With all the incredible things that modern technology is capable of, constructing a robotic hand capable of performing a simple act of dexterity such as grasping and using a pencil has remained a challenge – one that experts at the University of Washington may have finally solved.
Vikash Kumar, a doctoral student in computer science and engineering at the university, and his colleagues announced this week that they had successfully built a five-fingered robotic hand that can pivot, bend, and perform other feats of in-hand manipulation commonly limited to humans.
Furthermore, they said that the hand can sense friction and learn from its own experiences, which means that there is no need for humans to direct the unit. Kumar’s team will present their work at the IEEE International Conference on Robotics and Automation on Tuesday, May 17.
Self-learning algorithm eliminates the need for human input
“Hand manipulation is one of the hardest problems that roboticists have to solve,” Kumar said in a statement Monday. “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.” Conversely, his team’s new robotic hand utilizes an accurate simulation model that enables a computer to analyze its movements in real-time.
As a result, the robot hand is capable of multiple dexterous movements, and uses this model to improve its performance at tasks such as rotating long objects. This machine-learning algorithm allows the robot to study the basic physics associated with the task, and to plan which moves it should be performing in order to achieve the desired results, the study authors explained.
“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,” explained senior study author Emo Todorov, an associate professor of computer science, engineering and applied mathematics at UW.
“It’s almost like making an animated film,” added Todorov. “It looks real but there was an army of animators tweaking it. What we are using is a universal approach that enables the robot to learn from its own movements and requires no tweaking from us.” Their approach eliminates the need for every individual step of a task to be pre-programmed by a human controller.
So how did the UW team manage to pull off this feat?
It wasn’t easy, according to a UW press release. In fact, the university said that constructing a dexterous, five-fingered robotic hand posed significant “challenges” in terms of “both… design and control,” including making sure that it was quick, strong and flexible enough to mimic the basic behaviors that a human hand is capable of performing.
Furthermore, the robotic hand required tremendous resources as well. It cost Kumar, Todorov, and their colleagues approximately $30,000 to develop the system, which uses a Shadow Hand skeleton actuated with a custom pneumatic system. The device is capable of moving faster than human hands, but is too expensive for regular use in the commercial or industrial fields.
“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 hand while studying as a postdoc at the University of California, Berkeley. “The approach we took was quite different from a traditional controls approach.”
Thus far, the researchers have successfully demonstrated that the hardware system is capable of learning at the local level, which means that it can get better at tasks that require it to manipulate one object in similar ways. Their next step will involve demonstrating global learning, meaning that the hand could figure out how to manipulate unknown objects without human input.
—–
Image credit: University of Washington
Comments