[Edaily Reporter Shin Yeong-bin ] In logistics settings, robots handle over 200 types of fast-moving items using five fingers, just like humans. RealWorld is accelerating the development of a next-generation robot artificial intelligence (AI) model for this purpose.

Ryu Joong-hee, CEO of RealWorld, unveiled next-generation precision manipulation technology for applying five-fingered robotic hands to industrial settings, as well as the development roadmap for “RLDX-2,” at the “Humanoid Summit Seoul 2026” held at COEX in Seoul on the 22nd.

CEO Ryu identified logistics sites as a major challenge.

Citing the “ CJ LOGISTICS(000120) ” case, he explained that in actual field operations, conveyor speeds are much faster than in laboratory settings, and the number of item types to be handled exceeds 200. RealWorld is developing technology that uses a five-fingered hand with more than 20 degrees of freedom (DoF) to grasp and manipulate moving objects.

Learning Human Movements and Tactile Sensations in Their Entirety

RealWorld is preparing its next-generation model, the RLDX-2. CEO Ryu explained that while the existing RLDX-1 already demonstrates superior performance compared to open-source models, the company is extensively incorporating new technologies rapidly emerging from academia and industry into the RLDX-2.

The approach differs from conventional methods starting with the training data. Unlike many robotics companies that rely on teleoperation data—generated when a human remotely controls a robot—or first-person video footage, RealWorld has chosen to directly collect data on the actual movements of people working in real-world environments.

This is the so-called “4D+ data pipeline,” which adds temporal information to Saramin’s three-dimensional (3D) movements and incorporates physical signals such as tactile feedback. By using not only first-person cameras attached to the head or body but also external cameras, the system precisely reconstructs how Saramin grasps and moves objects. The company is also developing thin gloves to collect tactile information.

CEO Ryu noted in particular that it is difficult to achieve the level of precision required in industrial settings using only existing Vision-Language-Action (VLA) models. This is because factory work involves not only picking up and moving large parts but also tasks such as picking up small screws and assembling them in precise locations.

He explained that while VLA is strong at planning the entire workflow and generating general-purpose actions, such precision tasks require separate technology. Accordingly, RealWorld is developing technology that reconstructs the 3D shape of an object, identifies the target location, and generates hand movements.

Not Just a Single VLA, but “Multiple Brains”

The “4D+ World Action Model” is another key technological direction for RLDX-2. While conventional world models predict future states based on past information, RealWorld is developing a model that integrates 3D spatial information with physical data—such as tactile feedback—to generate the actual actions a robot should take.

Ultimately, this approach moves beyond a structure where a single massive AI model handles everything. Instead, it simultaneously utilizes the VLA—which plans overall actions like a human—along with precision manipulation models and world action models, with a higher-level AI selecting the model best suited to the situation.

CEO Ryu noted that large language models (LLMs) are currently evolving toward a model where agents select and use various models and tools, and he predicted that robotics foundation models will eventually evolve into a structure where multiple models operate in concert. He explained that these ideas are being incorporated into RLDX-2.

CEO Ryu stated, “Physical AI must not stop at simply creating scientifically superior models,” adding, “We aim to be a ‘field deployment-first’ research institute and are focused on solving problems in real-world work environments.”