<?xml version="1.0" encoding="UTF-8"?><rss xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" version="2.0" xmlns:cc="http://cyber.law.harvard.edu/rss/creativeCommonsRssModule.html">
    <channel>
        <title><![CDATA[NC AI - Medium]]></title>
        <description><![CDATA[Humanity can be limitless with AI - Medium]]></description>
        <link>https://medium.com/nc-ai?source=rss----83564b09dfb3---4</link>
        <image>
            <url>https://cdn-images-1.medium.com/proxy/1*TGH72Nnw24QL3iV9IOm4VA.png</url>
            <title>NC AI - Medium</title>
            <link>https://medium.com/nc-ai?source=rss----83564b09dfb3---4</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Tue, 17 Nov 2026 06:01:10 GMT</lastBuildDate>
        <atom:link href="https://medium.com/feed/nc-ai" rel="self" type="application/rss+xml"/>
        <webMaster><![CDATA[yourfriends@medium.com]]></webMaster>
        <atom:link href="http://medium.superfeedr.com/" rel="hub"/>
        <item>
            <title><![CDATA[What Is a VLA Model? From AI That Sees to Robots That Act]]></title>
            <link>https://medium.com/nc-ai/what-is-a-vla-model-from-ai-that-sees-to-robots-that-act-d037c57d48be?source=rss----83564b09dfb3---4</link>
            <guid isPermaLink="false">https://medium.com/p/d037c57d48be</guid>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[vision-language-action]]></category>
            <category><![CDATA[robotics]]></category>
            <category><![CDATA[physical-ai]]></category>
            <dc:creator><![CDATA[NC AI]]></dc:creator>
            <pubDate>Wed, 11 Nov 2026 13:55:29 GMT</pubDate>
            <atom:updated>2026-11-11T13:55:29.713Z</atom:updated>
            <content:encoded><![CDATA[<p>What is a VLA model? See how vision-language-action models turn images and instructions into robot actions, from action chunks to real-time control.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*vh1IqT4MQdaRSoaGwbApfw.png" /></figure><p>“Pick up the cup.”</p><p>It’s a simple request for a person, but a robot has to go through several steps to carry it out. It must locate the cup, decide which direction to move its arm, and close its gripper (the device it uses to grasp objects) at the right position. In addition to recognizing the cup and understanding the instruction, it has to generate an action it can actually execute.</p><p>A model trained to turn visual information and language instructions into physical actions is called a Vision-Language-Action (VLA) model.</p><p>So how do the words “Pick up the cup” become a robot’s movements? How does the model represent actions, and how does it keep the robot moving as the situation around it changes?</p><p>We’ll start by looking at how a VLA differs from a model that understands scenes, then follow the cup example to see how a VLA works and how it represents actions.</p><p>We’ll then look at why real-time operation and visual understanding matter for real-world robots, and explore recent work on predicting actions and future scenes together.</p><p><strong>1. How Do VLMs and VLAs Differ?</strong></p><p>There’s a cup on the left side of a table. If you show a Vision-Language Model (VLM) a photo of the scene and ask, “Where’s the cup?”, it can answer, “It’s on the left side of the table.” A VLM can recognize objects in a photo, describe their positions relative to one another, and answer questions about the scene.</p><p>In the same scene, a VLA given the instruction “Pick up the cup” outputs action values specifying things like which direction to move the arm, how far to move it, and when to close the gripper. The robot’s controller uses those values to move the arm and gripper.</p><p>One common way to build a VLA is to take a VLM trained to understand scenes and language and train it further on robot action data. RT-2, released by Google DeepMind in 2023, was built on existing vision-language models and trained on recordings of robot movements alongside vision-language data. It represented robot actions as text tokens, the symbolic units a model uses to process text.</p><p>VLAs also overlap with other concepts in robot AI. A <a href="https://medium.com/nc-ai/what-is-a-robot-foundation-model-what-you-need-to-know-before-adopting-physical-ai-dd80b3854c87">robot foundation model (RFM)</a> serves as a common foundation across tasks and robots. When an RFM is implemented as a VLA, the same model fits both descriptions. Researchers are also exploring architectures that connect an <a href="https://medium.com/nc-ai/why-do-robots-that-move-on-their-own-need-an-agent-a8cf3551abf0">agent</a> that plans the order of multiple tasks with a VLA that generates the movements for each task</p><p>Back to the cup. What inputs does a VLA receive, and what steps does it follow to produce action values?</p><p><strong>2. How Does “Pick Up the Cup” Become a Movement?</strong></p><p>Even when the cup stays in the same place, the movement required depends on where the robot’s arm is. If the arm is already near the cup, it only needs to move a little. If it’s farther away, it needs to reach farther.</p><p>The VLA first takes in camera images and the language instruction “Pick up the cup.” Depending on the model, it may also receive information about the robot’s own state, such as the current positions of its joints or whether the gripper is open.</p><p>Using this information, the model predicts the action values the robot will execute. Some models specify how far to move and how much to rotate the end effector (in this example, the gripper at the tip of the arm). Others output target joint angles or joint velocities. The form of the action values depends on the data the model was trained on and how the robot is controlled.</p><p>Once the robot executes an action and its arm changes position, the model observes the updated situation and predicts the next action. The robot keeps moving by repeating the cycle of observation → action prediction → execution → re-observation. Some models predict several upcoming action values at once. The resulting sequence of action values is called an action chunk.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*pwHx6NN0No1N5ZWVxheqww.png" /></figure><p>So how does the model generate these action values, whether they represent how far the end effector should move or what angle a joint should reach?</p><p><strong>3. How Do Robots Represent Actions?</strong></p><p>Two common approaches to producing action values are predicting tokens and generating continuous values.</p><p>The first approach represents actions as text tokens, as in RT-2 from Section 1. For example, we can divide end-effector displacement into 1 cm intervals, with the interval containing 0.5 cm labeled A and the interval containing 1.5 cm labeled B. Once movements are represented by symbols like these, the model predicts tokens much as it would when generating a sentence. The resulting tokens are converted back into action values the robot can execute.</p><p>Physical Intelligence, a U.S. robotics AI company, reported that this simple approach of dividing values into intervals and mapping them to symbols fell short in its experiments on dexterous robot manipulation. To address those limitations, the company introduced FAST in January 2025. FAST compresses patterns in upcoming movements to represent action chunks as tokens. π₀-FAST is a VLA model that uses this tokenization method.</p><p>In the continuous-value approach, the model directly generates numbers for displacement, rotation, and joint angles. Physical Intelligence’s π₀ uses a method called flow matching. It starts with random values and refines them over several steps into action values that match what the robot observes and the instruction it receives.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*3kPHtt7saSs97exJN3cntQ.png" /></figure><p>It takes time for a model to compute an action chunk. How can the robot keep moving in the meantime?</p><p><strong>4. How Quickly Does a Robot Need to Decide Its Next Action?</strong></p><p>The next action must be ready by the time the robot’s controller needs to execute it. As the arm reaches for the cup, the controller executes action values at short intervals. If the arm moves while the model is computing the next motion, that motion must continue smoothly from the arm’s new position.</p><p>In the model setup described by Physical Intelligence, each predicted action chunk contains action values for 50 time steps, covering one second of motion. To keep the robot moving, the next chunk must be ready while the current one is still being executed.</p><p>Inference is the process by which a model takes an input and computes an output. With synchronous inference, the system executes the planned segment of motion before computing the next chunk, so the arm may pause until the computation finishes. Even if computation starts earlier, the arm may jerk or stall at the switch if the new chunk doesn’t match the motion already in progress.</p><p>Real-Time Chunking (RTC), proposed by Physical Intelligence, can be applied to VLAs that refine action values over multiple steps. This covers models that use flow matching, like π₀ from Section 3, and diffusion-based models, which work on a similar principle. The robot executes previously planned motions while the model computes the next chunk. The model accounts for the movements in the overlap between the two chunks and generates action values that let the new motion continue smoothly.</p><p>The researchers compared performance by artificially adding up to 200 ms (0.2 seconds) to the existing inference latency. To calculate average task throughput, they divided the fraction of the task completed in each attempt by the time taken, then averaged the results. With synchronous inference, average task throughput fell as latency increased on the tasks tested. In the same experiments, real-time chunking maintained throughput close to the level measured before extra latency was added.</p><p>Fast inference helps ensure the next action is ready in time. To execute it smoothly, the system must also let the robot continue its current motion while the model runs and transition to the next motion when needed.</p><p>Before deciding how to move, what does the robot need to know about the object it’s working with?</p><p><strong>5. To Move Well, Robots First Need to See Well</strong></p><p>The robot needs to know the object’s position, shape, and current state. If the object is moving, the robot also needs to track how it changes across video frames. If the task involves a workpiece to be welded in an industrial setting instead of a cup, the robot must first identify where on the workpiece to perform the operation.</p><p>NC AI’s VAETKI Vision 3B plays this role. Think of it as “the eyes of a VLA.” It is a lightweight vision-language model (VLM) designed with the limited hardware resources of robots and industrial edge devices (equipment that runs AI directly on-site) in mind. For video inputs, it is designed to analyze relationships between frames to identify subtle movements and object trajectories.</p><p>The evaluation of a welding-specialized model shows the effect of additional training tailored to the industrial setting. Compared with the base model, the welding-specialized model’s bounding box accuracy score for position and shape recognition (RefWeld) rose from 5.3 to 72.5. This score measures only the model’s ability to identify the position and shape of the workpiece in an image.</p><p>NC AI plans to further develop VAETKI Vision 3B and deploy it on robots and edge devices as a core vision engine and a VLA-based model.</p><p>What if a robot could also predict its next action and how the scene would change as a result?</p><p><strong>6. AI That Predicts Both the Action and the Scene That Follows</strong></p><p>One example is GR00T 2, a new model in NVIDIA’s GR00T family that the company previewed at GTC 2026 and described in research released in November of the same year. NVIDIA calls it a World Action Model (WAM). It takes in recent camera images, a task instruction, and the robot’s current state, then predicts future video and an action chunk together. The robot executes the predicted actions and repeats the same process using newly observed information.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/905/1*yebc0LpbYLWuTgK4b3p7hg.png" /><figcaption><em>Source — Conference on Robot Learning, “CoRL 2026 Day 1,” GR00T 2 keynote by NVIDIA’s Jim Fan.</em></figcaption></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*wFfPUdzYqusJIcjlvgIg_A.png" /></figure><p>In other words, it takes the loop from Section 2 and adds a prediction of what the scene will look like after the action.</p><p>To move well, a robot must see its surroundings, understand the instruction, generate actions it can execute, and carry them out in sequence at the right time. Each technology plays a different role, and those roles are becoming more closely connected as the robot moves.</p><p>LLMs process language. VLMs understand scenes. VLAs generate actions. World models predict the outcomes of actions.</p><p>Following this progression, NC AI is building AI that sees, decides, and moves in the real world.</p><p><strong>Further Reading</strong></p><p><a href="https://medium.com/@nc-ai/how-do-robots-learn-to-move-building-the-training-data-behind-physical-ai-e750155f3dc5?utm_source=chatgpt.com">How Do Robots Learn to Move? Building the Training Data Behind Physical AI</a></p><p><a href="https://medium.com/nc-ai/what-is-sim-to-real-27dea4b8e1c9?utm_source=chatgpt.com">What Is Sim-to-Real?</a></p><p><a href="https://medium.com/nc-ai/what-is-a-world-model-key-concepts-to-know-before-adopting-physical-ai-a243ada494ee?utm_source=chatgpt.com">What Is a World Model? Key Concepts to Know Before Adopting Physical AI</a></p><p><a href="https://medium.com/nc-ai/beyond-virtual-worlds-nc-ais-vision-for-the-future-of-physical-ai-a2e7d13d3758">Beyond Virtual Worlds: NC AI’s Vision for the Future of Physical AI</a></p><p><strong>References</strong></p><p>Google DeepMind, <a href="https://deepmind.google/blog/rt-2-new-model-translates-vision-and-language-into-action/">“RT-2: New model translates vision and language into action”</a> (July 2023)</p><p>Physical Intelligence, <a href="https://pi.website/research/fast">“FAST: Efficient Robot Action Tokenization”</a> (January 2025)</p><p>Physical Intelligence, <a href="https://arxiv.org/abs/2410.24164">“π₀: A Vision-Language-Action Flow Model for General Robot Control”</a> (October 2024)</p><p>Physical Intelligence, <a href="https://pi.website/research/real_time_chunking">“Real-Time Action Chunking with Large Models”</a> (June 2025)</p><p>Black et al., <a href="https://arxiv.org/abs/2506.07339">“Real-Time Execution of Action Chunking Flow Policies”</a> (June 2025)</p><p>NC AI, <a href="https://nc-ai.com/ko/newsroom/press/nc-ai-3b-sota">“NC AI Develops VAETKI Vision 3B, Achieving State-of-the-Art Performance Among Comparable Models Globally”</a> (November 2026)</p><p>NVIDIA, <a href="https://research.nvidia.com/labs/gear/gr00t-2/">“GR00T 2: Scaling Robot Foundation Models for Generalization, Dexterity, and Adaptation”</a> (November 2026)</p><p>Conference on Robot Learning, <a href="https://www.youtube.com/watch?v=F_4AAOQsWlA&amp;t=20513s">“CoRL 2026 Day 1”</a> (November 9, 2026, keynote by NVIDIA’s Jim Fan)</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=d037c57d48be" width="1" height="1" alt=""><hr><p><a href="https://medium.com/nc-ai/what-is-a-vla-model-from-ai-that-sees-to-robots-that-act-d037c57d48be">What Is a VLA Model? From AI That Sees to Robots That Act</a> was originally published in <a href="https://medium.com/nc-ai">NC AI</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[What Should an Industrial Robot Look Like?]]></title>
            <link>https://medium.com/nc-ai/what-should-an-industrial-robot-look-like-f0b8b9721dec?source=rss----83564b09dfb3---4</link>
            <guid isPermaLink="false">https://medium.com/p/f0b8b9721dec</guid>
            <category><![CDATA[industrial-automation]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[humanoid-robot]]></category>
            <category><![CDATA[robotics]]></category>
            <category><![CDATA[physical-ai]]></category>
            <dc:creator><![CDATA[NC AI]]></dc:creator>
            <pubDate>Mon, 09 Nov 2026 04:00:54 GMT</pubDate>
            <atom:updated>2026-11-09T04:00:53.924Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*e-w36M0h8jNfhSjvciY2Zg.png" /></figure><p>Humanoid robots are attracting growing attention across the robotics industry. Robots that walk on two legs, grasp objects with their hands, and use tools as humans do are no longer confined to science fiction. As global companies such as Tesla, Hyundai Motor Group, and Unitree accelerate humanoid development, robotic movements are becoming more refined, and efforts to deploy these robots in real industrial settings continue.</p><p>But a broader look at robotics reveals far more variety. On automotive production lines, large robotic arms assemble components. In logistics centers, wheeled robots transport goods. Robots that walk on four legs move through industrial facilities to inspect equipment.</p><p>Why do robots take so many different forms? And how will they evolve in the future?</p><h3>Why Humanoids Are in the Spotlight</h3><p>Most of the spaces we live and work in are designed around the human body. The height of stairs, the placement of door handles and work surfaces, and even the way tools are used all reflect how people move.</p><p>This is a key part of the appeal of humanoid robots. Robots with arms, legs, and hands similar to ours may be better able to use spaces and tools designed for people. A humanlike structure can offer important advantages, particularly in environments where tasks change frequently or require a combination of different movements.</p><p>Global companies are expanding the potential applications of humanoids through their own technologies.</p><p>Tesla is developing Optimus, a general-purpose humanoid designed to perform a variety of tasks. Boston Dynamics, part of Hyundai Motor Group, is working to bring its next-generation electric humanoid Atlas, unveiled at CES 2026, onto factory floors.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*gaK3BAHN1lNdxVViRYSpLQ.png" /><figcaption><em>Source: </em><a href="https://www.youtube.com/watch?v=tv90GFM9RAo"><em>Figure on YouTube</em></a></figcaption></figure><p>Efforts to put humanoids to work on actual production sites are also underway. At its Spartanburg plant in the United States, BMW demonstrated logistics tasks using Figure AI’s next-generation humanoid, Figure 03. The robot picked up automotive components, placed them on a transport cart in a specified sequence, and moved the cart.</p><p>In November 2026, Xiaomi also released a demonstration video showing a humanoid sorting automotive components at its electric vehicle factory. According to Xiaomi, the robot performed the sorting task for approximately eight hours.</p><p>Humanoid movements are becoming more dynamic as well. Chinese robotics company Unitree showcased its new H2-X humanoid in a demonstration video featuring a flip from the roof of a car and a kick that dents a car door. The robot also jumped from a structure approximately four meters high and rolled as it landed, showcasing its ability to perform complex, dynamic movements.</p><p>From precise task execution to highly dynamic motion, humanoids are evolving in many different directions.</p><h3>How Human Does a Humanoid Need to Be?</h3><p>Being a humanoid does not necessarily mean reproducing the human body exactly. Even robots within this category take different forms depending on the tasks they are designed to perform.</p><p>Digit, developed by U.S. company Agility Robotics, is one example. Although Digit walks on two legs, it does not directly replicate human leg anatomy. Its legs and object-handling mechanisms are designed for picking up and moving totes in logistics centers.</p><p>U.S. logistics company GXO uses Digit in a logistics center for repetitive material-handling tasks. In October 2026, GXO announced plans to expand its use to additional logistics facilities starting in 2027.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*IZf5OPcM2tmg2RgDMJgb2Q.png" /><figcaption><em>Source: </em><a href="https://www.youtube.com/watch?v=oyq9BOwK5XI"><em>Agility Robotics on YouTube</em></a></figcaption></figure><p>Some robots use wheels instead of two legs. GRW, developed by Chinese robotics company Fourier, has a humanlike upper body and two arms but moves on wheels. It combines two arms for picking up and moving objects with the advantages of wheeled locomotion.</p><p>Because it combines mobility with object manipulation, this type of robot can also be considered a form of mobile manipulator. A mobile manipulator combines a mobile platform with a robotic arm, allowing it to move between workspaces and manipulate objects.</p><p>An autonomous mobile robot (AMR) equipped with a robotic arm also falls into this category. A robot can retain the ability to handle objects in humanlike ways without having to move like a person.</p><p>Robots borrow from human anatomy, but they select and combine only the functions a given task requires.</p><h3>Designing for the Task, Down to the Robot’s Fingertips</h3><p>While humanlike robots are evolving into a variety of forms, industrial environments also rely on robots designed around the functions required for specific tasks.</p><p>The articulated robotic arms commonly seen on automotive production lines are specialized for precise assembly, welding, and repetitive component handling. Because they often work from a fixed position, they may have no need for legs or wheels.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*XbsxZ3d7OSe09CPkA1BQJQ.png" /></figure><p>The tools attached to the ends of these arms matter just as much. Humans use five fingers to grasp and manipulate many different objects, but robots do not always need a five-fingered hand. The most suitable structure depends on the object being handled and the task at hand.</p><p>A device attached to the end of a robotic arm to grasp objects or perform a particular task is called an end effector. Common examples include mechanical grippers that clamp onto objects, vacuum grippers that use suction, and magnetic grippers that lift ferromagnetic components.</p><p>End effectors are selected based on factors such as an object’s shape, weight, material, and surface condition. Robotic hands with multiple joints are useful for manipulating a variety of objects, but a dedicated gripper may be more effective when repeatedly handling a specific component.</p><p>A robot’s performance depends not only on the structure of its arms and joints, but also on how well its tools suit the task.</p><p>Other robots focus on transportation or inspection rather than object manipulation. In large logistics centers, AMRs transport goods. Their design centers on moving items to their destinations while perceiving their surroundings and avoiding obstacles.</p><p>In industrial facilities with stairs, changes in floor height, or complex obstacles, quadruped robots can be useful. Boston Dynamics’ Spot moves through industrial sites to read gauges and collect thermal imaging and acoustic data.</p><p>Two legs, four legs, wheels, robotic arms, and a wide range of end effectors: a robot’s structure and tools are selected and combined according to the role it needs to perform.</p><h3>Adapt Robots to the Workplace, or Adapt the Workplace to Robots?</h3><p>There is another question to consider when deciding what form a robot should take: Is it more effective to introduce robots into environments designed for people, or to create environments that make it easier for robots to work?</p><p>Existing factories and buildings contain many tools and facilities designed for human use. When automation needs to expand while keeping these environments largely as they are, humanoids that make use of a humanlike body structure can be a promising option.</p><p>The approach can be different when building a new production line or modifying a workplace. A logistics center’s traffic routes might be designed around AMRs, for example, or a robotic arm and workstation might be configured together for a particular assembly task. In these cases, the environment is designed to help robots work efficiently, rather than requiring robots to work the way people do.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*UXuXVJK9fk9scjnMrunzTQ.png" /></figure><p>Productivity and economics also matter when evaluating the benefits of introducing robots. Even when robots perform the same task, their practical value can differ depending on processing speed, accuracy, uptime, and maintenance costs.</p><p>Humanoids have the potential to perform a wide range of tasks flexibly. Robots specialized for particular tasks can offer advantages in the efficiency of repetitive work and the consistency of day-to-day operations.</p><p>Ultimately, industrial applications call for technologies that fit the task and operating conditions. The choice of robot form follows from those needs.</p><h3>Robot Evolution Goes Beyond the Body</h3><p>The variety of robot forms and tools we have explored shows how specialized and advanced robotic hardware has become. Sophisticated joints and hands, mobile platforms, and end effectors provide the physical foundation for interacting with the real world.</p><p>But capable hardware alone does not complete a robot’s abilities. To perform effectively in real working environments, robots also need the software that drives them.</p><p>Imagine a robotic arm picking up components from a conveyor and moving them elsewhere. If each component is always in the same position, the robot may be able to perform the task simply by repeating a predefined motion. But what happens when the position or orientation changes, or a new product enters the line?</p><p>Using information collected by sensors, the robot must identify the object’s position and shape, decide where to grasp it, and adjust its movements. It must also respond to unexpected obstacles and changes in its working environment.</p><p>Software links what a robot senses to how it moves. It enables the robot to perceive its surroundings through sensors, plan the actions needed for a task, and control motors and joints to turn those plans into physical behavior.</p><p>Hardware provides the physical basis for movement, while software determines and controls how those capabilities are used in actual work. Software is also advancing to help robots learn new tasks and respond more flexibly to changing environments.</p><p>In particular, research into robot foundation models (RFMs) is actively exploring how learning from data collected across diverse tasks and environments can improve robots’ general-purpose capabilities.</p><p>At NC AI, we are also developing technologies to bring this robotic intelligence into real industrial settings. Together with POSCO DX, we are researching general-purpose RFMs and vision-language-action (VLA) models that can support a variety of industrial tasks. With Hanwha Ocean, we are developing robot AI models specialized for welding that recognize weld areas and act according to task conditions in complex shipyard environments.</p><p>We are also researching technologies that use digital twins and world models to recreate industrial task scenarios in virtual environments, where robots can learn and validate the actions they need to perform.</p><p>Robot evolution is a process in which hardware and software advance together. Sophisticated body structures and tools make new movements possible, while advances in software allow robots to use those capabilities across a wider range of situations.</p><p>As robots expand into more areas of work, the diversity of their hardware and the importance of the software supporting it will continue to grow. Bipedal robots, wheeled robots, and robots specialized for particular tasks will develop alongside one another, each fulfilling the roles needed in its own environment.</p><p>The future of robotics has no single shape. What matters is the problem a robot can solve and how it solves it. We look forward to seeing what becomes possible as diverse robotic forms come together with increasingly capable intelligence.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=f0b8b9721dec" width="1" height="1" alt=""><hr><p><a href="https://medium.com/nc-ai/what-should-an-industrial-robot-look-like-f0b8b9721dec">What Should an Industrial Robot Look Like?</a> was originally published in <a href="https://medium.com/nc-ai">NC AI</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[What Is Sim-to-Real?]]></title>
            <link>https://medium.com/nc-ai/what-is-sim-to-real-27dea4b8e1c9?source=rss----83564b09dfb3---4</link>
            <guid isPermaLink="false">https://medium.com/p/27dea4b8e1c9</guid>
            <category><![CDATA[robotics]]></category>
            <category><![CDATA[physical-ai]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <dc:creator><![CDATA[NC AI]]></dc:creator>
            <pubDate>Fri, 30 Oct 2026 10:45:32 GMT</pubDate>
            <atom:updated>2026-10-30T10:45:32.771Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*wkkZpykEGd_-mZh36WebDw.png" /></figure><p>A robotic arm picks up a part in simulation and places it in a designated location. It repeats the task without making a mistake. But what happens when we apply the behavior it learned in simulation directly to a robot on a real factory floor?</p><p>Even when the scene looks similar, the results can be different. Reflected light may cause the robot to misjudge the part’s position. A surface that is more slippery than expected may cause the part to slip from its grasp. The robot may also respond to commands at a different speed than it did in simulation.</p><p>These differences between a simulated environment and the physical world are known as the <strong>sim-to-real gap</strong>.</p><p>Simulation provides an important foundation for robots to learn and test a wide range of tasks through repetition. What does it take to bring that experience into the real world? In this article, we follow a robotic arm as it picks up a part, exploring where these differences arise, how to reduce them, and how to validate the results. We then look at how NC AI and Hanwha Ocean have compared and aligned simulated processes with actual operations.</p><h3>Where Does the Sim-to-Real Gap Come From?</h3><p>To move a single part, a robotic arm must locate it, grasp it with a gripper, and carry it to the target position. This may look like one continuous action, but each step creates an opportunity for simulation and reality to diverge.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*tieOVR5cc2jCcakes_wSzA.png" /><figcaption>AI-generated image</figcaption></figure><p><strong>The first difference is perception.</strong> The robot uses camera images to determine where the part is. In simulation, its outline may be clear. On a real factory floor, reflections from metal surfaces or shadows cast by nearby equipment can obscure it. If the camera is installed at a slightly different position or angle than in simulation, the robot’s estimate of the part’s position will also be off.</p><p><strong>The second difference is contact.</strong> Knowing exactly where the part is does not guarantee a successful grasp. The force needed to hold it depends on its weight and surface friction. If the real surface is more slippery than the simulated one, a grip that was strong enough in simulation may allow the part to slip as soon as the robot lifts it.</p><p><strong>The third difference is motion.</strong> There is a small delay between the robot receiving a command and its motors responding. If that delay differs from the one modeled in simulation, the gripper reaches its target later than expected. When picking up a moving part on a conveyor, even a small timing difference can lead to failure.</p><p>These differences can also combine. A small error in estimating a part’s position, together with a surface that is more slippery than expected, may be enough to make the robot lose its grip — even if each difference seems minor on its own.</p><h3>System Identification: Match the Conditions That Matter Most to the Task</h3><p>If a part that was easy to grasp in simulation keeps slipping in the real world, what should we check first? Start by examining how the conditions specified in the simulation differ from the actual ones.</p><p>Suppose the simulated part was set to be lighter and less slippery than the real part. The robot would have learned to grasp it with less force than it actually needs. Measuring the real part’s weight and friction and incorporating those values into the simulation can help reduce this mismatch. We can also compare how quickly the simulated and physical robots respond to the same command and adjust the simulation accordingly.</p><p>The process of using real measurements to estimate properties such as weight, friction, and motor response, then calibrating the simulation based on those estimates, is called <strong>system identification</strong>.</p><p>There is no need to match every value precisely from the start. <strong>It is more efficient to measure the conditions that have the greatest effect on the outcome.</strong> If the part slips, examine friction first. If the robot reaches its target too late, examine its response time. After calibration, repeat the same action to check whether the mismatch has decreased.</p><p>A digital twin — a digital representation of a real factory’s space, equipment, and their states — can provide an environment resembling the actual site in which to test these adjustments.</p><p>Still, a value measured once will not remain accurate under every condition. Even parts of the same type can vary slightly in weight, and friction changes with surface conditions.</p><h3>Domain Randomization: Train Robots to Handle the Differences That Remain</h3><p>One widely used way to train for variations around measured conditions is <strong>domain randomization</strong>. During training, conditions in the simulated environment are repeatedly varied within a defined range. <strong>If real operating conditions fall within the range encountered during training, the robot can treat reality as another variation of what it has already experienced.</strong></p><p>For a robotic arm picking up parts, this means varying the brightness and direction of lighting, as well as the position and orientation of the parts. The robot gains experience locating and grasping them under bright lights, in shadows, and when they are placed at an angle.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Pcofso4YwlPg3lMPK3-6Jw.png" /><figcaption>AI-generated image</figcaption></figure><p>The same approach can be applied to the part’s weight, surface friction, and the robot’s response time. Varying the lighting prepares the robot for differences in perception. Varying weight and friction prepares it for differences in contact. Varying response time prepares it for differences in motion.</p><p>However, the range of variation needs to be chosen carefully. If it is too narrow, the robot will not experience enough of the variation found in real operations. If it is too wide, training becomes more difficult because the robot must also learn to handle situations that would never occur in practice.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*9AMQsA-dWbFycQrOjRTcQQ.png" /></figure><p>Measurements gathered through system identification provide a starting point for setting this range. Another option is to begin with a narrow range and expand it gradually. In its research on solving a Rubik’s Cube with a robotic hand, OpenAI used an approach that automatically expanded the range of variation once the robot reached a certain level of performance.</p><h3>Real-World Validation: What Can a Robot’s Failures Tell Us?</h3><p>Even a robot that has experienced many different conditions in simulation can fail during actual operations. <strong>These failures offer clues about what needs to be checked and improved.</strong></p><p>Start with the basic setup. Are simulation and reality using the same coordinate frames and units? Are commands reaching the correct joints at the right time? If these settings are wrong, changing the training alone will not resolve the problem.</p><p>Next, investigate the cause using the three types of differences discussed earlier. Did the robot correctly locate the part (<strong>perception</strong>)? Did the part slip during grasping or lifting (<strong>contact</strong>)? Did the robot move at the intended time and along the intended path (<strong>motion</strong>)?</p><p>There may be more than one cause. In that case, review video and sensor logs from the operation, then retest by changing one suspected condition at a time while keeping the others fixed. This helps narrow down the cause.</p><p>Once the cause is identified, address it accordingly. If a simulated parameter does not match reality, use system identification to recalibrate the simulation. If failures recur under a particular condition, include that condition in the training range for domain randomization. Then test again.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*s-Xlm00sIklD4apqPK0HAg.png" /></figure><p>For this process to work, success must be defined in advance. Moving a part successfully once is not enough. Relevant criteria include:</p><ul><li>Success rate over repeated attempts at the same task</li><li>Performance when lighting or layout changes</li><li>Task completion time and quality</li><li>Whether human intervention was required</li></ul><p>Simulation and reality should then be evaluated against the same criteria.</p><p>A small performance gap is not sufficient on its own, either. If success rates are low in both simulation and reality, a narrow gap still does not make the robot ready for deployment. What ultimately matters is whether the physical robot delivers the performance and quality required on site.</p><h3>Autonomous Welding at Hanwha Ocean: Comparing and Aligning Simulation with Actual Operations</h3><p>If moving a single part requires this process, what about welding in a shipyard, where both the working environment and the workpieces change frequently?</p><p>As discussed in <a href="https://nc-ai.com/ko/blog/hanwha-ocean-nc-ai-autonomous-welding-physical-ai">our earlier interview with Hanwha Ocean</a>, welding involves many variables, including ambient temperature and dimensional tolerances — the allowable variation in the dimensions of the steel components that make up a ship. These variables are difficult to address with predefined rules alone. They have also often made it difficult to run simulations of the process in advance.</p><p>To address this, NC AI worked with Hanwha Ocean to build a virtual environment from 3D scans of the welding automation system, simulate the process in advance, and compare and reconcile the results with actual operations. This gave the team an opportunity to refine the model by exploring in simulation the kinds of issues that could arise on site.</p><p>A shipyard is a challenging environment even for perception. Intense light from the welding arc, sparks, welding dust, and contamination on camera lenses can all interfere with visual recognition. When NC AI announced the project’s launch in June 2026, it identified these conditions as key challenges. Research to address them is ongoing, incorporating data from the site and feedback from engineers.</p><h3>From Success in Simulation to Reliability in the Real World</h3><p>Even when world models predict the outcomes of actions and robot foundation models (RFM) learn a variety of tasks, as explored in our previous articles, performance in the real world still needs to be validated separately.</p><p>The goal of sim-to-real is not to create a perfect virtual replica of reality. It is to identify the differences that affect the task, calibrate what can be measured, and train the robot to handle the variation that remains. Problems found during real-world validation then guide the next round of calibration and training.</p><p>A robot that can be trusted on site is not one that performs perfectly under every possible condition. It is one whose operating conditions — and limits — have been verified. Turning success in simulation into reliability in the real world requires being able to explain what differs between simulation and reality and how far the system has been validated.</p><h3>References</h3><ul><li>NVIDIA, <a href="https://docs.nvidia.com/learning/physical-ai/getting-started-with-isaac-lab/latest/robot-learning-fundamentals/08-simulation-to-reality.html">Getting Started With Isaac Lab: Simulation to Reality</a></li><li>Tan et al., <a href="https://arxiv.org/abs/1804.10332">Sim-to-Real: Learning Agile Locomotion For Quadruped Robots</a> (2018)</li><li>OpenAI, <a href="https://openai.com/index/solving-rubiks-cube/">Solving Rubik’s Cube with a Robot Hand</a> (October 2019)</li></ul><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=27dea4b8e1c9" width="1" height="1" alt=""><hr><p><a href="https://medium.com/nc-ai/what-is-sim-to-real-27dea4b8e1c9">What Is Sim-to-Real?</a> was originally published in <a href="https://medium.com/nc-ai">NC AI</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Why South Korea Could Lead the World in Sovereign Physical AI: The Difference Industrial Data Makes]]></title>
            <link>https://medium.com/nc-ai/why-south-korea-could-lead-the-world-in-sovereign-physical-ai-the-difference-industrial-data-makes-6a30a0425d0c?source=rss----83564b09dfb3---4</link>
            <guid isPermaLink="false">https://medium.com/p/6a30a0425d0c</guid>
            <category><![CDATA[south-korea]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[physical-ai]]></category>
            <category><![CDATA[robotics]]></category>
            <category><![CDATA[manufacturing]]></category>
            <dc:creator><![CDATA[NC AI]]></dc:creator>
            <pubDate>Wed, 14 Oct 2026 08:52:17 GMT</pubDate>
            <atom:updated>2026-10-14T08:52:17.637Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*7twLphIT3y1oS0DdWY283g.png" /></figure><p>What does AI need to learn before it can perform tasks in factories and shipyards? It needs experience locating components, adjusting movements to working conditions, and checking whether the results meet quality standards.</p><p>This kind of experience is difficult to learn from text on the web alone. Physical AI refers to AI that understands its surroundings and acts in the real world through robots and industrial equipment.</p><p>From production lines assembling precision components to shipyards building massive vessels, South Korea has manufacturing environments with different operating conditions and process know-how accumulated over many years.</p><p>But having these facilities does not automatically give South Korea an advantage in AI. The capabilities built over time depend on which models learn from that experience and who can use and improve the results.</p><p>Having control over industrial data, the training process, and its outcomes is the core of what this article calls sovereign physical AI.</p><p><strong>What advantages could South Korea’s industrial experience offer in sovereign physical AI? And whose AI capabilities will that experience help build?</strong></p><p>This article explores the learning opportunities and know-how found in South Korea’s manufacturing environments, along with the conditions needed to turn them into a lasting competitive advantage.</p><h3>What Changes When Manufacturing Operations Meet AI Capabilities?</h3><p>Manufacturers know which tasks need improvement, what to observe during a process, and how to evaluate the results.</p><p>AI developers organize this information into a form models can learn from and build the technology that connects perception and decision-making to physical action. Together, these capabilities can turn industrial challenges into specific training objectives and put the results to use in manufacturing operations. For example, manufacturers can define the criteria for weld quality, while AI developers prepare the training data and technology that help models learn those criteria.</p><p>This collaboration does not end with a single transfer of data or the introduction of a model. As operating conditions change and new problems arise, the ability to respond improves, and the experience gained becomes an asset for technology development. How the collaboration is structured, and how its training results may be used, also determine which company’s development capabilities benefit from that asset.</p><p>South Korea has the AI development base to sustain this connection domestically. According to the Ministry of Science and ICT, Stanford HAI’s AI Index 2026 report lists eight “notable AI models” from South Korea. Based on models released in 2025, this is the third-highest count, after the United States and China. The measure does not directly assess physical AI’s performance in industrial environments, but it shows that multiple companies in South Korea are developing their own models.</p><p><strong>South Korea’s opportunity lies in its ability to connect manufacturing know-how with domestic AI development capabilities.</strong></p><h3>Diverse Manufacturing Processes Provide Learning Tasks for AI</h3><p>From a physical AI perspective, South Korea’s manufacturing sector offers more than scale. It also offers a wide range of processes. Industries such as automotive, electronics, semiconductors, batteries, and shipbuilding involve tasks that differ in the size and shape of the objects being handled, the precision required, and the surrounding conditions.</p><p>The country also has a strong base of industrial robot use. According to the International Federation of Robotics’ World Robotics 2025 report, South Korea had 1,220 industrial robots per 10,000 manufacturing employees in 2024 — the highest robot density in the world. Robot density does not directly indicate how much training data is available, but it shows a high level of robot use in production. These operating environments can provide a foundation for connecting physical AI training and deployment to production processes.</p><p>The variety of these processes becomes clearer when we look at specific tasks. Imagine fitting a small cable connector into its designated position. This requires locating the component, aligning it, and adjusting the movement based on contact. Moving a battery cell without damaging it requires assessing its condition and controlling the gripping force and movement path. Both tasks involve handling objects, but the requirements for success are different.</p><p>Inspecting a product’s surface for tiny defects presents a perception and decision-making challenge. It requires distinguishing normal features from defects and assessing what is observed against quality standards.</p><p>In a shipyard, the scale and complexity of the workspace change. Deciding where to weld and how to approach the target inside a ship structure requires understanding both the surrounding geometry and the condition of the components. Even within welding, the position, posture, and access path can vary from task to task, making it necessary to choose actions that fit the situation.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*44JG7o2CbAhFaA3xAnk1oQ.png" /></figure><p>Manufacturing processes require different capabilities, including precise positioning, force control suited to the object, quality assessment, and understanding complex spaces. <strong>This diversity broadens the range of tasks physical AI can learn.</strong></p><h3>What Experience Has Been Built Up in South Korea’s Manufacturing Environments?</h3><p>The depth of experience within these environments matters as much as their breadth. Production output and equipment operating hours are common indicators of how a process is running. To learn how to perform tasks, physical AI also needs to account for the conditions at the time and how workers responded. What did skilled workers observe? Which methods did they choose? How did they respond to the results? These decisions contain know-how built through years of manufacturing.</p><p>The autonomous welding project jointly pursued by South Korean shipbuilder Hanwha Ocean and NC AI illustrates why this judgment matters. Shipyard welding involves many variables, including temperature and component tolerances — the allowable variation in component dimensions. These variables are difficult to address through predefined rules alone, so the work relies heavily on workers’ experience and practical judgment. Alongside the movements used to perform a weld, understanding the surroundings and deciding how to proceed also become subjects for AI learning.</p><p>During the collaboration, Hanwha Ocean confirmed that AI could identify relevant process features in actual welding footage. Dong-young Kim, a principal researcher in Hanwha Ocean’s Welding Technology Research Team, described the moment this way:</p><blockquote><em>“I thought that if AI could recognize this complex environment, the next step could be to turn it into data and use it to train models.”</em></blockquote><p><em>(translated from Korean)</em></p><p>To translate this possibility into learning, the conditions during a task, the actions taken, and the resulting outcomes need to be recorded and organized together. Video and sensor information must be linked to the actions and outcomes at the corresponding points in time so that AI can learn the task as a connected process.</p><p><strong>The process know-how South Korean industry has accumulated can become a new technological asset when it is translated into a form AI can learn from.</strong> If the data captures which responses were effective under which conditions, judgment that once depended on individual workers’ practical experience can become knowledge an entire facility can continuously use and improve.</p><h3>Validation Across Virtual and Physical Environments</h3><p>Before a trained model is applied to a production process, it needs to be checked to see whether it can perform the intended task in that environment. It also needs to respond appropriately when the arrangement of objects or the state of the equipment changes. Repeatedly testing different conditions on physical equipment can face constraints involving time, cost, and safety. To address these constraints, development teams also test in virtual environments and compare the results with physical operations.</p><p>In the Hanwha Ocean example discussed above, the welding automation system was scanned in 3D to create a virtual environment. The process was simulated in advance, and the results were compared with actual operations to guide adjustments. This provided an opportunity to refine the model by examining potential issues virtually before encountering them in the field.</p><p>Validation involves comparing working conditions and movements in virtual and physical environments, then evaluating task completion, quality, and stability. The differences identified are used to improve the model and the virtual environment.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*4cE85N08789IAGR0FguC-A.png" /></figure><p>This requires collaboration between industry experts and AI developers. People who understand the process evaluate whether the results meet quality standards and identify operational constraints. Developers analyze whether a problem arose during perception, decision-making, or action, then use that analysis to guide additional data collection or model adjustments. Even when failures look similar, their causes need to be distinguished and recorded so that the next round of training has a clear focus.</p><p><strong>South Korea’s manufacturing environments can serve both as sources of data and as a foundation for validating and improving technology.</strong> If the companies operating production equipment and the AI developers establish an evaluation framework together, they can assess performance against industrial requirements and accumulate deployment experience.</p><h3>What Does It Take to Build Sovereign AI Capabilities from Industrial Experience?</h3><p>The results accumulated through repeated training and validation ultimately become someone’s AI assets. Control over how the data may be used, and who can apply and improve the training results, is necessary for industrial know-how to become part of South Korea’s own AI capabilities.</p><p>First, the partners need to agree on <strong>the purpose and scope of data use</strong>. This includes deciding whether the data may be used only to improve a particular process, whether training for other tasks is permitted, and where the data will be processed.</p><p><strong>Rights over the results of model training</strong> also need to be clearly defined. The partners must agree on who will hold the developed model, how it may be used, and whether further training or adjustments are permitted as the process changes. For example, using a welding model trained on one production line on another line requires the right to retrain it for the new conditions. Agreeing on the scope of use and improvement in advance makes it possible to use the project’s results in later development and operations.</p><p>Sustaining this collaboration also requires <strong>industry-specific data formats and quality standards</strong>. If companies record different information or use different criteria to judge task success and quality, it becomes harder to use their data together. Sharing the necessary standards and establishing access within agreed limits can support the accumulation of training assets while accounting for process characteristics and security.</p><p>The same applies when using external models or infrastructure. <strong>Who builds AI capabilities from South Korea’s industrial experience depends on how these arrangements are structured.</strong> Securing rights to use data and to apply and improve training results can turn the combination of a manufacturing base and AI development capabilities into a lasting competitive advantage.</p><h3>Taking Technology Validated in South Korea to Global Markets</h3><p>Expanding technology validated in South Korea into overseas markets requires adapting models to different equipment and environments. It is also important to build a solution that covers the full process, from data collection to training, validation, and operation.</p><p>The physical AI research and development initiative in the Jeonbuk and Gyeongnam regions, in which NC AI participates, aligns with this direction. Led by the Ministry of Science and ICT and the National IT Industry Promotion Agency (NIPA), the initiative focuses on autonomous factory technology that connects and operates the entire factory in Jeonbuk, and on precise control of manufacturing processes in Gyeongnam. It also envisions developing the outcomes into a “K-AI Factory Package” for expansion into global manufacturing markets.</p><p>As part of the initiative, NC AI has been tasked with creating 3D virtual environments of industrial facilities to establish a foundation for AI training and validation. It also plans to develop world models and action models for applications such as shipyard welding and semiconductor back-end processes, which involve chip packaging and testing. World models understand and predict changes in the environment, while action models plan and control robot actions. Addressing domestic industrial challenges could help build the technology and operating frameworks needed for deployment overseas.</p><p>South Korea’s industrial data can become a lasting competitive advantage when it is turned into assets for AI training and the results are validated in real production processes. That advantage also depends on retaining the ability to use and improve the results independently.</p><p><strong>If this framework expands across industries and into overseas environments, South Korea could build a foundation for global leadership in sovereign physical AI.</strong></p><p><strong>Would you like to turn your manufacturing know-how into lasting AI assets?</strong></p><p>NC AI researches physical AI using digital twins and AI models. This research draws on manufacturing data for training and validation, with the aim of connecting what AI learns to the actions of industrial equipment.</p><p><strong>Explore how physical AI could be applied to your manufacturing processes with NC AI.</strong></p><figure><img alt="https://nc-ai.com/en/solutions" src="https://cdn-images-1.medium.com/max/1024/1*nGRn6j-SkTnsVawiC17f4g.png" /></figure><p><strong>[ References ]</strong></p><p>Stanford HAI, <a href="https://hai.stanford.edu/ai-index/2026-ai-index-report">The 2026 AI Index Report</a> (April 2026)<br>Byline Network, <a href="https://byline.network/2026/04/23-579/">Stanford Revises the Number of Notable Korean AI Models from Five to Eight</a> (April 2026, in Korean)<br>International Federation of Robotics (IFR), <a href="https://ifr.org/ifr-press-releases/news/robot-density-surges-in-europe-asia-and-americas">Robot Density Surges in Europe, Asia, and Americas</a> (April 2026)<br>Ministry of Science and ICT, <a href="https://eiec.kdi.re.kr/policy/materialView.do?num=283906">Call for Proposals Opens for the Gyeongnam and Jeonbuk AX R&amp;D Programs</a> (July 2026, in Korean)</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=6a30a0425d0c" width="1" height="1" alt=""><hr><p><a href="https://medium.com/nc-ai/why-south-korea-could-lead-the-world-in-sovereign-physical-ai-the-difference-industrial-data-makes-6a30a0425d0c">Why South Korea Could Lead the World in Sovereign Physical AI: The Difference Industrial Data Makes</a> was originally published in <a href="https://medium.com/nc-ai">NC AI</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Hanwha Ocean–NC AI Autonomous Welding Project: The Beginning of Physical AI Transformation in…]]></title>
            <link>https://medium.com/nc-ai/hanwha-ocean-nc-ai-autonomous-welding-project-the-beginning-of-physical-ai-transformation-in-ad08b224dfac?source=rss----83564b09dfb3---4</link>
            <guid isPermaLink="false">https://medium.com/p/ad08b224dfac</guid>
            <category><![CDATA[robotics]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[physical-ai]]></category>
            <dc:creator><![CDATA[NC AI]]></dc:creator>
            <pubDate>Fri, 02 Oct 2026 02:00:07 GMT</pubDate>
            <atom:updated>2026-10-12T05:39:40.993Z</atom:updated>
            <content:encoded><![CDATA[<h3>Hanwha Ocean–NC AI Autonomous Welding Project: The Beginning of Physical AI Transformation in Shipyards</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*dwwyofjoy7DLKZjajoiVMw.jpeg" /></figure><p>Hanwha Ocean is a global shipbuilding and offshore company that builds everything from commercial vessels to naval ships. In recent years, the company has been advancing its transition toward a “Smart Yard” by applying AI-powered automation technologies to its production sites. One of the key processes Hanwha Ocean is focusing on is welding automation. Welding accounts for a significant portion of shipbuilding, but with so many variables, from on-site temperature to component tolerances, rule-based automation alone falls short. As a result, welding still relies heavily on the experience and judgment of skilled workers.</p><p>Hanwha Ocean has been collaborating with NC AI to apply autonomous welding technology to this process. Using real welding footage from harsh shop-floor conditions, the two companies have confirmed that domain-specific elements can be recognized, turned into data, and used to train models. They are now developing the technology step by step.</p><p>In this interview, we spoke with Dongyoung Kim, Principal Researcher on Hanwha Ocean’s Welding Technology Research Team, about challenges in the field, why Hanwha Ocean chose NC AI as a partner, the changes a VLA (Vision-Language-Action) autonomous welding solution could bring, and how Physical AI, AI that operates in the physical world, could transform shipyards.</p><h3>Only 40% of the Day Is Spent on Actual Welding</h3><p>Workers spend only about 40% of an eight-hour workday actually welding with the torch on. That, Kim says, is the reality of conventional welding. A welding process involves much more than simply producing an arc.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*beEon_fXwXYRNcIQlQtltg.png" /></figure><p>“Once you factor in prep work before and after welding, understanding the site and conditions, and deciding which process to use, actual welding takes up only about 40% of an eight-hour day.”</p><p>In other words, more than half of the day is spent not on welding itself, but on “observing the surroundings and making decisions in order to weld.” For Hanwha Ocean, true autonomous welding is not simply about having a robot perform the welding. It is about bringing intelligence to the entire process of how workers read and assess their surroundings.</p><h3>Why Did Hanwha Ocean Choose NC AI?</h3><p>Hanwha Ocean’s sites have accumulated a wealth of real manufacturing data over several decades. But welding varies so much with working conditions and the state of each component that running reliable simulations in advance has often been difficult.</p><p>This is where the collaboration with NC AI offered a way forward. Hanwha Ocean’s welding automation system was first 3D-scanned to build a virtual environment, where the process could be simulated in advance and then compared against and fine-tuned to real-world operations. Working through trial and error virtually, before it happens on the shop floor, made it possible to refine the model more reliably.</p><p>When asked why Hanwha Ocean chose to work with NC AI among the many AI companies available, Kim points to NC AI’s focus on manufacturing.</p><p>“There are many companies that build AI models or aim to specialize in specific industries, but NC AI stood out for its clear commitment to the manufacturing floor. It may have started by chance, but once we saw how well our strengths complemented each other, the partnership felt almost inevitable.”</p><h3>One Look at the Welding Footage Revealed the Potential</h3><p>During the collaboration, Hanwha Ocean used actual welding footage to examine whether domain-specific elements in the video could be recognized. Kim says the team confirmed that these elements could be meaningfully identified.</p><p>That finding sparked a bigger idea.</p><p>“I thought: if AI can make sense of an environment this complex, the next step is to turn it into data and use it to train models.”</p><p>Kim says this process gave him greater confidence in the project.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*xP92bOBN4T66K3p5SYAAqg.png" /></figure><h3>The Worker’s Role Expands into Data Management</h3><p>Hanwha Ocean does not expect the introduction of a VLA autonomous welding solution to cause a sudden, dramatic change on the production floor. Instead, Kim anticipates a gradual shift, with workers stepping in less and less over time.</p><p>In the long term, the nature of human work is also expected to change. Instead of continuously making manual adjustments to welding conditions, workers may take on broader roles such as monitoring the status of automated equipment and analyzing and managing quality data. Through this shift, Hanwha Ocean expects to improve workplace safety while gradually increasing the level of unmanned operation.</p><h3>From a Single Welding Cell to the Entire Shipyard</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*eEqSJjQht2CIoTZatXg2xQ.png" /></figure><p>The path toward unmanned operation begins with a single small work area, or cell. The area where Hanwha Ocean is currently seeking to apply autonomous welding represents only a small part of the overall shipyard. However, Kim believes that if models can be built up and proven within this small unit, they could later extend to other processes and serve as a starting point for the Physical AI transformation of the entire shipyard.</p><p>Kim says the project is about more than technical achievement. It is about using technology to offer a solution to manufacturing’s long-standing labor shortages and building a workplace where people and technology work side by side.</p><p>“Maintaining manufacturing competitiveness even when there are not enough people, passing the skills of experienced workers on to the next generation through data, and enabling workers to do more valuable work in a safer environment — that is what we see as the ultimate goal of autonomous welding.”</p><h3>Are You Considering a Physical AI Transformation for Your Manufacturing Site?</h3><p>NC AI is developing Physical AI technologies, including VLA, based on data from real industrial environments, and is working to apply them to manufacturing processes by bridging simulation and the real shop floor.</p><p>If you are looking for opportunities to automate complex manufacturing processes, start your journey with NC AI.</p><figure><a href="https://nc-ai.com/ko/solutions"><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*nGRn6j-SkTnsVawiC17f4g.png" /></a></figure><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=ad08b224dfac" width="1" height="1" alt=""><hr><p><a href="https://medium.com/nc-ai/hanwha-ocean-nc-ai-autonomous-welding-project-the-beginning-of-physical-ai-transformation-in-ad08b224dfac">Hanwha Ocean–NC AI Autonomous Welding Project: The Beginning of Physical AI Transformation in…</a> was originally published in <a href="https://medium.com/nc-ai">NC AI</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Why Do Robots That Move on Their Own Need an “Agent”?]]></title>
            <link>https://medium.com/nc-ai/why-do-robots-that-move-on-their-own-need-an-agent-a8cf3551abf0?source=rss----83564b09dfb3---4</link>
            <guid isPermaLink="false">https://medium.com/p/a8cf3551abf0</guid>
            <category><![CDATA[robotics]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[ai-agent]]></category>
            <category><![CDATA[physical-ai]]></category>
            <category><![CDATA[agentic-ai]]></category>
            <dc:creator><![CDATA[NC AI]]></dc:creator>
            <pubDate>Wed, 30 Sep 2026 05:59:39 GMT</pubDate>
            <atom:updated>2026-09-30T07:10:07.013Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*CXtuo3HTZFaokdSKypuVwA.png" /></figure><p>AI is rapidly expanding beyond digital environments and into the physical world. One of the most representative examples is <strong>Physical AI</strong> — AI that perceives its surroundings through cameras and sensors, makes decisions based on the situation, and translates those decisions into the physical movements of robots.</p><p>One of the characteristics often associated with Physical AI is <strong>autonomy</strong>. This means that AI can understand its surroundings and choose appropriate actions without a person having to define every movement in advance.</p><p>At the same time, <strong>Agentic AI</strong> has also been gaining attention in the robotics field. Agentic AI, too, is described as technology that understands a goal provided by a person, creates its own plan, and acts to achieve it.</p><p>This raises a natural question.</p><p><strong>If Physical AI can already make decisions and act on its own, why does it need an “agent” on top of that?</strong></p><h3>Physical AI Understands the Real World. Agentic AI Plans Around Goals.</h3><p>The difference between the two lies less in whether they are autonomous and more in <strong>what each technology focuses on</strong>.</p><p>Physical AI focuses on enabling AI to perceive and understand the physical world and translate that understanding into real-world action. Identifying where an object is located, deciding how to grasp it, and then moving a robot’s arm and hand are all part of this process.</p><p>Agentic AI addresses a slightly different question:</p><p><strong>“What needs to be done to achieve the given goal?”</strong></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*iF-fcqNYy36Ce8PZbBZf0g.png" /><figcaption><em>AI-generated image</em></figcaption></figure><p>An agent breaks a goal down into multiple tasks, determines the order in which they should be carried out, and selects the necessary tools or robots. It also checks whether the work is progressing properly and revises the plan if a problem occurs.</p><p>Put simply, if Physical AI gives a robot the <strong>ability to act in the real world</strong>, Agentic AI can be understood as organizing multiple actions into a single <strong>job</strong>.</p><p>Imagine telling a robot, “Put the cup on the table onto the shelf.” Identifying the locations of the cup and the shelf, deciding how to grasp the cup, and moving the arm are typical areas of Physical AI.</p><p>But the problem becomes more complex when the goal is, “Clean up this room.”</p><p>The robot must inspect the room, determine which objects need to be organized, divide the work into separate tasks, and decide the order in which to carry them out. Different objects may require different actions, and if an unexpected situation occurs during the process, the plan may need to change.</p><p>In other words, the robot must decide not only <strong>how to move</strong>, but also <strong>what to do and in what order</strong>.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*b6g5EViIgjH7Wn4fbdCyGg.png" /></figure><p>Of course, in real-world robot AI, <strong>the boundary between Physical AI and Agentic AI is not always clear-cut.</strong></p><p>Recent developments in <strong>VLA (Vision-Language-Action)</strong> models and robot foundation models are also moving toward incorporating reasoning and planning capabilities.</p><p>Rather than treating the two technologies as separate stages, <strong>it is closer to the truth to say that goal-oriented planning and execution management are increasingly being built into AI that acts in the physical world.</strong></p><h3>Robot AI That Understands Goals and Connects Them to Action</h3><p><strong>A good example of this relationship is a pair of models from Google DeepMind: Gemini Robotics 2 and Gemini Robotics ER 2.</strong> The two models serve different roles.</p><p>Gemini Robotics 2 is a VLA model that converts visual and language information into actual robot movements. It is responsible for generating physical actions such as picking up objects or using tools.</p><p>Gemini Robotics ER 2 operates at a higher level as an <strong>embodied reasoning model</strong> that understands goals and situations and plans tasks accordingly.</p><p>It observes the surrounding environment, plans the steps required to achieve a goal, and then delegates the actual execution of robot movements to a lower-level VLA model. It can also monitor task progress, retry failed steps, or adjust the next action when necessary.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*DNFX8FqAUEkAspoZtvMhsQ.jpeg" /><figcaption><em>Source: Google DeepMind, “Introducing Gemini Robotics ER 2”</em></figcaption></figure><p>By connecting the <strong>reasoning ability to understand goals and situations and plan tasks</strong> with the ability to generate actual physical actions, robots can carry out a single job made up of multiple steps.</p><h3>Orchestration: Connecting Multiple Robots Around a Single Goal</h3><p>The role of agents can become even more important when multiple robots and pieces of equipment need to work together.</p><p>Real industrial environments involve more than just a single robot. Robotic arms, AMRs, inspection equipment, and various production systems with different functions operate together.</p><p>For example, in a process where a logistics robot delivers required parts and a robotic arm assembles them, it is not enough to define the role of each robot. The order of tasks and the progress of the overall process must also be coordinated.</p><p>Connecting them toward a single goal requires deciding which robot handles which task and coordinating the next steps as each task progresses.</p><p>This kind of <strong>orchestration</strong> — understanding the capabilities and status of different robots and equipment and assigning and coordinating tasks under a common goal — is becoming increasingly important as Physical AI and agents come together.</p><p>The challenge is no longer only about how much one robot can do. It is also about how robots with different capabilities can be made to operate as part of a single working system.</p><h3>Safety and Execution Validation for Robots That Plan on Their Own</h3><p>As agents begin to plan and decide more tasks, another issue emerges.</p><p><strong>Should a robot simply execute whatever plan the AI creates?</strong></p><p>In digital environments, incorrect results can often be corrected or tasks can be run again. But the actions of robots operating in the physical world can directly affect equipment, products, and human safety.</p><p>Even if AI creates a logically appropriate plan, the action may not be something the robot can currently perform.</p><p>A person may enter the robot’s path during operation, or changes in equipment conditions may make the original plan difficult to execute.</p><p>For this reason, there needs to be a process between an agent’s planned action and a robot’s actual execution that verifies whether the action is possible under the current conditions and whether it falls within an allowed range.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Mvrbm7t0yPlYCVAwYw0BOw.png" /><figcaption>Xinyuan Liu 외(2026), “Physical Agentic AI”, arXiv:2608.22657v1.</figcaption></figure><p><em>Xinyuan Liu et al. (2026), “Physical Agentic AI,” arXiv:2608.22657v1.</em></p><p>A recent study titled <strong>“Physical Agentic AI”</strong> proposed an architecture that separates planning from execution.</p><p>In this structure, an AI-based <strong>Mission Planner</strong> breaks a goal into multiple steps and creates a task plan. A <strong>Robot Orchestrator</strong> then checks the robot’s capabilities, current state, and task conditions to verify whether the planned action can actually be performed before allowing execution.</p><p>Ultimately, what real-world agents need is not simply the ability to make more decisions on their own.</p><p>We also need to design <strong>how far they can make decisions autonomously, which actions must be validated before execution, and when they should stop or ask for human judgment if a problem occurs.</strong></p><h3>Autonomous Agents That Connect Perception to Execution in the Field</h3><p>NC AI is also developing <strong>VAETKI Agent</strong>, an autonomous agent designed to connect AI in digital environments with physical reality.</p><p>The agent perceives conditions in the field and plans the tasks needed to achieve a goal. At the same time, NC AI is building an orchestration structure that connects the AI’s decisions to the physical control of robots and equipment.</p><p>What matters is that the process does not end after a single decision and action.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*fcaaodDoCniCspLkclMEuA.png" /></figure><p>The system perceives the situation in the field, creates a plan, executes that plan through actual robots, and then checks the result.</p><p>If the outcome differs from what was expected, it updates the next plan based on the new situation.</p><p>As this <strong>closed loop of perception → planning → execution → result verification</strong> repeats, AI decisions and changes in the real world remain continuously connected.</p><p>To apply these autonomous agents in industrial environments, NC AI also considers it important to build a structure that ensures <strong>reliability and controllability alongside autonomy</strong>.</p><p>This includes recording AI decisions and actions so they can be traced, and defining the authority and scope within which an agent is allowed to make decisions on its own.</p><p>A control system is also needed to safely stop robot operations when a risk is detected or when the situation moves beyond the permitted range.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*3PFgBeOAFq2P9CQQOoF6DA.jpeg" /></figure><p><em>Source: NC AI YouTube</em></p><p>In other words, using autonomous agents in industrial environments requires more than simply allowing AI to make more decisions on its own.</p><p>It also requires a structure in which people can review those decisions and actions and intervene when necessary.</p><p>Physical AI provides the foundation that allows AI to understand and act in the real world.</p><p>When an agent’s planning and coordination capabilities are added, individual actions can be connected into a workflow designed to achieve a single goal.</p><p>As that scope expands to multiple robots and pieces of equipment, AI’s role also expands beyond deciding the actions of individual robots. It begins to plan and coordinate how work itself progresses across the industrial environment.</p><p>Ultimately, what will matter in the future of robot AI is not simply <strong>what movements a robot can perform</strong>.</p><p>The real question is whether it can understand a given goal, plan the necessary work, and carry it through to completion in the physical world in a safe and reliable way.</p><p>That is why the combination of Physical AI and Agentic AI is changing the way robots <strong>“work.”</strong></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=a8cf3551abf0" width="1" height="1" alt=""><hr><p><a href="https://medium.com/nc-ai/why-do-robots-that-move-on-their-own-need-an-agent-a8cf3551abf0">Why Do Robots That Move on Their Own Need an “Agent”?</a> was originally published in <a href="https://medium.com/nc-ai">NC AI</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[From the Virtual World to Reality: NC AI Goes to the Shipyard]]></title>
            <link>https://medium.com/nc-ai/from-the-virtual-world-to-reality-nc-ai-goes-to-the-shipyard-988b16bcc82f?source=rss----83564b09dfb3---4</link>
            <guid isPermaLink="false">https://medium.com/p/988b16bcc82f</guid>
            <category><![CDATA[physical-ai]]></category>
            <category><![CDATA[manufacturing]]></category>
            <category><![CDATA[robotics]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <dc:creator><![CDATA[NC AI]]></dc:creator>
            <pubDate>Wed, 30 Sep 2026 04:59:55 GMT</pubDate>
            <atom:updated>2026-09-30T08:17:49.774Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*zAdzJt4Mkow48j4dIxc1zA.png" /></figure><p>AI is expanding beyond the screen and into real-world industrial environments. Physical AI refers to technology that can interpret real-world environments and situations and translate that understanding into physical action.</p><p>One of the places where NC AI is exploring the potential of Physical AI is the <strong>shipyard</strong>, where enormous vessels are built. What technologies and preparations are needed for AI to understand complex welding processes and eventually bring them toward automation?</p><p>We spoke with <strong>Saebyeok Lee, Team Lead of the Multimodal AI Team at NC AI’s Physical AI Lab</strong>, about what Physical AI is, how AI can make sense of manufacturing environments, and what technologies NC AI is developing at shipyards.</p><h3>AI Is Beginning to Move Beyond the Screen</h3><p>The AI we are familiar with has mainly operated inside computers and smartphones. It finds information and answers questions, and more recently, AI has also evolved toward agentic systems that can use multiple tools to make plans and carry out complex tasks on their own.</p><p>Lee describes Physical AI as the next step. AI is expanding beyond digital environments and connecting with physical devices in real industrial settings to perform tasks in the real world.</p><p>When people think of Physical AI, humanoid robots may be the first thing that comes to mind, but its scope is much broader. Forklifts, cranes, and collaborative robots (cobots) that have traditionally been operated remotely by humans, as well as self-driving cars and drones, can all be areas where Physical AI is applied.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*PvfFDLWAUILuXglf_3bOpQ.png" /></figure><p>However, AI operating in the physical world needs more than the ability to simply perceive its surroundings. It must anticipate what may happen next, interpret the environment around it, and choose the appropriate action. For tasks where even a small difference in force could damage an object, its movements must also be controlled with precision.</p><p>In this sense, <strong>Physical AI can be understood as AI that interprets the real world, makes decisions, and connects those decisions to physical actions.</strong></p><p>The area NC AI is focusing on first in this process is the ability to <strong>understand the worksite</strong>. Before AI can actually move, it must first know what is happening in front of it.</p><h3>Before Physical AI Can Move, It Must First Understand the Environment</h3><p>For AI to operate properly in an industrial environment, the first thing it needs is the ability to <strong>accurately determine what is happening at that moment</strong>. NC AI’s Multimodal AI Team is researching Vision-Language Models (VLMs) as a technology capable of performing this role.</p><p>A VLM is a model that interprets objects and situations within a scene based on visual inputs such as images. For example, it can examine an image from a worksite and identify what equipment is present, determine the condition of the workpiece, or recognize what has changed between different images.</p><p>Since 2024, NC AI has released VLMs on three occasions, building up its capabilities in this area. The team has also gained experience preparing training data, designing model architectures, and conducting large-scale training in multi-node environments.</p><p>Now, the application of this technology is expanding beyond general image understanding into industrial environments. This includes processing and training on data generated at manufacturing sites for specific purposes, as well as optimizing models for edge-device environments so they can be used directly on-site.</p><p>Going forward, interpreting a single image will not be enough. To judge what is happening in an actual process more accurately, AI must be able to identify <strong>meaningful differences</strong> between multiple images and understand how objects change over time within video.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Xaq0xsOSSKSVUErtX_KvxA.png" /></figure><p>The problem NC AI is trying to solve first in its shipyard project also begins with this kind of <strong>understanding of the worksite</strong>. Rather than immediately making robots move, the first goal is to enable AI to <strong>observe a welding site and understand the state of the process</strong>.</p><h3>So NC AI Went to the Shipyard</h3><p>When ships are built at a shipyard, welding is a core task that plays a major role throughout the construction process. Because welding connects the structures that make up the hull and ensures their quality, both the speed and accuracy of the work directly affect the overall construction schedule and productivity.</p><p>At the same time, shipyards are facing a growing <strong>shortage of skilled workers</strong>. Welding requires years of experience and relies heavily on the expertise of skilled workers. As the workforce ages and workers leave the industry, finding ways to supplement this expertise has become an increasingly important challenge.</p><p>NC AI’s shipyard project begins with these real-world problems.</p><p>At the current stage, however, the goal is not to have robots immediately perform every welding task. The focus is first on developing <strong>AI that can understand the welding process itself and determine the work target and its current condition based on visual information from the site</strong>.</p><p>For example, the AI needs to determine which area it is looking at, which stage of the process is currently underway, and what condition the welding target is in before further automation becomes possible. In other words, this is the stage of first building <strong>intelligence capable of interpreting the process</strong> before moving on to actual equipment control.</p><p>The shipyard project is not simply about applying a single AI model to an industrial site. It is also a process of verifying <strong>how far Physical AI can go in interpreting conditions and making judgments within an actual manufacturing environment</strong>.</p><h3>Why an AI Engineer Started with Welding Terminology</h3><p>Manufacturing environments were not entirely new to Lee, who had previously worked on systems integration (SI) projects in the manufacturing sector. Shipbuilding, however, and welding processes in particular, were a different domain.</p><p>One of the first things Lee had to do when the project began was learn the procedures involved in welding and the terminology used on-site. Shipyards often use a mixture of Japanese-derived shop-floor jargon, English abbreviations, and Sino-Korean terminology, making it feel almost like a foreign language at first.</p><p>This process involves more than simply memorizing unfamiliar terms. Even when looking at the same scene, it is difficult to determine what information matters without understanding the process and the way people work on-site. Engineers need to know what constitutes a normal condition, what needs to be distinguished, and what criteria actual workers use to make decisions before they can clearly define the data and problems that AI needs to learn.</p><p>Lee emphasizes <strong>an understanding of the domain and the worksite</strong> as one of the factors that can determine the success or failure of a real-world project.</p><p>Machine learning engineers need to do more than focus on the model itself. They must meet customers directly, identify the problems occurring at the worksite, and understand the language and criteria used by the people performing the work.</p><p>Applying AI to an industrial environment does not end with simply bringing in a technology. It also requires <strong>translating problems from the worksite into the language of technology and connecting those solutions back into a form that can actually be used in the field</strong>.</p><h3>How Far Can Physical AI Go Beyond the Shipyard?</h3><p>The shipyard project is one application case, but NC AI sees the potential scope of Physical AI as much broader. The technologies and experience gained from understanding the shipyard environment can also be extended to industries with different environments and processes.</p><p>In manufacturing, the technology could be used to identify the condition of equipment and processes in greater detail and raise the level of automation in smart factories. In logistics, it could provide the underlying technology needed to quickly identify and sort large volumes of items or help autonomous mobile robots understand and navigate their surroundings.</p><p>It also has potential in environments that are difficult for people to access directly. In the defense sector, it could help unmanned robots operating in dangerous areas recognize their surroundings, while the range of applications for AI that can interpret visual information could continue to expand across broadcasting and other industries as well.</p><p>NC AI is currently collaborating with companies or exploring potential applications across a range of fields, including manufacturing, logistics, defense, and broadcasting.</p><p>What matters is not applying a single model to every industry in the same way, but <strong>redefining what AI needs to see and understand based on the data and ways of working specific to each environment</strong>.</p><p>As experience from the shipyard continues to accumulate, the range of real-world environments that Physical AI must learn to understand is expanding as well.</p><h3>The Goal Is 100% Automation</h3><p>The ultimate goal of Physical AI is not simply to assist with work currently performed by people. Lee says the long-term goal should be 100% automation.</p><p>Of course, real industrial environments cannot be transformed into fully automated systems all at once. Even in welding, some processes have already been automated or semi-automated, while other areas still require human judgment and intervention depending on the working environment and conditions.</p><p>The process will need to begin with a <strong>human-in-the-loop</strong> approach, in which people remain involved at intermediate stages, and gradually expand the scope of automation as technology improves and more real-world data is accumulated.</p><p>Above all, <strong>safety is critical throughout this process.</strong> Unlike AI that produces results on a screen, Physical AI moves actual equipment and robots. Even a small error in judgment can affect people or equipment, which means that an accurate understanding of the worksite and stable control are just as important as model performance.</p><p>For this reason, the performance of Physical AI is difficult to evaluate through benchmark scores alone. It is also necessary to determine <strong>whether the system can operate reliably in real-world environments and maintain safety even when unexpected situations occur</strong>.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*X2ocbeVd09vk7nXe2eAA7w.png" /></figure><p>Lee describes the current moment in the industry as a time when “a massive wave of Physical AI” is arriving. <strong>Riding that wave, NC AI is building the technologies needed for real industrial environments, one step at a time.</strong></p><p><strong>“In five years, perhaps we’ll have become a seasoned pro surfer. By then, I hope the wave we helped pioneer will have grown into a powerful current.”</strong></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=988b16bcc82f" width="1" height="1" alt=""><hr><p><a href="https://medium.com/nc-ai/from-the-virtual-world-to-reality-nc-ai-goes-to-the-shipyard-988b16bcc82f">From the Virtual World to Reality: NC AI Goes to the Shipyard</a> was originally published in <a href="https://medium.com/nc-ai">NC AI</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Why Sovereign AI Matters in the Physical AI Era: The Difference Action Data Makes]]></title>
            <link>https://medium.com/nc-ai/why-sovereign-ai-matters-in-the-physical-ai-era-the-difference-action-data-makes-e0cd38fffb54?source=rss----83564b09dfb3---4</link>
            <guid isPermaLink="false">https://medium.com/p/e0cd38fffb54</guid>
            <category><![CDATA[sovereign-ai]]></category>
            <category><![CDATA[manufacturing]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[physical-ai]]></category>
            <dc:creator><![CDATA[NC AI]]></dc:creator>
            <pubDate>Wed, 23 Sep 2026 06:46:43 GMT</pubDate>
            <atom:updated>2026-09-30T07:10:40.159Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*6zYnUY8HKZHfPI1GYdJZqg.png" /></figure><p>Generative AI has advanced by learning from vast amounts of text and images accumulated across the web. Until now, a major measure of AI competitiveness has been who can train on more data and secure larger models and greater computing infrastructure.</p><p>But the story changes when AI moves beyond generating answers on a screen and begins taking action in the real world through robots and machines.</p><p>For Physical AI to operate in real-world environments, it needs a different kind of data. This is where <strong>action data</strong> becomes important.</p><p>So, as AI enters the physical world, is having a strong model enough to stay competitive? If it also matters who can secure the data generated in the field, who can train on it, and who can use the resulting capabilities, shouldn’t the meaning of Sovereign AI evolve as well?</p><p>In this article, we look at the difference between text data and action data to understand why Sovereign AI becomes increasingly important for Physical AI.</p><h3>1. What Is Sovereign AI?</h3><p>Sovereign AI is often understood simply as “AI developed domestically.” But it does not necessarily mean developing every technology independently.</p><p>At its core, Sovereign AI is about having the ability to choose and control the data, models, and computing infrastructure required to develop and use AI.</p><p>As generative AI has rapidly expanded, countries around the world have been working to secure their own AI models, data centers, AI chips, and other foundational technologies. One reason is that excessive dependence on a particular company or foreign technology could narrow the range of choices a country has in deciding where and how AI is used.</p><p>The important point, then, is not to build everything independently. It is to be able to decide which critical assets should remain under your own control and where external technologies can be used.</p><p>As AI begins moving into the physical world, however, the scope of what needs to be considered under Sovereign AI is also expanding.</p><p>In Physical AI, it is no longer enough to think only about models and computing infrastructure. <strong>Data generated in real-world environments must also be part of the equation.</strong></p><h3>2. Physical AI Cannot Learn from Text Alone</h3><p>If generative AI has learned knowledge about the world through text, Physical AI needs to go one step further in order to operate in the real world.</p><p><strong>Knowing what to do is different from knowing how to actually do it.</strong></p><p>For example, simply telling a robotic arm positioned beside a conveyor belt to “pick up the incoming part and assemble it in the designated position” does not mean the robot can immediately perform the task.</p><p>To actually complete the job, the robot needs to adjust the direction and speed of its movement, the force used to grasp the object, and other parameters according to its surroundings.</p><p>In other words, Physical AI needs more than knowledge about the world. It also needs <strong>experience gained through acting in the real world.</strong></p><p>And action data is what captures those experiences in a form that AI can learn from.</p><h3>3. How Is Action Data Different from Text Data?</h3><p>Action data describes <strong>what action was taken in a given situation and what happened as a result.</strong></p><p>It is more than simply a video recording of a robot moving.</p><p>Take the robotic arm beside the conveyor belt again. A single experience could include the position of the part and the robot’s current state, as well as <strong>joint angles, movement speed, the opening and closing state of the gripper, the force or torque applied to the object</strong>, and whether the action ultimately succeeded or failed.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*pQS8qaIiktPDAEtbBX4RYw.png" /></figure><p>The difference becomes clearer when we compare this with text data.</p><p><strong>Text Data</strong></p><blockquote><em>“Pick up the part and assemble it in the designated position.”</em></blockquote><p><strong>Action Data</strong></p><blockquote><em>“When the part was in this position and orientation, the robotic arm moved in this way and grasped it with this amount of force, resulting in successful assembly.”</em></blockquote><p>Text data has accumulated across the web and in documents at enormous scale, making it relatively accessible to a wide range of users and organizations.</p><p>Action data, by contrast, is <strong>generated while actual robots and equipment are operating.</strong> Because every factory uses different equipment, parts, workflows, and surrounding environments, data created at one site may not always be directly applicable to another or easy to reproduce in exactly the same way.</p><p>As a result, action data tends to be <strong>more dependent on specific real-world environments, requires more time and cost to collect and accumulate, and cannot be easily substituted with other data.</strong></p><p>And it is this difference that connects action data directly to the growing importance of Sovereign AI in Physical AI.</p><h3>4. Why Does Sovereign AI Become More Important in Physical AI?</h3><p>If action data is created within the specific equipment and workflows of an industrial site, one question naturally follows:</p><p><strong>Whose AI assets will the experience generated in the field ultimately build up?</strong></p><p>Action data contains more than the movements of a robot.</p><p>It can also reflect <strong>process know-how, equipment operation experience, and working methods</strong> — such as how particular equipment is operated, how certain parts are handled, and which conditions need to be adjusted for a task to succeed.</p><p>This means that even when the same field data is used, the AI capabilities that accumulate over time can differ depending on which model is trained on the data, who owns the resulting capabilities, and who is able to continue improving them.</p><p>For example, <strong>if data generated in the field is provided or used for further training of an external model</strong>, the industrial experience contained within that data could also become a training asset for an external AI system.</p><p>On the other hand, <strong>if an organization can decide where its data is processed, which models are trained on it, and who can own and use the resulting capabilities, it can continue accumulating the experience generated in its own operations as part of its own AI assets.</strong></p><p>This is why, for Physical AI, simply possessing data is not enough. <strong>The ability to choose and control which data is used to train which models, and how the resulting capabilities are used, also becomes important.</strong></p><p>If models and computing infrastructure were major pillars of Sovereign AI in the era of text-based AI, Physical AI adds another important dimension: <strong>control over field data and the training process itself.</strong></p><h3>5. How Can Industrial Data Become an AI Training Asset?</h3><p>The fact that large amounts of data are generated in industrial environments does not mean that the data can immediately be used to train Physical AI.</p><p>Having data and <strong>having data in a form that AI can actually learn from are two different things.</strong></p><p>The robotic arm beside the conveyor belt is a good example. During operation, a wide range of data may continuously be generated, including camera footage, sensor information, the robot’s state, control commands, and the outcome of each task.</p><p>But if these records are scattered across separate systems, or if it is impossible to tell which action led to which result at a particular point in time, it becomes difficult for AI to learn from them as a single experience.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*_dhsW5mSMopHVCmpTsFZaA.png" /></figure><p>The data therefore needs to be aligned along the same timeline, with actions and their outcomes connected and organized together.</p><p>Data prepared so that it can actually be used for AI training is often referred to as <strong>AI-Ready Data</strong>.</p><p>Ultimately, what matters is not only how much data is available, but whether there is <strong>a system for turning data generated in the field into assets that AI can actually learn from.</strong></p><h3>6. What Do We Need to Secure in the Physical AI Era?</h3><p>Action data is an important training asset for Physical AI, but data alone is not enough to build a complete Physical AI system.</p><p>It also requires models capable of learning from that data, computing infrastructure, and an operational framework that can deploy those capabilities to actual robots and equipment and continuously improve them.</p><p>Ultimately, Sovereign AI in the Physical AI era is <strong>less about what has been developed domestically and more about how much control and choice an organization has over the learning assets created in the real world and the processes through which those assets are used.</strong></p><p>At NC AI, we also approach Physical AI from this perspective.</p><p>We are researching ways to connect data generated in industrial environments with technologies such as digital twins, world models, and Robot Foundation Models (RFMs), enabling AI to understand and act within real-world environments.</p><p><strong>The ability to build and sustain this virtuous cycle — turning field data into learnable assets, applying them back to real-world environments, and using the resulting experiences to generate new data — could become a new source of competitiveness in the Physical AI era.</strong></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=e0cd38fffb54" width="1" height="1" alt=""><hr><p><a href="https://medium.com/nc-ai/why-sovereign-ai-matters-in-the-physical-ai-era-the-difference-action-data-makes-e0cd38fffb54">Why Sovereign AI Matters in the Physical AI Era: The Difference Action Data Makes</a> was originally published in <a href="https://medium.com/nc-ai">NC AI</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Automating Work from the Office to the Factory: NC AI Ushers in the Era of Agentic AI]]></title>
            <link>https://medium.com/nc-ai/automating-work-from-the-office-to-the-factory-nc-ai-ushers-in-the-era-of-agentic-ai-fbce7e7b051f?source=rss----83564b09dfb3---4</link>
            <guid isPermaLink="false">https://medium.com/p/fbce7e7b051f</guid>
            <category><![CDATA[agentic-ai]]></category>
            <dc:creator><![CDATA[NC AI]]></dc:creator>
            <pubDate>Thu, 17 Sep 2026 05:52:07 GMT</pubDate>
            <atom:updated>2026-09-30T08:18:38.241Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*qa_9Nxom8BfenjjEPe63OQ.jpeg" /></figure><p>Sangha Nam, who leads the Agentic AI Lab at NC AI, works on <strong>Agentic AI</strong>, an AI that can make decisions on its own and carry out tasks directly.</p><p>He has researched AI technologies that enable natural interaction with people, ranging from conversational technology and personas to RAG and emotion-related decision-making technologies. Today, he is researching NC AI’s proprietary VAETKI LLM lineup and core Agentic AI technologies, while also working on a range of government-funded projects that apply these technologies to real-world problems.</p><p>How is Agentic AI different from conventional conversational AI? What technologies are needed for AI to perform actual work autonomously, and how will this change the way we work in offices and manufacturing environments?</p><p>We spoke directly with Sangha Nam to find out.</p><h3>From AI That Answers to AI That Works on Its Own</h3><p>Conventional conversational AI was a model designed to <strong>“give good answers”</strong> to people’s questions.</p><p>Agentic AI goes one step further.</p><p>According to Nam, the biggest difference lies in whether AI can actually <strong>“perform and complete”</strong> a task.</p><blockquote><em>“Before giving an answer, Agentic AI first thinks, assesses the situation, decides what action to take, creates a plan on its own, and then actually carries out the task.”</em></blockquote><p>In other words, instead of simply telling people what to do, it can directly perform the actions available to it in the digital world and complete the work itself.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*nz7RMtAzHqfsSFD5o2NnHA.png" /></figure><p>The idea of <strong>“AI that works on its own”</strong> has been a research topic for a long time. But with the recent emergence of high-performance, large-scale models, AI has gained powerful reasoning capabilities. It can now call external tools according to context and carry out multi-step tasks that cannot be completed in a single interaction.</p><p>As AI has become capable of actually <strong>“taking action,”</strong> attempts to apply Agentic AI to real work in companies and industrial environments are rapidly increasing.</p><h3>Why Does AI Lose Context and Give Irrelevant Answers?</h3><p>When the amount of information AI has to process becomes too large, it can miss important details or handle them less clearly.</p><p>Nam compares this to a person trying to read a 300-page report in one sitting.</p><p>Being able to input a large amount of information at once is not the same as being able to properly understand all of it from beginning to end.</p><p>Another problem arises when multiple contexts become mixed together within a single conversation.</p><p>As the topic changes more frequently, earlier and current contexts can interfere with one another. Information that people can naturally distinguish may become confusing for AI.</p><p>Nam explains that solving this problem requires <strong>three things to work together.</strong></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*tUTDXbF7jgwOtLqtYW2Osg.png" /></figure><p>The first is improving the model’s ability to handle multi-turn conversations and long contexts — in other words, strengthening its <strong>long-context reasoning capabilities</strong>.</p><p>The second is <strong>Context Engineering</strong>, which selects and provides the information the model actually needs to focus on in the current context.</p><p>The third is that it is not enough for AI alone to become smarter. <strong>The way people and organizations work with AI must evolve as well.</strong> NC AI refers to this as both a <strong>“digital coworker system”</strong> and a <strong>“co-evolution structure.”</strong></p><p>NC AI has also experienced these challenges while applying conversational AI agents to real services. As conversations became longer, context sometimes broke down, and when RAG was used to retrieve supporting information, the system occasionally brought back information that did not match the user’s actual intent.</p><p>Through the process of applying its proprietary language models to B2B and B2G AX projects, NC AI also learned that no matter how advanced the technology is, it is ultimately difficult to use if it does not align with an organization’s rules and data.</p><h3>AI as a “Coworker,” Not a “Tool”</h3><p>NC AI is currently carrying out the <strong>“Real-World Proactive Agentic AI Technology Development Project,”</strong> led by Korea’s Ministry of Science and ICT.</p><p>Nam describes the goal of the project as turning AI from a <strong>“tool” into a “coworker.”</strong></p><p>NC AI is leading the project together with researchers from five laboratories at Korea University and Yonsei University, as well as Gabia, the company that operates the groupware platform Hiworks.</p><p>The project is structured so that research and real-world application are connected from the very beginning, allowing the technology to be validated within actual enterprise work environments.</p><p>By 2028, the project aims to demonstrate its success through two quantitative targets.</p><p>The first is a <strong>75% “autonomous task completion rate,”</strong> meaning that AI can carry out multi-step tasks within actual enterprise systems from beginning to end without additional human intervention.</p><p>The second is an <strong>85% “proactive suggestion hit rate,”</strong> meaning that AI identifies and suggests necessary actions before a person asks, and that those suggestions are actually accepted by users.</p><p>The goal is for AI to complete a significant portion of work without requiring human involvement — and eventually reach a level where it can take initiative before being instructed.</p><h3>From the Office to the Factory</h3><p>Agentic AI is not limited to digital environments.</p><p>Nam explains the relationship between Agentic AI and Physical AI this way:</p><blockquote><em>“If the Agentic AI we have discussed so far is AI that gained hands and feet in the digital world, Physical AI is what happens when those hands and feet become robots and move out into the physical world.”</em></blockquote><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*7zx0xew_t3dpbE5hHrZMzw.png" /></figure><p>The fundamental process — perceiving, reasoning, planning, and acting — remains the same. What changes is the environment in which the AI operates.</p><p>And once AI moves into the physical world, the level of difficulty rises significantly.</p><p>In an office, if a document is created incorrectly, it can simply be deleted and recreated. But if a robot makes a wrong move in a factory, the consequences may be difficult to reverse and can directly affect human safety.</p><p>Equipment, robots, logistics systems, and workers all operate together in real time, meaning that a single decision can potentially create cascading effects across an entire factory.</p><p>For this reason, Agentic AI operating in physical environments requires <strong>more accurate situational awareness and stricter safety mechanisms</strong>.</p><p>NC AI is also participating in a project that validates these technologies in real manufacturing environments.</p><p>The project is led by Jeonbuk National University, with NC AI participating as the foundation model company alongside industrial AI companies and multiple universities and research institutions.</p><p>Jeonbuk is a region where 94% of Korea’s medium- and heavy-duty commercial vehicles are produced. The project plans to use an entire representative parts factory within the regional cluster as a demonstration zone, validating the technology across the factory — from pressing and assembly to logistics.</p><p>Nam explains that if the previous project is about creating a <strong>“digital coworker for the office,”</strong> this project is about creating a <strong>“digital coworker for the factory.”</strong></p><p>As part of the project, NC AI is responsible for specializing and advancing a foundation model called AMCore for the manufacturing domain based on its proprietary models.</p><h3>How Can We Make AI That Acts on Its Own Safe?</h3><p>Then how can the decisions and actions of autonomous AI be controlled safely?</p><p>Nam explains that NC AI is working to build <strong>three safeguards directly into the system</strong> so that even when AI makes decisions and takes action autonomously, those actions remain safe.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*xiqnnH03V8b3Qvh869GdZQ.png" /></figure><p>The first is to <strong>leave a record of the reasoning behind every AI decision</strong>.</p><p>The system records what the AI decided and why, both in natural language and in structured formats, so that any decision can later be traced and explained.</p><p>The second is to <strong>embed rules and permissions into the system so that AI understands the boundaries of its own decision-making</strong>.</p><p>AI must follow organizational rules and access permissions. In areas where rules are missing or conflict with one another, the system should prevent AI from making decisions arbitrarily.</p><blockquote><em>“The condition for trustworthy AI is not that it can ‘do everything well,’ but that it knows exactly where it can be confident — and where its boundaries are.”</em></blockquote><p>The third is to <strong>make sure the system can stop safely and recover when something goes wrong</strong>.</p><p>If a risk is detected during execution, the AI should safely stop the task and either correct actions that have already been performed or automatically restore the system to a previous state.</p><p>Nam summarizes these principles as: <strong>“leave evidence, know the boundaries of your own judgment, and stop safely.”</strong></p><p>Because actions in the physical world are more difficult to reverse, he emphasizes that the same safeguards must be applied even more strictly — first validating actions sufficiently in virtual environments, then gradually expanding the AI’s authority as trust is established.</p><h3>In an Era When AI Works on Its Own, What Kind of Coworker Should Humans Become?</h3><p>Nam describes the future of Agentic AI through a scene on a <strong>Monday morning</strong>.</p><p>Over the weekend, a digital coworker has already identified bottlenecks in pending approvals and classified attendance anomalies according to company policy. When employees arrive at work, the AI provides a morning briefing.</p><blockquote><em>“I handled this, and these three items require your decision.”</em></blockquote><p>At the same time, in a factory, AI may have detected signs of equipment abnormalities overnight, analyzed their causes, and prepared recommended actions for adjusting the production plan.</p><p>Whether in an office or a factory, the final scene is the same.</p><p><strong>AI may do the work overnight, but the first decision of the morning is still made by a person.</strong></p><p>Then what capabilities will people need in order to work alongside AI?</p><p>Nam says people must be able to <strong>accurately interpret the results produced by AI and identify what may be wrong and what risks may be involved.</strong></p><p>Even if people no longer perform every task themselves, they will need a deeper understanding of their own field in order to judge whether the results brought by AI are correct and to make the final decision.</p><blockquote><em>“I believe that being good at using AI ultimately means being good at verifying AI.”</em></blockquote><p>From a mindset perspective, people will also need to accept AI not as a competitor, but as a coworker — while at the same time avoiding blind trust and continuing to question and verify its results.</p><p>And if this is combined with a sense of ownership — the willingness to take responsibility for the outcome — as well as the willingness to keep learning, experimenting, and moving forward without being afraid of failure, we can become good coworkers capable of building this future together.</p><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fwww.youtube.com%2Fembed%2FdbuY50b6Prs%3Ffeature%3Doembed&amp;display_name=YouTube&amp;url=https%3A%2F%2Fwww.youtube.com%2Fwatch%3Fv%3DdbuY50b6Prs&amp;image=https%3A%2F%2Fi.ytimg.com%2Fvi%2FdbuY50b6Prs%2Fhqdefault.jpg&amp;type=text%2Fhtml&amp;schema=youtube" width="854" height="480" frameborder="0" scrolling="no"><a href="https://medium.com/media/bf3ebc01db4883d9ffc16d7bae7b455b/href">https://medium.com/media/bf3ebc01db4883d9ffc16d7bae7b455b/href</a></iframe><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=fbce7e7b051f" width="1" height="1" alt=""><hr><p><a href="https://medium.com/nc-ai/automating-work-from-the-office-to-the-factory-nc-ai-ushers-in-the-era-of-agentic-ai-fbce7e7b051f">Automating Work from the Office to the Factory: NC AI Ushers in the Era of Agentic AI</a> was originally published in <a href="https://medium.com/nc-ai">NC AI</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The Next Stage of AI at LEAP 2026: From Generative AI to Physical AI]]></title>
            <link>https://medium.com/nc-ai/the-next-stage-of-ai-at-leap-2026-from-generative-ai-to-physical-ai-d7d2eec86d5b?source=rss----83564b09dfb3---4</link>
            <guid isPermaLink="false">https://medium.com/p/d7d2eec86d5b</guid>
            <category><![CDATA[leap]]></category>
            <category><![CDATA[physical-ai]]></category>
            <dc:creator><![CDATA[NC AI]]></dc:creator>
            <pubDate>Tue, 15 Sep 2026 10:16:47 GMT</pubDate>
            <atom:updated>2026-09-15T10:16:47.390Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*nS4DUtQ9gvogGl8ASopKOw.png" /></figure><p>NC AI attended LEAP 2026, held in Riyadh, Saudi Arabia, from August 31 to September 3. As part of an initiative to demonstrate industrial AI services running on Korean-made AI semiconductors, we participated in the <strong>Korea Full-Stack AI Pavilion</strong>. Over four days, the event welcomed more than 200,000 visitors and saw over $15 billion in investments and partnerships announced.</p><p>What stood out most was how clearly the exhibition floor showed <strong>AI moving from inside the screen — LLMs and agents — to the world beyond it, through digital twins and robots.</strong></p><p>In this article, we take a look at the global AI trends we observed at LEAP 2026, the changing competitive landscape of Physical AI, and the key considerations for organizations exploring the adoption of digital twins or Physical AI.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/901/1*Stmnmzkb_3tOhfTCXN3Hrw.png" /></figure><h3>LEAP 2026: A Full-Stack View of Global AI</h3><p>LEAP is a major technology exhibition and conference hosted by Saudi Arabia’s Ministry of Communications and Information Technology (MCIT) and other organizations, and this year marked its fifth edition. With Saudi Arabia designating 2026 as the “Year of Artificial Intelligence,” much of the attention was focused on AI infrastructure.</p><p>From hyperscalers such as AWS and Microsoft, semiconductor companies including AMD, NVIDIA, and Qualcomm, and robotics companies such as Unitree and AGIBOT, to national pavilions like the Korea Full-Stack AI Pavilion, the event brought together the entire stack — from infrastructure to robotics — with more than 1,500 exhibitors participating.</p><p>Each hall had a different theme, but walking through them in sequence revealed one clear direction: <strong>AI that had operated on screens is now being deployed into the physical world.</strong></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*t6jQ7V_zrR56b8wwrEFIXA.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/901/1*y6JsVCvKuWA0P9JzL9wvzg.png" /></figure><h3>Three Key Takeaways from LEAP 2026</h3><h3>① AI Is Expanding into the Physical World</h3><p><strong>LLM → Agent → Digital Twin → Robot &amp; Physical System</strong></p><p>This sequence reflects both the progression of technological maturity and the order in which enterprises have adopted AI.</p><p>After using generative AI to work with documents and code, the next question companies are asking is, “Can we apply AI to our equipment and physical operations as well?” The exhibition floor at LEAP 2026 was filled with companies working to answer that question.</p><h3>② Digital Twins Are Evolving from a “Visualization Tool” into an “AI Context Layer”</h3><p>One of the clearest shifts was in the role of digital twins. If conventional digital twins were designed to <strong>visualize reality</strong>, they are now being used as a <strong>“Context Layer that enables AI to understand reality and make decisions.”</strong></p><p>For an agent or robot to make decisions in the field, it needs a machine-readable representation of “what the real world looks like right now.” When equipment locations, sensor values, work sequences, and logistics flows are structured within a single virtual environment, AI can understand the situation, simulate possible choices, and predict outcomes. In this sense, digital twins are moving beyond visualization to become part of the foundation for AI decision-making.</p><p>There is, however, one practical bottleneck: building the twin itself takes time. In older industrial plants, equipment information often remains in the form of 2D drawings, and converting it into 3D assets still largely requires manual work. For companies that already have digital twins, those assets can serve as a starting point for Physical AI. For sites that do not, the question becomes how quickly a digital twin can be built.</p><h3>③ The Competitive Edge in Physical AI Comes from Connecting Data, Simulation, Models, and Control — not Hardware Alone</h3><p>The competitive landscape of Physical AI we observed at the event was clear. More than the robot hardware itself, competitiveness increasingly depends on how effectively companies secure and connect <strong>Robot Data + Simulation/DT + AI Model + Robot Skill</strong>.</p><p>Even when using the same robot, outcomes can differ depending on what real-world data it learns from, what virtual environment it trains in, what model is used, and what task skills it acquires.</p><p>Physical AI is therefore shifting from competition around individual technologies to competition around an integrated structure:</p><p><strong>Digital Twin → Simulation &amp; World Model → VLA &amp; Robot Foundation Model → Robot &amp; Equipment Agent</strong></p><p>In other words, it is becoming a competition across the <strong>Full Stack</strong>.</p><h3>NC AI at the Korea Full-Stack AI Pavilion</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/903/1*Z-C-WzBPDdoFswt8d_Dc_g.png" /></figure><p>At the Korea Full-Stack AI Pavilion, NC AI showcased the Physical AI technology framework of its <strong>industry-specific AI, VAETKI</strong>.</p><p>The framework connects digital twins that recreate virtual environments, <strong>world models that understand the physical world, and robot foundation models that determine robot actions</strong> within a single structure.</p><p>We also presented demonstrations using the 3D generative AI solution <strong>VARCO 3D</strong> and the digital twin platform <strong>VAETKI Twin</strong>. Visitors were able to explore an animated digital twin of a logistics center, urban districts and buildings, and a 3D virtual environment recreating the LEAP 2026 exhibition venue and the Korea Full-Stack AI Pavilion.</p><p>NC AI will continue to validate Physical AI in industrial environments, connecting digital twins, robot learning, and real-world execution within a single end-to-end flow.</p><h3>Considering the Adoption of Physical AI?</h3><figure><a href="https://nc-ai.com/ko/solutions"><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*nGRn6j-SkTnsVawiC17f4g.png" /></a></figure><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=d7d2eec86d5b" width="1" height="1" alt=""><hr><p><a href="https://medium.com/nc-ai/the-next-stage-of-ai-at-leap-2026-from-generative-ai-to-physical-ai-d7d2eec86d5b">The Next Stage of AI at LEAP 2026: From Generative AI to Physical AI</a> was originally published in <a href="https://medium.com/nc-ai">NC AI</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
    </channel>
</rss>