AI researchers have demonstrated a self-learning algorithm that gives a robotic hand a remarkable new dexterity. Her creation taught herself how to manipulate a cube with uncanny skill by practicing in a computer simulation for an equivalent of a hundred years (though only a few days in real time).
The robot hand is far from being as mobile as a human, and far too clumsy to be used in a factory or warehouse. Nevertheless, research shows the potential for machine learning to unlock new robot capabilities. It is also suggested that robots could acquire new skills in virtual worlds, which could significantly speed up the programming or training process.
The robot system called Dactyl was developed by researchers of the nonprofit organization OpenAI in Silicon Valley. It uses a commercial robot hand from a British company called Shadow, a common camera and algorithm that already masters a sprawling multiplayer video game called DotA and uses the same self-learning approach (see "A Team of AI Algorithms") crushing people in a complex computer game "
The algorithm uses a machine learning technique known as" reinforcement learning. "Dactyl was tasked with maneuvering a cube to turn another face up, trying to find out which movements were moving
Dactyl videos show how the cube is turned with impressive agility, automatically finding out several handles that people use frequently, but research also showed how far AI still has to go: the robot could the cube after his hundred years of virtual trai ningszeit successfully manipulate 1
"It will not fit into an industrial workflow anytime soon," says Rodney Brooks, emeritus professor at MIT and founder of Rethink Robotics, a startup that makes smarter industrial robots. "But that's fine – research is a good thing."
Reinforcement learning is inspired by the way animals seem to learn through positive feedback. It was first proposed decades ago but, thanks to advances in artificial neural networks, has only come in handy in recent years (see "10 Groundbreaking Technologies 2017: Reinforcement Learning"). Alpha daughter DeepMind used reinforcements to develop AlphaGo, a computer program that appropriated itself to the devilishly complex and subtle board game Go with superhuman abilities.
Other robotics researchers have tested the approach for a while, but have been paralyzed by the difficulty of mimicking the complexity and unpredictability of the real world. The OpenAI researchers avoid this by introducing random variations in their virtual world so that the robot can learn to consider disturbances such as friction, noise in the robot hardware, and moments where the cube is partially hidden.
Alex Ray, one of the engineers behind the robot, says Dactyl could be improved by giving it more processing power and introducing more randomization. "I do not think we've reached the limit," he says. Ray adds that there is no plan to commercialize the technology. His team focuses exclusively on the development of the most powerful generalized learning methods.
"That's hard to do," says Dmitry Berenson, a roboticist at the University of Michigan who specializes in machine manipulation. Berenson says it's not clear how far the latest machine learning approaches will bring us. "There is a lot of human effort to find the right network for a specific task," he says. But he believes that simulated learning could be very useful: "If we can reliably cross the & # 39; reality gap, it will make learning easier exponentially."