The paper makes multiple relevant contributions. The overall concept is promising. The idea of decoupling the five-bar DOFs from the flexion DOFs by routing the flexor tendons through the linkage is great. The manipulation planning algorithm, based on lookups through pre-computed configuration tables, is also very well suited for the mechanism, and achieves good performance. (I suspect a kd-tree for nearest neighbors in a small 3D data set is overkill on most modern computers, but conceptually a good idea.) This level of in-hand manipulation, with robustness to object shape, is very rarely seen in the literature. The paper is generally clear and well-written, in both the description of the mechanism and its algorithmic components. The only aspect that I found a bit harder to understand was the spring mechanism that keeps all three fingers oriented such that they all flex towards a common point located between them. The authors use the term “finger direction” - perhaps use “finger orientation” or “finger flexion direction” instead? I did eventually figure it out, with the help of the video. While the paper is certainly interesting and opens up some very relevant avenues for research, I feel that it only scratches the surface of the complex behavior of these manipulation tasks. Assume that a grasp has been established, and the hand is reconfigured by the base motors. The bases of the fingers are fully constrained to move in a prescribed path, while the tips are partly constrained by the object itself. If these constraints are not mutually compatible (and there is no reason to expect them to be), the hand must adapt, either through finger reconfiguration in the null space of the flexor tendon, or through contact rolling or slipping. The spring mechanism also exerts moments on the fingers, further complicating the force equilibrium of the whole system. These are all very complex behaviors. I suspect that it is possible that during such a process the grasp could be lost. When does that happen? How can we guard against such behavior? I also suspect that such reconfigurations are the main reason for the experimental discrepancies between desired and achieved final object pose. Finally, some of the experiments in the video show out-of-plane object rotation, which I assume is not desired or accounted for, and likely stemming from the same source. As presented here, the hand is also primarily usable for planar (2D) manipulation. The fingers are indeed capable of producing enveloping (power) 3D grasps. However, if any of the fingers makes contact on multiple links, it is not clear if manipulation is still possible. It is likely that any finger with multiple contacts is over-constrained if its base moves (due to movement of the five-bar linkage). I suspect that is why all manipulation examples are essentially planar, with fingertip grasps only. Even for planar grasping, the hand has 2 manipulation DOFs to control 3 object DOFs, thus can only manipulate along a 2D manifold. It is not clear if and how such a method would apply to 3D grasping, creating a 6 DOF object movement space. The method is still highly compelling even given these considerations, and I am not asking the authors to solve all of 6DOF manipulation in a single paper - the work shown here represents important steps in that direction. However, I feel like these aspects should be explicitly mentioned and discussed here. In conclusion, I believe this is a very relevant paper, with strong ideas and good performance in practice on tasks rarely seen. It opens up some very interesting future work directions and questions, so the reader can be left hoping for a more in-depth analysis. Such an analysis could either be included here (taking the paper to a next level) or be carried out in future work.

Originality The paper is original, it presents a design that is novel and brings something new to the literature. However, I did find a major drawback that is only addressed at the very end as a small drawback, while I think it's actually a concept thing. The idea of the gripper is explained as the gripper grasps an object, and then this can be moved thanks to the base 5-bar linkage. At the same time, authors say the gripper can reconfigure into different grasp types. This is true, but what is happening is that the gripper reconfigures into different grasps types while it is grasping. Meaning the fingertips change orientation and move along the object surface while the in-hand manipulation is happening. In practice, it is a small motion probably due to high frictions, but it probably means a lot of force is needed to move the 5-bar linkage, and it may be even impossible to do so if the gripper is performing a power grasp with several contacts in each finger. This fact is minimized in the paper, as the only examples of in-hand manipulation that are shown are with objects held in the fingertips. But I was a bit disappointed, as it would have more value to be able to move an object that is actually strongly grasped, which is not doable in other gripper designs where the in-hand manipulation has to necessarily happen with precision grasps. In addition, the gripper is not compared to any other hand. I know it is difficult to compare a hand that is so particular. But there are other designs that also perform in-hand manipulation with little actuation. The claim of the paper is that they can do it with only 3 actuators and simple planning and control, but maybe it would be interesting to compare volumes of manipulation workspace for different gripper designs with low actuation and simple control, and in-hand manipulation capabilities. There are so many hands that with original and simplified designs achieve variate and large in-hand manipulations, for instance Ma, Raymond R., and Aaron M. Dollar. "An underactuated hand for efficient finger-gaiting-based dexterous manipulation." 2014 IEEE International Conference on Robotics and Biomimetics (ROBIO 2014). IEEE, 2014. or Amend, John, and Hod Lipson. "The JamHand: Dexterous manipulation with minimal actuation." Soft robotics 4.1 (2017): 70-80. These hands above are just a few, and one could argue that they achieved larger amounts of in-hand manipulation with similar or less actuation. I believe that this hand has a value, but I also missed this kind of discussion in the paper, to understand what this hand brings to the landscape and why we should care about considering yet a new design to the large amounts of existing ones. Another thing that could be easily compared to other designs is the error in positioning during in-hand manipulation planning and execution, at least compared to the reported ones in the papers such as Hang, Kaiyu, et al. "Hand–object configuration estimation using particle filters for dexterous in-hand manipulation." The International Journal of Robotics Research (2019). At least, to have an idea if the reported errors in position is something normal for in-hand manipulation. Quality and clarity In my opinion, the quality of the paper is good, but it could be improved making the text and figures a bit more clear in the following aspects: 1- There is not a clear diagram of the 5-bar linkage and the nomenclature used, the name of the joints, the location of the points P, which one is the link 1...to 5, etc... You can figure it out but you need to make your own drawing. It may be fixed adding labels to Fig.2, but remember to also add the points P1...P5 referred in Sec. III - C. Also, this figure needs to be cited at the very beginning of Section II.A. 2- There is a lot of confusion between what they call "grasping workspace", "manipulation workspace" and "manipulation map", which are basically the same thing. It seems like authors wanted to use different names so that it seemed they were doing more computations, but this just increases readability. There one single workspace in this design, which is composed of a continuous XY slide, and for each point, it corresponds to a different orientation of the object, (X,Y,Phi). So, the grasping workspace is the XY slide, and the manipulation workspace is the whole thing, and the manipulation map the correspondence (X,Y -> phi). This should be simplified for clarity. Significance The paper adds to the idea that underactuated and simple grasping can still achieve in-hand manipulation capabilities. It is a new design that could lead to variations, using n-bars mechanisms. I think it is a good idea, but it could have more significance if the design is shared in an open-hardware fashion (sharing building instructions and models) and if some of the comparisons I mentioned before are reported.