Physical AI customers are no longer only waiting for finished robots
Humanoid robotics is usually narrated from the supplier side: who has the better hand, the stronger foundation model, the smoother gait or the larger production plan. Recent Korean cases look different when viewed from the industrial customer side.
HD Hyundai is co-designing hardware around shipbuilding work, Samsung is turning affiliate workflows into validation cases with robotics companies, and CJ Logistics chose the first task for humanoids inside a live distribution center. These companies are moving from “we will buy a good robot when it exists” toward “this is the work, this is what success looks like, and this is where to start.”
Banseog calls this Task Specification Ownership. The industrial company often knows the task sequence, exceptions, quality boundary and failure cost more deeply than the robot supplier does.
At HD Hyundai, the task is shaping the robot hardware itself
HD Hyundai Robotics invested KRW 13 billion in AIDIN Robotics and plans to co-develop a five-finger robotic hand for shipbuilding and heavy-industry environments. Force-torque and tactile sensing are central because the system must understand contact and force distribution while controlling the hand precisely.
The target is not “manufacturing” in the abstract. The collaboration points toward surface-processing work such as polishing and grinding. Workpiece geometry, contact pressure, finish quality, tool wear and safety requirements can therefore flow backward into hand, sensor and control requirements.
In this case, industrial process knowledge is not merely a software configuration after the product is finished. It helps shape the product requirement itself.
Samsung is turning factory work into problems AI can train and evaluate against
At RX ART 2026, Samsung SDS described validation work around component handling and consumable replacement at Samsung Electro-Mechanics and SEMES, rough outdoor work for Samsung Heavy Industries and Samsung E&A, and dexterous manipulation for Samsung Display.
The notable point is the use of real industrial problems to secure data needed for AI learning and to validate applicability, rather than treating the event as a generic robot demonstration. Industrial work becomes part of the training and evaluation curriculum.
A general robot model still needs someone to define what counts as completion, which failures are unacceptable, what force or precision is sufficient, and what cycle time fits the line. The customer’s process knowledge can therefore become part of the training-and-evaluation infrastructure.
CJ Logistics owns the first task — and the order of the next tasks
CJ Logistics deployed two dual-arm humanoids in a live Olive Young logistics center in Yongin. Their first task is inserting cushioning material into boxes. CJ says it plans to expand step by step into picking, sorting, inspection and packing.
The point is not that cushioning insertion is technologically spectacular. It is that the operator chooses one bounded task, tests whether it survives live operations, then decides which adjacent task comes next.
A logistics center is therefore more than a place where a robot is used. It is a training and evaluation environment containing product shapes, box formats, conveyor speeds, exception paths, work sequence and quality rules.
The next advantage may come from owning a good task curriculum, not only a good robot
Agility’s public economics ask whether a particular workflow can clear an ROI gate. These Korean cases add a different question: who defines that workflow in the first place? A robotics supplier cannot realistically own every industry’s process rules and exception knowledge by itself.
Shipbuilding, electronics and logistics companies have accumulated quality standards, failure patterns, cycle-time constraints, safety rules, equipment limitations and operating data over decades. Those assets can become a task curriculum: what the robot should learn, in what order, and against which evaluation criteria.
If robot companies own models and hardware while industrial companies own tasks, environments and evaluation data, the relationship becomes more reciprocal. The “customer” is also a partner in creating capability.
The talent boundary expands from people who build robots to people who can define work for robots
Automating grinding requires more than control expertise. Someone must know what a good surface looks like, how pressure and vibration affect quality, when a tool is worn, and what safety limits matter. Logistics requires similar domain depth around picking, packing, WMS, equipment operations and process improvement.
That changes the talent question. Instead of asking only “has this person worked in robotics?”, employers can ask “can this person define the work to be automated, explain the exceptions and turn success into measurable criteria?”
Responsibilities scattered across application engineering, automation, production engineering, process engineering and quality may combine into new roles that translate between robots and industrial work. As in many emerging markets, the work can appear before the title stabilizes.
BANSEOG VIEW
In Physical AI, the customer can become a co-author of the robot’s work curriculum
The three cases are at different maturity levels: HD Hyundai is in co-development and planned validation/commercialization, Samsung SDS is validating real industrial use cases, and CJ Logistics has robots in a live operating process.
What they share is direct industrial involvement in choosing tasks and shaping hardware, data, success criteria and expansion order. The Physical AI customer can move from buyer to task-spec co-designer.
For talent, this raises the value of domain experts who can translate process quality, safety, failure modes and cycle time into trainable and measurable robot requirements — even when they did not begin their careers in robotics.
SOURCES
Primary sources and references
- HD Hyundai Robotics · KRW 13B investment in AIDIN Robotics
Published Sep. 14, 2026. HD Hyundai said it invested KRW 13 billion on Sep. 11 and will co-develop a five-finger hand for shipbuilding/heavy industry, expanding toward grinding and polishing automation.
- CJ Logistics · AI humanoids enter live logistics operations
Sep. 3, 2026. Two dual-arm humanoids were deployed in a live Olive Young logistics center, starting with cushioning-material insertion and planned expansion to picking, sorting, inspection and packing.
- Samsung SDS · RX ART 2026
Sep. 3, 2026. Samsung affiliates validated component handling, consumable replacement, rough outdoor work and dexterous manipulation with Physical AI companies, using real tasks to secure AI-training data and refine deployment plans.
The three cases are not at the same commercialization stage. “Task Specification Ownership” and “co-author of the work curriculum” are Banseog analytical frames connecting the public evidence. They do not imply that all three companies directly train AI models in the same way.