[Edaily Reporter Shin Yeong-bin ] The phase where people cheered at the sight of robots walking like humans is rapidly fading. Attention is now shifting to what those robots can actually do in the field and how long they can work reliably.

This trend was clearly evident at the “Humanoid Summit Seoul 2026,” held at COEX in Seoul on the 22nd and 23rd. Companies presenting at the event identified the robot’s “brain” and “hands,” real-world data, and reliability in the field as key factors for commercialization.

The Importance of the ‘AI Brain’ Over the Robot Body Is Gaining Prominence

Ali Agha, CEO of FieldAI, emphasized a “universal brain” that can be applied across multiple robots

rather than

specific hardware. He said, “If a humanoid is the right solution, we’ll use a humanoid; if a quadruped robot is the right fit, we’ll use that.”

FieldAI’s AI brain has been implemented in over 30 different robot platforms. It aims to operate without maps or GPS and move without predetermined paths; the company recently revealed a case where a large vehicle traveled over 35 km without human intervention.

Competition in the field of robotic hands is also fierce. Ryu Joong-hee, CEO of RealWorld, cited CJ LOGISTICS’ distribution center as an example, explaining that in real-world settings, there are over 200 types of items to handle, and the movement speed is much faster than in experimental environments. RealWorld is developing technology to handle these items using a five-fingered robotic hand.

RealWorld’s next-generation Robotics Foundation Model (RFM), “RLDX-2,” utilizes “4D+” data, which trains the model on a combination of physical signals—including human 3-dimensional (3D) movement, time, and tactile feedback. Precision tasks, such as picking up and tightening small screws, have limitations when relying solely on existing Vision-Language-Action (VLA) models; therefore, the company is also developing separate technology to identify an object’s 3D shape and determine where and how to grasp it.

ROBOTIS has adopted a platform strategy that facilitates the easy integration of various AI models. CEO Kim Byung-soo revealed that the company has tested approximately 20 types of RFMs on its proprietary software, “CYCLO.” CEO Kim predicted that humanoids would evolve in a similar way to early smartphones, which initially underperformed compared to dedicated cameras or music players but became universal platforms by incorporating a wide range of apps.

In industrial settings, how naturally a robot integrates into the existing work environment is also crucial. Michael Perry, Head of Commercial Strategy at Persona AI, emphasized, “Welders don’t just weld,” noting that humanoids must also perform a series of tasks—such as adjusting the position of steel plates, managing cables, and verifying work results—in addition to welding.

Perry added, “The reason humanoids are superior is not because the human form is perfect for all robots, but because they are a suitable solution for solving specific industrial problems.”

From the Exhibition Hall to the Factory… The Race for Commercial Deployment Numbers

This shift is also reflected in the numbers within the global humanoid industry.

A humanoid robot from U.S.-based Figure AI operated for over 1,250 hours over 11 months at the BMW Spartanburg plant, handling more than 90,000 parts and participating in the production of over 30,000 vehicles. This year, the company deployed a next-generation model, expanding its scope to include complex tasks such as handling parts while pulling a cart.

Agility Robotics’ “Digit” has also been operated for a cumulative total of over 65,000 hours in logistics settings. At the GXO logistics center, it moved more than 100,000 totes and achieved an operational accuracy rate of approximately 98%. Building on this success, the company announced plans this month to unveil its next-generation robot and move toward large-scale commercial deployment.

Aptronic is also deploying its humanoid “Apollo 2” at actual customer sites and its own “Robot Park” to collect large-scale data on manufacturing, logistics, and distribution operations. This data is being used in collaboration with Google DeepMind to train next-generation robot AI models.

Tesla is taking it a step further by preparing for mass production. It has designated the third-generation Optimus, unveiled this year, as its first mass-production design and aims to begin production by the end of the year. In Fremont, the company is preparing a first-generation production line with an annual capacity of 1 million units.

The industry’s competitive benchmarks are naturally shifting as well. Whereas hardware performance—such as walking and balance control—was the focus in the past, key challenges now revolve around how quickly robots can learn new tasks and whether they can achieve high enough utilization rates to recoup the actual investment costs.

Ryu Joong-hee, CEO of RealWorld, stated, “Physical AI should not stop at simply creating scientifically superior models,” adding, “It must focus on solving problems in actual work environments.”