IEEE Robotics & Automation Magazine - December 2022 - 133

for each category k, we calculate the Euclidean distances
between the category prototypes and normalize them. Therefore,
we define loss function 2 as
Loss = R == ! +
2
ij ij
N
11 (( ,))
,,
1 Dp p
1
ij
where pi and pj are the prototypes of two different categories.
D (.) represents calculating the Euclidean distance between
two prototypes pi and pj. The total loss function is defined as
Loss
total 12
=+a # Loss
Loss
(6)
where α is a hyperparameter; in this article, we choose 0.1. In
the process of meta training, the purpose is to minimize the
loss and optimize the parameters in the feature extraction
model. In the experimental part, we compare the performance
of Pronet with the optimized loss function (Pronet-L)
and the original Pronet.
Model Details
In this work, a simple end-to-end feature extraction network
is proposed. The input of this CNN model is a size of 60 (the
magnetic field intensity values from 20 sensors in the x-, y-,
and z-directions) × 10 (time step) magnetic signals. The
details of the CNN are shown in Table 1. The network has
three convolutional layers and maxpool layers. The batchnorm
layer is used to standardize raw data. In the end, the
network generates a feature vector for each sample. In addition,
the other details (hyperparameters and the metalearning
setting) used for the metatraining of our model are
shown in Table 2.
Performance Evaluation
In the stage of performance evaluation, we put all of the
support sets used for meta-validation into the trained
model. Then, the prototype of each class was computed,
and all of the unlabeled samples from the meta-validation
set were also similarly used to compute the feature. By
measuring the distance between the prototype of the support
set and the features of the unlabeled samples, we can
obtain the contact surface shape and contact pose classification
results.
Experiments
For this part, the details of the experimental setup are
described, and the tactile feature extraction method is evaluated
by a series of experiments. The process of data collection
and the method used for the object shape and contact mode
are all elaborated. We also show and discuss the experimental
results at the end.
Experimental Setup and Data Collection
The experimental device is a z-axis movement platform and a
digital force gauge. A series of resin measuring heads with different
shapes were designed for collecting the contact shape
information. The sensor is fixed on the platform and loaded
(a)
(5)
Table 1. Details of the cnn network.
Layer
Kernel
Size
Conv2D 1
Batchnorm2D 1
ReLU 1
MaxPool2d 1
Conv2D 2
ReLU 2
MaxPool2d 2
Conv2D 3
ReLU 3
MaxPool2d 3
(2, 2)
Channels
(In, Out)
(1, 16)
Stride Padding
(2, 2) Yes
by the measuring head, which is attached to a digital force
gauge in the process of data collection. In Figures 7 and 8, we
show some of the measuring heads. For the contact surface
(2, 1)
(2, 2)
(2, 1)
(2, 2)
(2, 1)
(16, 16)
(16, 32)
(32, 32)
(32, 64)
(64, 64)
(2, 1) No
(2, 2) Yes
(2, 1) No
(2, 1) Yes
(2, 1) No
ReLU: rectified linear unit; Conv2D: convolution 2-dimension;
Batchnorm2D: batch normalization 2-dimension; Maxpool2D: max
pooling 2-dimension.
Table 2. Details of the model setup.
Description
Optimizer
Value
Adam
Learning rate
Batch size
Maximum epochs
Distance metric
Tasks per epoch
0.0001
32
100
Squared Euclidean distance
100
Support samples per class (Ks) 1, 3, of 5
Query samples per class (Kq)
10-Ks
10 mm
(b)
(c)
(d)
Figure 7. (a) The contact location of the magnetic tactile sensor
in the process of data collection. (b)-(d) The measuring head,
which has different shape features: (b) triangle, (c) square,
and (d) circle.
DECEMBER 2022 * IEEE ROBOTICS & AUTOMATION MAGAZINE *
133
10 mm
IEEE Robotics & Automation Magazine - December 2022

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