Humanizing Artificial Touch

A lifelike glove that is stiff and cold. It looks human until someone shakes it, then it falls into the valley.

The advice in the 1970s was to stay on the near slope. The 2006 paper looked at what amputees say they want and argued the opposite: pick materials by how close they come to real skin, and cross the valley. The rest of the book works out how to make it.

The soft pulp lets the human pad spread out at the lightest touch. Most of its give is used up before the force reaches 1 N, then it stiffens and holds. Silicone and polyurethane fingertips of the same shape gave much less at first.

Each material was first squeezed and held in a test machine, and the curves were fitted to a model that has both springiness and slow give. With those numbers, the fingertip can be pressed in software as often as needed, and you can see inside it, where a sensor would sit.

At 1 N a human fingertip moved 1.83 mm, about four times as far as silicone and more than six times as far as polyurethane. On a sharp wedge the synthetic skins did fine for the first 2 mm from a wedge, but needed far more force to get there.

The human fingertip soaks up eight to thirteen times more energy per press than either synthetic skin.

Under 1 and 3 mm of skin the three shapes still left different marks on the film. Under 5 mm the marks blurred and the shapes could no longer be told apart.

Fast adapting, small field. Fires when contact starts and stops, and when something slides. The slip sensor in chapter 9 copies it.

Robots usually read touch as a stream of numbers. Nerves send spikes, and only when something changes. A 2013 study turned a fabric sensor array into spike trains like these, and a spike-timing classifier told local curvatures apart with more than 99% accuracy.

A grip has three contacts to keep track of: finger on object, object on the world, and the moment one hands over to the other. Each leaves a short buzz of vibration and a change in how the load is shared. A hand that can feel those moments knows when it is safe to let go.

The 2013 study measured normal force, shear force and acceleration while people passed an object from one hand to the other, to give prosthetic and robot hands a template for a natural handover.

A reflective optical sensor sees an object move across the fingertip. Clear surfaces reflect less, so the signal is weaker, and a fixed threshold set for ordinary objects can miss the slip. A threshold that adapts to the surface catches it, and each spike tells the fingers to squeeze.

The ridged silicone skin sat over four cells, and a support vector machine sorted the presses. It did better with ridges than with a flat skin. The squares show which cells feel the most; their values here are illustrative.

Fingertips feel local detail. The shape of a whole object comes from how the hand closes around it. A sensor glove recorded finger joint angles while people held boxes, cylinders and spheres, and a support vector machine learned the patterns. With one grip per shape it was always right. When people held each shape in more than one way, the patterns overlapped and accuracy fell.

Muscles in the forearm pull long tendons that cross the wrist and run to each finger bone, so a hand can be strong without being bulky. Artificial hands borrow this layout.

Step through the process. At the end the finished arm, socket and liner are shipped to the patient for fitting.

A broken pin far from a clinic means a hand that stays broken. The cost of soft joints is that the hand no longer rests flat open, so the 2019 study set the rest angles to match the relaxed pose of a human hand.

A 2025 case study built a child's hand with silicone joints and servo motors and found it followed the natural hand's movement in both the open and closed positions.

Pull the cable and the joints close one after another, each stopping when its link reaches the ball. With nothing to hold, the finger keeps curling until it reaches its limits.

What a touch feels like depends on the temperature where two surfaces meet, which sits between the two and leans toward the one that holds heat better. The paper measured each silicone's thermal properties and computed this value. Ecoflex at 27 °C gives 28.4 °C, close to a real fingertip.

The last chapter puts the layers back into one hand. A skin is chosen for how it gives, recovers and warms. Sensors sit under it at a depth that keeps detail, and send spikes when contact changes. A skeleton with soft joints and one cable closes the hand.

Back to skin, sensing or structures.