acceptodds
Under review as a conference paper at ICLR 2027

FabricBench: A Standardized Policy Benchmark for Bimanual Fabric Manipulation with a Contact-Driven Deformable Simulator

Abstract

Fabric manipulation challenges robot policies with changing geometry, self-occlusion, and contact, yet standardized evaluation remains focused on rigid objects. We introduce FabricBench, a benchmark for bimanual fabric manipulation with six tasks spanning garments, silk, and bags, 10,000 simulation-ready assets, and controlled tests of configuration and appearance generalization. Built on an open-source solver in Newton, our simulator combines redesigned cloth–rigid contact with bidirectional force coupling, enabling grasping, release, and load transfer without artificial attachment. It succeeds in all 45 grasp–hold–release validation trials, compared with none for the tested baseline configurations. Evaluating six representative policies, including vision-language-action models and world action models, reveals substantial remaining challenges: the best policy achieves 77% average success, while no policy exceeds 36% on long-horizon T-shirt folding. Under the most deformed initial T-shirt configurations, success falls to 6–14% across policies, with larger drops than under the evaluated appearance shifts. Zero-shot deployment of π₀.₅ on a real robot largely preserves task difficulty ordering across six tasks (Spearman ρ = 0.94). These results establish FabricBench as a reproducible testbed for measuring fabric-manipulation capabilities and exposing limitations in configuration generalization.

Primary area:
datasets and benchmarks
Keywords:
Deformable Object Manipulation, Bimanual Manipulation, Robotic Benchmarking, Contact Simulation, Sim-to-Real Evaluation

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_D1r2OpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposed the FabricBench benchmark for bimanual fabric manipulation. Proposed simulator built on Newton combines redesigned cloth-rigid contact with bidirectional force coupling, enabling several manuplation without artificial attachment. Benchmark included six representative policies reveals substantial remaining challenges in bimanual fabric manipulation. Overall, the paper is clearly written, easy to follow, and provides informative empirical findings.

Critical strengths
  • This paper proposed a contact-driven cloth simulator that avoids artificial attachment mechanisms. This makes grasping, holding, and release emerge from physical contact, reducing evaluation artifacts caused by heuristic attach detach rules.

  • This paper explicitly split the generalization problem into configuration shift and appearance shift. The experiments show that current policies are substantially more sensitive to changes in fabric configuration than to appearance changes, providing a useful diagnosis of current policy limitations.

Critical weaknesses
  • The physical fidelity of the simulator is not fully validated against real-world fabric dynamics. Current experiments in paper show that proposed benchmark is behaviorally suitable for manipulation evaluation, but quantitative sim-to-real comparisons of cloth deformation, trajectories, contact forces, or other physical quantities are limited. This makes it difficult to assess how accurately the simulator reproduces real fabric dynamics.
Questions for authors
  • Are the 10k assets actually utilized? The main experiments mostly study appearance and configuration shifts, while geometry and material generalization appear relatively limited.
  • Could the authors provide stronger quantitative validation of the simulator’s physical fidelity, for example by comparing simulated and real cloth trajectories, deformation, or interaction forces?
  • How extensively were the baseline simulators tuned? Since the comparison mainly uses released/default configurations, could task-specific tuning substantially improve their grasp/hold/release performance?
  • How representative are the three discrete configuration levels C0/C1/C2 of the broader fabric configuration space? Could the conclusions change under a denser or more continuous range of deformation levels?
3 comments

W1 & Q2: Quantitative physical fidelity. We add two matched sim–real comparisons with the same robot actions and initial fabric state. (i) A dual-arm grasp–hold–release trajectory is executed on the real robot and replayed in each simulator; only FabricBench reproduces the real grasp, hold and release (videos: https://anonymous.4open.science/w/Fabric-Anonymous-23A0/website/index.html). (ii) On RGBench (Hu et al., AAAI 2026), a public sim-to-real benchmark that replays real bimanual fold and grasp-and-lift trajectories and measures the per-frame Chamfer distance to real point clouds (lower is better):

SimulatorFoldGrasp-liftAvg
PyBullet0.0560.1990.128
Isaac Sim PBD0.0430.1340.088
Ours0.0270.0810.054

Both will be added to Sec. 5.1.


Q1 (a): Use of the 10k assets. All 10k assets pass the same stability check and have each been simulated (Sec. 4.2; App. A). In training, Carry bag uses 10 bag appearance variants (App. E). In evaluation, for consistency, the standard setting uses the canonical asset, and the appearance axis uses 50 variants sampled from the held-out assets, with only appearance changed (Table 4).

Q1 (b): Geometry and material generalization. We thank the reviewer for this interesting question. Beyond the randomized handle stiffness of the bag (App. E), we varied stretch stiffness, bending stiffness and mass over 0.5–2×; this check is not in the paper because the effect was small: grasp-and-place and folding were largely insensitive, while flinging was sensitive mainly to air damping rather than these parameters. Since each asset carries an explicit material profile (App. A), the material profile can represent fabrics as different as cotton and leather, and we will continue to study this question. For geometry, we thank the reviewer for the suggestion: morphology variants are included in the released assets, and we will explore them in future work.


Q3: Tuning of baseline simulators. The baselines start from tuned settings, e.g., the cloth parameters released by DexGarmentLab for Isaac Sim PBD, and we further tuned mass and friction for Isaac Sim and the step size, iterations and contact parameters for VBD and Style3D. All simulators use the same solver and contact settings across tasks, without task-specific tuning. In our sweeps, tuning did not remove these failures (App. B.7), because the failures come from how these simulators couple cloth and rigid bodies. Isaac Sim PBD resolves cloth–rigid contact only at particles padded by a rest offset, with dynamic friction only, so the closing jaws push the fabric's particles aside instead of pinching them; DexGarmentLab likewise reports that grippers "penetrate the garment without achieving effective lifting" and adds adhesion. VBD and Style3D detect gripper contact only at cloth vertices, so vertices sink into the fingers and the fabric is not released; coupled to an external rigid solver, they also apply contact forces to the cloth only, so a bag's contents are not carried. FabricBench instead fixes these failures in the contact mechanism itself, through triangle-level detection, anchored friction and bidirectional coupling (Sec. 3), so grasping, holding and releasing are ensured by the mechanism and are robust to its physical parameters within the tested ranges.


Q4: Representativeness of C0–C2. C0–C2 do not cover the full configuration space; they are three designed difficulty levels, each a seeded distribution of initial states (Fig. 3; App. C), spanning the range where policies remain distinguishable. On Flatten T-shirt, success falls from 70–92% (C0) to 16–62% (C1) and 6–14% (C2), with different policy rankings per level (Table 4). Harsher levels would not change this trend: success is already low at C2 (0–2% on Flatten silk), and more random initializations, such as drops from randomly chosen points, bring all policies to nearly zero. Sec. 6 already notes this limitation, and denser or continuous levels are future work.

We thank the reviewer again for the positive assessment and hope these additions address the remaining concerns.

Thanks for authors' response, clarification and additional experimental results.

For the Q1(b), I would encourage the authors includes more comprehensive evaluation of generalization in paper or supp, even the effect was relatively small, which would affect the reliability of proposed simulator.

Considering the exsiting evidence and additional experiments, I would not decline my rating, nor raise my rating.

Thank you for the follow-up and for your constructive suggestion. We agree that reporting these results strengthens the reliability of the simulator, and we will include the generalization evaluation in the revision. We appreciate your positive assessment of our work.

Best regards, The Authors

Reviewer_uXwxOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

FabricBench introduces a benchmark for bimanual fabric manipulation with six tasks involving garments, silk cloth, and bags. Building on an existing cloth solver, the authors improve cloth–rigid contact detection, friction handling, and bidirectional force coupling to support grasping, release, and load transfer without artificial attachment. The benchmark provides scripted demonstrations and separately evaluates changes in initial fabric configuration and appearance. Experiments include grasp–hold–release comparisons, component ablations, evaluations of six policies, and real-world deployment of simulation-fine-tuned π₀.₅. The results highlight difficulties in long-horizon folding and configuration generalization, while the real-world experiments show agreement in task difficulty ordering for the evaluated policy.

Critical strengths
  1. Contact modeling addresses consequential failure modes in fabric manipulation. The paper explicitly handles contact detection, frictional holding and release, and bidirectional load transfer. These capabilities matter for policy evaluation because failures caused by the simulator can otherwise be confused with failures of the policy. The component ablations provide evidence for the functional importance of triangle-level contact detection and bidirectional coupling.
  2. The benchmark provides a structured evaluation protocol. The six tasks cover several relevant manipulation behaviors, and policies are evaluated using shared demonstrations and matched evaluation seeds. Separating initial-configuration changes from appearance changes makes the results more interpretable than evaluations that change several object properties simultaneously.
Critical weaknesses
  1. The evidence for physical fidelity remains indirect. The simulator validation primarily measures successful grasping, holding, release, penetration events, and loaded lifting. These tests establish useful operational behavior, but do not directly quantify agreement with real fabric dynamics. The paper does not provide a matched simulation–real comparison of quantities such as slip thresholds, deformation, or motion under the same robot actions and initial conditions. Consequently, the reported improvements do not yet establish how accurately the contact model reproduces real interactions, which matters for trusting the benchmark’s physical failure modes.
  2. Insufficient validation of the practical utility of the simulation environment. Although the paper demonstrates zero-shot real-world deployment of π₀.₅ fine-tuned on FabricBench demonstrations, it remains unclear how much the proposed simulator contributes to improving real-world manipulation performance. In particular, the paper lacks downstream policy-learning evaluations comparing VLA performance when trained on demonstrations generated by existing fabric simulators or by ablated versions of the proposed simulator under comparable training settings. It also does not quantify the improvement over the pretrained VLA without FabricBench fine-tuning. Such experiments would help establish whether the proposed simulation improvements produce more effective training data and enable VLAs to acquire or improve real-world fabric manipulation skills. More importantly, it would be particularly compelling to demonstrate that training on FabricBench-generated data enables a VLA to achieve near-perfect success on a challenging real-world manipulation task that it previously struggled to perform. Such a result would provide stronger evidence of the simulator's practical value in transferring manipulation skills from simulation to reality.
Questions for authors
  1. Comparison of simulation fidelity. Could the authors provide more direct comparisons between FabricBench and existing fabric simulators, particularly in terms of visual fidelity, fabric deformation, and contact dynamics? Comparisons with real-world observations under matched manipulation conditions would also help establish whether the proposed simulator more faithfully reproduces real fabric behavior.
  2. Ablation of simulation data effectiveness on real robots. Could the authors evaluate the contribution of FabricBench-generated training data to real-world VLA performance? In particular, it would be valuable to compare policies fine-tuned on the proposed simulation data with those trained on data from other existing simulators, ablated versions of FabricBench, or without simulation fine-tuning. Such comparisons would help quantify how much the proposed simulation improvements contribute to real-world manipulation performance.
  3. Demonstrating the practical value of simulation-generated data. Could the authors provide a more compelling demonstration of how FabricBench-generated data can enable a VLA to solve challenging real-world manipulation tasks? For example, it would be particularly convincing to show that a task with a very low initial real-world success rate can achieve near-perfect success after fine-tuning on FabricBench-generated data. This would provide stronger evidence of the simulator's practical utility for transferring manipulation skills from simulation to reality.
5 comments

We thank the reviewer for recognizing our contact modeling and structured evaluation. The new results address both concerns directly: on the real robot, only FabricBench among the tested simulators reproduces the grasp–hold–release sequence; on RGBench, it reduces the sim–real error of Isaac Sim PBD and PyBullet by 39% and 58%; and simulation-only fine-tuning lifts π0.5 from 0% to 60% average real-world success.


W1 & Q1: Direct sim–real fidelity and comparison with existing simulators. We add two matched sim–real comparisons, in which the real robot and every simulator execute the same robot actions from the same initial fabric state.

(i) Matched sim–real replay of a contact trajectory. The same dual-arm grasp–hold–release trajectory is executed on the real robot and replayed in FabricBench, Isaac Sim PBD, VBD and Style3D, with the same T-shirt in the same initial pose. Side-by-side videos from aligned camera views show appearance, cloth deformation and contact behavior over the whole trajectory: https://anonymous.4open.science/w/Fabric-Anonymous-23A0/website/index.html. On the real robot the T-shirt is grasped, lifted without slipping and released. Only FabricBench reproduces this: Isaac Sim PBD fails to grasp the cloth, and in VBD and Style3D the fingers penetrate the cloth, so it is not released. The videos will also be included in the supplementary material.

(ii) Quantitative sim–real gap on a public benchmark. We further evaluate on RGBench (Hu et al., AAAI 2026), a public sim-to-real benchmark that replays real bimanual fold and grasp-and-lift trajectories on real garments and measures the per-frame Chamfer distance between simulated and real point clouds (lower is better). This directly quantifies deformation and motion under identical actions, and grasp-and-lift tests whether the grip holds without slipping as on the real robot. Baselines are the publicly available simulators, with the numbers reported by RGBench:

SimulatorFoldGrasp-liftAvg
PyBullet0.0560.1990.128
Isaac Sim PBD0.0430.1340.088
Ours0.0270.0810.054

With the same parameters as the benchmark, FabricBench reduces the error of Isaac Sim PBD by 39% and of PyBullet by 58%, with the largest gain on grasp-and-lift. We will add both experiments to Sec. 5.1 and the per-garment results to the appendix.


W2 & Q2: Contribution of FabricBench data to real-world performance.

Without FabricBench fine-tuning. Under the same success criteria, the pretrained π0.5 succeeds in no trial on any task, both in simulation and on the real robot (20 real trials per task). After simulation-only fine-tuning on FabricBench demonstrations, its real-world success rises to 35–85% (60% on average):

Real robot SR (%)Lift silkFlatten T-shirtFlatten silkCarry bagPack snackFold T-shirt
π0.5, no FabricBench000000
π0.5, FabricBench only705565508535

Other simulators or ablated variants. We attempted to generate the same demonstrations in existing simulators, starting from released settings (e.g., the DexGarmentLab parameters for Isaac Sim PBD) and further tuning mass, friction, step size and contact parameters. In our attempts, however, valid demonstrations with contact-driven grasping, as on the real robot, could not be collected: under the same grasp, Isaac Sim PBD fails to grasp the cloth, and in VBD and Style3D the fingers penetrate the cloth, so it is not released (Table 1). We attribute this primarily to how these simulators couple cloth and rigid bodies, rather than to parameter choices. Likewise, in the ablated variants release or load transfer fails (Table 2a). Neither therefore produces the contact-driven demonstrations that a matched comparison needs; providing them is precisely FabricBench's contribution.


Q3: A challenging real-world task. Bimanual fabric manipulation is beyond what current VLAs can do out of the box: the pretrained π0.5 succeeds in none of the six tasks on the real robot. After simulation-only fine-tuning, it performs all six tasks zero-shot sim-to-real (35–85%), reaching 85% on Pack snack, 70% on Lift silk and 65% on Flatten silk. FabricBench thus provides demonstrations in which grasping, holding and releasing arise from contact as on the real robot, offering a sim-to-real route for fabric skills that pretrained VLAs lack. These gains come entirely from simulation, without any real-robot data.

We hope these results address the reviewer's concerns.

Reviewer_uXwxOpenReview ↗

response

Thank you for the additional experiments and clarifications. The matched video comparisons and quantitative point-cloud measurements alleviate my concerns about the simulator’s physical fidelity. I also acknowledge FabricBench’s value as a policy benchmark: more faithful contact modeling can help distinguish policy failures from simulation artifacts. However, my concern about comparisons with other simulators and ablated variants remains only partially addressed. Although the authors adopted released parameters and performed further tuning, it is unclear whether they reproduced the baselines’ complete native configurations, including their grasping mechanisms, and verified their original successful demonstrations. Failure under the shared replay protocol does not by itself establish that these methods cannot generate useful training data. Furthermore, please clarify how the cross-simulator test trajectory was generated and selected, and whether it was adjusted based on FabricBench’s performance. Replaying identical actions is appropriate for comparing physical responses, but the trajectory’s provenance and selection procedure are needed to assess potential bias.

[1/2] We thank the reviewer for the follow-up and are glad the fidelity concern is alleviated.


Native configurations and grasping mechanisms. Our Isaac Sim PBD baseline follows GarmentLab and DexGarmentLab, both built on Isaac Sim PBD cloth: we use DexGarmentLab's released configuration with a parallel-jaw gripper. The Newton VBD and Style3D baselines use their official released configurations, with VBD at twice its released substep rate for stability (App. B.7). We also reproduced both frameworks natively, in their own environments and tasks: with a parallel-jaw gripper, their demonstrations succeed only with attachment enabled, and without it the gripper fails to grasp the cloth. Newton's official cloth–gripper examples likewise succeed only along their own tuned trajectory and closing command; once the trajectory changes, the fingers penetrate the cloth. On our shared cross-simulator test trajectory without attachment (Table 1; provenance in the next comment), Isaac Sim PBD forms a single-hand grasp in only 3/45 trials and a bimanual grasp in 0/45; VBD and Style3D can grasp, but the fingers penetrate the cloth, so it sticks to them and is not released. This is not specific to our test but a widely recognized problem: DexGarmentLab reports that grippers "penetrate the garment without achieving effective lifting", and NVIDIA staff confirm that "parts of the cloth can end-up penetrating into the finger and getting stuck" (NVIDIA Developer Forums, threads 332704 and 333193).

These failures are mechanistic. In PBD/XPBD, contact and friction are position-level corrections on particles, so the normal force is not well defined and the grip force varies with iteration count and time step; holding cloth by friction between two jaws is therefore unstable. In Newton VBD and Style3D, gripper contact is detected only at cloth vertices, so the fingers sink between them and the cloth sticks; coupled to an external rigid solver, contact forces also act on the cloth only. These limitations lie in the contact formulation itself and cannot be removed by parameter tuning. Existing frameworks therefore bypass contact with attachment or anchoring, acknowledging that this cannot represent real finger–cloth contact (Matas et al., 2018).


Usefulness of attachment-generated data. We agree that attachment can collect demonstrations that succeed in simulation. However, the successful data collection in these frameworks relies on attachment blocks (GarmentLab), which decide whether the cloth is grasped by a distance rule rather than by contact: they bind cloth vertices near the closing fingers, and DexGarmentLab notes that "even minimal contact—such as a single finger block touching the garment—can establish attachment". We replaced our contact model in FabricBench with GarmentLab's block and the block in DexGarmentLab's code, using the sizes and offsets from their code and keeping the cloth (silk), physics and trajectory identical; the same grasp trajectories, with the offsets below, were also executed on the real robot (10 trials each; entries: cloth lifted):

Grasp offsetRealFabricBenchAttach (GarmentLab)Attach (DexGarmentLab code)
None10/1010/1010/1010/10
10 mm above0/100/1010/1010/10
30 mm outward0/100/1010/1010/10

When the fingers close above or beside the cloth, the real robot and FabricBench miss, while both blocks still lift it. Because the blocks select vertices by distance, they also cannot tell layers apart: when several layers lie under the fingers, they bind all of them. In a test we designed, unfolding a sleeve folded onto the garment body, both blocks bind the body beneath together with the sleeve, so the sleeve is unfolded in 0/10 trials versus 10/10 with FabricBench. Collecting data this way is also laborious: for folded or crumpled cloth, as in our folding and flattening tasks, handling wrong bindings and tuning the blocks for each task is cumbersome and does not resolve the underlying problem, and wrong bindings produce erroneous demonstrations, i.e., grasps that would fail on the real robot.

On the evaluation side, the threshold-triggered attach–detach used in existing benchmarks misjudges policies: under closed-loop control, small fluctuations in the policy's gripper closing trigger spurious detachments that drop the fabric mid-grasp, so the same Diffusion Policy checkpoints succeed in 9/50 instead of 19/50 on Lift silk (Table 2b). These failures stem from the grasp model rather than the policy. In both cases, success does not reflect real grasping, which is exactly what a benchmark must measure. FabricBench avoids both errors, as grasping, holding and releasing arise from contact alone.

Reviewer_sH18OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper focuses on standardized policy benchmarking for bimanual fabric manipulation with a contact-driven deformable simulator. The authors argue for the need for a benchmark capable of controlled tests of configuration and appearance generalization. To this end, they introduce FabricBench, featuring six fabric-manipulation tasks spanning garments, silk and bags, 10 000 simulation-ready assets, and an improved open-source cloth solver that enables grasping, release, and load transfer without artificial attachment.

Critical strengths
  1. The proposed FabricBench benchmark enables controlled, separate tests of configuration generalization and appearance generalization for bimanual fabric-manipulation policies, and the open-source cloth solver achieves grasping, release, and load transfer without artificial attachment.
  2. The work provides 10 000 simulation-ready fabric assets and conducts zero-shot deployment.
Critical weaknesses
  1. The paper lacks a qualitative comparative evaluation of the proposed cloth-rigid body contact model, making the explanation of the improvements neither intuitive nor clear. The analysis relies on textual descriptions and binary success/failure outcomes, failing to provide visual comparisons, such as image demonstrating penetration or release failures in baseline simulations.
  2. The task suite is relatively small, consisting of only six bimanual fabric-manipulation tasks. As generalist robot policies develop rapidly, the limited task coverage restricts the benchmark's capacity to comprehensively test diverse bimanual fabric-manipulation capabilities.
  3. The paper mentions many relevant benchmarks, but does not provide a detailed comparison.
Questions for authors

Please see the "Critical Weaknesses".

2 comments

W1: Qualitative comparison of contact behavior. We thank the reviewer and add visual comparisons: https://anonymous.4open.science/w/Fabric-Anonymous-23A0/website/index.html. Under the same dual-arm grasp–hold–release trajectory, side-by-side videos, next to the real robot, show that Isaac Sim PBD fails to grasp the cloth, that in VBD and Style3D the fingers penetrate the cloth, so it is not released, and that only FabricBench grasps, holds and releases the cloth as on the real robot. We will add key frames from these videos as a figure in the appendix and the videos to the supplementary material.


W2: Task suite size. The six tasks are chosen to be representative of the physical challenges in fabric manipulation (Sec. 4.1; Table 10): long-horizon folding with self-contact (Fold T-shirt); recovering a crumpled, self-occluded configuration, including by dynamic flinging (Flatten T-shirt, Flatten silk); pinch-grasping thin, slippery fabric without it slipping out (Lift silk); and lifting a fabric bag by its handles, empty or with an item inside (Carry bag, Pack snack). All six tasks require grasping, holding and releasing fabric through contact, which the tested simulators (Isaac Sim PBD, VBD, Style3D) do not support with stable simulation and reliable evaluation (Sec. 5.1): none of them completes the grasp–hold–release cycle (Table 1; details in App. B.7), and attach–detach grasping underestimates policy success (Table 2b). The suite is far from saturated: the best success on Fold T-shirt is only 36%, and under C2 no policy exceeds 14%, leaving substantial headroom as generalist policies improve. We thank the reviewer for this suggestion; scaling up the suite with more tasks and fabric categories is part of our future work (Sec. 6).


W3: Comparison with existing benchmarks. We thank the reviewer and will add the following comparison to Sec. 2 of the revision.

Contact-driven grasp & releaseBag with contentsReported generalization axesPolicies
SoftGym✗ attach✗–RL
Def. Ravens✗ attach◐ attached bag–IL
DaXBench✗ primitive✗–RL, IL, planning
GarmentLab✗ attach✗–RL
DexGarmentLab✗ PBDᵃ✗unseen garmentsIL
LeHome✗ PBDᵃ✗–IL, VLA
FlatLab✗ PBDᵃ✗unseen objects / categoriesIL, VLA
Ours✓✓deformation state; appearanceIL, VLA, WAM

ᵃ Isaac Sim PBD: fails to grasp the cloth in our replay (Table 1).

These suites complete their tasks, but through attachment, adhesion, pick-and-place primitives or task-specific grasp strategies; under the same recorded parallel-jaw grasp, Isaac Sim PBD forms no grasp, and simulators that do hold fail to release (Table 1), so outcomes partly reflect the grasp model rather than the policy (Table 2b). FabricBench is the only suite in which grasp, release and bag load transfer arise from contact forces alone, and the only one that evaluates policies along deformation-state and appearance axes separately, reporting a stage-wise progress score in addition to success rate. Rigid-object suites such as RoboTwin 2.0, LIBERO and RoboDojo are not listed, as they contain no deformable task or only a single one (Sec. 2).

We hope these additions address the reviewer's concerns and would be happy to clarify any further questions.

Thank you again for your review. As the discussion period ends in a few hours, we would like to kindly remind you of our responses above: side-by-side videos comparing contact behavior across simulators and the real robot (W1), a clarification of the task suite's coverage (W2), and a detailed comparison table with existing benchmarks (W3). We would greatly appreciate it if you could let us know whether these address your concerns, and, if so, consider reflecting this in your assessment.

Best regards, The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.