In a small factory outside Shanghai, a new kind of worker is learning how to move. Its name is AgiBot, a humanoid robot that can be trained by people in real time to assemble electronics, test parts, and pass them down a production line.
AgiBot’s idea is simple but radical. Instead of teaching robots through endless simulation, the company pairs each machine with a human trainer who guides it through a task for about ten minutes. Then the robot learns to repeat it alone. This process, known as real-world reinforcement learning, blends human intuition with machine precision.
It is already being tested by Longcheer Technology, a major Chinese manufacturer that builds smartphones, VR headsets, and other electronics. The AgiBot system allows robots to take on repetitive but high-volume tasks, such as moving components from testing machines to assembly lines, while still adapting to shifting workflows.
Unlike traditional industrial robots that perform rigid motions, AgiBot’s machines learn through touch, vision, and trial. They are not coded to complete one routine forever. They evolve through repetition, much like human workers do.
Behind the system is Jianlan Luo, a UC Berkeley researcher turned entrepreneur. Luo helped pioneer human-in-the-loop robotics research in California before bringing the idea home to Shanghai. At AgiBot, his team of engineers and teleoperators trains robots for different factories across China, from electronics to consumer goods.
Training robots this way takes a surprising amount of human effort. In AgiBot’s training center, hundreds of operators guide robot arms through tasks, generating data that improves the company’s learning models. It is part of a growing trend in robotics where human labor fuels machine intelligence.
“Robots are not replacing workers,” said Yuheng Feng, an AgiBot representative. “They are learning from them.”
Each robot session creates more adaptable code, faster learning cycles, and smarter machines that can move to new production lines without weeks of reprogramming. For manufacturers, that flexibility is gold.
China’s government has made robotics a core focus in its latest five-year plan, alongside artificial intelligence and automation. The country already operates more industrial robots than the rest of the world combined, giving startups like AgiBot a vast playground for scaling quickly.
Experts say this fusion of human skill and robotic learning could define the next phase of manufacturing. “AgiBot is using some of the most advanced reinforcement learning seen outside a lab,” said Jeff Schneider, a Carnegie Mellon roboticist. “If it works as described, it could reshape how factories operate.”
Across the Pacific, startups in the United States are racing to catch up. Companies like Physical Intelligence and Skild are developing similar models that teach robots to adapt to new shapes, arms, and environments. But China’s scale and production speed may give AgiBot a lasting advantage.
AgiBot’s long-term goal is to create humanoid robots that can walk, handle tools, and work alongside people safely. For now, its focus remains clear: give robots a human touch and let them learn from the people who know the work best.
The quiet revolution is already underway, and it is not happening in a lab. It is happening on a factory floor where humans and machines are learning to build the future together.
Read the original coverage at WIRED.
Explore more Laterstack startup insights:
-
Cluely AI cofounder says startups fail because engineers cannot make viral content
-
Saquon Barkley joins Nvidia in $1.4 billion investment in data center startup Crusoe
[thb_gap height=”150″]