This Remarkably Agile Robot Hand Teaches Itself How to Handle Objects

Gif: YouTube

In a split second before you reach to pick up an object, your brain pre-calculates all the movements needed to safely reach and grasp it securely. It’s a subconscious approach that’s the result of years of childhood development and learning, and one that robotics researchers are now using for their own creations. Festo’s new BionicSoftHand is not only remarkably dextrous, but using AI, it figures out how to properly hold and manipulate an object before it makes any actual movements.

The BionicSoftHand is yet another creation that takes the ‘soft’ approach to robotics. The robots you see moving heavy parts around in a factory are made of steel and pneumatic components that make them strong and fast, but not very forgiving. Were a towering, industrial robot from the likes of FANUC to accidentally make contact with a human while going through its motions, it would result in serious injuries. Soft robots, by comparison, are built using pliable materials like smart fabrics and inflatable bladders. As a result, they have give and compliancy, so if they happen to make contact with a human while working, they won’t cause immediate injury or damage to themselves.

For maximum safety, Festo’s new BionicSoftHand doesn’t have a stiff skeletal structure inside. It instead features a series of inflatable bellows surrounded by a fabric skin knitted from elastic fibers that move and flex along with the hand’s motions as air is pumped in to create movement. It works similar to the muscle and tendon system the human hand uses, but the soft components mean it’s much safer for humans to directly interact with it.

Advertisement
Screenshot: YouTube

The BionicSoftHand’s digits are also equipped with inertial (movement) and force sensors, facilitated through the use of flexible circuit boards and wiring that won’t snap when deformed. These sensors provide feedback to the robot’s control systems about when the hand and digits are moving, but also when they’ve stopped, indicating that contact has been made with an object, or that an object is in a position where it can’t be further moved. They essentially provide a sense of touch, which is important given this robot hand’s other abilities.

Robot arms and manipulators in a factory setting are pre-programmed to endlessly repeat very specific movements. They can occasionally account for tiny amounts of variation in the tasks they’re performing, but mainly, they pick up the exact same object that will always be in a very specific spot in the hand and move it to another pre-defined location.

Photo: Festo
Advertisement

Like humans, the BionicSoftHand instead learns how to grasp and manipulate objects, and perform tasks, all by itself. But it does it through virtual trial-and-error, instead of risking an accident in real life. As a toddler, you probably dropped a cup of milk 100 times before you mastered gripping a cup and taking a sip; but the BionicSoftHand can make those mistakes billions of times using a digital twin powered by AI and machine learning—without ever spilling a single drop of milk in real life.

The robot is given a specific goal, such as re-positioning a 12-sided shape in its hand so that a specific corner points upwards, but it’s never told how to move its digits to manipulate that object. A depth-sensing camera creates a digital duplicate of a real-world object the hand needs to interact with, which allows countless virtual hands to manipulate the copy of the object until a solution is discovered, at which point the real BionicSoftHand takes over. It’s able to learn much quicker than a toddler can, and it’s better at using movements and techniques it’s already figured out on new challenges.

Advertisement

In this first look at the technology Festo has provided, a single BionicSoftHand is shown deftly moving a 12-sided cube around, but it doesn’t take much imagination to envision a pair of these hands quickly learning to solve a Rubik’s Cube: a skill I have yet to master. It costs millions of dollars to design, develop, and program an industrial robot to perform just a single task in a factory, but this self-learning approach means just a single robot could help out with countless tasks around the house.

Share This Story

About the author