Grasp2Grasp: Vision-Based Dexterous Grasp Translation via Schrödinger Bridges

We sincerely thank the reviewer for their time and insightful feedback. We are encouraged that the reviewer found our manuscript "insightful" and our method "novel and well-motivated." We address each point below.

W1: On VAE Evaluation and Latent Space Visualization.

Thank you for these concrete suggestions to improve the paper's completeness.

VAE Evaluation

We agree that an evaluation of the VAE is important. We have conducted an additional ablation study to validate our design choices. We compare our full VAE against three alternatives:

• A PointNet VAE baseline based on [1] that reconstructs the hand point cloud via Chamfer distance and decodes hand parameters in a separate branch.
• Ours w/o Mesh Loss: Our VAE trained only on the hand parameter reconstruction loss, without the FK layer and mesh vertex loss.
• Ours w/o Param Loss: Our VAE trained only on the mesh vertex loss, without the direct parameter loss.

We evaluate performance using the Hand Parameters L2 Error on a held-out set of objects. We choose this metric for evaluating reconstruction quality because our VAE ultimately decodes the hand parameters. As shown in the table below, our proposed VAE architecture significantly outperforms all alternatives.

Table 1: Evaluation for VAE: Hand Parameters L2 Error

Our analysis indicates that the PointNet VAE struggles because the Chamfer distance loss is often indiscriminative, as evidenced by many works [1]. It can achieve a low value even when the reconstructed grasp pose is incorrect but has large overlapping regions. Ablating the mesh loss harms performance on fine-grained details, while ablating the parameter loss can lead to memorization issues given the high capacity of the latent space, as the size of the hand configuration space (typically ~20) is relatively small compared to the size of the latent space (e.g., 768 in our case). We will include this table and analysis in the appendix of the revised manuscript.

Latent Trajectory Visualization

This is an excellent idea. Direct visualization of the latent trajectory by decoding intermediate points (i.e., for $t$ between 0 and 1) is unfortunately not straightforward. While the latent space is shared, the VAE decoder is hand-specific due to the use of the forward kinematics (FK) layer. Decoding an intermediate latent $z_t$ would be ill-defined, as it is unclear which hand's FK layer should be used for visualization.

However, to provide the intuition the reviewer requested, we will add a new figure to Section 4 showing the learned transport process on a 2D toy example (e.g., translating from a two-moons distribution to an 8-Gaussians distribution). This will visually illustrate the latent evolution concept that underpins our method.

W2: Presentation of core technical contribution.

We appreciate this valuable feedback on improving the paper's structure and clarity. We will implement the following changes:

1. We will streamline Sections 3.1 and 3.2, moving the general background on Schrödinger Bridges (SB) and the derivation of the SF$^2$M to the appendix. We will retain only the key equations and concepts in the main text that are directly referenced in our methodology (e.g., the entropic OT plan and the final conditional loss function).
2. We will use the freed-up space to expand Section 4.1, providing a more detailed explanation of our core technical contribution: modeling the SB between latent grasp spaces. This new text will elaborate on how the neural networks $v_{\theta}(t,z)$ and $s_{\theta}(t,z)$ approximate the conditional flow and score, how they are trained using paired samples drawn from the OT plan, and how they are used at inference time to evolve a source latent code to a target one.
3. The new 2D visualization figure mentioned above will be included in this revised section to make the explanation more intuitive.

Q1: Remaining three transfer settings (A→H, S→H, and A→S)

Yes, this is a great suggestion to make our evaluation more comprehensive. We have run experiments on the remaining three transfer settings: Allegro→Human (A→H), Shadow→Human (S→H), and Allegro→Shadow (A→S).

We must note a technical limitation: we were unable to report the IsaacGym success rate for settings where the target is the human hand (A→H and S→H). This is because our human hand model is an articulated version of the MANO model [2] with non-watertight link meshes, which causes instability in the IsaacGym simulator. Porting our evaluation to a simulator capable of handling such assets (e.g., IsaacSim) is beyond the current scope of this work and is planned for future development.

However, we are able to report the success rate for A→S and the other metrics (Diversity, 6D GWH IoU) for all three new settings. The results are presented below.

Table 2: Comparison on Success Rate (%)↑

Table 3: Comparison on Diversity (rad)↑

Table 4: Comparison on 6D GWH IoU (%)↑

Our method consistently outperforms both baselines in preserving functional intent, achieving significantly higher 6D GWH IoU (Table 4), particularly when transferring to the more complex human hand. We also generate a more diverse set of plausible grasps (Table 3), highlighting the flexibility of our learned transport plan. Finally, in the A→S setting where a direct stability comparison was possible, our method achieved the highest success rate (Table 2), demonstrating superior physical plausibility. These findings show that the advantages of our approach are robust, even in these challenging transfer settings.

We will add these tables, along with the analysis and corresponding qualitative results, to the appendix to provide a more complete picture of our method's performance.

Reference:

[1] Achlioptas et al., "Learning representations and generative models for 3d point clouds." ICML 2018.

[2] Romero et al., "Embodied hands: Modeling and capturing hands and bodies together." SIGGRAPH ASIA 2017.

Thank you for your clear and detailed rebuttal. With the revised manuscript incorporating more comprehensive evaluations—such as latent visualizations, additional baselines, extended experimental setups, and a discussion on inference speed—I believe the paper is of high quality. I have no further questions.

We sincerely thank the reviewer for their time and insightful feedback. We are encouraged that the reviewer found our approach "well-formulated" and "interesting." We address each point below.

W1: On Experimental Quality, Plausibility, and Analysis

Physical Plausibility and Validation

We acknowledge that static images can sometimes appear unstable due to challenging viewpoints or occlusion. However, we want to emphasize that all generated grasps are quantitatively validated for physical stability in the high-fidelity IsaacGym simulator. We followed the rigorous protocol from GenDexGrasp [1], which we detail on Lines 292 - 299 in our manuscript:

"We report the grasp success rate using IsaacGym, following the GenDexGrasp [1] evaluation protocol. A grasp is considered successful if it maintains a stable hold on the object under six perturbation trials, where external forces are applied along $\pm x$, $\pm y$, and $\pm z$ axes. We apply each force as a uniform acceleration of 0.5 m/s$^2$ for 60 simulation steps and measure whether the object translates more than 2 cm. A grasp passes if it withstands all six trials. Simulation parameters include a friction coefficient of 10 and an object density of 10,000. We use IsaacGym’s built-in positional controller to achieve the desired joint configurations."

Our experimental results show that our method generates grasps that are not only stable but are more stable than all baselines. In response to other reviewers' feedback, we have also compared against two additional baselines, Dex-Retargeting [2] and CrossDex [3]. Our method continues to achieve the highest success rate, as shown below.

Table 1: Comparison on Success Rate (%)↑

Qualitative Results and Visual Similarity

We accept the criticism that the qualitative results in the initial submission were limited to two objects. Unfortunately, we are not allowed to present any multimedia results via external links as per the rebuttal guideline. We promise to include a more diverse set of qualitative examples from our 34 unseen test objects in the final version.

Regarding the "generated poses for the cup differ significantly from the source poses," we wish to reiterate the core goal of our work. Our formulation of grasp translation is object-centric and aims to transfer the functional intent of a grasp, which allows for morphologically and posturally different but functionally equivalent outcomes.

For example, to turn on a TV remote, a source grasp might be from below, palm up, with the thumb on the power button. A teleoperation system would try to replicate this exact pose. Our method, however, recognizes the functional goal (applying force to the power button while supporting the remote) and could generate a target grasp from above, palm down, with the index finger on the button and other fingers providing support. We consider both grasps functionally equivalent though they might appear very different visually.

We believe the qualitative examples for the cup in Figure 3 all share a similar functional intent: the hand approaches from the bottom, with fingers touching the bottom of the cup, and the thumb positioned near the handle for stability. Our method successfully transfers this intent, while the baseline method RobotFingerPrint [4] fails to generate a stable grasp for this specific example.

W2: Regarding real-world experiments.

We agree that real-world validation is crucial. Given the short rebuttal period, we were unable to conduct these experiments. Our work currently focuses on the core algorithmic contributions, validated in simulation. We are committed to including real-world experiments in future work.

However, to strengthen the paper now, we have conducted extensive new simulation experiments, including comparisons to two additional strong baselines (Dex-Retargeting [2], CrossDex [3]), evaluation with a new pose alignment metric, a detailed efficiency report, and evaluation on the remaining three transfer settings (A→H, S→H, and A→S). Please refer to Table 1-5 of our response to reviewer AEMy and Table 2-4 of our response to reviewer 4u79 if interested. We believe these additions strengthen our empirical results and address key questions about performance and applicability.

W3 & L2: On Methodological Clarity

Thank you for pointing out areas where the presentation can be improved. We will revise the manuscript to be more self-contained. To clarify the link between Sections 3 and 4, we will add the following at the beginning of Section 4.1:

"We now apply the general Schrödinger Bridge framework from Section 3 to our specific problem of grasp translation. The abstract distributions $q_0$ and $q_1$ become the latent distributions of source and target grasps, respectively. The transport process between them is learned by optimizing the SF$^2$M objective from Eq. 7 in a latent space."

To clarify the integration of OT costs, we will add the following at the beginning of Section 4.2:

"To guide the Schrödinger Bridge, we must first compute an entropic OT plan $\pi_{\varepsilon}^*$ (Eq. 2), which is defined by a ground cost $d(\cdot, \cdot)$. We design several physics-informed costs for this purpose. For a given cost, we compute the OT plan between minibatches of source and target latent samples. We then draw pairs $(z_0, z_1) \sim \pi_{\varepsilon}^*$ and use these pairs to construct the conditional training targets for the SB model as described in Section 3.2 and used in the loss function of Eq. 7."

We will also move some of the general formulation details from Section 3 to the appendix to improve the main manuscript's flow. To provide the intuition the reviewer requested, we will add a new figure to Sec 4.1 showing the learned transport process on a 2D toy example (e.g., translating from a two-moons distribution to an 8-Gaussians distribution). This will visually illustrate the latent evolution concept that underpins our method.

Q1 & Q2: Role of Grasp Pose $g_s$ and Encoding Input

This is a great catch. The $o_\mathrm{hand}$ in Figure 2 represents the visual observation of the hand, which is used as the input to the VAE encoder. The grasp configuration $g$ is used during VAE training as part of the reconstruction loss in Eq. 8, but it is not an input to the SB model. In Algorithms 2 and 3 (SB training and inference), $g_s$ is not used. The process is conditioned only on the encoded visual observation $o_s$. We will correct the algorithms to remove $g_s$ from the requirements for Latent SB training and inference to avoid confusion.

Q3: Metric on SO(3).

We apologize for the lack of clarity. We use the squared Frobenius norm between the two rotation matrices in SO(3). The 6D continuous representation is first converted to a $3\times 3$ matrix representation via Gram-Schmidt orthogonalization. The term is actually $\left\lVert R_{\mathrm{source}} - R_{\mathrm{target}} \right\rVert^2_F$ in Eq. 9. We will add this explicit definition to the paper.

Q4: Are the generated grasping poses validated in a physics simulator?

Yes. As detailed in our response to W1, all grasps are rigorously tested for stability in the IsaacGym physics simulator using a 6-axis perturbation protocol.

L1: Thorough discussion and analysis of the experimental outcomes.

We agree that a more thorough analysis would strengthen the paper. We will add the following discussion to Section 5 to better interpret the results from Table 1:

"Our results in Table 1 reveal a key trade-off between the different physics-informed objectives. The Jacobian-based cost yields the highest mean success rate (67.45%), suggesting it is most effective at capturing the kinematic relationship for producing stable grasps. However, the GWH and Contact costs achieve superior functional alignment, measured by 6D GWH IoU (14.63% and 14.54%, respectively). This indicates that directly optimizing for force-exertion capabilities (GWH) or contact patterns (Contact) is more effective for transferring the grasp's underlying mechanical intent, even if this sometimes results in a slightly lower stability score compared to a pure manipulability-based objective (Jacobian)."

Reference:

[1] Li et al., "Gendexgrasp: Generalizable dexterous grasping." ICRA 2023.

[2] Qin et al., "Anyteleop: A general vision-based dexterous robot arm-hand teleoperation system." RSS 2023.

[3] Yuan et al., "Cross-embodiment dexterous grasping with reinforcement learning." ICLR 2025.

[4] Khargonkar et al., "RobotFingerPrint: Unified Gripper Coordinate Space for Multi-Gripper Grasp Synthesis and Transfer." arXiv 2024.

Dear Authors and AC,

Thank you to the authors for the additional clarifications and to the AC for the follow-up questions. I also apologize for not articulating my concerns clearly earlier. I will lay out my concerns more clearly, taking the original paper, the rebuttal and clarifications, the authors’ stated plans into consideration.

Transfer quality and fidelity

• Functionality not convincingly preserved & Balancing functionality and visual similarity: While I understand that the method prioritizes preserving functionality when transferring grasps, the transferred poses often diverge substantially from the source. In Figure 1, the source grasp contacts the lid, body, and bottom of the bottle; the transferred grasp contacts only the lid and body, and the global hand orientation changes markedly. In Figure 4 (supp), the grasp pose of the third object shifts from one end to the other end. -- This degree of deviation harms not only visual similarity, but also functional similarity. -- How to use a hammer if the transferred pose has shifted from grasping its handle to grasping its head?
• Value of transfer vs. generic synthesis: If the goal is solely to achieve functionality, a generic grasp-synthesis method that produces physically valid grasps with functionality guidance or filtering might suffice. A transfer method should preserve at least coarse visual/affordance similarity to the source.
• Comparison to Dexonomy (RSS 2025; https://pku-epic.github.io/Dexonomy/): It synthesizes physically plausible grasps for diverse objects from human-demonstrated templates. Although it is not a transfer method per se, it seems capable of addressing the Grasp2Grasp problem by treating the source grasp as a template. Its results appear functionally and visually similar to the original grasps and physically plausible, whereas the Grasp2Grasp results lag in these aspects.

Physical plausibility

• The Shadow Hand example in Figure 1 appears physically unstable.

Limited evaluations

• Real-world evaluation: Real-world experiments are important and non-trivial. I cannot base my assessment on a promise to add them in the camera-ready version.

Metrics

• I am not convinced by the superiority claimed on the contact alignment metric. As seen in Figure 4 (supp.) and Figure 1, the contact regions shift substantially.

I appreciate the additional evaluations conducted during the rebuttal period. However, I remain unconvinced that the transfer quality sufficiently improves over prior work. Accordingly, I cannot raise my rating.

We sincerely thank the reviewer for their time and insightful feedback. We are encouraged that the reviewer found our formulation "innovative" and our contribution "valuable." We address each point below.

W1 & W2: On Claims, Datasets, and Generalization

We thank the reviewer for highlighting a potential ambiguity in our paper's motivation versus its direct claims. We agree that the introduction could be framed more precisely, and we will revise it accordingly.

Our Core Contribution: Translation without Paired Data

The central motivation of our work is to overcome the need for paired demonstration data for grasp translation. While single-hand grasp datasets can be generated efficiently with modern simulators (e.g., GraspIt! [1], GenDexGrasp [2]), creating a large-scale dataset of corresponding functionally equivalent grasps across two different hands is extremely expensive and often intractable.

Our method addresses this bottleneck. We use the MultiGripperGrasp dataset [3] for its rich collection of single-hand grasps, but our framework does not use or require any pre-existing pairings between them. Instead, the Schrödinger Bridge formulation, guided by our physics-informed OT designs, learns to discover these functional correspondences automatically from unpaired sets of grasps. This is analogous to machine translation: while monolingual text data is abundant, high-quality parallel corpora (paired sentences) are scarce and expensive. Our work provides a new perspective on how to scale up grasp translation by leveraging readily available, unpaired grasp examples.

Clarification on Unseen Hand Generalization

We also wish to clarify that generalizing to unseen hands is a long-term motivation for this research area, not a claimed contribution of our current model. We believe our approach is a meaningful step towards this motivation. The current implementation does not generalize to unseen hands at inference time because our VAE decoder learns the hand's kinematic structure implicitly from the training data. To generate a grasp for a new hand, the decoder would need to be retrained or fine-tuned on data from that hand. We acknowledge this limitation and view the development of learning-based methods that can generalize to new hand kinematics on-the-fly as an important and open research direction.

To address these points and prevent misunderstanding, we will revise the introductory section (Lines 16-22) as follows:

"A key challenge in data-driven grasp synthesis is the heavy reliance on large-scale datasets, which are often curated for a specific hand through computationally expensive simulation or real-world trials. Transferring grasp knowledge to a new robotic hand is non-trivial, as models trained for one hand often require complete retraining or significant fine-tuning for another, which poses a practical bottleneck. A particularly challenging aspect is the acquisition of paired demonstration data, where a grasp for a source hand is mapped to a corresponding functional grasp for a target hand. Manually creating or algorithmically discovering such pairings at scale is extremely difficult. This motivates the need for a framework that can learn to translate the intent of a grasp from a source hand to a target hand using only unpaired, hand-specific grasp datasets while leveraging the underlying object geometry and physics to bridge the morphological gap."

W3: Tables and figures are not self-contained.

Thank you for this feedback. We will improve the clarity of our figures. For Figure 3, the labels "Pose," "Contact," "GWH," and "Jacobian" refer to variants of our model trained with the different OT cost functions described in Section 4.2. We will add the following sentence to the caption to make it self-contained:

"The 'Pose', 'Contact', 'GWH', and 'Jacobian' columns show results from our method when trained using the respective OT cost functions from Sec. 4.2."

We will perform another round of proofreading on all figures and tables to ensure they are self-contained and easy to understand.

W4: Lack real-world experiments

We agree that real-world validation is a crucial step. Given the rebuttal timeframe, we were unable to conduct and document physical experiments. Our current research has focused on the core algorithmic contributions and their validation in a high-fidelity simulator. We fully recognize the importance of this aspect and are committed to including real-world experiments in future work.

However, to further strengthen the paper's evaluation based on the reviewers' feedback, we have conducted extensive new simulation experiments, including comparisons to two additional strong baselines (Dex-Retargeting [4], CrossDex [5]), evaluation with a new pose alignment metric, a detailed efficiency report, and evaluation on the remaining three transfer settings (A→H, S→H, and A→S). Please refer to Table 1-5 of our response to reviewer AEMy and Table 2-4 of our response to reviewer 4u79 if interested. We believe these additions strengthen our empirical results and address key questions about performance and applicability.

Q1: Human hand as source and extension to real human data.

This is an excellent question. The human hand in our experiments is a fully articulated 3D model derived from MANO [6]. Specifically, we use distinct part meshes for each link, which are kinematically chained. While this means the complete mesh surface can appear discontinuous at the joints, this has little negative impact on our method. The input to our VAE is a point cloud sampled from the surfaces of these parts, which is an inherently discrete representation. Thus, the visual difference between a point cloud sampled from our articulated model versus one from a continuous, skinned MANO model is not significant.

Therefore, our method can be directly applied to real human hand data, assuming that a system can provide a full, segmented point cloud of the hand grasping the object. We will provide some thoughts on how to obtain that in our response to your next question.

Q2: Input modality to VAE

We apologize for the confusion. The input to the VAE is a full hand point cloud, as shown in the architecture diagram (Figure 2). We will clarify this in the text.

We acknowledge that obtaining a complete, segmented point cloud in the real world is challenging due to occlusion. This is indeed a common problem in robotic perception. Standard solutions include:

1. Multi-camera systems to fuse views and reconstruct a more complete point cloud, which is often seen in real-world experimental setups.
2. Pose estimation: If 3D models of the object and hand are known, the problem can be reduced to a 3D hand-object pose estimation (HOPE) task, which is generally well studied with 3D inputs [7]. Once the poses are estimated, complete point clouds can be generated from the models and used as input to our system.

Q3: Regarding contact anchor

Yes, your understanding is correct. Contact anchors are required at inference. They are derived directly from the assumed input: the segmented source hand and object point clouds. The procedure is deterministic and computationally inexpensive:

1. For each point on the object, we calculate its minimum distance to the source hand point cloud.
2. Points with a distance below a small threshold (e.g., 0.005m in our experiments) are identified as the contact map on the object.
3. We then apply Farthest Point Sampling (FPS) to this contact map to select the 5 contact anchors.

This process requires no additional information beyond the input point clouds that are already conditioned upon. We will clarify this procedure in the implementation details.

Reference:

[1] Miller and Allen. "Graspit! a versatile simulator for robotic grasping." RA-M 2004.

[2] Li et al., "Gendexgrasp: Generalizable dexterous grasping." ICRA 2023.

[3] Casas et al., "Multigrippergrasp: A dataset for robotic grasping from parallel jaw grippers to dexterous hands." IROS 2024.

[4] Qin et al., "Anyteleop: A general vision-based dexterous robot arm-hand teleoperation system." RSS 2023.

[5] Yuan et al., "Cross-embodiment dexterous grasping with reinforcement learning." ICLR 2025.

[6] Romero et al., "Embodied hands: Modeling and capturing hands and bodies together." SIGGRAPH ASIA 2017.

[7] Hampali et al., "Honnotate: A method for 3d annotation of hand and object poses." CVPR 2020.

We sincerely thank the reviewer for their time and insightful feedback. We are encouraged that the reviewer found our formulation "novel and interesting" and our presentation "of high quality." We address the identified weaknesses below.

W1-3: Regarding pose alignment metric, efficiency report, and comparison with Dex-Retargeting and CrossDex.

We appreciate the reviewer's suggestions regarding evaluation metrics, efficiency, and comparisons to other relevant work. We group these points as they all relate to the application context of our method.

Clarification: Grasp Translation vs. Teleoperation

While our work is technically related to teleoperation, we wish to clarify that our formulation of grasp translation has a fundamentally different goal.

• Teleoperation is typically manipulator-oriented, often relying on a human-in-the-loop for perception and feedback, and aims to mimic the source agent's actions as closely as possible.
• Our formulation of grasp translation is object-centric and aims to transfer the functional intent of a grasp, which does not necessarily require a human source and allows for morphologically and posturally different but functionally equivalent outcomes.

For example, to turn on a TV remote, a source grasp might be from below, palm up, with the thumb on the power button. A teleoperation system would try to replicate this exact pose. Our method, however, recognizes the functional goal (applying force to the power button while supporting the remote) and could generate a target grasp from above, palm down, with the index finger on the button and other fingers providing support. We consider both grasps functionally equivalent.

Back to your question about metrics, in our problem setting, we can imagine that two functionally equivalent grasps are likely to induce similar contact patterns (like in the remote example, contact at the bottom to fight gravity and at the on/off button for functionality) and/or wrench profiles (force and torque the grasp can exert). Similar contact patterns often lead to similar wrenches, as one can only exert force during contact, but similar wrenches do not necessarily indicate a similar contact pattern. You can imagine that flipping a water bottle requires torques about any axis on the horizontal plane, where any pinches on the side of the object are likely to induce these types of torques.

This distinction motivates our choice of the 6D Grasp Wrench Hull (GWH) IoU as a primary alignment metric. Functionally equivalent grasps should be able to exert similar forces and torques on the object, which is precisely what GWH measures.

Additional Experiments: Baselines and Metrics

As per the reviewer's suggestion, we have conducted new experiments to include the positional retargeting module from Dex-Retargeting [1] as baselines. We also tweak the retargeting neural network from CrossDex [2] to be trained with generated data from positional retargeting to output both wrist pose and finger joint configuration as an additional baseline. Further, we introduce a new Contact Alignment metric as a proxy to pose alignment metric, inspired by Dex-Retargeting's positional optimization objective, which measures the squared Chamfer distance between source and target grasp contact maps.

A note on baselines: Both Dex-Retargeting and CrossDex require pre-defined link-to-link correspondences between the source and target hands. This information is not typically available in large-scale grasp datasets, and we had to manually annotate it for these comparisons. This requirement limits their generality. For instance, they cannot retarget to a hand with more fingers than the source. Our method has no such constraints.

The results, shown in the tables below, demonstrate that our method significantly outperforms these new baselines across all key metrics.

Table 1: Comparison on Success Rate (%)↑

Table 2: Comparison on Diversity (rad)↑

Table 3: Comparison on 6D GWH IoU (%)↑

Table 4: Comparison on Contact Alignment (cm^2)↓

Our analysis suggests the baselines struggle because their IK-style optimization focuses only on select distal links, ignoring potential collisions from other parts of the hand (e.g., the palm), which leads to physically invalid grasps. The CrossDex model, trained on this suboptimal data, inherits these limitations and fails to generalize due to its simple architecture and small size (4-layer MLP). We will integrate these results into the final paper and add this discussion to the related works.

Efficiency Report

We provide a more detailed inference time report below. While our method is designed primarily as an offline planner, its amortized inference time is on the same order of magnitude as the baselines and significantly faster than the optimization-based RobotFingerPrint [3]. CrossDex Retargeting network is significantly faster due to its simple architecture, and it only needs to do the forward pass once.

Table 5: Average Inference Time per Grasp, Amortized (s)

W4: Regarding real-world experiments.

We agree that real-world validation is a crucial step to demonstrate the full impact of our work. Given the short rebuttal period, we were unable to conduct these experiments. However, our current work focuses on the core algorithmic contributions for grasp translation, which we believe are sound and well-supported by our extensive simulation results. We are fully committed to incorporating real-world experiments in future work.

Reference:

[1] Qin et al., "Anyteleop: A general vision-based dexterous robot arm-hand teleoperation system." RSS 2023.

[2] Yuan et al., "Cross-embodiment dexterous grasping with reinforcement learning." ICLR 2025.

[3] Khargonkar et al., "RobotFingerPrint: Unified Gripper Coordinate Space for Multi-Gripper Grasp Synthesis and Transfer." arXiv 2024.