The smart, agile hand from British firm Shadow Robot Company looks like a prop from a science fiction movie. It is also super fast, extremely strong, and quite large – twice the size of a human hand. It cannot be damaged by either a press or a hammer.
That strength is needed for the hand to survive the rigorous, often destructive process of training artificial intelligence — learning through interaction with the world.
This robotic hand has only three fingers. Unlike many other robotic hands the company has made, it is less similar to a human hand.
However, Rich Walker, director of Shadow Robot Company, says the bulkiness of the design is a necessary compromise.
What is this three-pronged hand for?
Roboticists have been developing the arm for more than four years in collaboration with Google DeepMind, a company that researches and builds safe artificial intelligence systems. The primary purpose of the arm is to train AI.
While robots in factory workshops perform simple tasks (such as moving sheets of metal from one place to another), the real world demands more complex devices.
Robots need AI to handle more complex tasks or act autonomously. But training AI can be costly and damaging to the robot.
Walker told the Daily Mail, “If you’ve ever seen a child learning to walk, at the beginning of the process, they move around clumsily before they start to connect their movements with what they see in front of them.”
If a robot is continuously training AI, the process is similar, but robots aren’t as resilient as humans. To solve even a simple task like picking up objects, AI has to go through thousands of trial-and-error attempts. The early stages of learning are often so chaotic that robots get damaged.
Last year, researchers from the University of California, Berkeley used robots to train AI by having them repeatedly collect laundry.
Super strength is a key requirement for the hand.
Walker explained that what people usually call touch is called a collision in the robotics community. “Robots typically avoid collisions because collisions are bad; they cause damage,” he said.
To prevent collisions that could crush the robot during training, the company made its new arm extremely rigid and durable. Engineers used a combination of heavier, more resilient components and motor-control elements that let the fingers move during impacts to avoid damage. Each finger of the robotic arm weighs about 1.2 kilograms.
Each finger is modular, so if it gets damaged it can be quickly disabled.
Sensitivity is another property.
But to be useful for machine learning, the robotic hand also needs to be sensitive. Beneath each fingertip’s surface are hundreds of tiny cameras that look at the inner side of the artificial skin.
When the hand grabs an object, these cameras capture how the skin deforms and can determine how hard or soft the object is.
“Three-degrees-of-freedom sensors on the middle phalanges provide additional information during grasping and manipulation,” the Shadow Robot Company website says.
Walker said the hand is currently in production, so it does not yet have a confirmed price. He added, “This will not be a cheap robot, but it will be cheaper than other robots we have created in the past.”
The lower price reflects that the robotic arm is cheaper to produce than some others, since it was designed with spare parts and ease of repair in mind.
Although the hand was initially developed for Google DeepMind, Rich Walker says other companies and research institutions will want the robot. He added that robotics researchers he has spoken to are excited about the emergence of new robots “that can be used right now.”
