Shadow Hand withstands the rigours AI research

A company known for making highly anthropomorphic robot hands has developed Shadow Hand, a dextrous three-fingered robotic hand made to withstand the rigours of real-world machine learning research.

 The New Shadow Hand is a hardware platform for dexterous manipulation research, delivering dynamic and controlled motion in a robust and reliable package, enabling long-running experiments without interruptions due to hardware failure
The New Shadow Hand is a hardware platform for dexterous manipulation research, delivering dynamic and controlled motion in a robust and reliable package, enabling long-running experiments without interruptions due to hardware failure - Shadow Robot Company

Around five years in the making, the new Shadow Hand was built by Shadow Robot Company for Google DeepMind.

“They found that all robot hardware used for machine learning was not really up to the job,” Rich Walker, director of London-headquartered Shadow Robot Company, told a press briefing.

Walker added that robots developed for AI research can literally fall apart when used beyond their design parameters, fail to acquire useful data because of the way they are designed for control and operation, cannot operate for long periods without needing a restart or reset, or are not built with the assumption that interaction and collision are normal for it, which leads to damage.

“And lastly, almost all the robots that they've looked at, even the ones that were pretty robust and pretty reliable, didn't have the kind of dexterity and manipulation capabilities that DeepMind really cared about,” said Walker.

Known for Dextrous Hand, Shadow Robot’s five-digit flagship product, the company’s new solution ‘is a lot chunkier’ with each easily swappable finger weighing 1.2kg and the total unit weighing 4.1kg.

The hand measures 350mm in length, and 165mm x 160mm in width and height. According to Walker, it behaves much more like an optimised manipulator rather than a human-like hand. It can also withstand being struck by a hammer.

“We think this is a kind of a new level of robustness and performance in mobile hardware,” said Walker.

The kinematics of each finger - containing 155 individual sensor channels, plus video from the distal tactile sensor - are similar to a human finger, with an ad-abduction joint at the base, and three flex/extend joints along its length.

The four joints of each finger are driven by five motors housed in the base of the finger in an N+1 configuration.

“The reason for this is that if you simply drive each joint directly when you change direction, there's a point where you have no force on the tendon and therefore you have backflash, a point where the joint doesn't move,” said Walker. “If you have an N+1 configuration you can maintain tension in all the tendons all the time."

Each motor unit executes a 10kHz force control loop, giving active compliance to every joint.

Shadow Hand goes from fully open to closed in 500 milliseconds and can give a 10N fingertip pinch. Each finger contains dozens of 3-DOF taxels (sensors that recognise the pressure of contact with a physical object) on the middle and proximal phalanges and hundreds on the fingertip, with ‘a significant dynamic range’ of 0.01N to 80N.

Shadow Hand, which is available for purchase, will be demonstrated to the public for the first time at ICRA in Japan, which runs from May 13-17 2024.