Junghee Ryu, founder and CEO of RLWRLD, talks with John Koetsier on NEXT about AI models for high-DOF robot hands and the data, hardware, and benchmarks needed to deploy them.
Ryu says RLWRLD is building RealDex One, a foundation model for multi-DOF robot hands, and another model with NVIDIA’s Office of Intelligence. The company chose five-finger hands after hearing from more than 200 industry leaders who wanted human-like dexterity. Ryu says those hands can have 22 to 24 active degrees of freedom, which requires a different model design.
Ryu describes a multi-stream action transformer that uses visual and physics information. Inputs can include tactile sensors and actuator impedance signals, which can indicate contact or resistance. RLWRLD is testing different hardware, including tendon-driven hands and direct-drive designs, because their mechanical characteristics differ.
Koetsier asks how two hands working together change the problem. Ryu says RLWRLD uses an end-to-end model trained on teleoperation demonstrations, with a system built around Manus gloves and Vive trackers to capture five-finger movements. He says the model can mimic recorded actions, while basic layers may help prevent collisions.
Ryu says RLWRLD targets factories, logistics, and hospitality, where customers have different needs and may not require a bipedal robot. The model sits above lower-level software such as ROS and generates action signals that can be passed to a robot body. He names KDDI and Lotte among customers, while noting that most customer relationships are under NDA.
Ryu identifies hardware and a lack of standards for dexterity benchmarks and human data pipelines as key challenges. RLWRLD is working with NVIDIA on a benchmark and has created DexBench to test hand designs against workplace skills. The company also plans human and synthetic data pipelines, using workplace video and glove-based tactile capture, and has published hand information and a mini simulator on All Hands Up.
Briefs summarise talks and interviews; timestamps open the video at that moment. Claims in a talk are the speaker's own.
Junghee Ryu of RLWRLD discusses building dexterous humanoid hands and deployment-focused AI at the Humanoids Summit, including work with NVIDIA and industry partners.
