Physical AI company RLWRLD (RLWRLD) said on Wednesday it unveiled its in-house robotics foundation model (Robotics Foundation Model, RFM) 'RLDX-1'.

RLDX-1 consists of 3 variants, including a pre-training checkpoint (RLDX-1-PT) and 2 platform-specific mid-training checkpoints (RLDX-1-MT-ALLEX, RLDX-1-MT-DROID), each with 8.1B parameters. Model weights, training code and technical documents were also made available to external researchers via GitHub and Hugging Face.

The company said RLDX-1 is a Dexterity-First foundation model designed to give a high-degree-of-freedom 5-finger robot hand human-level manipulation capability. It processes not only vision and language but also torque, touch and task memory in a single model, differentiating it from existing general-purpose VLA (Vision-Language-Action) models centered on vision and language.

RLDX-1 outperformed existing models such as NVIDIA (NVIDIA) GR00T (GR00T) and Physical Intelligence (Physical Intelligence) pi0 (π0) on 8 global public benchmarks.

RLWRLD points to a Multi-Stream Action Transformer (Multi-Stream Action Transformer, MSAT) architecture as a core of its technology. Unlike existing VLA models that process different signals such as vision, language, action, touch and memory in a single stream transformer, MSAT assigns an independent stream to each modality and then integrates them through joint attention across modalities.

The company explained that physical signals not captured by vision, such as torque and touch, and long-term memory are also handled in separate modules (Physics Module, Memory Module). It said this is designed to enable a single model to see, feel, remember and adapt.

RLWRLD CTO Jae-kyung Bae (배재경) said, "The core of RLDX-1 is separating the structure so that each modality can be sufficiently represented in its own place." He added, "The ability to precisely capture moments of contact through torque signals and infer dynamic changes over time was an area that existing VLA models were structurally unable to handle."

RLWRLD has attracted investment from large South Korean and Japanese companies including SK Telecom, LG Electronics, CJ Logistics, Lotte, KDDI and ANA Holdings. It is also working with about 10 large South Korean and Japanese companies on joint benchmark development, proof-of-concept projects and RX (Robotics Transformation) projects.

At the 'Dexterity Night' launch event in the United States on May 13, humanoid hardware companies from South Korea, the United States and Japan are set to participate. A panel discussion is also scheduled under the theme of "Why the next inflection point in the robotics industry is the hand."

RLWRLD CEO Joong-hee Ryu (류중희) said, "Information not contained in pixels does not appear no matter how much video you collect." He added, "RLDX-1 is only the first milestone in the direction we are heading. Today is the starting point of a long roadmap toward a 4D+ world model with global humanoid partners, based on data and technology verified at industrial sites in Korea and Japan."

RLWRLD plans to hold RLDX-1 launch events sequentially in Japan and South Korea, starting with this U.S. launch event.