Pith. sign in

REVIEW 3 major objections 6 minor 35 references

Success-only scores hide unsafe grasps of soft objects; tactile feedback raises safe completion and lowers deformation.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · grok-4.5

2026-07-11 20:45 UTC pith:ZDGOH2UU

load-bearing objection Useful safety-aware soft-object benchmark with a real Goal–Safety gap, but the tactile claim is confounded by continuous vs binary gripper actions. the 3 major comments →

arxiv 2607.04234 v1 pith:ZDGOH2UU submitted 2026-07-05 cs.RO cs.AIcs.CV

SoftVTBench: A Safety-Aware Visuo-Tactile Benchmark for Physically Constrained Robotic Manipulation of Deformable Objects

Bowen Jing , Mingxin Wang , Ruiyang Hao , Chenchen Ge , Hanwen Shen , Junjie He , Yang Cui , Yiming Hou
show 10 more authors
Weitao Zhou Jiawei Wang Minglei Li Dandan Zhang Ding Zhao Houde Liu Xiaofan Li Si Liu Ping Luo Haibao Yu
This is my paper
classification cs.RO cs.AIcs.CV
keywords deformable object manipulationvisuo-tactile sensingsafety-aware evaluationSoftVTBenchFEM deformationGoal SuccessSafety Successrobotic grasping
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

SoftVTBench argues that finishing a pick-and-place task on a deformable object is not enough: the robot must also keep a stable grasp without drop or slip and keep peak object deformation under an object-specific limit. Existing robot benchmarks mostly score only whether the object ends up in the right place, so they can treat over-squeezed or barely-held rollouts as full successes. SoftVTBench is a closed-loop Isaac Sim suite with finite-element soft bodies, multi-view RGB, dual-finger tactile RGB and marker motion, proprioception, and language instructions. It reports Goal Success and a stricter Safety Success that uses hidden FEM states to reject drops and over-deformation. Matched rigid and soft suites separate ordinary manipulation skill from safe soft-body contact. Baseline policies show that many goal-complete episodes still fail safety, and that adding tactile sensing lifts Safety Success while Goal Success stays comparable and continuous deformation falls.

Core claim

On deformable grasp-and-place tasks, success-only evaluation substantially overstates policy quality because a large share of goal-completing rollouts still drop, slip, or over-deform the object. Under the same protocol, a visuo-tactile policy raises Safety Success relative to a vision-only policy (Object-Soft 21.4% to 35.6%; Spatial-Soft 32.6% to 44.6%) while Goal Success stays comparable or improves modestly, and the full FEM deformation distribution shifts downward.

What carries the argument

Safety Success: Goal Success and no drop and peak object-size-normalized FEM-RMS deformation D_peak ≤ τ_o, where τ_o is an offline-calibrated object-specific threshold from a compression sweep (default κ = 0.5 times the max stable reference deformation), measured only from policy-hidden privileged FEM states.

Load-bearing premise

That a one-time offline compression-sweep threshold on simulated peak FEM deformation is a faithful stand-in for real physical safety of soft objects under closed-loop robot contact.

What would settle it

Run the same vision-only and visuo-tactile policies on SoftVTBench deformable suites and check whether Safety Success remains far below Goal Success and whether tactile feedback still raises Safety Success and lowers the FEM deformation distribution under the reported protocol.

Watch this falsifier — get emailed when new claim-graph text bears on it.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 6 minor

Summary. SoftVTBench is a simulation benchmark (Isaac Sim / PhysX FEM) for physically constrained deformable-object manipulation with visuo-tactile observations. It defines four matched task suites over object type (soft vs. rigid) and variation axis (object vs. spatial), and reports Goal Success separately from Safety Success, where the latter requires no drop/slip and peak object-size-normalized FEM-RMS deformation below an offline-calibrated object-specific threshold (Eqs. 3–4; §3.4). Using π0.5 LoRA baselines, the paper claims that success-only metrics substantially overstate performance because many goal-completing rollouts violate safety, and that adding tactile RGB and marker-motion history improves Safety Success (e.g., Object-Soft 21.4%→35.6%; Spatial-Soft 32.6%→44.6%) while keeping Goal Success comparable and shifting the deformation distribution downward (Tables 3–4, Fig. 4).

Significance. If the Goal–Safety gap and the tactile-safety benefit hold under fair controls, the work fills a clear gap at the intersection of deformable manipulation, visuo-tactile sensing, and process-level physical safety—areas that existing suites (LIBERO, SoftGym, MoDeSuite, UniVTAC, SoGraB, etc.) address only partially (Table 1). Strengths that should be credited: the matched 2×2 suite design, policy-hidden privileged FEM evaluation, offline per-object calibration of τ_o, fixed-seed closed-loop protocol, deformation distribution analysis (not only threshold rates), and public code/website. These make the benchmark a useful, falsifiable testbed for contact-aware policies even if some baseline claims need tightening.

major comments (3)
  1. §4.1 and Tables 3–4: the headline causal claim that “incorporating tactile sensing improves Safety Success … while maintaining comparable Goal Success” is confounded by the gripper action interface. VO uses binary open/close; VT uses continuous gripper-width. Continuous closure is precisely the degree of freedom that regulates compression against τ_o = κ D_ref_o (Eq. 4; Appendix C.3, κ=0.5). The paper withholds continuous width from VO to avoid “implicit interaction information,” yet grants it only to VT. Safety Success and the downward shift in FEM-RMS (Table 4) therefore cannot be attributed cleanly to tactile RGB/marker history versus finer gripper control. An action-matched ablation (binary–binary and continuous–continuous, with and without tactile) is load-bearing for the tactile claim; without it the claim should be restated as a joint modality+action effect.
  2. §4.1–4.2: all reported policies are LoRA fine-tunes of a single architecture (π0.5) with a fixed training recipe (8×A100, batch 256, 7k steps, action horizon 50, execute 10). The abstract and conclusion phrase results as properties of “policies” and of “tactile sensing” in general. With only one backbone, architecture- or training-specific effects cannot be separated from the observation modality. At minimum, the manuscript should either (i) add one independent baseline family under the same protocol, or (ii) explicitly scope all causal language to π0.5-style VLA policies and treat broader claims as hypotheses for future work.
  3. §3.4 and Appendix C.3: Safety Success hinges on peak FEM-RMS after rigid-motion removal and on τ_o = κ D_ref_o from a compression sweep (default κ=0.5; sensitivity mentioned for {0.3,0.7} but not shown in the main tables). The Goal–Safety gap is large enough that the qualitative overstatement claim is robust, but ranking of methods and the reported absolute Safety Success rates are threshold-dependent. Main-text results should include the κ sensitivity (or an equivalent continuous deformation metric already partially in Table 4) so that the safety ranking is not tied to a single ad-hoc scale factor.
minor comments (6)
  1. Abstract and §1: “nearly 2,000 collected episodes” vs. Table 6 total of 2,000 demos / 200 val episodes—align the wording with the table.
  2. Table 3: Safety Success for rigid suites is marked “–” in the table but described as reducing to Goal Success ∧ NoDrop in §3.4; either report the NoDrop-conditioned rate or state N/A consistently with Fig. 4.
  3. §4.1: tactile encoding (8-frame history, 4×4 grid, same visual encoder) is underspecified relative to marker-motion history length and normalization; a short appendix table would aid reproducibility.
  4. Fig. 3 caption and body: useful qualitative evidence, but no quantitative link (e.g., correlation of marker shear with D(t)) is given; a brief note would strengthen the interpretation.
  5. Typos / polish: “π0.5” vs. “pi0.5” in the abstract; “V ariation” spacing artifacts in Tables 2 and 6; ensure arXiv author list and affiliation markers match the PDF header.
  6. Limitations (§5) correctly note sim-to-real and asset diversity; consider also stating that NoDrop includes “workspace escape” and “loss of stable containment,” which may mix kinematic failure modes with contact safety.

Circularity Check

0 steps flagged

No circularity: SoftVTBench is an empirical benchmark-plus-ablation paper whose Safety Success metric and VO/VT comparisons are not forced by construction from their inputs.

full rationale

The paper does not claim a first-principles derivation or a prediction that reduces to fitted inputs. Goal Success and Safety Success are evaluation predicates defined from policy-hidden privileged FEM states (Eqs. 3–5; §3.4): Safety Success = Goal Success ∧ NoDrop ∧ (D_peak ≤ τ_o), with τ_o = κ D_ref_o obtained from an offline compression-sweep calibration (Appendix C.3). Those thresholds are fixed evaluation knobs applied uniformly to all methods; they are not fitted to the reported VO–VT gap and do not make the measured Safety Success rates tautological. The main experimental claims (Table 3–4: Goal Success comparable, Safety Success higher and deformation lower for VT on soft suites) are closed-loop rollout measurements under a shared protocol, not quantities recovered by construction from the training objective or from self-cited uniqueness theorems. π0.5 is an external baseline (Physical Intelligence et al., 2025); related-work citations are positioning, not load-bearing premises that force the result. Author-chosen κ (default 0.5) and the binary-vs-continuous gripper encoding are design choices that may confound causal attribution of the tactile benefit, but that is an experimental-validity concern, not circularity of a derivation chain. No self-definitional loop, fitted-input-as-prediction, or self-citation uniqueness import is present.

Axiom & Free-Parameter Ledger

5 free parameters · 5 axioms · 3 invented entities

The central claims rest on standard robotics/simulation practice plus a few author-chosen evaluation knobs (κ, FEM-RMS definition, NoDrop semantics) and the assumption that the Isaac/PhysX/TacEx stack is adequate for ranking visuo-tactile policies on soft contact. No new physical entities are postulated; SoftVTBench and Safety Success are methodological constructs. Free parameters are evaluation and sim hyperparameters, not fitted physical constants claimed as discoveries.

free parameters (5)
  • safety_threshold_scale_kappa = 0.5 (default)
    τ_o = κ D_ref_o with default κ=0.5 (also report 0.3, 0.7); directly decides which goal-success rollouts count as Safety Success.
  • D_ref_o_per_object = per-asset offline calibration
    Max stable FEM-RMS deformation from offline compression sweep; object-specific and defines the absolute safety scale.
  • FEM_solver_settings = hex=6, iters=64, damping=2.5
    Hex resolution 6, 64 position iterations, damping 2.5, corotational linear elasticity—control simulated compliance and thus D(t).
  • finger_friction = μ_s=1.5, μ_d=1.2
    Static μ_s=1.5, dynamic μ_d=1.2 affect slip/drop rates and thus Safety Success.
  • training_and_control_hyperparameters = 7k steps, horizon 50, exec 10
    LoRA fine-tune steps (7k), batch 256, action horizon 50, execute 10 steps before replan—affect absolute success levels of baselines.
axioms (5)
  • domain assumption PhysX GPU FEM with corotational linear elasticity plus TacEx/Taxim/FOTS tactile rendering is a valid closed-loop testbed for ranking visuo-tactile soft-manipulation policies.
    Invoked throughout §3 and Appendix A/C; no real-robot transfer study is provided.
  • ad hoc to paper Object-size-normalized FEM-RMS deformation after removing rigid motion is the right scalar for “excessive deformation.”
    Defined in §3.4 Eq. (3); alternative stress/strain or contact-force metrics are not compared as primary safety labels.
  • domain assumption NoDrop_episode (no transient drop, workspace escape, or loss of stable containment) is necessary for physical safety of the task class.
    Part of Safety Success definition Eq. (4); standard for grasp-and-place but still a modeling choice for what counts as unsafe.
  • domain assumption Matched rigid suites isolate soft-object safety from basic manipulation competence.
    Design claim in §3.1–3.3 and Table 2; assumes rigid LIBERO-style objects are adequate controls.
  • ad hoc to paper Withholding continuous gripper calibration from VO while giving continuous width to VT is a fair modality comparison.
    Stated in §4.1 as avoiding implicit contact info for VO; still confounds gripper action representation with tactile input.
invented entities (3)
  • SoftVTBench no independent evidence
    purpose: Unified Isaac Sim benchmark environment, assets, observation API, and evaluation protocol for safety-aware visuo-tactile soft manipulation.
    The paper’s primary artifact; independent evidence is the public code/website claim, not external prior existence.
  • Safety Success metric no independent evidence
    purpose: Binary label requiring Goal Success ∧ NoDrop ∧ (D_peak ≤ τ_o) to expose unsafe goal completions.
    Defined in §3.4; methodological construct rather than a physical discovery.
  • Goal-Safety Gap no independent evidence
    purpose: Interpretive quantity (Goal Success − Safety Success) highlighting overstatement by success-only evaluation.
    Named in Fig. 2 and §4.4; derived from the two metrics, not an independent physical entity.

pith-pipeline@v1.1.0-grok45 · 20483 in / 3946 out tokens · 44166 ms · 2026-07-11T20:45:09.152050+00:00 · methodology

0 comments
read the original abstract

Deformable object manipulation poses challenges beyond task completion: successful execution must also maintain safe physical interaction, holding the object stably without slip or drop while avoiding excessive deformation. However, existing manipulation benchmarks are predominantly success-oriented and rarely evaluate whether a policy remains physically safe throughout execution. We present SoftVTBench, a safety-aware visuo-tactile benchmark for physically constrained deformable object manipulation. Built in Isaac Sim with finite-element-simulated deformable objects, SoftVTBench provides multi-view RGB observations, RGB tactile sensing with marker motion, proprioception, and language instructions, and defines four matched task suites over object type (deformable vs. rigid) and variation axis (object vs. spatial). It separately reports Goal Success and Safety Success; the latter additionally requires no drop and peak deformation below a calibrated object-specific threshold, measured from policy-hidden privileged Finite Element Method (FEM) states. We implement pi0.5-based baselines under this protocol. Experiments show that success-only evaluation substantially overstates policy performance, as a large fraction of goal-completing rollouts still violate physical safety. Furthermore, incorporating tactile sensing improves Safety Success (e.g., from 21.4% to 35.6% on object-centric deformable tasks) and reduces object deformation during execution, while maintaining comparable Goal Success. SoftVTBench provides a reproducible benchmark for studying visuo-tactile deformable manipulation under physical interaction constraints.

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

35 extracted references · 1 canonical work pages

  1. [1]

    Rika Antonova, Peiyang Shi, Hang Yin, Zehang Weng, and Danica Kragic Jensfelt

    doi: 10.1109/TRO.2025.3547267. Rika Antonova, Peiyang Shi, Hang Yin, Zehang Weng, and Danica Kragic Jensfelt. Dynamic environments with deformable objects. InThirty-fifth Conference on Neural Information Processing Systems Datasets and Benchmarks Track (Round 2),

    doi:10.1109/tro.2025.3547267 2025

  2. [2]

    Vla-touch: Enhancing vision-language-action models with dual-level tactile feedback.arXiv preprint arXiv:2507.17294,

    Jianxin Bi, Kevin Yuchen Ma, Ce Hao, Mike Zheng Shou, and Harold Soh. Vla-touch: Enhancing vision-language-action models with dual-level tactile feedback.arXiv preprint arXiv:2507.17294,

    Pith/arXiv arXiv

  3. [3]

    Gr00t n1: An open foundation model for generalist humanoid robots.arXiv preprint arXiv:2503.14734,

    Johan Bjorck, Fernando Castañeda, Nikita Cherniadev, Xingye Da, Runyu Ding, Linxi Fan, et al. Gr00t n1: An open foundation model for generalist humanoid robots.arXiv preprint arXiv:2503.14734,

    Pith/arXiv arXiv

  4. [4]

    Kevin Black, Noah Brown, Danny Driess, Adnan Esmail, Michael Equi, Chelsea Finn, Niccolo Fusai, Lachy Groom, Karol Hausman, Brian Ichter, et al.π0: A vision-language-action flow model for general robot control.arXiv preprint arXiv:2410.24164,

    Pith/arXiv arXiv

  5. [5]

    Univtac: A unified simulation platform for visuo-tactile manipulation data generation, learning, and benchmarking.arXiv preprint arXiv:2602.10093,

    Baijun Chen, Weijie Wan, Tianxing Chen, Xianda Guo, Congsheng Xu, Yuanyang Qi, Haojie Zhang, Longyan Wu, Tianling Xu, Zixuan Li, et al. Univtac: A unified simulation platform for visuo-tactile manipulation data generation, learning, and benchmarking.arXiv preprint arXiv:2602.10093,

    arXiv

  6. [6]

    Daxbench: Benchmarking deformable object manipulation with differentiable physics.arXiv preprint arXiv:2210.13066,

    Siwei Chen, Yiqing Xu, Cunjun Yu, Linfeng Li, Xiao Ma, Zhongwen Xu, and David Hsu. Daxbench: Benchmarking deformable object manipulation with differentiable physics.arXiv preprint arXiv:2210.13066,

    Pith/arXiv arXiv

  7. [7]

    Robotwin 2.0: A scalable data generator and benchmark with strong domain randomization for robust bimanual robotic manipulation.arXiv preprint arXiv:2506.18088,

    Tianxing Chen, Zanxin Chen, Baijun Chen, Zijian Cai, Yibin Liu, Zixuan Li, Qiwei Liang, Xianliang Lin, et al. Robotwin 2.0: A scalable data generator and benchmark with strong domain randomization for robust bimanual robotic manipulation.arXiv preprint arXiv:2506.18088,

    Pith/arXiv arXiv

  8. [8]

    Om- nivtla: Vision-tactile-language-action models with semantic-aligned tactile sensing.arXiv preprint arXiv:2508.08706,

    Zhengxue Cheng, Yiqian Zhang, Anni Tang, Keyu Wang, Wenkang Zhang, Haoyu Li, Hengdi Zhang, and Li Song. Om- nivtla: Vision-tactile-language-action models with semantic-aligned tactile sensing.arXiv preprint arXiv:2508.08706,

    arXiv

  9. [9]

    Safevla-bench: A benchmark for the success-safety gap in vision-language-action models.arXiv preprint arXiv:2606.00773,

    Jialiang Fan, Weizhe Xu, Oleg Sokolsky, Insup Lee, and Fanxin Kong. Safevla-bench: A benchmark for the success-safety gap in vision-language-action models.arXiv preprint arXiv:2606.00773,

    Pith/arXiv arXiv

  10. [10]

    Greenland, Josh Pinskier, Xing Wang, Daniel Nguyen, Ge Shi, Tirthankar Bandyopadhyay, Jen Jen Chung, and David Howard

    Benjamin G. Greenland, Josh Pinskier, Xing Wang, Daniel Nguyen, Ge Shi, Tirthankar Bandyopadhyay, Jen Jen Chung, and David Howard. Sograb: A visual method for soft grasping benchmarking and evaluation.arXiv preprint arXiv:2411.19408,

    Pith/arXiv arXiv

  11. [11]

    Maniskill2: A unified benchmark for generalizable manipulation skills.arXiv preprint arXiv:2302.04659,

    Jiayuan Gu, Fanbo Xiang, Xuanlin Li, Zhan Ling, Xiqiang Liu, Tongzhou Mu, Yihe Tang, Stone Tao, Xinyue Wei, Yunchao Yao, et al. Maniskill2: A unified benchmark for generalizable manipulation skills.arXiv preprint arXiv:2302.04659,

    Pith/arXiv arXiv

  12. [12]

    3d-vitac: Learning fine-grained manipulation with visuo-tactile sensing.arXiv preprint arXiv:2410.24091,

    Binghao Huang, Yixuan Wang, Xinyi Yang, Yiyue Luo, and Yunzhu Li. 3d-vitac: Learning fine-grained manipulation with visuo-tactile sensing.arXiv preprint arXiv:2410.24091,

    Pith/arXiv arXiv

  13. [13]

    Safemanip: A property-driven benchmark for temporal safety evaluation in robotic manipulation.arXiv preprint arXiv:2605.12386,

    Chengyue Huang, Khang Vo Huynh, Sebastian Elbaum, Zsolt Kira, and Lu Feng. Safemanip: A property-driven benchmark for temporal safety evaluation in robotic manipulation.arXiv preprint arXiv:2605.12386,

    Pith/arXiv arXiv

  14. [14]

    Tactile-vla: Unlocking vision-language- action model’s physical knowledge for tactile generalization.arXiv preprint arXiv:2507.09160,

    11 Jialei Huang, Shuo Wang, Fanqi Lin, Yihang Hu, Chuan Wen, and Yang Gao. Tactile-vla: Unlocking vision-language- action model’s physical knowledge for tactile generalization.arXiv preprint arXiv:2507.09160,

    Pith/arXiv arXiv

  15. [15]

    Tenenbaum, and Chuang Gan

    Zhiao Huang, Yuanming Hu, Tao Du, Siyuan Zhou, Hao Su, Joshua B. Tenenbaum, and Chuang Gan. Plasticinelab: A soft-body manipulation benchmark with differentiable physics.arXiv preprint arXiv:2104.03311,

    Pith/arXiv arXiv

  16. [16]

    Dreamgen: Unlocking generalization in robot learning through video world models

    Joel Jang, Seonghyeon Ye, Zongyu Lin, Jiannan Xiang, Johan Bjorck, Yu Fang, Fengyuan Hu, Spencer Huang, Kaushil Kundalia, Yen-Chen Lin, et al. Dreamgen: Unlocking generalization in robot learning through video world models. arXiv preprint arXiv:2505.12705,

    Pith/arXiv arXiv

  17. [17]

    Openvla: An open-source vision-language-action model.arXiv preprint arXiv:2406.09246,

    Moo Jin Kim, Karl Pertsch, Siddharth Karamcheti, Ted Xiao, Ashwin Balakrishna, Suraj Nair, Rafael Rafailov, Ethan Foster, Pannag Sanketi, Quan Vuong, et al. Openvla: An open-source vision-language-action model.arXiv preprint arXiv:2406.09246,

    Pith/arXiv arXiv

  18. [18]

    At-vla: Adaptive tactile injection for enhanced feedback reaction in vision-language-action models.arXiv preprint arXiv:2605.07308,

    Xiaoqi Li, Muhe Cai, Jiadong Xu, Juan Zhu, Hongwei Fan, Yan Shen, Guangrui Ren, and Hao Dong. At-vla: Adaptive tactile injection for enhanced feedback reaction in vision-language-action models.arXiv preprint arXiv:2605.07308,

    Pith/arXiv arXiv

  19. [19]

    Evaluating real-world robot manipulation policies in simulation.arXiv preprint arXiv:2405.05941,

    Xuanlin Li, Kyle Hsu, Jiayuan Gu, Karl Pertsch, Oier Mees, Homer Rich Walke, Chuyuan Fu, Ishikaa Lunawat, Isabel Sieh, Sean Kirmani, et al. Evaluating real-world robot manipulation policies in simulation.arXiv preprint arXiv:2405.05941,

    Pith/arXiv arXiv

  20. [20]

    Softgym: Benchmarking deep reinforcement learning for deformable object manipulation.arXiv preprint arXiv:2011.07215,

    Xingyu Lin, Yufei Wang, Jake Olkin, and David Held. Softgym: Benchmarking deep reinforcement learning for deformable object manipulation.arXiv preprint arXiv:2011.07215,

    Pith/arXiv arXiv 2011

  21. [21]

    Libero: Benchmarking knowledge transfer for lifelong robot learning.arXiv preprint arXiv:2306.03310,

    Bo Liu, Yifeng Zhu, Chongkai Gao, Yihao Feng, Qiang Liu, Yuke Zhu, and Peter Stone. Libero: Benchmarking knowledge transfer for lifelong robot learning.arXiv preprint arXiv:2306.03310,

    Pith/arXiv arXiv

  22. [22]

    Manifeel: Benchmarking and understanding visuotactile manipulation policy learning.arXiv preprint arXiv:2505.18472,

    Quan Khanh Luu, Pokuang Zhou, Zhengtong Xu, Zhiyuan Zhang, Qiang Qiu, and Yu She. Manifeel: Benchmarking and understanding visuotactile manipulation policy learning.arXiv preprint arXiv:2505.18472,

    arXiv

  23. [23]

    Masoud Moghani, Mahdi Azizian, Animesh Garg, Yuke Zhu, Sean Huver, and Ajay Mandlekar

    doi: 10.1109/LRA.2022.3180108. Masoud Moghani, Mahdi Azizian, Animesh Garg, Yuke Zhu, Sean Huver, and Ajay Mandlekar. Softmimicgen: A data generation system for scalable robot learning in deformable object manipulation.arXiv preprint arXiv:2603.25725,

    doi:10.1109/lra.2022.3180108 2022

  24. [24]

    Robotwin: Dual-arm robot benchmark with generative digital twins.arXiv preprint arXiv:2504.13059,

    Yao Mu, Tianxing Chen, Zanxin Chen, Shijia Peng, Zhiqian Lan, Zeyu Gao, Zhixuan Liang, Qiaojun Yu, Yude Zou, Mingkun Xu, et al. Robotwin: Dual-arm robot benchmark with generative digital twins.arXiv preprint arXiv:2504.13059,

    Pith/arXiv arXiv

  25. [25]

    Robocasa: Large-scale simulation of everyday tasks for generalist robots.arXiv preprint arXiv:2406.02523,

    Soroush Nasiriany, Abhiram Maddukuri, Lance Zhang, Adeet Parikh, Aaron Lo, Abhishek Joshi, Ajay Mandlekar, and Yuke Zhu. Robocasa: Large-scale simulation of everyday tasks for generalist robots.arXiv preprint arXiv:2406.02523,

    Pith/arXiv arXiv

  26. [26]

    Tacex: Gelsight tactile simulation in isaac sim: Combining soft-body and visuotactile simulators.arXiv preprint arXiv:2411.04776,

    12 Duc Huy Nguyen, Tim Schneider, Guillaume Duret, Alap Kshirsagar, Boris Belousov, and Jan Peters. Tacex: Gelsight tactile simulation in isaac sim: Combining soft-body and visuotactile simulators.arXiv preprint arXiv:2411.04776,

    Pith/arXiv arXiv

  27. [27]

    Physical Intelligence, Kevin Black, Noah Brown, James Darpinian, Karan Dhabalia, Danny Driess, Adnan Esmail, Michael Equi, Chelsea Finn, Niccolo Fusai, Manuel Y. Galliker, Dibya Ghosh, Lachy Groom, Karol Hausman, Brian Ichter, Szymon Jakubczak, Tim Jones, Liyiming Ke, Devin LeBlanc, Sergey Levine, Adrian Li-Bell, Mohith Mothukuri, Suraj Nair, Karl Pertsch...

    Pith/arXiv arXiv

  28. [28]

    The colosseum: A benchmark for evaluating generalization for robotic manipulation.arXiv preprint arXiv:2402.08191,

    Wilbert Pumacay, Ishika Singh, Jiafei Duan, Ranjay Krishna, Jesse Thomason, and Dieter Fox. The colosseum: A benchmark for evaluating generalization for robotic manipulation.arXiv preprint arXiv:2402.08191,

    Pith/arXiv arXiv

  29. [29]

    Zilin Si, Gu Zhang, Qingwei Ben, Branden Romero, Zhou Xian, Chao Liu, and Chuang Gan

    doi: 10.1109/LRA.2022.3142412. Zilin Si, Gu Zhang, Qingwei Ben, Branden Romero, Zhou Xian, Chao Liu, and Chuang Gan. Difftactile: A physics-based differentiable tactile simulator for contact-rich robotic manipulation.arXiv preprint arXiv:2403.08716,

    doi:10.1109/lra.2022.3142412 2022

  30. [30]

    Demavla: A vision-language-action foundation model for generalizable deformable manipulation.arXiv preprint arXiv:2605.31286,

    Taiyi Su, Jian Zhu, Tianjian Wang, Youzhang He, Zitai Huang, Jianjun Zhang, Chong Ma, Hanyang Wang, Tianjiao Zhang, Munan Yin, et al. Demavla: A vision-language-action foundation model for generalizable deformable manipulation.arXiv preprint arXiv:2605.31286,

    Pith/arXiv arXiv

  31. [31]

    Tabero: Learning gentle manipulation with closed-loop force feedback from vision, touch, and language.arXiv preprint arXiv:2605.27886,

    Qiwei Wu, Rui Zhang, Xin Xiang, Tao Li, Weihua Zhang, Junjie Lai, and Renjing Xu. Tabero: Learning gentle manipulation with closed-loop force feedback from vision, touch, and language.arXiv preprint arXiv:2605.27886,

    Pith/arXiv arXiv

  32. [32]

    Borong Zhang, Yuhao Zhang, Jiaming Ji, Yingshan Lei, Juntao Dai, Yuanpei Chen, and Yaodong Yang

    doi: 10.3390/s17122762. Borong Zhang, Yuhao Zhang, Jiaming Ji, Yingshan Lei, Juntao Dai, Yuanpei Chen, and Yaodong Yang. Safevla: Towards safety alignment of vision-language-action model via constrained learning.Advances in Neural Information Processing Systems, 38:153335–153373, 2025a. Chaofan Zhang, Peng Hao, Xiaoge Cao, Xiaoshuai Hao, Shaowei Cui, and ...

    Pith/arXiv arXiv doi:10.3390/s17122762

  33. [33]

    13 Yuhang Zheng, Songen Gu, Weize Li, Yupeng Zheng, Yujie Zang, Shuai Tian, Xiang Li, Ce Hao, Chen Gao, Si Liu, et al

    doi: 10.1109/LRA.2024.3396665. 13 Yuhang Zheng, Songen Gu, Weize Li, Yupeng Zheng, Yujie Zang, Shuai Tian, Xiang Li, Ce Hao, Chen Gao, Si Liu, et al. Omnivta: Visuo-tactile world modeling for contact-rich robotic manipulation.arXiv preprint arXiv:2603.19201,

    work page doi:10.1109/lra.2024.3396665 2024

  34. [34]

    Taco: Benchmarking tactile sensors for object manipulation.arXiv preprint arXiv:2605.21976,

    Anya Zorin, Zilin Si, Myungsun Park, Junsung Park, Alexiy Buynitsky, Sachin Bhadang, Taejun Park, Sohee John Yoon, Yong-Lae Park, Oliver Kroemer, et al. Taco: Benchmarking tactile sensors for object manipulation.arXiv preprint arXiv:2605.21976,

    Pith/arXiv arXiv

  35. [35]

    The simulator runs physics at 60Hz with a control decimation of 3, resulting in a 20Hz control and logging rate

    14 A Implementation Details SoftVTBench is implemented in Isaac Sim 4.5.0 with Isaac Lab 0.41.3 and the GPU-accelerated PhysX 5 pipeline. The simulator runs physics at 60Hz with a control decimation of 3, resulting in a 20Hz control and logging rate. All visual, tactile, proprioceptive, action, and privileged-state streams are synchronized at this rate. T...

    2000