�鶹��ý

Artificial intelligence guiding a robotic limb can now solve a . The task requires so much dexterity that even humans find the movements difficult.

The system was developed by researchers at OpenAI, a technology firm that has previously created an AI that could outplay humans at the video game Dota 2.

The team taught an AI to control a commercially available robotic hand developed by the Shadow Robot Company. The AI learned using a technique called reinforcement learning, which involves trial and error. “It starts from not knowing anything about how to move a hand or how a cube would react if you push on the sides or on the faces,” says Peter Welinder, one of the researchers.

The AI scored points when it successfully performed a manoeuvre like flipping the Rubik’s cube around or rotating a face of the cube and was programmed to try to maximise its score. It trained in a simulation for the equivalent of 13,000 years before being tested in the real world.

Teaching the AI to do fine manipulation was challenging because of the number of simultaneous points of contact between an object and the robotic hand, says Lilian Weng, also part of the team.

The AI wasn’t responsible for figuring out how to solve the Rubik’s cube by itself. Visual sensors and a dedicated cube-solving algorithm gave the AI instructions about what moves to make, so that the AI could concentrate on the physical movements needed.

The AI learned to correct for mistakes like accidentally rotating a face of the Rubik’s cube too far.

How long the robot hand took to solve the Rubik’s cube depended on how jumbled the cube was, with the best attempt taking around 3 minutes, says Welinder.

The team hopes the AI could eventually control the hand to do general purpose tasks, such as painting or making origami.