Task-Specific Grasp Planning for Robotic Assembly by Fine-Tuning GQCNNs on Automatically Generated Synthetic Data

Round 1

Reviewer 1 Report

The paper proposes a learning-based method to synthesize a grasp of an object to be used for a specific task. I definitely agree that the “quality” of the grasp to be selected should be related to the whole task the robot has to perform. The paper is generally well-written, however, I suggest explaining in more detail the whole pipeline, e.g., in a more schematic way, by using an item list or better a pseudo-code, or a flow-chart. In fact, it is not fully clear if the training scene in Blender should be similar to or exactly the same as the actual task to be executed. This might be a severe limitation since the training burden could be too high in case the task is subject to small changes. Another aspect that is not clear in the pipeline is the use of the motion planner. It seems that it is used only after training. If so, then it is not clear how the place grasp is considered valid without taking into account the collisions of the whole robot arm in the task execution. A similar problem is tackled in a logistic scenario in DOI: 10.1109/MRA.2021.3064754, where it has been demonstrated that fixed grasps might make the task unfeasible and only in-hand manipulation manoeuvres, such as those described in DOI: 10.1016/j.mechatronics.2021.102545, allow the motion planner to find a solution. I expect that, in a similar way, if the GQCNN outputs a valid grasp couple (pick-place) they could be unfeasible once the motion planning problem is tried to be solved. The authors should better discuss this aspect and highlight the limits of the proposed approach.

A minor remark concerns the statement in lines 35-36 about the complexity of grasp quality. There only the geometry aspect is cited, while dynamic (e.g., friction forces) aspects are equally relevant to grasp quality assessment. It is clear only later in the paper that it does not consider dynamic aspects; I suggest clarifying this from the beginning, even in the abstract.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 2 Report

1. The research is to the scope of the journal.

2. The paper writing is somehow subjective, i.e., too many "we" and "our" in the context. Please consider re-write and presenting findings/results in a more objective way.

3. It is suggested that the authors offer more details about GQCNN.

4. Is there any conjecture about why the success rate of "place" are worse than "pick" for both models? (shown in Table 1 and 2)  

Author Response

Please see the attachment.

Author Response File: Author Response.pdf