Junghee Ryu of RLWRLD discusses building dexterous humanoid hands and deployment-focused AI at the Humanoids Summit, including work with NVIDIA and industry partners.
Ryu says RLWRLD focuses on humanoid hands because many jobs require human-like dexterity. After meeting more than 200 companies, the team defined 18 atomic tasks based on real work pipelines, spanning industries such as manufacturing, logistics, retail, and hospitality. The public taxonomy, NextBench, is being developed with NVIDIA and partners as a standard for varied real-world tasks.
Ryu identifies compact precision, temporal precision, and contact context awareness as gaps in existing models. RLWRLD combines tactile and other physical signals, tracking of moving objects, and memory of up to 20 seconds. Its multi-stream action transformer keeps physics, motion, and memory information in separate streams.
Ryu describes a three-phase business path: map customer workflows, collect workplace data through pilots and meet customer performance standards, then offer “Labor as a Service” through a monthly AI subscription and hardware lease. Examples include a robot cleaning rooms at Lotte Hotel and data collection and pilots at Lawson stores. RLWRLD is also addressing parcel induction and picking at CJ Logistics, where workers handle more than 200 parcel types.
For pretraining, Ryu says RLWRLD targets human data with physical sensing; for fine-tuning, it combines human and local workplace data. Its 4D pose pipeline captures human motion over time using webcams and four action cameras, with finger gloves planned to capture tactile signals. The team is also developing 3D action policies that retarget human motions to different hand configurations, alongside a 4D pose world action model.
Ryu says RLWRLD is combining physical signals with visual data and using optical flow to adjust robot actions to changing situations. The company expects robotics foundation systems to use an agent layer that selects among multiple independent models and pipelines. Ryu’s closing emphasis is on solving practical problems in real workplaces through deployment.
Briefs summarise talks and interviews; timestamps open the video at that moment. Claims in a talk are the speaker's own.
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.
