artificial intelligence research institute opena a (switch to another service) has managed to build a robot hand that is capable of independently solving a rubik’s cube. Dactyl-the name of the robot hand movements are not as smooth as people, but haparoinnista in spite of it performing the task that the fifth time drop cubes.

Opena’s robot hand is not the first artificial intelligence to take advantage of the machine, which manages to solve a rubik’s cube. It is also not the fastest or the most powerful of the cube solver. The mit researchers were able last year to build a device that solves a rubik’s cube to 0.38 seconds (switch to another service).

in Less than 0.4 seconds rubik’s cube crucial robot is, however, not usable for any other task. Opena’s Dactyl instead of a capable human hand of the law to learn new tasks.

Opena’s researchers didn’t program the individual hand movements to the 24-joint for robot arm, but the artificial intelligence of the background pulsing of the neural network created the movements themselves. Rubik’s cube solution the necessary transfers to the artificial intelligence didn’t need to learn. They have been developed for already several algorithms.

Hard to jump in the virtual environment to the real world

artificial intelligence applications find it astonishing accomplishments have recently heard a lot. Self-learned ai you have won people so in the game of chess (you move to another service), Gossa (switch to another service) as a computer game in Dota 2 (you move to another service).

All of these applications operate in virtual worlds. The ai does not move pieces on a physical chess board or press keyboard Dota 2 tournament.

the artificial intelligence applications development virtual environment is easier and faster than in the real world. For example, computer games in the teaching process can be speed up to match the thousands of years of practice. The physical device teaching similar to the fast forward is not possible.

dactyl’s training, scientists managed to exploit the simulated environment and the physical hand training. First artificial intelligence practicing the Rubik’s cube handling and resolution in a variety of simulated environments. Within a few months of practice accumulated for thousands of years in front.

researchers created thousands of simulations, which the ai learned to solve the rubik’s cube. Every time, when the algorithm learns to perform the task well, the researchers made it difficult to do.Opena a

When the neural network built by the model was collected enough experience with the simulations, it was moved to control the physical arm. Although artificial intelligence capable of in a virtual environment to solve the rubik’s cube, it took years before it got the robot hand even spin the cube.

the Difficulties were due to the real-world characteristics. Friction, flexibility and dynamics, such as characteristics of the modeling virtual environments is difficult, when the artificial intelligence is not immediately able to take the physical world over.

Now, over two years for the whole process of onset, artificial intelligence succeeds, the task also in the physical world.

self-learning robots to adapt to unexpected situations

Opena’s researchers didn’t program the robot hand solving the rubik’s cube. Rather, they taught it to learn the song processing in different environments.

Traditionally, the robots will be programmed to accurately perform certain task in a certain way. Because of this, they are unable to adapt to unforeseen situations or to perform new types of tasks.

for Example, a lot of attention received the Boston Dynamics robots (switch to another service) is programmed really far, and they are unable to learn on our own new moves, even if the environment it requires.

Dactyl-a robot hand instead is able to adapt to new situations. Researchers, inter alia, taped two fingers together and disturbed it by poking me with a pencil. These do not mess up the ai performance, but it to adjust to shocks.

robotic arm was interrupted, inter alia, the giraffe toy.Opena a

This is just the right way to train the algorithm, i.e. let’s think about surprising events and implemented them, says university of Helsinki’s computer science assistant professor Teemu Roos .

Roos pointed out that self-learning neural networks have been utilized in the past of robotics. Opena’s robot hand is not a huge leap in development, but it is another step in the right direction.

– This is a question of strictly limited issue, namely the rubik’s cube, Roos points out.

rubik’s cube solution, however, gives promise that in the future self-learning robots can actually control our cars and work alongside us. Still Dactyl-like self-learning of robots is not afraid to let hover around to factories or transport, because in the real world, there are still too many surprises for an ai.

Subscribe Yle newsletters!

Proceed to order