A Robot That Won’t Even Lift a Finger

The importance of replicating the human hand in robotics lies in its potential to revolutionize a wide range of industries. With robotic hands that can match the dexterity and versatility of their human counterparts, tasks that were previously impossible for machines to perform could become routine. For example, robots with human-like hands could help surgeons perform delicate surgeries, assist with manufacturing tasks that require a high degree of precision, or even aid in disaster relief efforts by navigating complex terrain.

However, recreating the human hand is an incredibly challenging task for roboticists. The human hand is capable of a wide range of movements and is composed of dozens of bones, muscles, and tendons, all working in tandem to produce fine motor control. Designing an artificial hand that can match this level of complexity requires a deep understanding of human anatomy and physiology, as well as advanced engineering techniques.

In an effort to create a simpler robotic hand that still maintains much of the functionality of a human-like hand, researchers at the University of Cambridge have developed a highly energy-efficient soft, passive robot hand that can learn to pick up a wide range of objects using only wrist movements. The learning algorithm is informed only by a sparse array of sensors embedded within the hand’s artificial skin.

The underlying structure of the hand was 3D-printed, from the skeleton to artificial ligaments and tendons. These components give the hand a wide range of motion, and also a measure of stability at each joint. Certain elements, like volar plates, which prevent overextension of joints, were omitted from the design to maintain the team’s goal of keeping the device as simple as possible.

To serve as the skin, a silicone mixture was cast in a mold, and after hardening was fit over the skeleton like a glove. Embedded within the skin were 32 NXP MPXH6300AC6U absolute pressure sensors. The hand was finished off by the addition of a UR5 robot arm mounting plate that allows it to be attached to a standard robot arm. It is this robotic arm that enables the wrist movements, which is the only form of actuation available to the hand.

A recurrent neural network was developed and trained, with the goal being to predict when the hand was likely to drop an object that it was grasping. With that information, the team would be able to adjust the hand’s grip, via wrist movements, to get a better grip before the object actually did fall.

The learning algorithm was trained with the help of small 3D-printed balls. Pressure sensor measurements from 650 trials were fed into the model during training to help it learn what a good grip of an object looks like. Next, the finalized model was tested in a series of experiments to assess how well it generalized to understand when it had a good grip on any arbitrary object.

Throughout the course of these experiments, fourteen different types of objects were tested. It was found that the hand could successfully grip eleven of them, even though the model had not seen any similar objects during training. This suggests that there are good and bad patterns of pressure that apply widely to gripping many types of items. That could mean that a very simple, inexpensive hand might be capable of taking on a wide range of tasks that are now limited to being done by much more complex equipment.

The savings in cost, energy, and computation demonstrated by the researchers in their work is very impressive. However, the researchers are not done with their work yet — they plan to collect a larger, more diverse dataset to train their neural network with the hopes that the hand will be more capable under an even wider range of circumstances. They hope that improvements such as this will help them to make an impact in flexible industrial and logistic robotic applications where the environment often changes, and complex, expensive robotic systems are currently the norm.