Researchers at the University of California, Berkeley have developed a rapid, AI-driven training system that teaches robots intricate tasks with near-perfect precision. In a mere 1.25 hours, a robotic arm learnt “Jenga whipping,” a dextrous feat requiring pinpoint accuracy. Using reinforcement learning, the new system refines movements through trial and error, human corrections, and real-world feedback.

The robots didn’t stop at Jenga—the team also trained them to assemble motherboards, flip eggs, and manipulate fragile timing belts. Unlike traditional automation, which relies on repetitive tasks, this AI allows machines to adapt to unpredictable situations. The team hopes their open-source approach will democratise AI-powered robotics, making adaptive, high-precision machines as accessible as smartphones.