acceptodds
Under review as a conference paper at ICLR 2027

DittoHand: Zero-Shot Grip-Force Control Across Hand Morphologies with Surface-Aware Actuator Tokens

Abstract

To a typical cross-embodiment policy, a hand is its kinematic structure; two hands that differ only in finger thickness are the same hand. Yet the same actuator command meets the object earlier on the thicker finger, at a different point and along a different normal. Existing cross-hand approaches either leave the hand–object geometry implicit or recover it with a retargeting step trained per hand. A policy that leaves this collision geometry implicit can often find a grasp, but it keeps that grasp less reliably and barely tracks the commanded grip force. We propose DittoHand, which treats a hand as a set of actuators and gives each an actuator token carrying the collision surface it moves and the closing rate at which that surface approaches the object as the actuator turns, so an action means the same thing on every hand. Trained on 24 hands, one policy is tested on them and zero-shot on 648 unseen hands, with 20 objects of which 17 are unseen in training. On both hand sets it leads two cross-hand baselines, most widely in keeping the object (47.5% vs. at most 34.3% on unseen hands) and in tracking the grip command, where it beats each baseline on at least 99.8% of the unseen hands. In a virtual reality (VR) study on three hands DittoHand had never seen, 12 participants rated it significantly more controllable and realistic overall than ForceGrip policies trained for each hand.

Primary area:
other topics in machine learning (i.e., none of the above)
Keywords:
cross-embodiment learning, dexterous grasping, grip force control, zero-shot morphology generalization, reinforcement learning, VR avatar hands

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_fE7iOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes DittoHand, a cross-morphology policy for zero-shot grip-force control. Each actuator is represented by a token encoding its kinematic state, the collision surface it moves, local hand-object geometry, and an analytic closing rate. A shared Transformer then produces per-actuator actions for hands with different topologies and shapes. Trained on 24 simulated hands, DittoHand is evaluated on 648 unseen hands and 20 objects, outperforming adapted cross-hand baselines in grasp stability and force tracking. Ablations show the importance of surface information, and a 12-participant VR study reports improved perceived controllability and realism over per-hand ForceGrip policies.

Critical strengths
  1. A physically motivated, interpretable representation. Binding each actuator to the surface it moves and giving it an analytic closing rate tells the policy how actuator motion changes contact geometry and force transmission, which representations based only on joint topology leave implicit. This is the paper's main conceptual contribution.
  2. Broad controlled simulation evaluation. The paper evaluates 24 seen hands, 648 unseen hands, topology changes, thickness and length changes, 20 objects, surface/closing-rate/descendant ablations, and a VR user study. The large surface ablation gap supports the importance of the proposed representation within the simulator.
Critical weaknesses
  1. The method has not yet been evaluated on any existing hand. All 864 hands, including those in the VR study, are virtual edits of the MANO human-hand model. This is reasonable for the VR avatar setting, but the paper also describes collision geometry as "a necessary axis of embodiment" and compares itself with GET-Zero, which was tested on physical LEAP Hand variants. The broader claim would therefore benefit from evidence beyond MANO-derived hands, together with a discussion of how the per-phalanx contact-force, grip-force and object-velocity inputs could be obtained on hardware. This is my main concern; I would be glad to raise my score if the claim were scoped to VR avatars, or if grasping and force tracking are also evaluated on existing robot hands, even in simulation.
  2. The VR study relies mainly on subjective ratings and only partly tests the central claims. Objective measures such as task success, object drops, completion time and how closely F_grip followed the trigger would strengthen it, and could be reported if they were recorded. The pick-and-place task uses the three training objects and mostly requires holding the object, so it does not test how closely the grip force tracks the command, which is the paper's main quantitative advantage.
Questions for authors

How many training seeds were used? Can you report mean ± std or CIs over seeds for Tables 1–2? What is the wall-clock cost per control step of the surface-point queries and the network on the VR hardware, and how does it scale with object mesh complexity or with two hands?

7 comments
AuthorsOpenReview ↗

Response to Reviewer fE7i (1/4)

Thank you for your feedback, which greatly helped improve our paper. Thanks to it, we scoped our claims to VR avatar hands and strengthened them with evaluations on non-MANO avatar hands and a new force-control user study. Below, we respond to each of your questions.

1. Scope of the claim and evidence beyond MANO (W1)

We agree with your point. This study focuses on VR avatar hands with varying geometries, aiming to enable each actuator to regulate grip force by accounting for the geometry of the surface it controls. Transferring the method to physical robotic hands or arbitrary articulated grippers would require additional consideration of actuator dynamics, physical constraints, and the sim-to-real gap, and is therefore beyond the scope of this work.

Accordingly, we will revise "a hand" to "an avatar hand" in the Introduction and modify expressions such as "a necessary axis of embodiment" that could imply a broader generalization scope than our results support. To ensure consistency throughout the paper, we will also revise the relevant statements in the Abstract, the third contribution (Contribution 3), the Conclusion, and the Limitations section. In addition, our comparison with GET-Zero concerned the methodology for representing different hand structures, not performance on physical robotic hands.

However, we believe it is important to demonstrate that our method is not limited to MANO within this scope. We therefore evaluated five publicly available non-MANO VR avatar hands (Rocketbox male and female, Ultraleap, and two SteamVR glove hands) without retraining. We preserved their original meshes, joint positions, sizes, and masses while applying the same 21-DoF configuration as the training hands. We used the same joint evaluation as in Tables 1 and 3 (Section 3 of this response). Across the five avatar hands, Ours achieved 44.8% Success, comparable to unseen MANO hands (42.3%), with FTE and Gain also at similar levels (Table R3-1). The hand shapes are shown in Appendix B.6.

Table R3-1. VR avatar hands not derived from MANO (Ours without retraining, in the same combined experiment as Tables 1 and 3. The first two rows repeat the MANO-derived Seen and Unseen hands. All values are averaged over three training seeds; ± is the standard deviation over seeds for the grouped rows.)

HandsSuccess (%) ↑FTE (kgf) ↓Gain → 1
Seen MANO hands (24)62.9 ± 2.60.51 ± 0.070.52 ± 0.13
Unseen MANO hands (648)42.3 ± 4.50.68 ± 0.070.46 ± 0.12
Non-MANO avatar hands (unseen)
Rocketbox, male37.00.660.42
Rocketbox, female41.20.660.42
Ultraleap57.20.530.54
SteamVR glove41.50.910.65
SteamVR glove, slim47.00.800.69
Avatar hands (5)44.8 ± 6.40.71 ± 0.090.54 ± 0.13
AuthorsOpenReview ↗

Response to Reviewer fE7i (2/4)

2. Objective measures and force tracking in the VR user study (W2)

As you said, the previous VR user study focused on subjective ratings, and the pick-and-place task alone was not sufficient to verify force-command tracking. In fact, most participants performed the task with the trigger pulled nearly all the way (trigger ≥ 0.95: DittoHand 97%, ForceGrip 91%). To address this, we additionally analyzed objective measures from the previous study and conducted a new user study that directly evaluates force-command tracking.

Objective measures in pick-and-place. Analyzing the records of the original 12 participants, DittoHand had a higher success rate than ForceGrip (98.1% vs. 92.5%), a shorter completion time (11.2 s vs. 12.9 s), and a lower timeout rate (0.9% vs. 5.9%). All three measures were statistically significant (Table R3-2).

Table R3-2. Objective measures in pick-and-place (12 participants, three hands pooled, mean ± standard deviation, Wilcoxon signed-rank test, Holm correction)

Success rate (%) ↑Drops per trial ↓Completion time (s) ↓Timeouts (%) ↓
ForceGrip (trained per hand)92.5 ± 9.60.0912.9 ± 3.55.9
Ours (unseen hands)98.1 ± 3.30.0611.2 ± 3.20.9
p0.0350.4730.0020.035

The per-hand results are added to Table 15 in Appendix C of the revision.

Additional Force-Control Experiment. Inspired by ForceGrip's Can Squeeze experiment, we designed a task in which users match target grip forces using the trigger. Unlike the original experiment, we directly specified target forces and evaluated force controllability across different hand structures. The same 12 participants returned and held a capsule object (5 cm diameter, 10 cm length, 0.1 kg mass), matching displayed target forces of 0.8, 1.2, 1.6, and 2.0 kgf for 2 seconds each.

Table R3-3. Force-control task (12 participants. Mean ± standard deviation over participants, each participant averaged over the three hands; paired Wilcoxon signed-rank test. ForceGrip is trained on each hand; our policy saw none of them.)

MeasureForceGrip (per hand)Ours (unseen)p
Command following (primary)
Trigger–grip-force r (ForceGrip's method) ↑0.58 ± 0.090.81 ± 0.05< 0.001
Gain → 10.40 ± 0.090.78 ± 0.15< 0.001
Relative tracking error |F − C| / C ↓0.77 ± 0.260.29 ± 0.07< 0.001
Task measures
Time to reach the target (s) ↓0.95 ± 0.390.69 ± 0.220.021
Hit rate within half a level (%) ↑89.1 ± 5.694.0 ± 4.00.016
Drops per block ↓9.1 ± 2.42.4 ± 1.7< 0.001
Questionnaire (−3 to +3)
Realism Questionnaire total ↑−0.04 ± 0.520.78 ± 0.800.001
Grip-force items G1–G4 ↑0.09 ± 0.731.22 ± 0.970.005

Table R3-4. Force-control task per hand (12 participants; mean ± standard deviation over participants)

HandGain → 1, ForceGripGain → 1, OursDrops per block ↓, ForceGripDrops per block ↓, Ours
Mean MANO (β = 0)0.69 ± 0.180.80 ± 0.262.4 ± 2.41.0 ± 1.3
Thickness ×1.150.40 ± 0.080.75 ± 0.126.1 ± 3.42.7 ± 2.7
Three-fingered0.10 ± 0.150.79 ± 0.2118.8 ± 5.23.7 ± 2.3
  • Force-command tracking. The Gain of DittoHand is 0.78, much closer to 1 than that of ForceGrip (0.40), and its error relative to the command is also smaller, 0.29 vs. 0.77 for ForceGrip.
  • Differences across hand structures. Even on the mean MANO hand, on which ForceGrip was trained alone, DittoHand has a higher Gain of 0.80 than ForceGrip (0.69). Although ForceGrip was trained separately for each hand, it drops to 0.40 on the hand with changed thickness and to 0.10 on the three-fingered hand. DittoHand never saw any of the three hands but maintains 0.75–0.80 on all three. In the questionnaire as well, it received significantly higher scores on the RQ Total Score (0.78 vs. −0.04) and all three subscores (Figure 5b in the revision), as well as on the grip-force items (1.22 vs. 0.09).

Thus, DittoHand demonstrated advantages not only in stable object transport but also in accurately tracking user force commands. In particular, it maintained high force-tracking performance across different unseen hand structures, supporting our claim of consistent force control across diverse avatar hands. We will add this experiment to Section 4.5 and report the objective metrics in Appendix C.

AuthorsOpenReview ↗

Response to Reviewer fE7i (3/4)

3. Number of training seeds and mean ± standard deviation (Q1–2)

While the original submission reported results from a single seed for each method, we have now independently trained Ours, DGZ, and IBS with three seeds each, and updated Tables 1 and 3 to report mean ± standard deviation. We also made two changes to the evaluation conditions to better reflect real-world VR environments.

  • Joint Transport and Force Tracking: We previously evaluated transport and force tracking separately because objects slip when transported without sufficient grip force. This is also intended behavior in VR, where objects slip if users grip them too weakly. However, the separate evaluations did not sufficiently support our claim of transporting objects while tracking varying force commands, so we added a joint experiment.
  • Grasp Criterion: We replaced the previous static grasp test with an evaluation of whether the object remains held for 0.5 seconds after support removal under the actual simulation friction coefficient (0.5).

Table 1 of the revision. Comparison with the baselines (Mean ± standard deviation over three training seeds; Grasp and Success in %, FTE in kgf. Rows below Unseen hands differ from the Seen hands in one aspect; Topology means fewer fingers or phalanges. Bold: best per row and metric.)

(a) Keeping the object

HandsGrasp DGZGrasp IBSGrasp OursSuccess DGZSuccess IBSSuccess Ours
Seen hands (24)79.1 ± 3.575.1 ± 4.783.2 ± 3.453.3 ± 1.442.4 ± 1.462.9 ± 2.6
Unseen hands (648)61.6 ± 2.259.6 ± 3.370.6 ± 4.633.9 ± 4.428.0 ± 3.742.3 ± 4.5
β only (24)78.5 ± 3.074.8 ± 4.282.3 ± 3.652.1 ± 1.541.2 ± 1.960.4 ± 2.9
Thickness only (96)74.8 ± 4.168.2 ± 4.678.1 ± 3.946.6 ± 3.635.5 ± 1.154.7 ± 3.5
Length only (48)76.6 ± 6.670.1 ± 4.679.2 ± 1.456.4 ± 8.744.6 ± 6.564.4 ± 5.0
Topology only (48)53.0 ± 0.550.8 ± 4.464.8 ± 5.322.6 ± 3.318.2 ± 3.532.4 ± 5.0

(b) Tracking the grip command

HandsFTE DGZFTE IBSFTE OursGain DGZGain IBSGain Ours
Seen hands (24)0.57 ± 0.030.71 ± 0.100.51 ± 0.070.48 ± 0.010.21 ± 0.210.52 ± 0.13
Unseen hands (648)0.76 ± 0.020.86 ± 0.090.68 ± 0.070.42 ± 0.040.19 ± 0.170.46 ± 0.12
β only (24)0.57 ± 0.030.71 ± 0.100.52 ± 0.060.48 ± 0.020.23 ± 0.200.53 ± 0.13
Thickness only (96)0.68 ± 0.040.90 ± 0.090.64 ± 0.060.47 ± 0.020.23 ± 0.220.49 ± 0.13
Length only (48)0.64 ± 0.030.92 ± 0.190.68 ± 0.110.62 ± 0.020.33 ± 0.270.64 ± 0.15
Topology only (48)0.82 ± 0.020.88 ± 0.070.69 ± 0.080.30 ± 0.020.12 ± 0.130.37 ± 0.12

Table 3 of the revision. Ablations (Grasp and Success in %, FTE in kgf. Mean ± standard deviation over three training seeds. Bold: best per column.)

MethodSeen Grasp ↑Seen Success ↑Seen FTE ↓Seen Gain → 1Unseen Grasp ↑Unseen Success ↑Unseen FTE ↓Unseen Gain → 1
Ours83.2 ± 3.462.9 ± 2.60.51 ± 0.070.52 ± 0.1370.6 ± 4.642.3 ± 4.50.68 ± 0.070.46 ± 0.12
−κ76.2 ± 1.854.9 ± 3.40.65 ± 0.050.36 ± 0.0360.4 ± 2.134.3 ± 1.00.82 ± 0.030.30 ± 0.01
−desc78.5 ± 3.655.6 ± 5.00.71 ± 0.110.16 ± 0.2767.7 ± 4.538.7 ± 5.90.85 ± 0.090.15 ± 0.26
−surf69.3 ± 1.533.5 ± 1.50.78 ± 0.020.20 ± 0.0455.6 ± 10.722.2 ± 7.00.89 ± 0.070.20 ± 0.02

Thus, Ours maintained its performance advantage across three seeds and under the revised evaluation conditions, supporting our claim that the surface-based representation, including κ and descendant, contributes to generalization to unseen hands.

Reviewer_BsqTOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies force control for grasping in simulation. It proposes a representation that represents each actuator using it's joint angle and collision surface so that hands with different morphology can be trained together with a shared policy. The authors show quantitative experiments on object transportation task and show better task success rate and force tracking results than baselines. They also conducted user-study on how easy the method can be controlled by using VR teleoperation.

Critical strengths

The representation addresses a concrete problem. First, only using joint states as the representation cannot describe the surface shape when contacting with the object. Second, this representation enables policy training on different embodiments. And the authors also show improved success rate on unseen hands, which is quite interesting.

The ablation experiments are comprehensive. The authors clearly demonstrate the improved task sucess and force-tracking accuracy, which demonstrates the usefulness of force tracking.

The VR study also tests whether the controller is useful for its stated application. The comparison on unseen hands provides supporting evidence about perceived controllability and realism.

Critical weaknesses

My main concern is the gap between the task definition and the quantitative evaluation. The intended behavior is to follow a variable force command. However, the transport experiment instead fixes the command at its maximum action after the initial ramp. On the other hand, the force-tracking test keeps the wrist stationary and object supported. These are useful experiments but does not fully show the controller can satisfy both constraints together.

The baseline comparion mixes representation and adaptation choices. It's useful to see the comparisons with IBS and DGZ, but they are less useful than the within-model ablations. The adapted methods differ in observations and action space, and IBS-adapted requires an adaptation that changes how point displacements become actuator commands. Its very low success on seen hands also raises a practical concern.

The generalization and precision claim might need careful wording. The 648 unseen hands are structured variants of 36 underlying MANO shapes, rather than 648 independently designed embodiments. The evaluation also uses precomputed feasible object placements. These choices are reasonable for testing transfer under controlled changes, but they make the result conditional on a reachable initial grasp and a particular family of hand constructions. The paper should not imply unrestricted transfer to arbitrary articulated grippers or arbitrary object placements.

In addition, there are no real-world results.

Questions for authors

The most useful additional result would be a combined transport-and-force experiment. Can the authors vary the requested force while moving an unsupported object and report both tracking error and task success rate?

Can the authors show the learning curves and variation across different runs for Tables 1 and 2? In particular, what explains IBS-adapted's low seen-hand success, and how sensitive is it to its adaptation and tuning choices? This would help distinguish a representation limitation from a difficulty in training the adapted controller.

4 comments
AuthorsOpenReview ↗

Response to Reviewer BsqT (1/3)

Thank you for suggesting the combined transport-and-force experiment. Thanks to it, we could directly test our claim in a single experiment, and it is now the main evaluation in the revision. Below, we respond to each of your questions.

1. An experiment that satisfies transport and force tracking together & seeds (W1, Q1, Q2 run-to-run variation)

You are right. We separated the two tasks because transporting an object without sufficient grip force causes it to slip. In VR, objects also slip when users grip them too weakly, and we intended this behavior. Nevertheless, the separate experiments did not sufficiently support our claim of transporting objects while tracking varying force commands, so we conducted the suggested joint experiment.

  • Lower Bound of 0.79 kgf: During transport, the grip must remain strong enough to prevent slipping, regardless of how low the command is. Under friction, the grip force must satisfy μFgrip≥m∣g−aobj∣\mu F_{\mathrm{grip}} \geq m|\mathbf{g}-\mathbf{a}_{\mathrm{obj}}| (m=0.1m=0.1 kg, effective simulation friction μ=0.5\mu=0.5). Holding the object stationary requires 0.20 kgf, while wrist motion increases the required force depending on object acceleration. We sampled 200,000 random wrist motions from training Stage 4 (50–100 cm/s, 270–540°/s) and calculated the required grip force for each. The maximum was 0.79 kgf. Since commands below this threshold would effectively require releasing the object, we used 0.79 kgf as the lower bound. The rationale and calculation of this lower bound are added to Section 4.1 and Appendix B.4 of the revision.

Table 1 of the revision. Comparison with the baselines (Mean ± standard deviation over three training seeds; Grasp and Success in %, FTE in kgf. Rows below Unseen hands differ from the Seen hands in one aspect; Topology means fewer fingers or phalanges. Bold: best per row and metric.)

(a) Keeping the object

HandsGrasp DGZGrasp IBSGrasp OursSuccess DGZSuccess IBSSuccess Ours
Seen hands (24)79.1 ± 3.575.1 ± 4.783.2 ± 3.453.3 ± 1.442.4 ± 1.462.9 ± 2.6
Unseen hands (648)61.6 ± 2.259.6 ± 3.370.6 ± 4.633.9 ± 4.428.0 ± 3.742.3 ± 4.5
β only (24)78.5 ± 3.074.8 ± 4.282.3 ± 3.652.1 ± 1.541.2 ± 1.960.4 ± 2.9
Thickness only (96)74.8 ± 4.168.2 ± 4.678.1 ± 3.946.6 ± 3.635.5 ± 1.154.7 ± 3.5
Length only (48)76.6 ± 6.670.1 ± 4.679.2 ± 1.456.4 ± 8.744.6 ± 6.564.4 ± 5.0
Topology only (48)53.0 ± 0.550.8 ± 4.464.8 ± 5.322.6 ± 3.318.2 ± 3.532.4 ± 5.0

(b) Tracking the grip command

HandsFTE DGZFTE IBSFTE OursGain DGZGain IBSGain Ours
Seen hands (24)0.57 ± 0.030.71 ± 0.100.51 ± 0.070.48 ± 0.010.21 ± 0.210.52 ± 0.13
Unseen hands (648)0.76 ± 0.020.86 ± 0.090.68 ± 0.070.42 ± 0.040.19 ± 0.170.46 ± 0.12
β only (24)0.57 ± 0.030.71 ± 0.100.52 ± 0.060.48 ± 0.020.23 ± 0.200.53 ± 0.13
Thickness only (96)0.68 ± 0.040.90 ± 0.090.64 ± 0.060.47 ± 0.020.23 ± 0.220.49 ± 0.13
Length only (48)0.64 ± 0.030.92 ± 0.190.68 ± 0.110.62 ± 0.020.33 ± 0.270.64 ± 0.15
Topology only (48)0.82 ± 0.020.88 ± 0.070.69 ± 0.080.30 ± 0.020.12 ± 0.130.37 ± 0.12

Table 3 of the revision. Ablations (Grasp and Success in %, FTE in kgf. Mean ± standard deviation over three training seeds. Bold: best per column.)

MethodSeen Grasp ↑Seen Success ↑Seen FTE ↓Seen Gain → 1Unseen Grasp ↑Unseen Success ↑Unseen FTE ↓Unseen Gain → 1
Ours83.2 ± 3.462.9 ± 2.60.51 ± 0.070.52 ± 0.1370.6 ± 4.642.3 ± 4.50.68 ± 0.070.46 ± 0.12
−κ76.2 ± 1.854.9 ± 3.40.65 ± 0.050.36 ± 0.0360.4 ± 2.134.3 ± 1.00.82 ± 0.030.30 ± 0.01
−desc78.5 ± 3.655.6 ± 5.00.71 ± 0.110.16 ± 0.2767.7 ± 4.538.7 ± 5.90.85 ± 0.090.15 ± 0.26
−surf69.3 ± 1.533.5 ± 1.50.78 ± 0.020.20 ± 0.0455.6 ± 10.722.2 ± 7.00.89 ± 0.070.20 ± 0.02

Ours is best in 23 of the 24 cells of Table 1, and in Table 3 Ours is best on all four metrics for both the seen and unseen hands.

AuthorsOpenReview ↗

Response to Reviewer BsqT (2/3)

2. Questions on the baselines (W2, Q2)

Common Experimental Conditions. DGZ-adapted and IBS-adapted were trained and evaluated in the same VR grasping environment as Ours, with identical training and evaluation hands and objects, episode configurations, curricula, reward functions, training budgets, and evaluation metrics. To match our task, where users directly control wrist motion and specify target grip force through a trigger, we excluded wrist control from policy outputs and trained both baselines to track commanded grip force. Both baselines received identical global observations and differed only in their core representations and network architectures.

IBS-adapted. IBS represents spatial relationships between the hand and object, but directly using its original policy is unsuitable for our zero-shot evaluation on unseen hands. To compare the geometric representations themselves, we replaced only the surface representation with IBS. Thus, this baseline compares IBS and our proposed surface representation under the same policy architecture, rather than reproducing the entire original IBS policy.

  • As you noted, IBS-adapted shows low Success even on seen hands. We believe one reason is that different hands require different joint motions to produce the same change in the surface–object relationship. IBS provides the current hand–object spatial relationship but does not explicitly represent how much that relationship changes when each joint rotates. Thus, when training a shared policy across multiple hands, similar spatial relationships may require different joint rotations for each hand, increasing the burden of learning these mappings separately. This can affect performance on both seen and unseen hands.
  • In contrast, our κ provides the rate of change in surface–object distance with respect to joint rotation, representing the effects of different joint motions across hands on a common basis. Removing κ reduced Success from 62.9% to 54.9% on seen hands and from 42.3% to 34.3% on unseen hands (Table 3). These results support our claim that explicitly representing how joint motion affects contact relationships benefits shared-policy learning and transfer to unseen hands.
  • To examine whether IBS's low performance resulted from training instability, we added multi-seed training curves to Section 4.3 of the revision (Figure 4). IBS showed low variation in Success across seeds (42.4 ± 1.4%) and converged stably at a lower level than Ours. This suggests that the performance gap cannot be explained solely by a failed seed or training instability.

Additionally, the submitted IBS Gain of 0.00 resulted from a single seed that learned a policy unresponsive to force commands. IBS in Table 1 is the mean of three runs excluding this seed, and IBS underperformed Ours across all metrics on both seen and unseen hands.

AuthorsOpenReview ↗

Response to Reviewer BsqT (3/3)

3. Scope of generalization and object placement (W3, W4)

We agree with your assessment of the generalization scope. This paper focuses on VR avatar hands with different geometries. Our goal is to enable each actuator to account for its own geometry when controlling grip force. Transfer to arbitrary articulated grippers is beyond our scope. We will therefore narrow the wording in the Introduction from "a hand" to "an avatar hand" and leave transfer to physical robots as a limitation and future work. (reflected in Sections 1 and 5 of the revision)

Initial Wrist Pose. Since our study focuses on finger grasp control rather than wrist control, we used FSG to initialize the wrist in a grasp-suitable pose during simulation evaluation. This reduces the effect of initial wrist position and allows consistent comparison of finger policy performance. However, FSG is not required for VR grasping. In actual VR, users can directly move their wrists to approach objects and adjust their poses. Thus, while FSG helps control evaluation conditions, it does not establish robustness to arbitrary initial wrist poses. To address this, we also evaluated our method in a VR user study where users directly controlled their wrists, achieving a 98.1% success rate (ForceGrip: 92.5%).

Specific Hand Family. As you noted, all 648 hands are variants derived from MANO. However, this was to provide a common basis for training and evaluation (e.g., length and thickness changes), and our method is not limited to MANO hands, as the policy takes tokens computed directly from the hand assets' joints and collision geometry, rather than MANO shape parameters. To verify this, we evaluated five publicly available non-MANO VR avatar hands (male and female Rocketbox hands, Ultraleap, and two SteamVR glove models) using the same combined evaluation protocol, without retraining (Table R2-1).

Table R2-1. VR avatar hands not derived from MANO (Ours without retraining, in the same combined experiment as Tables 1 and 3. The first two rows repeat the MANO-derived Seen and Unseen hands. All values are averaged over three training seeds; ± is the standard deviation over seeds for the grouped rows.)

HandsSuccess (%) ↑FTE (kgf) ↓Gain → 1
Seen MANO hands (24)62.9 ± 2.60.51 ± 0.070.52 ± 0.13
Unseen MANO hands (648)42.3 ± 4.50.68 ± 0.070.46 ± 0.12
Non-MANO avatar hands (unseen)
Rocketbox, male37.00.660.42
Rocketbox, female41.20.660.42
Ultraleap57.20.530.54
SteamVR glove41.50.910.65
SteamVR glove, slim47.00.800.69
Avatar hands (5)44.8 ± 6.40.71 ± 0.090.54 ± 0.13

We preserved the original meshes, joint positions, sizes, and masses while applying the same 21-DoF configuration as the training hands. Across the five avatar hands, Ours achieved 44.8% Success, comparable to unseen MANO hands (42.3%), with similar FTE (0.71 kgf) and Gain (0.54). These results demonstrate that DittoHand can generalize to VR avatar hands with diverse geometries beyond MANO.

AuthorsOpenReview ↗

Dear Reviewer BsqT

Thank you again for your time and constructive feedback. As the discussion period nears its end, we wanted to follow up to see whether our response and revisions, in particular the combined transport-and-force experiment you suggested, have addressed your concerns.

We would be grateful for any further thoughts and are happy to provide additional clarification if needed.

Thank you for your time and consideration:>

Reviewer_zTiHOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2(+1)
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes DittoHand, a single reinforcement-learning policy that lets a VR user grasp with any avatar hand and set the grip force with the controller trigger, without retraining for each hand. Each actuator becomes a token that carries the surface it moves, that surface's relation to the object, and a "closing rate" (the rate at which each surface point approaches the object as the actuator turns). A Transformer with GET-Zero's graph-distance attention bias maps the tokens to per-actuator actions. Trained on 24 MANO hands and three primitive objects, the policy is tested on 648 unseen MANO-derived hands and 20 objects. It beats adapted DexGrasp-Zero and IBS baselines on transport success (47.5% vs. at most 34.3% on unseen hands) and force-tracking gain (0.59 vs. at most 0.25), and 12 VR participants rated it more realistic than per-hand ForceGrip policies. All results are in simulation or VR; there are no hardware experiments.

Critical strengths

The method is also explicitly force- and contact-aware, which most cross-embodiment grasping work is not. Grip force is defined physically as the compressive part of the internal contact forces, computed through the grasp matrix, so it ignores forces that merely push or turn the object, and the same quantity serves as the command, the reward, an observation and every force metric. Each actuator token sees the contact force on its phalanx and the gap, direction and normals from its surface to the object, and the closing rate tells the policy both how fast a surface approaches the object before contact and how hard the actuator presses once it touches. Most cross-hand policies optimize only grasp success; this paper instead targets a grip force the user commands, which makes the results more relevant to contact-rich interaction. The actuator token is a clean, hand-agnostic interface, and computing the closing rate explicitly is simple and well justified. Thickness and length are held out of training entirely, so the geometric generalization test is genuine. The per-hand analysis (Figure 4) shows that the force-tracking advantage holds on almost every unseen hand rather than on average only. The appendix is detailed and code is provided. The user study is a further strength: it tests the policy with 12 different people driving it in real time, in counterbalanced and blinded blocks, which shows that the controller holds up under varied, unscripted human input rather than only under scripted commands. Its impact is limited, however, because the participants controlled simulated hands. The same study with real manufactured robot hands, for example in a teleoperation setting, would have shown that the policy transfers to physical contact and sensing and would have made this result far more significant.

Critical weaknesses
  1. The baselines appear weakened by their adaptation. DexGrasp-Zero reports 92% success on training hands and 85% on unseen hands in its own paper, but only 54.8% and 34.3% here. The task is harder, but the authors also replaced its node inputs (Table 10) and removed its wrist head. The IBS-adapted baseline has a force-tracking gain of 0.00 in every row of Table 1, which indicates a failed training run rather than a meaningful comparison. No GET-Zero-style baseline and no per-hand expert in simulation is reported, so there is no upper bound.

  2. Absolute performance is low relative to the literature. Unseen-hand success of 47.5% compares with 85% for DexGrasp-Zero and 81.9% for One Hand to Rule Them All (Wei et al., RSS 2026), though the tasks differ. Force tracking remains loose: a gain of 0.59 and an error of 0.46 kgf, about 23% of the 2 kgf range. ForceGrip used a 10 kgf ceiling and reported a trigger-to-force correlation of 0.884; DittoHand reports no comparable figure.

  3. The gains are concentrated where prior methods already generalize. DittoHand keeps 88% and 101% of its seen-hand success under thickness and length changes but only 58% under finger or phalanx removal, where all methods grasp at nearly the same rate (49.0–50.3%). This mirrors GET-Zero (97% on new geometry, 62% on new graph structure). The ablations also show that removing the closing rate or the descendant terms does not reduce unseen-hand success (48.2% and 47.5% vs. 47.5%); these components help force tracking only.

  4. The generalization space is narrow. All hands are MANO-derived, fingers are only removed (three to five), the thumb is always intact, joint axes are preserved, and the 21-token, five-chain template cannot represent more than five fingers or non-human hands such as Allegro or LEAP. Objects are rescaled to a 4.5 cm width with one mass and friction. Start poses come from an offline grasp search run for every hand–object pair, with infeasible pairs dropped, and the Grasp metric uses friction 1.5 while the simulation uses 0.5.

  5. There are no real-world results. Every number comes from Isaac Sim, and the VR study uses simulated hands. This is defensible for a VR paper, but the comparison set (DexGrasp-Zero, GET-Zero, CrossDex) all report real-hand experiments, and the paper does not list the omission as a limitation.

  6. The statistical support is thin. No seeds, variances or confidence intervals are reported, and each method appears to be a single training run. The user study has 12 participants (ForceGrip used 20), and the thicker-hand condition is not significant (p = 0.32). The ForceGrip baseline also differs in inputs and network, not only in being trained per hand.

  7. The contribution is mainly a combination of existing components. VR grasp controllers such as VR-HandNet (Han et al., 2024) and ForceGrip (Han et al., 2025) already track a commanded grip, and DittoHand reuses ForceGrip's reward and extends its curriculum; their only missing piece is transfer across hands. Cross-embodiment policies such as MetaMorph (Gupta et al., 2022), GET-Zero (Patel & Song, 2025) and DexGrasp-Zero (Wu et al., 2026) already share one per-joint or per-part policy across hands, and DittoHand adopts GET-Zero's graph-distance attention bias directly. Geometric representations such as the interaction bisector surface (She et al., 2022; 2024) and DexRepNet (Liu et al., 2023) already give the policy hand–object distances and normals. DittoHand's new step is to attach these geometric features to individual actuators and to add the closing rate. That is a reasonable but incremental step, and the closing rate is itself a standard quantity (see point 8).

  8. Related work is incomplete. The most important omission is House of Dextra (Fay et al., ICLR 2026), published at this venue, which should be cited and ideally compared against. It trains a single morphology-conditioned PPO policy across generated dexterous hands with three to five fingers, two or three actuated joints per finger, and varied segment lengths, fingertip shapes and palm thickness. These are almost exactly the axes DittoHand varies (finger count, phalanx count, length and thickness). Each House of Dextra hand also carries explicit collision geometry (CoACD convex decomposition), each family is trained on 2,000–8,000 hand variants rather than 24, and the resulting policies transfer zero-shot to four fabricated robot hands. House of Dextra conditions its policy on a one-hot morphology code rather than on surface geometry, so it is a natural kinematics-only baseline for DittoHand's central claim that collision geometry must be represented explicitly. Without it, the paper's statement that existing cross-hand policies leave hand geometry implicit is not tested against the closest prior system. If such a policy, trained on many more hands, matched DittoHand, the benefit of the surface tokens would reduce to data efficiency rather than a necessary representation. The closing rate is the normal component of the standard contact Jacobian; this should be stated, and Neural Jacobian Fields (Li et al., Nature 2025), which learns per-point Jacobians for arbitrary robots, should be discussed. The internal-force definition of grip force builds on Yoshikawa & Nagai (1991), Bicchi (1995) and Buss et al. (1996), none of which are cited. Also missing are Won & Lee (SIGGRAPH Asia 2019) on single controllers across body shapes; geometry-aware cross-embodiment grasping (GeoMatch, CoRL 2023; RobotFingerPrint, IROS 2025; UniGrasp, RA-L 2020); embodiment scaling (Ai et al., CoRL 2025); and avatar-embodiment studies such as Schwind et al. (CHI PLAY 2017), which found that removing fingers from realistic avatar hands lowers presence.

Questions for authors
  1. What does DexGrasp-Zero, with its original inputs and wrist head, and a retuned IBS-adapted baseline (currently gain 0.00 in every row) achieve on your hands?

  2. Can you report means and standard deviations over at least three seeds for Tables 1 and 2?

  3. How close is the shared policy to a per-hand expert, and to ForceGrip trained per hand, on the same simulated metrics?

  4. Removing κ does not lower unseen-hand success (48.2% vs. 47.5%). Does κ improve anything besides force tracking?

  5. What is the trigger-to-force correlation (ForceGrip reports r = 0.884), and does tracking hold at ForceGrip's 10 kgf range?

  6. How do results change with start poses not drawn from FSG, and with Grasp computed at the simulated friction of 0.5?

  7. Can the 21-token template represent non-MANO hands such as Allegro or LEAP, or a six-finger hand? If not, can this limit be stated?

  8. Have you tested the policy on any physical robot hand, for example through teleoperation?

9 comments
AuthorsOpenReview ↗

Response to Reviewer zTiH (1/6)

Thank you for carefully reading our paper and raising specific questions. We have grouped your questions and concerns into eight topics based on shared evidence.

To address your questions, we added more random seeds and conducted joint transport and force-tracking experiments, additional grasping experiments, retraining of the revised DGZ-adapted, training of DittoHand-expert, evaluations on five non-MANO VR avatar hands and 24 six-fingered hands, objective metric analysis of VR study recordings, and a force control task.

1. Scope of the study and generalization across hand structures (W4, W5, Q7, Q8)

Research Scope (W5, Q8). This study focuses on VR avatar hands controlled by users through wrist movements and trigger signals. We did not evaluate performance on physical robotic hands. Real-world deployment would require additional consideration of actuator dynamics, link mass and inertia, joint constraints, and sim-to-real gaps. We will therefore explicitly acknowledge transfer to physical robots as an unverified limitation.

Hand Structure Scope (W4, Q7). This study covers VR avatar hands with different numbers of fingers, joints per finger, joint positions, link lengths and thicknesses, and collision geometries. However, joint-axis orientations remain constrained. The training hands share similar rotational-axis configurations for finger flexion and abduction, and the policy was trained within this range. Although the policy receives joint-axis directions as input, performance on unseen axis configurations cannot be guaranteed. Moreover, our study focuses on grasps in which the thumb opposes at least two other fingers to squeeze an object. We therefore evaluate hands with at least three fingers. This is a requirement of the grasping task, not an upper limit on finger count.

Importantly, 21 tokens are not a fixed architectural constraint. Each actuator corresponds to one token, so the token count varies with the number of joints. Training used hands with 17 and 21 tokens, while the original evaluation hands have 11–21 tokens. The policy uses joint structure and surface information rather than joint indices or finger names.

We added two experiments to test whether this architecture generalizes beyond the training range.

  • Six-fingered hands: We evaluated 24 unseen six-fingered hands with 25 actuators without retraining. Their success rate was 68.6%, comparable to 68.4% for the corresponding five-fingered hands (Table R1-1). This shows that the policy can generalize to hand structures exceeding the maximum training token count of 21.

Table R1-1. Six-fingered hands (no retraining; mean ± standard deviation over three runs)

24 six-finger handsOriginal five-finger hands
Transport Success (%) ↑68.6 ± 4.168.4 ± 3.6
Force FTE (kgf) ↓0.51 ± 0.150.50 ± 0.14
Force Gain → 10.51 ± 0.150.53 ± 0.14
  • Non-MANO Avatar Hands (Appendix B.6): We evaluated five avatar hands from Rocketbox, Ultraleap, and SteamVR without retraining. We preserved their original meshes, joint positions, sizes, and masses while applying the same 21-DoF configuration as the training hands. The average Success was 44.8%, comparable to 42.3% for unseen MANO hands, with similar FTE and Gain (Table R1-2). This result is added to Table 2 in Section 4.3 of the revision, and the per-hand results and hand renderings to Appendix B.6 (Table 14, Figure 8).

Table R1-2. VR avatar hands not derived from MANO (Ours without retraining, in the same combined experiment as Tables 1 and 3. The first two rows repeat the MANO-derived Seen and Unseen hands. All values are averaged over three training seeds; ± is the standard deviation over seeds for the grouped rows.)

HandsSuccess (%) ↑FTE (kgf) ↓Gain → 1
Seen MANO hands (24)62.9 ± 2.60.51 ± 0.070.52 ± 0.13
Unseen MANO hands (648)42.3 ± 4.50.68 ± 0.070.46 ± 0.12
Non-MANO avatar hands (unseen)
Rocketbox, male37.00.660.42
Rocketbox, female41.20.660.42
Ultraleap57.20.530.54
SteamVR glove41.50.910.65
SteamVR glove, slim47.00.800.69
Avatar hands (5)44.8 ± 6.40.71 ± 0.090.54 ± 0.13

These results confirm that our policy is not tied to a fixed 21-token structure or MANO geometry and can generalize to unseen finger counts and avatar geometries.

AuthorsOpenReview ↗

Response to Reviewer zTiH (2/6)

2. Choice of baselines and validity of their implementation (W1, Q1, Q2, comments on GET-Zero)

Baseline Selection. Prior cross-embodiment studies can be divided into shared policies using joint structures (MetaMorph, GET-Zero, DexGrasp-Zero) and representations using hand–object geometric relationships (IBS, DexRepNet). We selected DGZ and IBS as representative baselines from each category.

GET-Zero also uses a joint graph-based shared policy, but its original task is in-hand cube rotation, rather than grasping. We therefore selected DGZ, which directly addresses grasping and uses object distance information. We separately evaluated the contribution of joint tokens and graph-distance attention bias, the core components of GET-Zero, through the surface-removal ablation (-surf).

Common Experimental Conditions. DGZ-adapted and IBS-adapted were trained and evaluated in the same VR grasping environment as Ours, with identical training and evaluation hands and objects, episode configurations, curricula, reward functions, training budgets, and evaluation metrics. To match our task, where users control wrist motion and specify target grip force through a trigger, we excluded wrist control from policy outputs and trained both baselines to track commanded grip force. Both baselines received identical global observations and differed only in their core representations and network architectures.

DGZ-adapted. We added previously missing global observations (previous command, object velocity, and gravity) and retrained the baseline. Success on unseen hands remained nearly unchanged, from 34.2% to 33.9%. We updated Table 1 accordingly.

IBS-adapted. IBS represents spatial relationships between the hand and object, but directly using its original policy is unsuitable for our zero-shot evaluation on unseen hands. To isolate the effect of geometric representation, we replaced only the surface representation with IBS. Thus, this baseline compares IBS and our proposed surface representation under the same policy architecture, rather than reproducing the entire original IBS policy.

  • As you pointed out, the Gain of 0.00 in the original submission came from a failed training run. We therefore trained two more IBS runs with different random seeds, and the results in Table 1 now report the average of three newly trained runs. Even with these updated results, IBS consistently underperformed Ours across all metrics (28.0% vs. 42.3% in success rate on unseen hands, and 0.19 vs. 0.46 in Gain).
  • IBS Performance Gap (W2, Q2). IBS represents the current surface–object relationship but does not directly provide distance changes induced by joint rotation. Since these changes differ across hand structures, a policy trained on multiple hands struggles to learn consistent joint control. In contrast, our κ represents the distance changes of all surfaces moved by each joint. We believe this representational difference explains the performance gap even on seen hands (Success: 42.4% vs. 62.9%).
AuthorsOpenReview ↗

Response to Reviewer zTiH (3/6)

3. Evaluation conditions and statistical reliability (W4, W6, Q2, Q6)

The submission reported single-seed results for each method. We have now independently trained Ours, DGZ, and IBS with three seeds each, and updated Tables 1 and 3 to report mean ± standard deviation. We also made two changes to the evaluation conditions to better reflect real-world VR environments.

Evaluation Metrics. We revised two issues in the original evaluation based on your comments.

  • Joint Transport and Force Tracking: The submission evaluated transport (fixed maximum force command) and force tracking (with support and a fixed wrist) separately. Since this did not sufficiently support our claim of transporting objects while following force commands, we added an experiment combining both tasks.
  • Grasp: The submitted Grasp metric was a static force-closure (wrench-space) test assuming a friction coefficient of 1.5, which, as you noted, did not match the simulation friction of 0.5. We therefore re-measured Grasp as whether the object is actually held for 0.5 seconds after the support is removed in simulation.

Initial Pose (W4, Q6). Since our study focuses on finger grasp control rather than wrist control, we used FSG to initialize the wrist in a grasp-suitable pose during simulation evaluation. This reduces the effect of initial wrist position and allows consistent comparison of finger policy performance. In actual VR, users can directly move their wrists to approach objects and adjust their poses. Our VR user study achieved a 98.1% success rate (ForceGrip: 92.5%).

Object Specifications (W4). The 4.5 cm width allows even hands with the shortest fingers or only three fingers to enclose the object. Since our study focuses on hand morphology, we prioritized object shape diversity over mass or friction variation. Accordingly, 17 of the 20 evaluation objects have shapes unseen during training. In the VR study, we varied object size by 0.75–1.25×. We will explicitly state generalization to object physical properties, such as mass and friction, as future work.

Table 1 of the revision. Comparison with the baselines (Mean ± standard deviation over three training seeds; Grasp and Success in %, FTE in kgf. Rows below Unseen hands differ from the Seen hands in one aspect; Topology means fewer fingers or phalanges. Bold: best per row and metric.)

(a) Keeping the object

HandsGrasp DGZGrasp IBSGrasp OursSuccess DGZSuccess IBSSuccess Ours
Seen hands (24)79.1 ± 3.575.1 ± 4.783.2 ± 3.453.3 ± 1.442.4 ± 1.462.9 ± 2.6
Unseen hands (648)61.6 ± 2.259.6 ± 3.370.6 ± 4.633.9 ± 4.428.0 ± 3.742.3 ± 4.5
β only (24)78.5 ± 3.074.8 ± 4.282.3 ± 3.652.1 ± 1.541.2 ± 1.960.4 ± 2.9
Thickness only (96)74.8 ± 4.168.2 ± 4.678.1 ± 3.946.6 ± 3.635.5 ± 1.154.7 ± 3.5
Length only (48)76.6 ± 6.670.1 ± 4.679.2 ± 1.456.4 ± 8.744.6 ± 6.564.4 ± 5.0
Topology only (48)53.0 ± 0.550.8 ± 4.464.8 ± 5.322.6 ± 3.318.2 ± 3.532.4 ± 5.0

(b) Tracking the grip command

HandsFTE DGZFTE IBSFTE OursGain DGZGain IBSGain Ours
Seen hands (24)0.57 ± 0.030.71 ± 0.100.51 ± 0.070.48 ± 0.010.21 ± 0.210.52 ± 0.13
Unseen hands (648)0.76 ± 0.020.86 ± 0.090.68 ± 0.070.42 ± 0.040.19 ± 0.170.46 ± 0.12
β only (24)0.57 ± 0.030.71 ± 0.100.52 ± 0.060.48 ± 0.020.23 ± 0.200.53 ± 0.13
Thickness only (96)0.68 ± 0.040.90 ± 0.090.64 ± 0.060.47 ± 0.020.23 ± 0.220.49 ± 0.13
Length only (48)0.64 ± 0.030.92 ± 0.190.68 ± 0.110.62 ± 0.020.33 ± 0.270.64 ± 0.15
Topology only (48)0.82 ± 0.020.88 ± 0.070.69 ± 0.080.30 ± 0.020.12 ± 0.130.37 ± 0.12

Table 3 of the revision. Ablations (Grasp and Success in %, FTE in kgf. Mean ± standard deviation over three training seeds. Bold: best per column.)

MethodSeen Grasp ↑Seen Success ↑Seen FTE ↓Seen Gain → 1Unseen Grasp ↑Unseen Success ↑Unseen FTE ↓Unseen Gain → 1
Ours83.2 ± 3.462.9 ± 2.60.51 ± 0.070.52 ± 0.1370.6 ± 4.642.3 ± 4.50.68 ± 0.070.46 ± 0.12
−κ76.2 ± 1.854.9 ± 3.40.65 ± 0.050.36 ± 0.0360.4 ± 2.134.3 ± 1.00.82 ± 0.030.30 ± 0.01
−desc78.5 ± 3.655.6 ± 5.00.71 ± 0.110.16 ± 0.2767.7 ± 4.538.7 ± 5.90.85 ± 0.090.15 ± 0.26
−surf69.3 ± 1.533.5 ± 1.50.78 ± 0.020.20 ± 0.0455.6 ± 10.722.2 ± 7.00.89 ± 0.070.20 ± 0.02

The conclusion that Ours outperforms the baselines remains the same.

Comments on the paper

2 comments
AuthorsOpenReview ↗

General Response(1/2)

We thank the three reviewers for their careful comments. As strengths, the reviewers noted the representation that binds each actuator to the surface it moves, capturing contact geometry that joint states alone cannot describe (BsqT, fE7i; fE7i called it the paper's main conceptual contribution); the physically defined grip force and direct handling of force and contact (zTiH); the performance on unseen hands and the test with thickness and length held out of training (BsqT, zTiH); the comprehensive ablations and controlled simulation evaluation (BsqT, fE7i); and the real-time VR user study (zTiH, BsqT).

We have uploaded a revised paper, with changes marked in blue. The new experiments support one main finding: in a combined experiment that transports an unsupported object while following a changing force command, DittoHand outperforms the baselines and ablations on both seen and unseen hands, averaged over three seeds. Without retraining, it performs on avatar hands not derived from MANO about as well as on unseen MANO hands, and in a new VR force-control user study, the grip force followed users' commands more closely with DittoHand than with ForceGrip.

Below, W# and Q# refer to each reviewer's weaknesses and questions, in order.

zTiH

  • Baselines (W1, Q1): We retrained DGZ-adapted with the previously missing global observations (previous command, object velocity, and gravity), and its unseen-hand Success changed from 34.2% to 33.9%. IBS-adapted now reports the average of three newly trained runs. → Table 1, App. B.5
  • Per-hand experts (W1, W6, Q3): Experts trained on each hand with the same inputs and architecture outperformed ForceGrip on all four metrics for all three hands. The shared policy is close to the experts (three-hand average Success 56.1% vs. 62.5%). → Table R1-3 in our response to zTiH
  • Trigger–force correlation (W2, Q5): In the VR force-control experiment, r was 0.81 (ForceGrip 0.58) and Gain was 0.78 (ForceGrip 0.40). → Table 4 in Sec. 4.5, App. C
  • Absolute performance (W2): Prior papers differ from ours in friction, wrist motion, and force commands, so a direct comparison is difficult. On our task, the re-evaluated DGZ reaches 33.9% unseen-hand Success, below Ours (42.3%). → Table 1
  • Topology changes (W3): Under the new Grasp criterion, Grasp on the topology-only hands is 64.8% for Ours, 53.0% for DGZ, and 50.8% for IBS. → Table 1
  • κ and descendant (W3, Q4): Over three seeds, removing either lowers unseen-hand Success from 42.3% to 34.3% and 38.7%, respectively. → Table 3 in Sec. 4.4
  • Seeds and evaluation conditions (W4, W6, Q2, Q6): All methods are trained with three seeds and reported as mean ± standard deviation. Grasp is re-measured as whether the object is held for 0.5 s after the support is removed, at the simulation friction of 0.5. → Sec. 4.1, Tables 1 and 3
  • Hand structure (W4, Q7): The number of tokens follows the number of actuators. Without retraining, the policy also works on 24 six-fingered hands unseen in training (Success 68.6%, vs. 68.4% for the matched five-fingered hands) and on five non-MANO avatar hands (44.8%). → Table 2, App. B.6, Table R1-4 in our response to zTiH
  • Physical robots (W5, Q8): We state that our scope is VR avatar hands and list transfer to physical robots as an unverified limitation. → Secs. 1 and 5
  • User study sample size (W6): In the second experiment (force control), DittoHand was also rated significantly higher on all four RQ measures (Holm-corrected p ≤ 0.027). → Fig. 5b
  • Novelty and closing rate (W7, W8): We state that κ is the normal component of the contact Jacobian and describe how our approach differs from Neural Jacobian Fields. → App. A.1
  • Related work (W8): We now cite all the suggested works. The −surf ablation, which removes only the surface, reaches only half of our Success even on the seen hands (33.5% vs. 62.9%), so we do not regard this as a gap that training on more hands would close. → Secs. 2.1–2.3 and 3.5, Apps. B.1 and B.4, Table 3
AuthorsOpenReview ↗

General Response(2/2)

BsqT

  • Combined experiment (W1, Q1): The main evaluation of the revision is now an experiment that transports an unsupported object while following a changing force command (lower bound 0.79 kgf). Ours is best in 23 of the 24 entries of Table 1. → Sec. 4.1, Table 1, App. B.4
  • Learning curves and run-to-run variation (Q2): We added learning curves over three seeds per method and report mean ± standard deviation. → Fig. 4, Tables 1 and 3
  • IBS-adapted (W2, Q2): Its Success varies little across seeds (42.4 ± 1.4%), and its curves converge stably below Ours, so we attribute the gap to the representation rather than to a training failure. In the same architecture, removing only κ lowers seen-hand Success from 62.9% to 54.9%. → Fig. 4, Table 3
  • Generalization scope (W3, W4): We narrowed "a hand" to "an avatar hand" in the Introduction and list transfer to physical robots as a limitation. On five avatar hands not derived from MANO, Success is 44.8% without retraining (unseen MANO hands: 42.3%). → Secs. 1 and 5, Table 2, App. B.6
  • Initial poses (W3): FSG is an evaluation setting for comparing finger policies under the same conditions. In the VR study, where users move the wrist themselves, the success rate was 98.1% (ForceGrip 92.5%). → App. B.3, Table 15 in App. C

fE7i

  • Scope of the claim (W1): The Abstract, Introduction, third contribution, and Conclusion now scope the claim to VR avatar hands. → Abstract, Secs. 1 and 5
  • Evidence beyond MANO (W1): We evaluated five public avatar hands not derived from MANO without retraining. Success is 44.8%, comparable to unseen MANO hands (42.3%). → Table 2 in Sec. 4.3, App. B.6
  • Objective VR measures (W2): In pick-and-place, DittoHand was significantly better in success rate (98.1% vs. 92.5%), completion time (11.2 s vs. 12.9 s), and timeouts (0.9% vs. 5.9%). → Table 15 in App. C
  • Force-tracking user study (W2): The same 12 participants took part in a new experiment in which they matched target forces. Gain was 0.78 vs. 0.40, and 0.79 vs. 0.10 on the three-fingered hand. → Table 4 in Sec. 4.5, Tables 16 and 17 in App. C
  • Seeds (Q1–2): We trained with three seeds and report mean ± standard deviation. → Tables 1 and 3
  • Wall-clock time (Q3): A control step takes 4.8 ms (14% of the 30 Hz period), and 10.9 ms even with a 500K-triangle mesh and two hands. → App. D

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.