IEEE Robotics & Automation Magazine - December 2022 - 131

by utilizing the multipole magnetic layer, the collected magnetic
field information contains more contact features than
other magnetic surfaces. When the sensor contacts the
object, the response is also obvious. It is conducive to processing
the raw data for tactile recognition. In the next section,
we introduce the details of our method for contact
surface and pose recognition.
In addition, we tested the actual response of the sensor
using the same experimental platform. The experimental
results are shown in Figure 5. The test results
show that the multipole magnetic layer tactile sensor
shows a greater variation of the magnetic strength. It
means that the sensors are more sensitive in measuring
contact with objects.
Method
When objects contact the magnetic tactile sensor, the collected
magnetic data include the shape and pose information
of the contact surface. The existing magnetic tactile
sensor process methods usually direct the contact position
and force recognition through a combination with deep NN
(DNN) methods and obtain good results [14]. However,
related works are few for the contact surface shape and pose
recognition using magnetic tactile sensors. To calibrate the
sensor, the DNN methods always need many labeled samples
(e.g., in [14], they used 40,500 samples to train the
DNN model to achieve the calibration of only an object). In
practice, this approach is a limitation for the popularization
of the use of magnetic tactile sensors. In this part, we introduce
the contact surface recognition method, which uses
prototypical networks (Pronets) based on limited samples
and obtains good results.
Few-Shot Metalearning
It is easy for a human to learn to recognize a new object from a
few instances because the human can use prior knowledge to
learn. Inspired by this ability, the metalearning method, a flexible
framework that can learn prior knowledge from performing
related tasks, is presented. At present, there are two main
kinds of metalearning methods: metric-based metalearning
(e.g., the matching network [25], Siamese network [26] and
Pronet [27]) and optimization-based metalearning (e.g., model-agnostic
metalearning [28]). Based on the metalearning
method, researchers have solved some problems: few-shot
fault diagnosis [29], image classification [27], and speech recognition
[30]. For this article, combined with metric-based
metalearning methods (Pronet), we extract and recognize the
contact features based only on magnetic field information.
Problem Definition
For few-shot shape and contact status classification,
we define this problem following [27]. We organize the
Loading
Position
Loading
Position
300
250
200
150
100
50
∆B for M1
∆B for M1
∆B for M2
∆B for M2
∆B for M3
∆B for M3
∆B for M4
∆B for M4
300
350
250
200
150
100
50
0 0.2 0.4 0.6 0.81 1.2
Displacement (mm)
(a)
1.4 1.6 1.8 2
∆B for M1
∆B for M1
∆B for M2
∆B for M2
∆B for M3
∆B for M3
∆B for M4
∆B for M4
0 0.2 0.4 0.6 0.81 1.2
Displacement (mm)
(b)
Figure 5. The results of the change in magnetic strength when in contact with the measuring head. (a) The actual result measured
by sensor array A1. (b) The actual result measured by sensor array A2. TB is the total magnetic field change value measured by the
sensor array. The magnetic field distribution on the surface of sensor arrays A2-A5 is similar, so the experimental results are similar,
and only the result of A2 is shown.
DECEMBER 2022 * IEEE ROBOTICS & AUTOMATION MAGAZINE *
131
1.4 1.6 1.8 2
∆B (µT)
∆B (µT)

IEEE Robotics & Automation Magazine - December 2022

Table of Contents for the Digital Edition of IEEE Robotics & Automation Magazine - December 2022