TOKYO - Japan’s government is now weighing plans to develop humanoid machines capable of such roles by 2030, as breakthroughs in artificial intelligence make it possible for robots to undergo years of accelerated learning in virtual environments, mastering movement, reflexes, and adaptability that allow them to operate in unpredictable real-world conditions without relying on cameras or sensors.
According to Chiba Institute of Technology’s Future Robotics Technology Center director Takayuki Furuta, researchers have trained a Chinese-made quadruped robot weighing around 16 kilograms that can always get back on its feet no matter how it falls. Although the base machine is a low-cost mass-produced model with limited functions, once fitted with the center’s AI, it is able to move freely under simple commands such as “forward” or “to the side.” The robot has no camera or tactile sensors, yet can navigate stairs and other obstacles by drawing on “muscle memory” developed in virtual simulations.
The training method involved creating about 4,000 digital copies of the robot and running them in a fast-forwarded virtual world. Years’ worth of simulated walking, climbing, tripping, and falling were condensed into just four to five hours. The results were consolidated into a single physical unit, enabling it to automatically react to situations such as being pushed, dropped, or rolling down stairs. Over time, the robot even developed protective reflexes, similar to an animal’s ability to land safely.
Furuta likened the process to abandoning a cat in the wild and finding, years later, that it had learned survival skills never imagined. Unlike traditional “robot walks” engineered through equations, this model acquired movements spontaneously, from rolling recovery to turning motions, through countless trial-and-error cycles in simulation.
Asked whether such robots might one day attack people, Furuta dismissed the idea. He explained that the AI installed in this prototype only governs movement and adaptability to changing environments, not higher-level decision-making. Aggressive behavior would stem from a “cerebrum” layer of intelligence, which must be carefully designed to prioritize safety. “A robot that meets a good engineer will become a good robot,” he said, stressing that education and programming determine whether advanced machines will serve people beneficially.
While acknowledging concerns about potential misuse, Furuta maintained that humanoid robots trained in virtual worlds represent a step toward real-world adaptability. By 2030, he suggested, such robots may not only work alongside humans in shops and factories but even take the role of customers themselves.
Source: TBS
