Systems powered by machine learning routinely incorporate visual and audible data into their algorithms. By doing so, these systems are able to interact with the world in ways similar to those of humans. When it comes to the crucial sense of touch, however, machine learning is lagging behind. This presents challenges in accomplishing a whole host of useful, everyday activities.
In an effort to fill this gap, researchers at Meta AI and Carnegie Mellon University have developed an open source, touch-sensing artificial skin called ReSkin. It was designed to help researchers quickly advance the tactile-sensing skills of their artificial intelligence projects.
The artificial skin is composed of a deformable elastomer with embedded magnetic particles. When the shape of the skin is deformed, it causes a change in the surrounding magnetic fields. Magnetometers can measure these changes, which can be algorithmically transformed into measures of contact location and applied force.
ReSkin is inexpensive, costing about six dollars per unit in quantities of one hundred. The 2-3 millimeter thick patches have been proven to last for over 50,000 interactions, and with 90% accuracy at one millimeter of spatial resolution. These properties make ReSkin suitable for applications ranging from robot hands to tactile gloves and arm sleeves.
One of the key problems with existing tactile-sensing devices is that they need to be retrained each time the skin is replaced, which is both inefficient and impractical. For this reason, these devices are limited to a single sensor — the data is not as rich, but it makes the relearning task more reasonable.
ReSkin was engineered in such a way that it can sidestep this issue with a generalizable design. To accomplish this, the patches were made such that they do not need to have an electrical connection between the soft materials and the electronics, which makes replacing skins easy. Further, multiple sensors were incorporated into ReSkin such that the higher diversity of data can produce a generalizable output. Lastly, advances in self-supervised learning were leveraged to automatically fine-tune sensors with small amounts of unlabeled data, rather than retraining models from scratch.
To show off ReSkin’s capabilities, the team set up an experiment in which a robot arm picked up grapes and blueberries. With ReSkin applied to the grippers, the arm was able to gently pick the fruit up with no damage. Using only the robot's built-in force sensing, the fruit was crushed. They also showed that they were able to replace the soft portion of ReSkin and continue the experiment with no manual tuning or relearning required.
By leveraging advances in machine learning, ReSkin offers a more powerful, and more practical option for tactile-sensing when compared with existing approaches. The novel, low-cost, compact, and long-lasting sensor design makes it a viable option for a wide range of future use cases.