[ad_1]
Researchers sponsored by Elon Musk and investor Sam Altman of Silicon Valley have found a new way to use software to teach new tasks to a real robotic hand, which could reduce the costs of Training robots to work that are easy for humans.
Developers of OpenAI, a group dedicated to the creation of non-profit artificial intelligence, said that they had taught a robotic hand to spin a block of different colors up to The desired side is lifted.
"We are now aiming to conquer more complex tasks," said Lilian Weng, a member of the OpenAI technical team, who worked on the research.
Although it was a simple task, the new fact was that all learning took place in a simulation program and later it was transferred to the physical world with relative ease. Robotic hands have been commercially available for years, but engineers have difficulty programming them.
Source: Courtesy of OPENAI
A major breakthrough was to overcome the "reality gap" between simulations and physical tasks. The OpenAI researchers have injected random noise into the simulation program, making the virtual world of the robotic hand confusing enough not to be overwhelmed by unexpected elements in the real world.
Engineers can write specific computer codes for each new task, which requires a new, expensive program every time. Or they can train machines with software that allows them to "learn" through physical training.
Physical training takes months or years and has its own problems, as if a robotic hand loosens a piece, a human must pick it up and give it to him. Researchers have sought to reduce these times and distribute them to multiple computers for a simulation program that can do training in hours or days, without human help.
"It's the beauty of having many computers working on that," said Ken Goldberg, professor of robotics at the University of Berkeley who was not involved in OpenAI research. "No robot is needed, it only requires a lot of simulation," he added.
Source link
