American engineers have developed an inexpensive and compact touch sensor for robots / robotic arms. It consists of an elastomer with magnetic particles and magnetometers, which makes it possible to reconstruct the deformation of the elastomer upon contact with objects and thereby determine its strength and location. The article will be presented at Conference on Robot Learning 2021, and the library code for using sensors is published on GitHub.
Capturing objects is one of the most important tasks in the field of robotics, which is solved by a huge number of research groups and companies. For example, Amazon has been running competitions among engineering teams for several years dedicated to this particular problem. Most of the attention in this area is focused on computer vision algorithms that allow a robot to recognize the type of an object, reconstruct its shape, and even predict stiffness in order to grab it well. This is an extremely important operation, since the massive introduction of robots is closely related to their, including cognitive abilities. The more developed the robot is, the more likely it is that when choosing the employees that a particular company wants to hire, the choice will be made in favor of a particular model.
The grips themselves are also gradually improving, but less attention is paid to this aspect, although it also has important unresolved tasks, including reliable gripping of fragile objects. Engineers offer different designs for neat and precise grips, often they themselves are made of soft material, an analogue of human skin, by deformation of which the robot understands the force of compression and even the shape of an object. For example, last year Toyota showed a prototype of a gripper with inflatable bladders and dots on the inner surface to reconstruct the curvature of the surface. But he, like many analogues, was quite massive.
Engineers at Carnegie Mellon University and Meta AI (formerly Facebook AI), led by Abhinav Gupta, have created a new compact sensor for robots (robotics tech) with a soft material tip that allows you to track touch parameters with high precision. The sensor for robots consists of two parts. On top, that is, on the touching side, there is an elastomer with a thickness of 2-3 millimeters and a width of a one-ruble coin – 20 by 20 millimeters. The elastomer contains ferromagnetic particles. at the stage of creation, the liquid precursor is poured into the mold and 16 magnets are installed above and below it. This gives the elastomer a predetermined magnetization after hardening.
A microcircuit with five magnetometers is located under the elastomer. The principle of operation of the sensor for robots is that when it comes into contact with any object, the elastomer is deformed and the magnetometers register the change in the magnetic flux. The developers taught the sensor to convert the change in magnetic flux into coordinates, direction and magnitude of the applied force by training a multilayer perceptron.
Accuracy tests showed that the root mean square error for the contact location was 0.514 ± 0.601 square millimeters and for the force 0.142 ± 0.025 square newton. The engineers demonstrated the real application of the development by installing two sensors on a robotic gripper and using it to pick up blueberries and grapes without flattening them. Robots equipped with new generation sensors will be a good buy for the company that bought them, and will be able to perform more delicate and complex operations than they do now.