We all just wondered the possibilities of robot company Boston Dynamics. However, recent developments of the nonprofit organization OpenAI, which occupies research in the field of artificial intelligence technology, even against Robo-dogs Boston Dynamics looks like a completely different, more advanced level. OpenAI presented Dactyl – s system, trained robotic arm. It would seem that what is there is impressive? Explain. Dactyl – is a system that allows the robotic arm to manipulate physical objects at a level previously unattainable any AI system. In the future the system will be able to give the robot physical parameter of dexterity.
About the full level of dexterity of a robotic body, speech does not go yet, but engineers OpenAI finally laid the Foundation.
Sleight of hand and no fraud
Perhaps one of the finest instruments given to us by mother nature are hands. Human hands, nor any other, including the limbs of the most advanced apes, are capable of so finely controlled with your fingers. For robotic engineers are struggling to adapt with the dexterity of human hands in their iron protégé, but some obvious and important progress in this direction was not long ago.
Developed OpenAI system Dactyl is designed for, it would seem that the simplest tasks which even a child will cope. Is it difficult for us to pick up, say, a cube marked on each of its sides with drawings and turn it over to the right side? For us humans, this task is not difficult. For robot is a real torture. But thanks to powerful computers, the researchers were able to teach it that trick in just 50 hours.
It’s all about the practice
Training system Dactyl took place in a simulated environment. To do this, engineers created a digital copy hands and put it in the computer environment, working on the principle of randomization. Engineers were asked a specific set of parameters for the environment (for example, the size of the cube, gravity), and then accidentally change these variables. For faster tutoring system researchers have created several of these virtual hands. Teaching Dactyl to adapt to a wide variety of possible scenarios in a virtual environment, the researchers gave the AI system experience required to adapt to the different scenarios of tasks in the real world.
After 50 hours of training inside among the virtual AI system learned to control the real robotic arm and performed 50 tasks set for it: to rotate a cube on one side or the other, not dropping it and completed the task in 80 seconds to solve a single task. In order to enable the system to “understand” which way to rotate the cube, it uses a set of cameras.
On its official website, the developers Dactyl explain that as the basis for a new system they applied the algorithm OpenAI Five, previously developed to create a team of 5 neural networks, capable of team play in DOTA 2. This digital brethren we will be able to see on the main cybercorner this year — the 2018 International held annually by the company Valve.
The effectiveness of Dactyl proves the possibility of creating a universal algorithm that can train an AI system on the performance of multiple tasks. The key benefit of this possibility is that in the future it will simplify and speed up the process of learning artificial intelligence systems, since developers do not have to create and train a new AI for solving various new problems.
OpenAI specialists are trained AI system dexterity of human hands
Nikolai Khizhnyak