Robots still have trouble gripping and manipulating objects. It’s one of the biggest challenges that needs to be solved.
But it appears the University of Washington (UW) has found a way to help robots get a better grip. UW built a five-fingered robot hand that teaches itself how to grasp and manipulate objects on its own, and it gets better with more practice.
The robot hand uses machine learning algorithms to model both the physics involved and to plan its course of action. At the 1:47 mark of the video above, for example, the robot hand gets better at spinning a tube. As the robot hand performs different tasks, the system collects data from various sensors and motion capture cameras, using machine learning algorithms to continually refine and develop more realistic models.
“A lot of robots today have pretty capable arms but the hand is as simple as a suction cup or maybe a claw or a gripper,” said lead author Vikash Kumar, a UW doctoral student in computer science and engineering.
UW’s hand is quite expensive at roughly $300,000, so don’t expect to see it in real-world applications anytime soon. It uses a Shadow Hand skeleton actuated with a custom pneumatic system and can move faster than a human hand. The UW team plans to use the robot hand “to push core technologies and test innovative control strategies.”
UW’s robot hand has 40 tendons, 24 joints and more than 130 sensors. (Credit: University of Washington)
The team emphasized how different its autonomous learning approach is to dexterous manipulation. “Usually people look at a motion and try to determine what exactly needs to happen – the pinky needs to move that way, so we’ll put some rules in and try it and if something doesn’t work, oh the middle finger moved too much and the pen tilted, so we’ll try another rule,” said senior author and lab director Emo Todorov, UW associate professor of computer science and engineering and of applied mathematics.
At this point, UW has tested the robot hand’s ability to improve its manipulation of the same object. The next step will tackle global learning – it’s ability “to manipulate an unfamiliar object or a new scenario it hasn’t encountered before.”
“There are a lot of chaotic things going on and collisions happening when you touch an object with different fingers, which is difficult for control algorithms to deal with,” said co-author Sergey Levine, UW assistant professor of computer science and engineering who worked on the project as a postdoctoral fellow at University of California, Berkeley. “The approach we took was quite different from a traditional controls approach.”