learning recognition, reinforcement learning-based control, and manipulation. If process automation and solution specialists like T-Robotics and Raontech can connect the layers above that, the picture becomes quite compelling. The upside for Korea’s robot industry depends less on cobot strengths themselves and more on how much intelligent software and operational data can be built on top. The risks are clear. A gap of 4 in precision reducers, a gap of 4 in VLA and robot foundation models, and a gap of 4 in manipulation AI are not just uncomfortable numbers. If those gaps are not closed, Korean companies may continue to remain “well-made robot equipment companies,” but they will not stand at the center of the industry’s value re-rating. The same goes for service and delivery robots. It is encouraging that LG Electronics, Woowa Brothers, and Bear Robotics are building real-world operating experience. But if operational data is not converted into training assets, they will remain stuck in a repeat-deployment business. In the end, there is only one question: will they be companies that sell robots, or companies that own systems through which robots learn? ※ This article is an analysis compiled from public data and industry materials, and some figures and assessments are estimates. It is not intended as a direct basis for investment decisions. (Written: 2026. Ealexandro) | ||||||||||||||||||||||||||||||
댓글
댓글 쓰기