Humans take touch for granted, and it is one of the most important types of input we get in some situations. Our sense of touch allows us to do things, such as feeling how we grasp an object so that we can easily pick it up without crushing it. Whenever we pick up something fragile, such as an egg, we all know how difficult it is to squeeze it so as not to crush it and make it a mess. For robots and artificial intelligence that do not have similar haptics, it is difficult for a robot to collect information about the object it is trying to process.
Facebook AI is working on a new project called ReSkin, which aims to provide additional input to AI devices. Today, artificial intelligence can combine the senses, including sight and sound, but getting the sense of touch is still an elusive thing. Part of the reason for the limited touch information of today's AI systems and robots is limited access to tactile sensor data.
Researchers related to AI hope to integrate touch data into AI models, but it is difficult to provide AI systems with the same touch sensing capabilities as humans. Facebook AI's ReSkin project is an open source touch-sensitive "skin" created by Meta AI researchers and Carnegie Mellon University scientists. ReSkin uses advanced machine learning and magnetic induction to help researchers quickly and on a large scale improve the tactile perception capabilities in their AI systems.
The system is inexpensive, versatile, durable and replaceable, making it very suitable for long-term use. ReSkin uses self-supervised learning algorithms to automatically calibrate the sensors and make the data versatile so that it can be shared between different sensors and systems. Facebook AI will release the design, documentation, code, and basic model, allowing researchers around the world to use ReSkin without training with their own data sets. This will allow tactile perception to be quickly integrated into various artificial intelligence systems.
In addition, generalized tactile sensing will allow researchers to collect data that will help advance the use of artificial intelligence in a range of touch-based tasks. It will help object classification, proprioception and robot grasping. Once artificial intelligence systems are trained in tactile perception, they will be able to perform new tasks, including the ability to work in healthcare environments and manipulate small, soft and sensitive objects. The ability to manipulate a wider range of objects can make robots more suitable for packaging orders in factories or very sensitive products (including agricultural products).
ReSkin costs less than $6 for mass production of 100 units, and is even cheaper in mass production. The sensor itself is 2 to 3 mm thick and can withstand 50,000 interactions. Once worn, it is easy to replace and provides high temporal resolution of up to 400 Hz, as well as a spatial resolution of 1 mm and an accuracy of 90%.
ReSkin also provides researchers with high-frequency, three-axis haptic signals, allowing the system to adjust during fast operating tasks such as throwing, slipping, catching, and clapping. The sensor itself uses a deformable elastomer in which magnetic particles are embedded. The elastic body deforms in any direction, allowing the magnetic signal to change with the deformation. A magnetometer can be used to measure changes in the magnetic signal, which allows conversion of information such as the contact position and the magnitude of the applied force. The researchers created a universal skin to eliminate the need to train a new skin every time the system is changed.
However, each sensor undergoes an initial thorough calibration procedure to determine its individual response. The calibration procedure can adapt to the time-varying characteristics of soft materials. Researchers overcome the challenge with ReSkin. They do not need to rely on the proximity of magnetic signals to establish an electrical connection between the soft material and the measurement electronics, so the electronics only need to be nearby.
The model mapping function does not rely on data from a single sensor, but uses data from multiple sensors. This allows researchers to train models on more diverse data, resulting in more general and effective data. Instead of collecting calibration data on each new sensor, a self-supervised learning system is used to automatically fine-tune the sensor with a small amount of unlabeled data. Facebook AI stated that ReSkin's work is part of its commitment to advancing tactile perception in the field of AI research.