UniContact:A Basic Model for Robotic Manipulation of Contact Synthesis on Rigid and Articulated Rigid Bodies with Arbitrary Manipulators

The paper lacks sufficient comparative experiments, emphasizing the complexity of robotic hand inverse kinematics. However, recent studies have proliferated in generating desired human hand postures based on object models. There is a certain similarity between the collision-free generation of robotic gripper configurations on objects and that of human hands, which calls for comparative experiments to demonstrate the superiority of the proposed method.

Moreover, concerning novelty, the paper employs the Artificial Potential Field method, a classic collision avoidance algorithm, though there are more advanced methods that might perform better. This traditional approach, while reliable, is not necessarily at the forefront of current technological advancements in robotics.

The author could strengthen the paper by comparing the proposed method with contemporary strategies that also aim for collision-free grasping. This comparison could highlight specific advantages in efficiency, accuracy, or applicability in diverse scenarios.

Additionally, while the paper highlights the intricacies of inverse kinematics in robotic grasping, a more in-depth exploration and comparison with human-like grasping techniques are advisable. These comparisons could offer insights into natural and intuitive grasping postures, potentially improving the robotic system's performance and versatility. It would be compelling to see if incorporating advanced techniques could further optimize collision avoidance and gripping efficiency, making a stronger case for the proposed method's applicability and superiority.

1, In inference stage, how can users obtain the target wrench of an object with specific task description? For the same object but different task requirement, the desired wrench might be different. The efforts to generate the target wrench should be discussed further in the paper to evaluate whether is realistic for robot manipulation.

2, It is not clear how this new formulation benefits autonomous robot manipulation. I have doubt that the delicate calculation of required wrench as in input for a manipulation task may make the task complex, which contradicts the purpose of learning-based approach.

3, Another concern about the paper is its technical novelty. The paper is mainly inspired by UniGrasp. As much as I appreciate the novel improvements including contact optimization and IK solver, I had a hard time justifying the technical novelty of this paper given that the main focus of the ICLR conference is on learning methods themselves.

4, The experiment part is weak and insufficient, where the baseline is only UniGrasp and there is no ablation study or detail analysis provided.

5, Some key information and experiments are stated in the manuscript to be provided in the website link. But until now, the website is still empty. With these, it makes the paper hard to understand.

6, There are some grammar errors in the paper, including the figure index in the first line of 7th page, the last sentence in RELATED WORK etc.

This paper seems to be incomplete; the provided website is empty, and the supplementary PDF is the main paper.

The writing of this paper also needs to be polished. For example, the sentence "Based on the predefined number of contact points K, ..., . Here each contact region set contains K regions. Each region contains one contact point, thus one region may appear more than once inside the contact region set." is unclear and hard to follow.

Furthermore, the authors did not introduce what clustering method was used. And if I understand correctly, clustering divides the point cloud into a finite number of regions. From these regions, K regions can be sampled to form a candidate region set (repeated sample is allowed).

"We gather the segmented point cloud leveraging the image segmentation from SAM Kirillov et al. (2023)." This is unclear. It seems that SAM can only segment 2D image, it seems the authors use 2D masks to obtain the object's point cloud, then use clustering methods to divide it into different regions. Here's a further question: the point cloud obtained from an RGB-D camera will be incomplete, how did the authors address this?

There are no particularly novel ideas in the approach, and the training data was also generated through a sampling-based method, make it less interesting.

The proposed methodology makes some severly constraining assumptions:

-The examined setting seems to assume that no external forces are applied to the manipulated object at any time, other than the manipulation forces from the robot. In the vast majority of real-world manipulation tasks this assumption does not hold, as friction and gravity play an important role on the required manipulation forces (for complex tasks other than grasping, where the submissions aims to focus).

-The proposed Inverse Kinematic solution does not consider collisions with background (support) objects (e.g. a table), or clutter from other objects in the scene. This limits the applicability of the proposed approach in real-world environments and generic tasks.

-Being mostly applied in simulation and pre-segmented object point cloud data, the resilience of the proposed approach to sensor noise, segmentation errors, partial observability etc is not evaluated. These factors can hinder its applicability to the real world. Additionally, the need for a segmented object pointcloud contradicts the motivation of manipulation arbitrary (previously unseen) objects.

Additionally, In my opinion, the contributions of the proposed methodology are best suited to the robotics community, where the proposed framework can be most appreciated. In the context of an ML-venue, the representation learning-related contribution appears minimal.