Strain Gauge Neural Network-Based Estimation as an Alternative for Force and Torque Sensor Measurements in Robot Manipulators

Round 1

Reviewer 1 Report

Major Comment:

1.      The caption of figures needs to be enlarged for better visibility. Also, the cation of the vertical axis in Figure 5 needs revision as it does not seem reasonable.

2.      References 16 and 20 are the same. One of them should be eliminated.

Minor Comments:

1.      All acronyms should appear in their full form first. Then, the authors can use their acronyms in the paper. Currently, 1D, 3D, 6D, and LSTM appear in the abstract without their full forms. After the abstract, their appearance in the body should be reset such that the first appearance should be again in the full form, after which the authors can use the abbreviated forms. In the body, apply the same routine to A/D in line 197, DoF in line 208, and FC in Table 1.

2.      All references should appear in the same format. References #30 and #36 need revision.

3.      The language of the manuscript needs some minor revisions. For example:

a.      “stain guage” in line 156, “anand” in line 242, “when” in line 243, “terms of” in line 306, and “haven’t” in line 320 need revision.

b.      The use of commas is also redundant in multiple cases; lines 142, 142, 280, 305.

 

c.      Lines 268 to 270 need revision for a better understanding of the readers.

Refer to the last comment above.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 2 Report

This article proposes a strategy that combines time series neural networks to efficiently estimate the 3D interaction forces of end effectors and robot joint torque. Using a 1D strain gauge array instead of the robot's built-in sensors resulted in comparable results, greatly saving resources.In this sense, this paper is a high-quality work. However, before publication, I have some comments and hope the authors can address:

1.In lines 96 what do you mean "it was argued"? Please be more specific and provide specific references.

2.Can the constructed LSTM network structure diagram be provided for easy reproduction?

3.The references are generally too old. It is recommended to quote the references of nearly three years.

4.In line 247-248, you said “they proved the most accurate for the task”. Can you verify it in the form of experiment？

 

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 3 Report

The authors propose an interesting approach that measures the the forces at the end-effectors of a robot by mounting strain gauges at the base. It is a novel idea of force measurement in robotics. However, there are some issues to be carefully considered. 

1. In the experiment, it seems that the robot base and the strain gauge platform is not firmly fixed on the experimental platform, which does not meet the standard of force tests. Meanwhile, it seems that the experimental platform is made of wood, whose hardness is low and not suitable for robot force tests, since significant distortions will occur on the wooden surface when heavy load is applied. 

2. There is a gap between the robot base and the strain gauge platform. The connection method between them should be clearly presented. The platform showed in Figure 1(b) is not as the same as the one applied in Figure 2.

3. The testing condition, i.e., the load, the trajectory, etc., of the robot should be clearly presented. Generally speaking, the authors should apply the rated load, relatively high speed and acceleration to the robot in the experiment. This is the major issue that the authors should carefully consider and explain. The measurement method of the ground-truth should be introduced.

4. It is also should be discussed that whether the precision of the proposed method will change after the robot has undergone a long time operation.

The quality of English is good. 

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 4 Report

In this paper, two ways to provide accurate force estimates for a robotic arm are presented. The main novelty is the usage of fairly inexpensive strain gauges at the base of the robotic arm to obtain forces applied at the end-effector or at the powered joints. At the same time, a more conventional, industrial-grade measurement cell is applied in order to compare the results. Both measurements are further processed by trained LSTM neural networks which supply the estimated forces at the end-effector and torques at the powered joints of the robotic arm.

Experiments were made to obtain the final results which are presented in the paper. The results from the two measurement systems are similar, but the strain gauges proved to provide slightly better results.

Comments to the paper

1. The strain-gauge platform shown in fig. 1 differs from the one shown in figs 2 and 3. I assume the intention is for fig. 1 to present the measurement principle and not to accurately show the measurement setup. Please explain this in the text (or change the picture in fig. 1 to coincide with figs 2 and 3.)

2. How do you provide the accurate force/torque data (“measured” in figs 4 and 5)? Accurate data must have been used to train the neural network and to evaluate the final results. Please, explain this clearly in your paper.

3. If accurate data force/torque data is provided by simulation, how was this simulation exactly conducted? Which software was used (if relevant) and how the model parameters were obtained or identified? Please, explain this in your paper.

4. In fig. 5 all the torque results are expressed as motor currents (which can be considered equivalent as discussed in the paper). Does it mean that these currents measured by the UR5e control system are actually considered as the accurate data the network has been trained against and serve as the basis for the evaluation of the results? Please, explain this in your paper.

Spelling/language

1. (line 169) toll – change to tool

 

Author Response

Please see the attachment.

Author Response File: Author Response.pdf