Insider Brief
- RLWRLD unveiled a robotics foundation model called RLDX-1 designed for dexterous humanoid robot tasks such as grasping, pouring and tool use, with the company positioning the system around contact-rich physical manipulation.
- According to RLWRLD, the model was developed using Nvidia’s robotics and AI stack and demonstrated benchmark gains across tabletop manipulation, kitchen robotics and real-world coffee-pouring tasks while operating across multiple robot platforms including WIRobotics’ ALLEX humanoid and Franka Research 3 systems.
- RLWRLD said the model’s architecture combines vision, language, action, tactile and memory signals to improve robot interaction with dynamic physical environments, while the company expands partnerships with industrial and robotics firms across South Korea and Japan.
PRESS RELEASE — RLWRLD, a physical AI company developing robotics foundation models for dexterous manipulation, unveiled RLDX-1 at “Dexterity Night in SF,” introducing a model designed to help humanoid robots perform contact-rich tasks such as grasping, pouring and tool use. The company also reported benchmark results across humanoid tabletop, kitchen manipulation and real-world coffee-pouring evaluations, and said the model runs across multiple robot embodiments including WIRobotics’ ALLEX humanoid, Franka Research 3 and OpenArm.
At the launch event, Amit Goel, Head of Robotics Ecosystem and Edge AI Product at NVIDIA, took the stage and said: “RLWRLD is one of the core partners in the physical AI ecosystem we are building at NVIDIA.” RLDX-1 was developed on the NVIDIA stack — including NVIDIA Isaac GR00T, NVIDIA Isaac Lab, NVIDIA Isaac Sim and cuRobo for simulation; NVIDIA AI infrastructure with Hopper GPUs for training; and NVIDIA Jetson AGX Thor with NVIDIA TensorRT for inference.
RLWRLD’s debut event brought together senior leaders from NVIDIA alongside top humanoid hardware and AI infrastructure companies — including WIRobotics, Enactic, Origami Robotics and Proception AI — signaling the formation of a new alliance built around dexterous manipulation. RLDX-1 is the world’s first foundation model designed from the ground up with a “Dexterity-First” philosophy, a direct response to what RLWRLD CEO Junghee Ryu calls the real bottleneck of the humanoid era.
“Robot AI so far has been stuck on ‘seeing’ and ‘talking,’” Ryu said during his keynote. “If robots are going to do real work in factories, kitchens and warehouses, they need to grasp, feel and hold on. RLDX-1 was built from day one to fill that gap.”
- Outperforming Isaac GR00T N1.6 and breaking the 70-point barrier: In a technical session led by RLWRLD Chief Scientist Jinwoo Shin, also a professor at KAIST, the company demonstrated RLDX-1 across three high-bar evaluations:
- GR-1 Tabletop (humanoid-specific): RLDX-1 outperformed Isaac GR00T N1.6 by 10.7 percentage points.
- RoboCasa Kitchen: RLDX-1 scored 70.6 — the first Vision-Language-Action (VLA) model in the world to break the 70-point mark on the long-horizon, contact-rich benchmark.
- Coffee-pouring on WIRobotics’ ALLEX humanoid: RLDX-1 achieved a 70.8% success rate — roughly double that of competing models.
The model’s core innovation is a Multi-Stream Action Transformer (MSAT) architecture that gives vision, language, action, tactile and memory signals their own independent streams before fusing them through joint attention — a design Shin says is essential for tasks involving dynamic weight shifts, such as pouring liquid from a pot into a cup.
A live-recorded demo of the “Pot-to-Cup Pouring” task — in which the robot hand sensed the changing weight of the vessel in real time — drew audible applause from the audience.
Humanoid CEOs converge: “The hand is the next inflection point”
The latter half of the evening featured product showcases and a candid panel discussion with humanoid startup CEOs Hiroto Yamamoto of Enactic, Quanting Xie of Origami Robotics and Jay Li of Proception AI.
The panel converged on three themes shaping the next phase of humanoid robotics: the importance of cross-embodiment architectures not locked to a single hardware vendor; the structural moat created by real-world industrial data partnerships; and the emerging global standards race forming around RFMs.
RLDX-1 already runs across multiple embodiments from a single backbone — including the WIRobotics ALLEX humanoid, the Franka Research 3 collaborative robot and the open-source OpenArm platform — a capability the panelists called “a compelling collaboration model from a hardware company’s perspective.”
Toward a 4D+ world model: “Seeing what pixels can’t capture”
Closing the event, Ryu framed RLDX-1 as a starting point rather than a destination.
“The information that isn’t captured in pixels will never appear in your dataset, no matter how much video you collect,” he said. “Today is the starting point of a long roadmap — one we’re walking together with the global humanoid partners in this room — toward a 4D+ world model.”
RLWRLD’s next-generation 4D+ world model goes beyond vision, language and action to jointly predict and generate contact, torque and robot state on a temporal axis — directly simulating physical information that conventional video-based world models cannot capture.
Backed by major global corporates; Japan and Korea events to follow
RLWRLD has raised investment from a roster of leading enterprises across multiple sectors, including SK Telecom, LG Electronics, CJ Logistics, Lotte, KDDI, ANA Holdings, Mitsui Chemicals and Shimadzu Corporation. The company is currently running joint benchmark development, proof-of-concept (PoC) and Robotics Transformation (RX) projects with more than ten large enterprise partners.
Following the U.S. debut, RLWRLD plans to host RLDX-1 launch events in Japan and Korea in the coming weeks.
