IEEE Robotics & Automation Magazine - December 2022 - 135

the tactile features in the raw magnetic data are complicated
(they include the force feature, contact mode feature,
location feature, and shape feature), it is difficult to extract
the contact surface shape feature by kNN or CNN methods.
The former has a limited ability to extract features, while
the latter requires more datasets. However, in the case of
limited data, the Pronet model can also obtain the ability to
perform shape feature extraction better than the kNN and
CNN methods.
In addition, we used five new measuring heads (rectangle:
10 × 15 mm and 8 × 12 mm; circle: diameter of 12 mm;
square: 12-mm side; and triangle: 12-mm side) to collect data
for meta-validation. In the meta-validation process, we split
the newly collected data into support sets and query sets.
Then, the support sets were put into the trained model to calculate
the corresponding prototype. The classification was
obtained by calculating the distance between the data in the
query set and the prototype. The classification accuracy of the
best model (Pronet-L) could still be maintained at 82.88 ±
2.1%. This shows that the model has the ability to extract new
contact surface shape features.
Experiment 2: Contact Pose Classification
The pose of objects relative to the tactile sensor is an important
feature for robotic hands to perform manipulation. For
example, to achieve the peg-in-hole task, the robotic hand
should know the attitude of the peg relative to the finger.
However, the manipulated object is always changing in actual
manipulation tasks. Therefore, the contact pose recognition
model should have the ability to adapt to different manipulated
objects.
For our experiments, we defined six contact poses
between the bar-type measuring head and tactile sensor, as
shown in Figure 7(a). We list the few-shot contact mode classification
of our dataset under the different measuring head
conditions, which have six contact poses, in Table 4.
For contact status classification, we analyzed three scenarios-one-shot,
three-shot, and five-shot learning-among
different measuring head conditions. The results are shown
in Table 5. All of these methods perform well on these tasks
but are still not as accurate as the one-shot learning methods.
These results indicate that the features of the contact
pose under different measuring head conditions are similar.
However, the metric-based metalearning method can
achieve the best performance in a five-shot learning setting.
Pronet-L has the best performance for addressing this issue,
and the optimized loss function can improve the performance
of the model.
We randomly select the contact points on the tactile sensor
beside the 25 points shown in Figure 7 to collect some
unlabeled samples. All of these samples are used for metatesting.
The results are shown in Figure 9. Even under the
condition that the measuring heads are in contact with tactile
sensors at other positions, the few-shot learning methods
still guarantee good recognition accuracy. The other two
methods perform worse than the state of meta-validation
(accuracy reduced by approximately 10%). This finding suggests
that few-shot learning used for contact pose recognition
has a better generalization ability compared with the
other two methods.
Discussion
The core task of few-shot tactile feature recognition is effectively
extracting and representing features from magnetic
data. Based on the few-shot contact information with the
label, the proposed model considers both the similarity
Table 4. The few-shot orientation recognition of
the tactile dataset.
Different Contact Orientations
Metatraining Set
Measuring heads 1 and 2
Measuring heads 1 and 3
Measuring heads 2 and 3
Metatesting Set
Measuring head 3
Measuring head 2
Measuring head 1
Table 5. The few-shot contact status classification
of the tactile dataset.
Accuracy (%)
Pronet (one shot)
Measuring
Heads 1
and 2 → 3
91.25
Pronet-L (one shot) 92.23
Pronet (three shot) 94.64
Pronet-L (three shot) 95.78
Pronet (five shot)
97.88
Pronet-L (five shot) 98.44
KNN
CNN
92.85
89.15
Measuring
Heads 1
and 3 → 2
91.86
92.14
95.12
95.34
98.14
98.26
93.54
87.58
Measuring
Heads 2
and 3 → 1
91.78
93.07
94.76
95.14
98.28
98.45
93.07
90.47
Accuracy (%)
100
90
80
70
60
50
40
30
20
10
Figure 9. The results of the unlabeled sample classification. L
means that the model used optimized loss function.
DECEMBER 2022 * IEEE ROBOTICS & AUTOMATION MAGAZINE *
135
One Shot
One Shot L
Three Shot
Three Shot L
Five Shot
Five Shot L
kNN
CNN
IEEE Robotics & Automation Magazine - December 2022

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