Researchers from University of California, Berkeley have made a major advance in the artificial intelligence field. They claim that they now have a robot that can learn new things and perform new tasks just like humans do – through trial and error.
Though robots can store a lot of information inside their “brains,” it is extremely hard to make them perform simple tasks such as opening a water bottle or folding a napkin. But UC Berkeley researchers say that they have developed some algorithms that may help robots learn new things through practice, just like we do.
Usually, robots are programmed beforehand with a series of limited scenarios that they later repeat without chances of improvement. This method is suitable in well-controlled environments such as a laboratory, but fails miserably out in the real world.
UC scientists struggled for years to find a way of improving robots so that they can quickly adapt to new life scenario. If all goes well, the new advance may become a critical step toward integrating intelligent machines into our daily lives. Rosie, the robotic maid from the 1960s’ animated television series The Jetsons, may soon become reality.
The learning algorithms that would help robots acquire new knowledge through trial and error were dubbed “deep learning,” which is a relatively new branch of artificial intelligence. Deep learning scientists usually try to copy the way human brains perceive and interact with the surrounding environment.
Sergey Levine, the lead researcher in Berkley’s “People and Robotics Initiative,” underscored how humans learn new things. Toddlers do not need a pre-programmed software in their brains to perform new tasks. Instead, humans acquire new skills and learn to perform new tasks in a life-long process based on experience and interaction with other humans, Levine said.
“This learning process is so deeply rooted in our nervous system, that we cannot even communicate to another person precisely how the resulting skill should be executed,”
the researchers explained. They also said that we only need guidance and pointers, but the rest of the process is learning on our own.
Nevertheless, the AI domain have recently experienced a series of game-changing advances including facial and voice recognition, and attempts to program a robot to conduct an intelligent and interactive conversation with a human.
But the UC Berkeley researchers went way beyond simple passive tasks of voice, image, and text recognition. They trained the “Berkeley Robot for the Elimination of Tedious Tasks” (BRETT) to perform simple tasks such as building a LEGO wall on a new algorithm that evaluates and rewards the machine every time a move is performed. The closer to completing a task the move is, the higher the score on an evaluation scale gets.
The Berkeley robot can perform new tasks within ten minutes. But if it needs to find new objects or learn new moves the process may take several hours. Researchers hope that robots would be able to learn complex tasks on their own such as cleaning our house in less than a decade.
Image Source: The Wire