Meet the Chinese Startup Using AI—and a Small Army of Workers—to Train Robots (www.wired.com)

Shanghai startup AgiBot is deploying two-armed humanoid robots on a Longcheer Technology production line using a hybrid of teleoperation and reinforcement learning. Workers first guide the robot through tasks via teleoperation to seed behavior, then the robot refines skills through what AgiBot calls Real-World Reinforcement Learning. The company says this loop can train a robot on a new, non-delicate assembly step in roughly ten minutes. To generate the real-world training data the algorithms need, AgiBot runs a robotic learning center that pays people to teleoperate robots—echoing a broader industry trend of human-in-the-loop data collection. The approach matters because it tackles a central bottleneck in industrial robotics: teaching machines dexterous, adaptive manipulation that simulations alone struggle to produce. While AgiBot currently targets pick-and-place-style tasks (moving tested components onto a line rather than fine, flexible or fragile assembly), its rapid on-site learning could let robots adapt to shifting production runs and reduce reliance on low-wage labor. Technical caveats remain—reinforcement learning still demands extensive real-world data and isn’t a silver bullet for delicate manipulation—but AgiBot’s blend of human-guided bootstrapping and RL, plus China’s vast manufacturing base and policy support, could accelerate the deployment of more capable factory robots and reshape manufacturing competition globally.