Comments to author (Associate Editor)
=====================================

Dear authors, thanks for your submission.
Please address the minor constructive revisions proposed by
two independent reviewers.



----------------------------------------
Comments on Video Attachment:

[None found]





Reviewer 3 of IROS 2022 submission 623

Comments to the author
======================

The paper presents an interesting work on a challenging
real-life scenario for a robotic manipulator.
In the focus of this paper, the authors mainly address the
manipulation of single or multiple layers of clothing using
the tactile information collected from the ReSkin sensors
and a Knn classifier. The scenario seems realistic, and the
experiments well conducted.
The work in overall is of good quality, is well written and
the thesis is well explained. I will recommend for
acceptance.


Comments on the Video Attachment
================================

The video attachment is clear and illustrate well the
experimental procedure used.
The quality of the video is good, and the captions are
clear.







Reviewer 4 of IROS 2022 submission 623

Comments to the author
======================

The paper introduces the use of a relatively new tactile
sensor based on an array of magnetometers interacting with
a conformable magnetic skin for robotic dexterous
manipulation of textiles. The challenge was to determine
the number of layers of textile grasped from a stack. This
was achieved using a simple k-means classifier from a
limited training set of data captured during real-world
interaction. Performance of the new sensor was compared
against random, open loop and vision based classifiers, of
which it performed better.
The paper is nicely positioned against other works with the
novelty and contribution clearly presented. The data
collection and experimental procedures are clearly
described with informative figures and concise captions.
The results are positive, statistically significant, and
are a valuable contribution to the community.

Minor criticisms to improve the paper:

There is mention of 350 x 15-D measurements taken over the
5 second interaction episode, which presumably implies that
the grippers are closed for ~1 second (400Hz sample rate).
Does this mean that the 1 second or so of tactile data is
presented as a single input example during training and
subsequently during validation? This seems to contradict
what is implied in section IV.B that point samples rather
than time series are used here. I would read this as time
series data and therefore somewhat richer in information
that a static image, for example. Clarity in such detail
would make for a stronger paper. Similarly, the vision
baseline approach is not clearly described. ResNet is
referred to but no mention of data set capture, image
quality, perspective view, etc. A reference to methods used
elsewhere in this regards should have been cited.

Comments on the Video Attachment
================================

Very nice video that supports the paper