RLWRLD Unveils RLDX-1 Robotics Foundation Model, Surpasses Benchmarks

SEOUL, South Korea — RLWRLD released its robotics foundation model, RLDX-1, on Thursday.

The model uses a "Dexterity-First" architecture intended to provide human-level manipulation capabilities for five-finger robot hands.

Unlike conventional Vision-Language-Action (VLA) models, which process primarily visual and language inputs, RLDX-1 also processes torque, tactile feedback, and working memory within a single model.

This is accomplished through a Multi-Stream Action Transformer (MSAT) structure, which assigns independent streams to different modalities and integrates them using joint attention.

The model uses separate modules to process physical signals and to maintain long-term memory, enabling it to adapt to dynamic environments.

According to the company, RLDX-1 outperformed existing models, including NVIDIA's GR00T and Physical Intelligence's π0, across eight open benchmarks. 

In the RoboCasa Kitchen evaluation, RLDX-1 scored 70.6 points, becoming the first VLA model to exceed 70 points.

On the GR-1 Tabletop benchmark, designed for humanoid robots, RLDX-1 scored 58.7 points, surpassing NVIDIA's GR00T N1.6 by 10.7 percentage points.

The model recorded an 86.7% rating in the LIBERO-Plus evaluation, which tests performance across seven variables, including lighting and camera angles.

In physical hardware testing using WIRobotics' ALLEX humanoid, RLDX-1 completed a "Pot-to-Cup Pouring" task with a 70.8% success rate. RLWRLD said this is approximately double the success rates of comparable models, which it placed in the high-30% range.

RLWRLD released three versions of the model, each with 8.1 billion parameters: a pre-trained checkpoint (RLDX-1-PT) and two mid-training checkpoints (RLDX-1-MT-ALLEX and RLDX-1-MT-DROID).

Model weights, learning code, and technical documentation have been made available to external researchers through GitHub and Hugging Face.

The model operates on a single backbone across multiple hardware platforms, including WIRobotics' ALLEX, Franka Research 3, and OpenArm.

RLWRLD developed the model using NVIDIA infrastructure, including Isaac GR00T, Isaac Lab, Isaac Sim, cuRobo, H100, and A100 GPUs, Jetson AGX Thor, and TensorRT.

RLWRLD said it plans to continue working with cloud infrastructure partners, including AWS and Microsoft.

The company also introduced DexBench, a proprietary benchmark that evaluates five manipulation domains: grasp diversity, spatial precision, temporal precision, contact precision, and context awareness.

RLWRLD said it has secured investments and initiated conceptual verification projects with companies in South Korea and Japan, including SK Telecom, LG Electronics, CJ Logistics, Lotte, KDDI, ANA Holdings, Mitsui Chemicals, and Shimadzu Corporation.

RLWRLD will hold a launch event, "Dexterity Night," in San Francisco on May 13 to announce additional collaborations with humanoid hardware startups from the United States, Japan, and South Korea.

RLWRLD said it subsequently plans to develop a "4D+ World Model" to simulate physical information, such as contact, torque, and robot state, over time.