Learning Sequential Force Interaction Skills
Round 1
Reviewer 1 Report
The paper is interesting. However, it requires a major revision as there is not scientifically solid.
Introduction: This section needs to be re-written. It only describes the approach and the main contributions are the same as for any other paper regarding robot learning from demonstrations (including forces). Why is it important to learn the force? The authors need to clearly state what is the problem they try to solve, what other approaches are there to solve the problem and what is their contribution.
Related work: It is well explained. However, it misses the comparison with the presented work. What are the limitations of related work that your approach will improve?
Equation (2): It is very complicated to follow what the symbols mean. Consider changing to a bullet format or a table. Also, is I the identity matrix? Also, for the J^T do not use the same formating as joint torque T. Consider using non-italic to symbolize the transpose.
Figure 1: the authors do not explain what is A_P/F, A_O/T, A_H, and a_P/F, a_O/T, a_H. Why for the reproduction are the objects not taking into account? How you can have a reproduction just by trained data and features F? This diagram needs to be improved.
Section 1.3: The features F needs to be explained in more detail. It is an important part of the reproduction phase.
Section 2.1: The segmentation is performed with the zero-velocity crossing. However, what happens when during the demonstration a teacher stops temporarily for some reason not related to the task? Or do you assume your demonstrations are perfectly executed?
Section 3: The authrors are describing a directional normal distribution. However, the equation is the same as the normal distribution. Please provide references for what is the directional normal distribution. If it is terminology you have developed, then clearly explain the difference with the normal distribution. Also in equation 3, there is x+tv: x is a data point and then v is velocity and t is time. It is not clear why the authors are using this and what they calculate.
The paper needs to be consistent with the symbols. x in section 2 is task space coordinates and in section 3 it is a vector of size dx1. However, d is not defined.
The equation (5) is not cleared of how it is generated.
In equation (7,8) not clear what is the capital Theta and what the capital E symbolizes. Also, it is not clear why the authors show c1, c2, c3, c4, and then d2, d3, d4 (equation 9). If thay are just constraints, why there is no d1?
Equation (11): Are the d the same constraints are calculated before?
Equation (12): Are a and b same as in section 1.2?
Section 5 in general: What are your evaluation metrics? There is a need to define the evaluation metrics first and then explain how the proposed method improves results over other methods. What about user studies? There is no information how many persons provided demonstrations, their experience with robots, etc.
Section 5.1: Why there is no change in z dimension? Also, how come that all the demonstration are so similar? Are the demonstrations from the same person?
Section 5.4: The authors state: "The reproduction showed that our system can learn to perform the tasks from scratch in an unsupervised way." I do not see this from the paper.
Conclusion: what do you define as few demonstrations?
Author Response
Please see the attachment.
Author Response File: 
Author Response.pdf
Reviewer 2 Report
The authors proposed a skill learning approach for robot to reproduce the task from kinesthetic demonstration. The demonstration was encoded to several movement primitives. The activation of the desired MP sequence was determined by using directional normal distribution. Experiments and comparisons with other approaches showed that the robot successfully reproduced the demonstrated tasks. The idea is sound with good novelty. The paper is well-written. The structure is clear and easy-to-follow. The reviewer has the following suggestions.
- Several variables are not well defined, e.g., I and q_{dot} in equation (2).
- In Line 312, Page 12, the author stated the uncertainty can be neglected, the author should justify this point.
- In the References, there are several long blanks, e.g., Ref [5], [11], [18], [28], [34].
- Some typos. Line 11, Page 1, “are are”. Line 107, Page 3, “suppport”. Line 356, Page 14, “are are”.
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
Reviewer 3 Report
This paper presents a novel approach to learn from demonstrations. In particular, the proposed approach is based on the realization of different methodologies to learning force interaction skills. First, a decomposition method is used to split the demonstrations into a set of Movement Primitives. A novel probability distribution called Directional Normal Distribution (DND) is used to estimate the composition of the MPs and the number of MPs for a specific task. Then, the system learns to sequence these MPs and, hence, to reproduce the task.
Overall, the novelty of the work is clearly stated, and the article is well-written and interesting for the scientific community. The approach is explained using a step-by-step toy example which improves the readability of the paper. Also, the whole system has been tested and validated with a real robot in three different tasks. Besides, the main limitation of the proposed method are clearly stated in the discussion (i.e., the teacher should pause between consecutive motions, and they should try to perform a point-to-point trajectory).
In general, I don’t have any major concerns about this work. In my opinion, the paper would be ready for publication, although I would suggest addressing the following minor comments that will not affect the overall good quality of the work:
1. In equations (1) and (2), the symbol “g” has a different meaning.
2. Section 1.3 describes the overview of the method. Here, it was hard for me to understand how the proposed approach works. That might not happen to the authors as they are already familiar with the methodology, but for a new reader, it could be difficult to understand. I think that simplifying/clarifying a little bit of this section would make it more understandable.
4. The method ZVC appears for the first time in the paper in Figure 2’s caption, on page 6, but there is no reference to [44] until line 220 on page 7. I would suggest adding the reference to Figure 2’s caption.
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
Round 2
Reviewer 1 Report
The authors addressed all my comments and concerns. The paper has been improved a lot and it can be published in its current form. The appendixes are very helpful.
