A robot hand solves a Rubik's Cube in San Francisco on Oct. 11, 2019. Researchers at the OpenAI lab in San Francisco spent months training their robotic hand to solve the cube. Image: Matt Edge/The New York Times
SAN FRANCISCO — Last week, on the third floor of a small building in San Francisco’s Mission District, a woman scrambled the tiles of a Rubik’s Cube and placed it in the palm of a robotic hand.
The hand began to move, gingerly spinning the tiles with its thumb and four long fingers. Each movement was small, slow and unsteady. But soon, the colors started to align. Four minutes later, with one more twist, it unscrambled the last few tiles, and a cheer went up from a long line of researchers watching nearby.
The researchers worked for a prominent artificial intelligence lab, OpenAI, and they had spent several months training their robotic hand for this task.
Though it could be dismissed as an attention-grabbing stunt, the feat was another step forward for robotics research. Many researchers believe it was an indication that they could train machines to perform far more complex tasks. That could lead to robots that can reliably sort through packages in a warehouse or to cars that can make decisions on their own.
“Solving a Rubik’s Cube is not very useful, but it shows how far we can push these techniques,” said Peter Welinder, one of the researchers who worked on the project. “We see this as a path to robots that can handle a wide variety of tasks.”
The project was also a way for OpenAI to promote itself as it seeks to attract the money and the talent needed to push this sort of research forward. The techniques under development at labs like OpenAI are enormously expensive — both in equipment and personnel — and for that reason, eye-catching demonstrations have become a staple of serious AI research.
The trick is separating the flash of the demo from the technological progress — and understanding the limitations of that technology. Though OpenAI’s hand can solve the puzzle in as little as four minutes, it drops the cube eight times out of 10, the researchers said.
“This is an interesting and positive step forward, but it is really important not to exaggerate it,” said Ken Goldberg, a professor at the University of California, Berkeley, who explores similar techniques.
A robot that can solve a Rubik’s Cube is not new. Researchers previously designed machines specifically for the task — devices that look nothing like a hand — and they can solve the puzzle in less than a second. But building devices that work like a human hand is a painstaking process in which engineers spend months laying down rules that define each tiny movement.
The OpenAI project was an achievement of sorts because its researchers did not program each movement into their robotic hand. That might take decades, if not centuries, considering the complexity of a mechanical device with a thumb and four fingers. The lab’s researchers built a computer system that learned to solve the Rubik’s Cube largely on its own.
“What is exciting about this work is that the system learns,” said Jeff Clune, a robotics professor at the University of Wyoming. “It doesn’t memorize one way to solve the problem. It learns.”
Development began with a simulation of both the hand and the cube — a digital re-creation of the hardware on the third floor of OpenAI’s San Francisco headquarters. Inside the simulation, the hand learned to solve the puzzle through extreme trial and error. It spent the equivalent of 10,000 years spinning the tiles up, down, left and right, completing the task over and over again.
The researchers randomly changed the simulation in small but distinct ways. They changed the size of the hand and the color of the tiles and the amount of friction between the tiles. After the training, the hand learned to deal with the unexpected.
When the researchers transferred this computer learning to the physical hand, it could solve the puzzle on its own. Thanks to the randomness introduced in simulation, it could even solve the puzzle when wearing a rubber glove or with two fingers tied together.
At OpenAI and similar labs at Google, the University of Washington and Berkeley, many researchers believe this kind of “machine learning” will help robots master tasks they cannot master today and deal with the randomness of the physical world. Right now, robots cannot reliably sort through a bin of random items moving through a warehouse.
The hope is that will soon be possible. But getting there is expensive.
That is why OpenAI, led by Silicon Valley startup guru Sam Altman, recently signed a billion-dollar deal with Microsoft. And it’s why the lab wanted the world to see a demo of its robotic hand solving a Rubik’s Cube. On Tuesday, the lab released a 50-page research paper describing the science of the project. It also distributed a news release to news outlets across the globe.
“In order to keep their operation going, this is what they have to do,” said Zachary Lipton, a professor in the machine learning group at Carnegie Mellon University in Pittsburgh. “It is their life blood.”
When The New York Times was shown an early version of the news release, we asked to see the hand in action. On the first attempt, the hand dropped the cube after a few minutes of twisting and turning. A researcher placed the cube back into its palm. On the next attempt, it completed the puzzle without a hitch.
©2019 New York Times News Service