A ro­botic hand can jug­gle a cube – with lots of train­ing

The Charlotte Observer (Sunday) - - Business - BY RYAN NAKASHIMA As­so­ci­ated Press

How long does it take a ro­botic hand to learn to jug­gle a cube?

About 100 years, give or take.

That’s how much vir­tual com­put­ing time it took re­searchers at OpenAI, the non­profit ar­ti­fi­cial in­tel­li­gence lab funded by Elon Musk and oth­ers, to train its dis­em­bod­ied hand. The team paid Google $3,500 to run its soft­ware on thou­sands of com­put­ers si­mul­ta­ne­ously, crunch­ing the ac­tual time to 48 hours. Af­ter train­ing the ro­bot in a vir­tual en­vi­ron­ment, the team put it to a test in the real world.

The hand, called Dactyl, learned to move it­self, the team of two dozen re­searchers dis­closed this week. Its job is sim­ply to ad­just the cube so that one of its letters – “O,” “P,” “E,” “N,” “A” or “I” – faces up­ward to match a ran­dom se­lec­tion.

Ken Gold­berg, a Univer­sity of Cal­i­for­nia, Berke­ley robotics pro­fes­sor who isn’t af­fil­i­ated with the project, said OpenAI’s achieve­ment is a big deal be­cause it demon­strates how robots trained in a vir­tual en­vi­ron­ment can op­er­ate in the real world. His lab is try­ing some­thing sim­i­lar with a ro­bot called Dex-Net, though its hand is sim­pler and the ob­jects it ma­nip­u­lates are more com­plex.

“The key is the idea that you can make so much progress in sim­u­la­tion,” he said. “This is a plau­si­ble path for­ward, when do­ing phys­i­cal ex­per­i­ments is very hard.”

Dactyl’s real-world fingers are tracked by in­frared dots and cam­eras. In train­ing, ev­ery sim­u­lated move­ment that brought the cube closer to the goal gave Dactyl a small re­ward. Drop­ping the cube caused it to feel a penalty 20 times as big.

The process is called re­in­force­ment learn­ing. The ro­bot soft­ware re­peats the at­tempts mil­lions of times in a sim­u­lated en­vi­ron­ment, try­ing over and over to get the high­est re­ward. OpenAI used roughly the same al­go­rithm it used to beat hu­man play­ers in a video game, “Dota 2.”

In real life, a team of re­searchers worked about a year to get the me­chan­i­cal hand to this point. Why?

For one, the hand in a sim­u­lated en­vi­ron­ment doesn’t un­der­stand fric­tion. So even though its real fingers are rub­bery, Dactyl lacks hu­man un­der­stand­ing about the best grips.

Re­searchers in­jected their sim­u­lated en­vi­ron­ment with changes to grav­ity, hand an­gle and other vari­ables so the soft­ware learns to op­er­ate in a way that is adapt­able. That helped nar­row the gap be­tween real-world re­sults and sim­u­lated ones, which were much bet­ter.

The vari­a­tions helped the hand suc­ceed putting the right let­ter face up more than a dozen times in a row be­fore drop­ping the cube. In sim­u­la­tion, the hand typ­i­cally suc­ceeded 50 times in a row be­fore the test was stopped.

OpenAI’s goal is to de­velop ar­ti­fi­cial gen­eral in­tel­li­gence, or ma­chines that think and learn like hu­mans, in a way that is safe for peo­ple and widely dis­trib­uted.

Musk has warned that if AI sys­tems are de­vel­oped only by for-profit com­pa­nies or pow­er­ful gov­ern­ments, they could one day ex­ceed hu­man smarts and be more dan­ger­ous than nu­clear war with North Korea.

ERIC LOUIS HAINES AP

This ro­botic hand de­vel­oped by OpenAI, called Dactyl, has a sin­gle job, and that’s to ro­tate a cube un­til the let­ter fac­ing up matches a ran­dom se­lec­tion.

Newspapers in English

Newspapers from USA

© PressReader. All rights reserved.