This five-fingered robot hand is close to human in functionality

Machine learning algorithms allow it to master new tasks autonomously
May 10, 2016

This five-fingered robot hand developed by University of Washington computer science and engineering researchers can learn how to perform dexterous manipulation — like spinning a tube full of coffee beans — on its own, rather than having humans program its actions. (credit: University of Washington)

A University of Washington team of computer scientists and engineers has built what they say is one of the most highly capable five-fingered robot hands in the world. It can perform dexterous manipulation and learn from its own experience without needing humans to direct it.

Their work is described in a paper to be presented May 17 at the IEEE International Conference on Robotics and Automation.

“Hand manipulation is one of the hardest problems that roboticists have to solve,” said lead author Vikash Kumar, a UW doctoral student in computer science and engineering. “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.”

The UW research team has developed an accurate simulation model that enables a computer to analyze movements in real time. In their latest demonstration, they apply the model to the robot hardware and to real-world tasks like rotating an elongated object.

Autonomous machine learning

With each attempt, the robot hand gets progressively more adept at spinning the tube, thanks to machine learning algorithms that help it model both the basic physics involved and plan which actions it should take to achieve the desired result. (This demonstration begins at 1:47 in the video below.)


University of Washington | ADROIT Manipulation Platform

This autonomous-learning approach developed by the UW Movement Control Laboratory contrasts with robotics demonstrations that require people to program each individual movement of the robot’s hand to complete a single task.

Building a dexterous, five-fingered robot hand poses challenges, both in design and control. The first involved building a mechanical hand with enough speed, strength, responsiveness, and flexibility to mimic basic behaviors of a human hand.

The UW’s dexterous robot hand — which the team built at a cost of roughly $300,000 — uses a Shadow Hand skeleton actuated with a custom pneumatic system and can move faster than a human hand and with 24 degrees of freedom (types of movement). It is too expensive for routine commercial or industrial use, but it allows the researchers to push core technologies and test innovative control strategies.

The team first developed algorithms that allowed a computer to model highly complex five-fingered behaviors and plan movements to achieve different outcomes — like typing on a keyboard or dropping and catching a stick — in simulation. Then they transferred the models to work on the actual five-fingered hand hardware. As the robot hand performs different tasks, the system collects data from various sensors and motion capture cameras and employs machine learning algorithms to continually refine and develop more realistic models.

So far, the team has demonstrated local learning with the hardware system, which means the hand can continue to improve at a discrete task that involves manipulating the same object in roughly the same way. Next steps include beginning to demonstrate global learning, which means the hand could figure out how to manipulate an unfamiliar object or a new scenario it hasn’t encountered before.

The research was funded by the National Science Foundation and the National Institutes of Health.


Abstract of Optimal Control with Learned Local Models: Application to Dexterous Manipulation

We describe a method for learning dexterous manipulation skills with a pneumatically-actuated tendon-driven 24-DoF hand. The method combines iteratively refitted timevarying linear models with trajectory optimization, and can be seen as an instance of model-based reinforcement learning or as adaptive optimal control. Its appeal lies in the ability to handle challenging problems with surprisingly little data. We show that we can achieve sample-efficient learning of tasks that involve intermittent contact dynamics and under-actuation. Furthermore, we can control the hand directly at the level of the pneumatic valves, without the use of a prior model that describes the relationship between valve commands and joint torques. We compare results from learning in simulation and on the physical system. Even though the learned policies are local, they are able to control the system in the face of substantial variability in initial state.