South Korean robotics AI startup RLWRLD has released RealDex1 (internally RLDX-1), a foundation model specifically built for high-degree-of-freedom (high-DOF) robot hands — positioning itself as a hardware-agnostic intelligence layer rather than a robot hardware maker.
The announcement came in a detailed interview with RLWRLD Founder & CEO Junghee Ryu on the NEXT with John Koetsier podcast. Named a World Economic Forum Technology Pioneer for 2026, the company argues that dexterous manipulation remains the single greatest bottleneck and last-mile barrier in industrial automation.
Why High-DOF Hands Are So Hard to Control
For years, industrial grippers and simple two-finger designs dominated factory automation. But as humanoid robots move from labs to production floors, the demand for five-finger hands with 20+ active actuators is exploding — and no existing AI system has been able to control them reliably across different hardware platforms.
Ryu says the team met with over 200 industry leaders across East Asia, and "almost every company said that we need humanlike dexterity." The problem? Existing AI approaches were designed for 1–2 DOF grippers and fall apart when scaling to 22–24 DOF five-finger hands.
RealDex1 Architecture: Multi-Stream Action Transformer
RealDex1 is an upper-body foundation model that takes a fundamentally different approach to hand control:
- Multi-stream action transformer architecture — integrates not just visual data but multiple physics streams simultaneously
- Tactile + force + torque sensing fusion — combines fingertip and palm tactile sensors with torque/force data in a single pipeline
- Motion-awareness modules — tracks object motion in real time during manipulation
- Extended memory capacity — required for the complexity of high-DOF coordination
Unlike companies that develop AI for a single robot platform, RLWRLD says its model works across a wide variety of robotic hands, supporting both tendon-driven designs (like Tesla Optimus and Figure) and direct-drive hands (like Shadow or Wuji).
Hardware-Agnostic Strategy
RLWRLD is currently testing hands from major providers across the US, China, Korea, and Japan. The company co-designed a tendon-driven lower arm and full body with Korean humanoid maker WIRobotics, but is positioning itself primarily as an AI supplier to the broader industry.
This hardware-agnostic approach mirrors NVIDIA's strategy with GR00T and Physical Intelligence's Pi-0 model — but RLWRLD is narrowing its focus exclusively on the hand/dexterity sub-problem, which many in the field consider the hardest part.
Benchmark Collaboration with NVIDIA
Ryu confirmed that RLWRLD is working with NVIDIA to define dexterity measurement standards and establish a benchmark for dexterity testing. This follows a broader industry trend toward standardized benchmarks as manipulation models proliferate.
The company is also involved in the OpenHand initiative, pushing for open standards in robotics — a notable contrast to the closed ecosystems increasingly common in the humanoid space.
Why This Matters
If RealDex1 delivers on its promise of cross-platform, human-level hand dexterity, it could:
- Shorten humanoid time-to-market — hand control is widely cited as the #1 engineering bottleneck
- Reduce integration costs — robot makers could buy dexterity as a service rather than building it in-house
- Accelerate factory deployment — dexterous manipulation unlocks assembly, packaging, and quality inspection tasks currently off-limits to robots
- Create an alternative to vertically integrated players — like Tesla Optimus and Figure, who build both hardware and AI
The bigger question is whether a specialist AI model focused only on hands can outperform generalist VLA models (like GR00T N2 and Pi-0) that handle full-body control. The "narrow vs. general" debate is playing out across every layer of the embodied AI stack — and dexterous hands are the most critical proving ground.
