Researchers have created a robot hand that can flip a cube into specific positions, using human-like techniques it learned on its own over the course of 100 simulated years of training.
Why it matters: The algorithm behind the feat has a surprising backstory: It previously trained AI agents to play Dota 2, a complex multiplayer video game. Using it again for a very different task is a leap over today’s algorithms, which can generally only do one thing well.
The two challenges — playing a video game and spinning a cube with robot fingers —
are very different, to be sure. But OpenAI found overlapping characteristics, said Jack Clark, OpenAI’s communications and strategy director.
- They both involve a quickly shifting environment. In Dota 2, AI players collaborate to outsmart a team of humans; with the cube, the robot hand uses 20 motors to control its fingers in a complex show of dexterity.
- Independent researchers suggest it's a breakthrough. "I think this is a very exciting result," said Blake Richards, a professor at the University of Toronto and fellow at the Canadian Institute for Advanced Research, who was not involved in OpenAI's research. He tells Axios:
"This is big, because it shows that real-world applicable AI should be possible using only a few small tweaks on already developed systems."
How it works: The algorithm controlling the hand learned how to move cubes around in a simulated environment, and then — in a surprisingly successful leap — transferred its experience to the real world.
- To prepare for the uncertainties of grasping real objects, the researchers introduced random variations to the virtual worlds the hand was trained in, changing parameters like gravity, friction, and the cube's weight and dimensions.
- This method also helped make OpenAI's artificial Dota 2 players good enough to beat amateur humans.
- This weekend, the AI team will take on a team of top Dota 2 players.
The sheer scale of the training regimen was the main reason the algorithm was able to solve these two different tasks, said Jonas Schneider, OpenAI’s head of robotics engineering.
- Huge amounts of compute power allowed the hand to accumulate 100 years of experience in about two days. An even more powerful setup helped the Dota 2 system play 180 years of games daily.
- "We see this trend continuing in the future," Schneider said. "Lots of new companies are building hardware accelerators for deep learning that have the potential of speeding up training by several orders of magnitude within the next four years."
This new robot hand isn’t ready to be deployed to factories: Many of its movements are still awkward, and it lacks the precision that a touch-sensitive hand might possess. Instead, the hand’s performance is hard proof of the algorithm's flexibility.
"It’s hard to know what kind of progress you’re making if you’re just making progress in simulators," Clark told Axios. "But if I make progress on a real robotics task, I've done a no-bullshit reality thing."
Go deeper: A look at the tasks today’s robot hands are capable of (NYT).