Physical artificial intelligence company Real World said data gathered from actual industrial sites is essential to improving the hand manipulation capabilities of robots. Rather than perfecting technology in a research setting before putting it to work, the company argues that technology should first be applied at industrial sites and then refined using the data accumulated there.
Ryu Jung-hee, chief executive of Real World, said in a keynote speech at the Humanoid Summit Seoul 2026, held recently at COEX in Seoul, that "Real World's core strategy is a structure in which three elements — field deployment, industrial data collection and improved hand manipulation — work together in tandem."
Real World is pursuing this under what it calls a "deployment first" strategy. The approach applies technology at real industrial sites first, accumulating data that is difficult to obtain in a research environment, and uses that data to train robots to handle objects more skillfully. The goal is to create a virtuous cycle in which the improved technology is then deployed at a wider range of sites to gather still more data.
An Ji-yoon, chief strategy officer at Real World, said the company "will use industrial site data to improve manipulation capabilities and extend the technology it develops to a range of robot hardware and work environments." The CSO added that "implementing the dexterity required in real industrial environments and extending it to a variety of sites is important for the practical industrial adoption of physical AI."
To that end, Real World is developing precise hand manipulation technology that allows robots to handle a wide range of objects, along with a robotics foundation model (RFM). A proof-of-concept project recently launched with CJ Logistics at its distribution facilities is another application of the deployment first strategy. The two companies have selected processes for robot deployment and are verifying core movements, and they plan to move step by step toward integrating robot hands with sensors and training models on site data.
