IEEE Robotics & Automation Magazine - September 2023 - 77

rate of 47 Hz. We observed that point cloud
filtering took almost half the time because
the MLS and voxel filters were run on the
CPU. An implementation on CUDA cores
could likely speed up these operations significantly.
The results are promising for
running the pipeline on an online system.
ONLINE RECOGNITION
A separate experiment was performed to
measure recognition accuracy of the system
during an online exploration of an
unknown object, using our 30 unpartitioned
explorations
T . We wished to
"
TO COMPARE THE
PROPOSED APPROACH
WITH RESPECT TO THE
STATE OF THE ART, WE
IMPLE MENTED TWO
DESCRIPTOR-BASED
METHODS: AN ESF
AND A CLUE.
compute and improve the objects' predictions
as the data became available. To this
extent, we treated the tactile data, collected
as described in the " Experiments " section, as an ordered
sequence of samples, defining the exploration completion as
follows: a 10% exploration contains the first 10% of points
collected for that object as sequenced in time. The point filters
were throughout and trained with both the vanilla and CL
schemes, utilizing the V and V2000 datasets, respectively.
The accuracy results from testing on our tactile dataset of
30 samples are presented in Figure 9. The curriculum boosted
performance by an average of 11% and outperformed the
vanilla-trained network for all exploration completions excluding
5%. It is not unsurprising that at the extremely low exploration
completion of 5%, the curriculum training did not make a
difference; the data may not hold sufficient information for one
object to be distinguished from another, no matter the recognition
method used. For a 30% explored object, the curriculum
improved recognition accuracy by 23%. Overall, the curriculum
proved effective for online recognition.
Compared to the results shown in Figure 8, where, with
only 20% of the samples the system was able to predict the
classes with an accuracy of 80%, here, the network starts
to provide acceptable results, i.e., an accuracy greater than
80%, with an exploration of 60%. This apparent drop in performance
is an artefact of the exploration strategy used to
collect our data. As explained in the " Experiments " section,
the data were collected by moving the robot's end effector
in small steps, thus simulating an incremental exploration
strategy. This meant that a low percentage of points representing
the object could only capture local information. On
the contrary, the random sampling procedure used in the
" Partial Point Cloud Recognition " section, which resembles
a random exploration strategy, obtains a broader geometric
representation of the object shape using a low percentage of
the points, easing the recognition process.
We remark that this is not a flaw of the proposed approach.
These two experiments show that this method can be used
effectively for cross-modal recognition of partial tactile point
cloud. Furthermore, they make it clear that a proper exploration
strategy is vital to efficient online recognition. This aspect will
be investigated in a future extension of this work by consider„
0.2
0.4
0.6
0.8
1
02040
60
80
100
Exploration Completion (%)
FIGURE 9. Performance of the classification pipeline when
considering an ordered sequence of samples. The plot shows how
the classification accuracy changes, as long as more tactile data
become available.
TABLE 3. Cross-modal results using handcrafted
descriptors computed on the full point cloud. (a) The
results obtained with the filtering parameters used in
[5] and (b) in this article.
(A) MLS RADIUS=60 MM;
LEAF SIZE=5 MM
THREE NN
ESF 40%
CLUE 20%
SVM
(B) MLS RADIUS=5 MM;
LEAF SIZE=3.5 MM
THREE NN
53.34% 30%
60%
SVM: support vector machine.
23.34%
SVM
73.34%
46.67%
ing methods based on information gain to
work alongside the proposed DL and CL
methods and minimize the number of contacts
required to recognize the object.
COMPARISON WITH DESCRIPTORBASED
APPROACHES
To compare the proposed approach with
respect to the state of the art, we implemented
two descriptor-based methods: an
ESF and a CLUE [5]. As discussed previously,
these descriptors are not suitable for
partial point cloud recognition, and therefore,
we evaluate them on the dataset of
the full point cloud i.e., p = 1. The system
was trained on the visual point clouds
without data augmentation, as described
in [5], and tested on the 30 tactile point clouds. We performed
the classification with a three-NN classifier and an SVM (with
a radial basis function kernel) in two different cases. In the
first experiment, we applied the same values for the MLS and
voxel filters used in [5]. In the second experiment, we applied
Vanilla
Curriculum
SEPTEMBER 2023 IEEE ROBOTICS & AUTOMATION MAGAZINE
77
Accuracy
IEEE Robotics & Automation Magazine - September 2023

Table of Contents for the Digital Edition of IEEE Robotics & Automation Magazine - September 2023
