A huge, super-fast indestructible robot hand might seem like a terrifying prop from a science-fiction film.
But this hefty 4.1kg (9.9lbs) hand is very real and is already being used to develop the next generation of AI robots.
Designed by UK-based Shadow Robot Company, this three-fingered claw can go from fully open to closed in just 500 milliseconds.
However, the robot hand is still tough enough to resist being bashed with hammers or pounded by pistons.
That toughness is designed to help the hand survive the rigorous and often destructive process of teaching AI how to interact with the world.
Indestructible machines might seem like a terrifying notion straight from Terminator (pictured), but this sturdy robot fist could be key to helping train future AI
At about 50 per cent larger than a human hand and with only three fingers, this robot isn’t quite as human-like as many others out there.
However, according to Shadow Robot Company director Rich Walker, that chunky design is a necessary trade-off.
Designed over more than four years with AI company Google Deepmind, the hand’s main purpose is to help teach AI.
While robots on factory floors are able to follow simple routines like moving sheets of metal from one place to another, the real world requires more complex movement.
To have robots be able to perform more complex tasks or to act autonomously, they will need AI to power their motor controls.
But the process of teaching AI to control a robot is one that can be particularly expensive and destructive for the robots involved.
The latest robot hand from UK-based Shadow Robot Company has been designed to hold up to the rigorous and often destructive process of training AI
Mr Walker told MailOnline: ‘If you’ve ever seen a baby learning, in the beginning of the process they move around randomly before it starts to connect those movements with what it can see in front of it.’
When you put AI in control of a robot, the process is very similar, the only difference being that robots aren’t quite as resilient as humans.
To learn a simple task like picking up objects, AIs need to make thousands of repeated trial and error attempts.
Last year, researchers from UC Berkeley used this technique to teach an AI how to pick up laundry through repetition.
However, those early stages of the learning process are often so erratic that it is not uncommon for robots to shake themselves apart.
Shadow Robot Company director Rich Waker (pictured) says that training AI leads to rapid random movements that can often shake robots to pieces. This meant they had to design a far bigger more sturdy option
Mr Walker explained: ‘What you and I call touching something, people in the robotics community call collision.
‘Robots typically avoid collisions because collisions are bad, they cause damage.’
To avoid robots simply smashing themselves to pieces as they learn, Shadow Robot Company has had to make its hand extremely tough.
Shadow Robot Company uses a combination of larger more resilient parts and motor controls which allow the fingers to move with impacts and avoid damage.
Each finger is also a modular component so it can be switched out quickly if there is damage.
However, to be useful in machine learning development, the hand also needs to be sensitive.
Each fingertip has hundreds of tiny cameras beneath the surface which focus on the inside of its skin.
This hand weighs 4.5kg (9.9lbs) with each finger coming in a hefty 1.2kg (2.6lbs), this allows it to withstand being pounded with hammers or slammed by pistons
As the hand grasps an object, these cameras detect how the skin is deformed and can work out how hard or soft the object might be.
Mr Walker says that the hand is currently in manufacturing, so there is no current confirmed price.
However, he told MailOnline: ‘This isn’t going to be a cheap robot, but it is going to be cheaper than other robots we’ve made in the past.’
He explains the robotic hand is actually cheaper to manufacture than some others because it is designed with replacement parts and easy repair in mind.
Although the hand was developed for and with DeepMind, Mr Walker says it will also be sold to other companies and research institutions.
He adds that robotics researchers he has spoken to are excited to have ‘robots we can hit now’.
