Nimble robot manipulates new objects with 99 per cent success rate
Image credit: UC Berkeley
Researchers at the University of California (UC) Berkeley, in the US, have used deep learning to teach a robot to pick up and move unfamiliar objects reliably and with no prior real-world practice.
While tying shoelaces, picking strawberries or cracking an egg comes naturally to humans, machines have struggled to achieve comparable levels of dexterity. Teaching robots to apply a grip has proved a headache to engineers for years.
However, UC Berkeley roboticists have developed a highly dexterous robot, DexNet 2.0, which can do just that.
DexNet 2.0 was taught to grasp new objects by practising with virtual, rather than real objects. This reduces the training time from months to just a day.
Deep learning is a computationally intense process, a complex, multi-layered version of machine learning in which a “neural network” detects patterns by processing human amounts of data in a rough approximation to the working of the human brain.
“We’re producing better results, but without that kind of experimentation,” said Professor Ken Goldberg of UC Berkeley’s automation science and engineering laboratory (AUTOLAB). “We’re very excited about this.”
Professor Goldberg worked with Jeff Mahler, a PhD student, to build a huge database of 6.7 million virtual 3D shapes. These were used as a library from which the neural network learned to grasp and move objects with irregular shapes. The researchers plan to make the data set used to train the neural network public in order to help other roboticists.
Once taught, the neural network was connected to a moving robotic arm and 3D sensor. The robot uses the sensor to capture an image of the object placed in front of it. It then selects the appropriate grasp for lifting and moving the object.
When the DexNet 2.0 was confident of being able to grasp an object, it succeeded in lifting the object and moving it 98 per cent of the time. When the robot was less confident about whether an object was graspable, it would poke the object before deciding on the right grasp. After this, it proved successful in 99 per cent of attempts.
As well as being a better manipulator of new objects, DexNet 2.0 is three times faster than its predecessor, DexNet 1.0. Such reliable dexterity in machines could lead to robots finding new jobs requiring fine motor skills in manufacturing, agriculture, hospitality, service and other sectors.