[Edaily Reporter Han Kwangbeom ] SamsungElectronics(005930)Samsung Electronics has officially unveiled a large-scale robot commercialization roadmap centered on its newly established RX (Robot eXperience) Business Promotion Office, combining next-generation Robot Foundation Model (RFM) technology with World Model technology. The company plans to prioritize the automation of its more than 100 manufacturing plants by applying its proprietary intelligence architecture to general-purpose robot bodies, with the ultimate goal of making a full-scale entry into the home assistant robot market.

SamsungElectronics unveiled its commercialization strategy for robot foundation models and research achievements in next-generation physical AI during a presentation at the “Samsung AI Forum 2026” held on the 30th at the Samsung headquarters in Seocho-dong, Seoul. Chris Hauser, Head of the Robotics Lab at SamsungElectronics’ RX Business Promotion Office, stated, “The RX Business Promotion Office was established to deeply explore industrial and consumer applications of humanoids and robots,” adding “We will prioritize the automation of Samsung’s more than 100 factories and over 100 production lines, including our investment in Rainbow Robotics, and use this as a springboard to provide home helper robots that assist with household chores and caregiving.”

Lab Director Hauser also highlighted the economic barriers hindering the commercialization of humanoids with specific figures. He explained that with hardware costs (CapEx) of around $100,000 per unit, initial engineering costs (NRE) of approximately $100,000, and annual maintenance costs (OpEx) of about $20,000, humanoids currently qualify as “the most expensive workers on Earth.” He specified, “To narrow the gap with human workers and existing industrial robots, whose annual labor costs range from $4,000 to $40,000, we are simultaneously working to reduce hardware unit costs and cut AI training expenses.”

◇Introduction of a 3-layer “Robot Foundation Model”… Aiming for a 99.99% Success Rate on the Factory Floor

Lab Director Hauser presented the “3-layer Robot Foundation Model (RFM) behavior stack”—comprising △ full-body control (System 0, 100 Hz), △ precision manipulation and VLA (System 1, 10 Hz), and △ high-level agent (System 2, 1 Hz)—as the core technology for robot commercialization. However, he pointed out that robotics requires a “success rate of 99.99% or higher” and millimeter (mm)-level precision on the factory floor, noting that no existing layer has fully addressed these challenges.

He emphasized “customer-oriented robot AI” and “multifunctional AI” as key R&D directions to overcome these challenges. Lab Director Hauser explained, “The approach used by existing startups, where engineers spend months on custom training, has its limitations,” adding, “We must reach a stage of personalization where customers can train robots themselves in just a few hours to enable the deployment of millions of units.” He went on to stress, “Rather than relying on half-baked general-purpose models, we must prioritize multifunctional robot AI that gradually adds verified functions—such as folding laundry and washing dishes—based on a solid performance guarantee, much like smartphone updates.”

In addition, alongside plans to operate a large-scale “Robot Data Factory” and “Real2Sim2Real” technology—which enables robots to learn precise manipulation skills from a single human video—he presented a “hybrid three-layer cognitive architecture” that links “pixels-to-concepts-to-torques” by adding spatial sensing and knowledge inference to the behavior stack. He added, “Rather than robots that perform a variety of tasks but break down when faced with environmental changes, we will first deliver reliable robots that achieve ‘environmental-level versatility’—performing tasks with 99.99% accuracy without variables, even within a narrow scope.”

◇ Overcoming the Challenge of Long-Term Robot Control with ‘World Models’… Equipped with a Commercialization Engine

The next-generation “World Models” research achievements from the Samsung Research Europe AI Center were introduced in detail as the core physical intelligence underpinning these robot foundation models. Timothy Hospidal, Executive Vice President and Head of the Samsung Research Europe AI Center, pointed out, “Existing observation-based models, such as LLMs or VLAs, are trained primarily on passively observed data, which limits their ability to directly intervene in the physical world and perform quantitative physical reasoning regarding cause and effect.”

Drawing on neuroscience, he emphasized, “Just as a mouse standing at a fork in a maze imagines the outcomes of future paths in its brain to choose the optimal route, robots in the physical world must also evolve beyond a simple, reactive ‘System 1’ to ‘System 2’ control, where they anticipate the consequences of various actions and deliberate carefully.”

Director Timothy specifically cited “the accumulation of errors and hallucinations that occur during long-term planning” as the most critical challenge in implementing world models. He explained that controlling a robot in real time requires a recursion depth of over 1,000, which drastically slows down processing speed and leads to the accumulation of errors.

To tackle this head-on, SamsungElectronics announced a new technology called “Metro-WM,” which was first published on arXiv that same day. This technology overcomes the limitations of existing methods—where robots fall into errors by arbitrarily generating virtual states they have not directly experienced—by creating a map using only verified, actual past movement paths and then navigating the shortest distance.

Center Director Timothy explained, “To overcome the vulnerability of existing planners, where hallucinated ‘ghost’ targets are generated in latent space, we imposed a constraint requiring that only states directly observed by the robot in past training sets be designated as subgoals.” He added, “By applying a shortest-path search algorithm based on a network of verified paths, we significantly improved computational efficiency.”

He noted that this approach significantly improved the task success rate by more than 32% compared to existing methods across various robot control and planning tasks, while reducing training costs by up to 56 times and cutting inference and planning times by up to 76%. In fact, the system demonstrated a high success rate in the “Lights Out” game environment and in the physical robot (F3) block-pushing task, which involves complex friction dynamics.

Center Director Timothy stated, “When evaluating the predictions of a world model, we should focus not on visual realism, but on ‘practical usefulness’—that is, its ability to help robots successfully complete tasks in dynamic environments,” and He added, “By continuing to refine the hierarchical structure and conduct research on properly incorporating inductive biases, we will fully integrate a world model—which flexibly switches between System 1 and System 2 control—as the next-generation control engine for humanoids and robots.”