AnySkin Senses a Disturbance in the Force

To be effective, robots need to have an awareness of their surroundings. As humans we may be partial to our five senses, but it is hard to deny that they give us rich information about the world around us. As such, many efforts have been made to reproduce these senses in artificial systems. For the majority of applications the most important senses are those of sight, touch and hearing. Given that fact, it may seem strange that the sense of touch is so often ignored in robotic systems. It sure comes in handy in just about everything we do, so why do we not give robots this same capability?

It is a complex issue, but the high cost and lack of versatility and consistency in existing tactile sensing systems make them impractical for many use cases. High-end components and complex fabrication procedures are common in the world of tactile sensing, driving costs sky-high. Furthermore, these systems are generally built for a specific platform, so swapping them from one robot to another, or even using them for a different purpose on the same robot, is not possible. Perhaps the worst issue of all is that from one device to the next, the sensors may produce very different results due to inconsistent manufacturing processes that are used in small production runs.

Contrast that with, for example, a camera or a microphone. These components are cheap, easy to integrate with any hardware, and they are highly consistent in the data they produce. Toward the goal of making tactile sensing as easy as other types of sensing, a team led by researchers at New York University has developed what they call AnySkin. It is a plug-and-play artificial skin designed for tactile sensing in robots. But unlike existing technologies, it is inexpensive, versatile, and consistent.

AnySkin builds on an existing technology called ReSkin, which uses magnetic fields for touch sensing. The key improvements include separating the sensing mechanism from the interaction surface and incorporating a self-adhering, self-aligning attachment mechanism. These innovations result in stronger magnetic fields for better sensor response, make AnySkin easy to fabricate for various surface shapes, and allow the sensor to be replaced without disrupting data collection or affecting models trained on previous sensors.

The new system has proven itself to be versatile and can be used on different robotic systems such as xArm, Franka, and Leap. It supports machine learning applications like slip detection and visuo-tactile policy learning, and it only takes an average of twelve seconds to replace while maintaining functionality after replacement. Furthermore, models trained on one AnySkin sensor transfer efficiently to another with minimal performance loss, outperforming the ReSkin system.

Nothing in this world is perfect, and AnySkin is no exception. The magnetic touch sensing system, for example, is susceptible to interference from any nearby magnetic or ferromagnetic objects. Incorporating a Faraday cage into the design could alleviate this problem, but of course that is not appropriate for every application. Moving forward, the researchers are hoping to improve upon AnySkin with the goal of making tactile sensing more practical in robotic systems.