What does a robot feel when it touches something? Little or nothing until
now. But with the right sensors, actuators and software, robots can be given
the sense of feel—or at least the ability to identify materials by touch.
Researchers at the University
of Southern California’s
Viterbi School of Engineering published a study in Frontiers in Neurorobotics
showing that a specially designed robot can outperform humans in identifying a
wide range of natural materials according to their textures, paving the way for
advancements in prostheses, personal assistive robots, and consumer product
testing.
The robot was equipped with a new type of tactile sensor built to mimic the
human fingertip. It also used a newly designed algorithm to make decisions
about how to explore the outside world by imitating human strategies. Capable
of other human sensations, the sensor can also tell where and in which
direction forces are applied to the fingertip and even the thermal properties
of an object being touched.
Like the human finger, the group’s BioTac sensor has a soft, flexible skin
over a liquid filling. The skin even has fingerprints on its surface, greatly
enhancing its sensitivity to vibration. As the finger slides over a textured
surface, the skin vibrates in characteristic ways. These vibrations are
detected by a hydrophone inside the bone-like core of the finger. The human
finger uses similar vibrations to identify textures, but the BioTac is even
more sensitive.
When humans try to identify an object by touch, they use a wide range of
exploratory movements based on their prior experience with similar objects. A
famous theorem by 18th century mathematician Thomas Bayes describes how
decisions might be made from the information obtained during these movements.
Until now, however, there was no way to decide which exploratory movement to
make next. The article, authored by Professor of Biomedical Engineering Gerald
Loeb and recently graduated doctoral student Jeremy Fishel, describes their new
theorem for this general problem as “Bayesian Exploration.”
Built by Fishel, the specialized robot was trained on 117 common materials
gathered from fabric, stationery, and hardware stores. When confronted with one
material at random, the robot could correctly identify the material 95% of the
time, after intelligently selecting and making an average of five exploratory
movements. It was only rarely confused by a pair of similar textures that human
subjects making their own exploratory movements could not distinguish at all.
So, is touch another task that humans will outsource to robots? Fishel and
Loeb point out that while their robot is very good at identifying which
textures are similar to each other, it has no way to tell what textures people
will prefer. Instead, they say this robot touch technology could be used in
human prostheses or to assist companies who employ experts to judge the feel of
consumer products and even human skin.