IEEE Robotics & Automation Magazine - December 2022 - 134

shape and pose feature classification, we use three different
shapes (square, triangle, and circle) of the measuring head
[Figure 7(b)-(d)] to collect the shape information and three
bar-type measuring heads to collect the information for six
different contact poses.
In the process of contact surface shape classification data
collection, we select 25 evenly distributed points on the tactile
sensor, as shown in Figure 7. For each contact point, we
collected 30 samples for each measuring head at 20 Hz. The
raw magnetic data were recorded, and then we manually
labeled these data. For contact status data collection, the
overall process was similar. The only difference is that we
needed to collect the contact mode information for each of
the 25 points in six postures, which are shown in Figure 8, by
using measuring heads 1-3.
Experiment 1: Contact Surface
Shape Classification
Researchers have used magnetic-based tactile sensor for
measuring the contact force and position [14]. Usually, they
used the deep learning method to train a model that can
establish the relationship between the raw magnetic data
and contact position. A traditional supervised learning
method for contact surface shape recognition usually needs
1
5
6
4
3
2
2
1
3
to prepare a large amount of labeled data to train the model.
When the sensor is used for measuring other types of contact
objects, researchers need to re-collect a large amount of
labeled data on this kind of object. However, there may be
only a few data samples of a new shape when the sensor
contacts a new object in the actual environment. Therefore,
this limits the use and spread of magnetic tactile sensors.
The raw magnetic data include both the contact surface
shape and pose feature in addition to the location and force
feature. For an identical object, when it contacts the tactile
sensor at different locations, the response magnetic signal is
different. Therefore, it is difficult to collect the contact
information of an object at each location of the magnetic
sensor for contact surface shape recognition, and it is timeconsuming
to collect many actual contact samples of a new
object. Therefore, in this article, we utilize a metric-based
metalearning method for solving the few-shot contact surface
shape and pose feature recognition problem. The purpose
of our method is to identify contact features based on
a few samples.
In the experimental state, the dataset of contact surface
shape recognition is collected from the magnetic tactile sensor
under different measuring head load conditions at 25
positions, which are shown in Figure 7. There are four different
shapes of measuring heads: triangle, square, circle,
and rectangle. The first three shapes include three different
sizes: 10, 15, and 20 mm. Each measuring head has 750 contact
samples, and each sample is a 3D magnetic signal with
20 points.
We set 80% of the collected data as a metatraining set, and
Figure 8. The six contact statuses we define and three measuring
heads used for data collection.
the other 20%, for which there is contact at other locations, is
set as a metatesting set. Based on the few-shot learning setup,
for five-way, one-shot contact surface shape recognition, we
took 10 samples from each category: one of the samples in
each category was used for the support set, and the other nine
samples were used as the query set. The settings for the fiveway,
three-shot and for five-way, five-shot approaches were
similar to this.
We used a Pronet model and improved Pronet-L to assess
Table 3. The few-shot shape classification
accuracies of different methods.
Model
Pronet (one shot)
Pronet-L (one shot)
Pronet (three shot)
Pronet-L (three shot)
Pronet (five shot)
Pronet-L (five shot)
KNN
CNN
Accuracy (%)
78.26
81.33
79.18
84.18
80.31
84.6
34.59
26.76
the performance for contact surface shape recognition. In
addition to few-shot learning methods, we also used two
other methods for comparison: a machine learning method
[k-nearest neighbor (kNN)] and a deep learning method
(CNN). The contact shape classification accuracy results in
the metatesting process are given in Table 3. Accuracy is
defined by the number of correct classifications divided by
the number of all samples in the query set. To compare these
models, the results we present show a better performance
than each other model on the same contact surface shape recognition
dataset. All experiments were implemented on a
computer with one Nvidia GeForce RTX 2060 GPU and one
Intel Core i7-9700F CPU.
The experimental results show that the Pronet-L is better
than the Pronet model in terms of recognition accuracy.
The kNN method and CNN method show poor performance
in the contact surface shape recognition. Because
134 * IEEE ROBOTICS & AUTOMATION MAGAZINE * DECEMBER 2022
IEEE Robotics & Automation Magazine - December 2022

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