Pith. sign in

REVIEW 4 major objections 5 minor 39 references

A single non-destructive stretch-and-release of an unseen rubber band suffices to identify its five elastic parameters and train a slingshot policy that transfers zero-shot to a real robot arm, hitting targets at 1.7–2.1 m more accurately t

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 →

A one-shot Real2Sim2Real framework fits five elastic parameters from a single non-destructive interaction and zero-shot transfers a slingshot RL policy to a real Franka arm.

T0 review reviewed 2026-07-31 challenge →

load-bearing objection Solid, well-executed engineering with real robot data, but the one-shot transfer claim rests on an untested premise: the pull-release calibration never validates that its five parameters hold for projectile-loaded launches. the 4 major comments →

arxiv 2607.23268 v1 pith:TCEVAWDJ submitted 2026-07-25 cs.RO

Sling2Sim2Real: One-Shot Elastic System Identification for Non-Destructive Slingshot Policy Learning

classification cs.RO
keywords elastic object manipulationsystem identificationsim-to-real transferslingshotdifferential evolutionCMA-ESreinforcement learningphysics simulator calibration
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.

The reading

This paper tries to establish that a robot can learn to use an unfamiliar elastic slingshot without any trial-and-error practice on the real device. The proposed pipeline performs one 'stretch and release' of the band by hand, records force, motion, and point-cloud data for about nine seconds, then searches a simulator's five elastic parameters until the simulated stretch matches the real one. With the calibrated simulator, a reinforcement-learning policy is trained to choose a pull position for a given target, and that policy is deployed directly on the real robot arm. Across three rubber bands with different stiffness and three target distances, the method achieves the lowest projectile landing errors among the compared calibration procedures, and the paper shows the calibration loss is lower for the proposed method as well. The reason this matters is that elastic objects look similar but feel different, so visual-only calibration fails; the paper's claim is that one cheap, non-destructive interaction carries enough information to close the sim-to-real gap.

Core claim

The central discovery is that a single interaction episode—grasp, 40 cm pull, pause until tension saturates, release, all without a projectile—provides enough constraint to identify the five parameters that dominate elastic-band behavior in simulation: Young's modulus, Poisson's ratio, elasticity damping, dynamic friction, and damping scale. The paper argues that matching the real and simulated interaction through a composite loss (visual Chamfer distance on the band's point cloud, per-step force error, and accumulated mechanical work from the force-velocity product) produces a simulator whose transition model is accurate enough that an RL policy trained solely in that simulator transfers to

What carries the argument

The load-bearing object is the covariance-informed multi-start optimizer over a five-parameter elastic model. Differential evolution first explores globally and keeps a search history; non-maximum suppression picks the most distinct low-loss regions; for each region, the empirical covariance of its nearest neighbors defines an anisotropic Gaussian, which is injected into a CMA-ES solver through a Cholesky reparameterization so that sampling starts already aligned with local parameter correlations. Five such solvers run in parallel and the best result is kept. The objective being optimized is a composite loss with three terms: a point-cloud Chamfer distance between real and simulated band geo

Load-bearing premise

The load-bearing premise is that one non-destructive 40 cm stretch-and-release without a projectile reveals the same elastic behavior that governs a launch with a 39 g projectile riding the band, and that five parameters are identifiable from that single episode; the paper offers no launch-vs-pull consistency check, and its own ablation shows the refined parameters can slightly lose accuracy at the farthest target distance.

What would settle it

Take the softest band, perform the identification twice on different days, train two policies, and launch at a fixed pull; if the two identified parameter sets differ enough to change the optimal pull by more than 1 cm, or if the calibrated simulator's predicted impact point for that pull deviates from the real impact by more than the reported error bars, the one-shot transfer premise fails.

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

If this is right

  • If the claim holds, calibrating an elastic-object simulator for a new band costs one roughly nine-second non-destructive interaction instead of repeated real launches, removing the main data bottleneck for slingshot-style tasks.
  • Zero-shot transfer becomes practical for tasks where the physics parameters are visually indistinguishable but haptically recoverable: the robot measures force during a manipulative test that is also the first step of the real task.
  • The reduction of the task to a one-step MDP (choose a pull position; release) is itself a design choice that the paper identifies as reducing sensitivity to imperfect physics, suggesting that similar single-decision elastic tasks can inherit the same pipeline.
  • The reported error pattern—stiffer band, higher error—implies that calibration accuracy, not policy learning, is the remaining bottleneck; improving the simulator's high-strain behavior should directly improve transfer on stiff materials.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Not stated in the paper: the single pull may not excite the same deformation modes as a launch carrying a 39 g projectile; a direct check is to compare the calibrated simulator's predicted impact point and flight time against a real launch at the same pull position.
  • A second, untested consequence is that the same covariance-informed multi-start recipe applies to any black-box simulator calibration problem with correlated, multimodal parameters—not just elastic bands—so the contribution may be reusable beyond slingshots.
  • The paper's own ablation notes that at the farthest target distance the covariance-informed refinement slightly degraded performance relative to a simpler search, which suggests the calibrated parameters are not uniformly valid across the operating range and that target-specific calibration would be a natural extension.
  • Since the method's gains come from force-domain information, a cheaper variant that estimates force from vision or motor current could be tested; if the work-loss term is the key, the method's usefulness depends on how cheaply that signal can be obtained.
Share X Bluesky LinkedIn Reddit HN

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

4 major / 5 minor

Summary. The paper proposes Sling2Sim2Real, a one-shot Real2Sim2Real framework for slingshot manipulation. A Franka arm performs a single non-destructive stretch-and-release of an unseen rubber band, with synchronized force, end-effector velocity, and point-cloud observations. The method then identifies five simulation parameters (Young's modulus, Poisson's ratio, elasticity damping, dynamic friction, damping scale) by minimizing a composite loss consisting of Chamfer-distance geometric error, pointwise force error, and accumulated work error (Eqs. 2-5). The optimization is a two-stage hierarchical scheme: differential evolution for global exploration, followed by multiple parallel CMA-ES solvers initialized from covariance-informed regions. A PPO policy is trained in the calibrated Isaac Sim simulator for a one-step MDP slingshot task and deployed zero-shot on the real robot. Experiments with three rubber bands and three target distances report landing errors for the proposed method, two literature baselines, and several ablations; the authors report the lowest average landing errors at the aggregate level and claim successful zero-shot transfer.

Significance. If confirmed, the result is valuable: it addresses an underexplored aspect of real-to-sim calibration for elastic objects (damping), uses a genuinely non-destructive interaction, and evaluates on a real robot rather than only in simulation. The inclusion of force/work terms in the SI loss and the covariance-informed multi-start optimization are reasonable contributions. The real-robot landing errors are independent evidence in favor of the method, because they are not directly optimized by the SI objective. The main value is the possibility of reducing real-world interaction for EOM policy learning. However, the statistical evidence is thin and the central regime-transfer assumption is not explicitly validated, so the contribution is currently a promising demonstration rather than a fully supported claim.

major comments (4)
  1. [III-B.1, IV-A, Eqs. (2)-(5)] The central claim rests on an untested regime-transfer assumption. The SI loss L is minimized on D^real, a single non-destructive pull-release episode recorded without a projectile (Sec. III-B.1), while the policy is trained and deployed with a 39 g projectile attached to the band (Sec. IV-A). Table II's lower L values are in-sample: they show the fit to the calibration episode itself, not that the identified parameters reproduce launch dynamics. The paper's own Table III, Band 1 at g=212.5 cm, shows that the full method (9.50 cm) is worse than DE_L (3.00 cm), which is exactly the signature of an objective that is not consistently predictive of deployment performance. A launch-vs-pull consistency check is needed: for example, use the calibrated simulator to predict projectile landing positions for several pull displacements and compare directly with real launches, or at least evaluate th
  2. [IV-B, Table I] The statistical support is thin: five trials per band-target-method condition, with no repeated runs of the stochastic SI (DE/CMA-ES) and no confidence intervals or distributions for the identified parameter vector. Several headline comparisons are within one reported standard deviation or equal in mean (e.g., Band 1 at g=172.5 cm: 11.00±2.00 vs 13.00±1.87; Band 1 at g=212.5 cm: 9.50±4.30 vs 9.50±2.50). To support 'consistently outperforms,' the authors should repeat the SI procedure multiple times, report parameter dispersion, and either increase trials or use a formal significance test. This also directly addresses whether a single interaction reliably determines the five parameters.
  3. [Eqs. (3)-(5), Algorithm 1] The loss weights w_geom, w_force, w_work and the optimization hyperparameters K_top=5, K_NN=15, δ=0.1 are never given values or justified. Since L is the selection criterion for the calibrated simulator, these choices are load-bearing: different weightings would produce different parameter sets and, potentially, different transfer outcomes. Report the weight values, parameter bounds, solver budgets, and at least one sensitivity check (e.g., varying the weights by an order of magnitude) so the reader can assess the stability of the reported results.
  4. [III-B, Table II] The physical identifiability of the five parameters is not addressed. No ground-truth or independent measurement of Young's modulus, Poisson's ratio, or damping is provided, and the 'damping scale' parameter is not physically defined in the text. It is possible that the method finds a parameter set that is effective for landing-error minimization without uniquely identifying the true properties, which can be acceptable if deployment performance is the goal; but with only three bands, five trials, and no identifiability analysis, the paper should at least report parameter bounds, the covariance structure used, and the range of values explored. This would clarify whether the reported numbers are meaningful physical estimates or merely simulation tuning parameters.
minor comments (5)
  1. [V, Table II discussion] The sentence 'Lower landing distance errors correspond to lower composite loss' is too strong. It is an observed correlation over a small set of methods, and the Table III g=212.5 cm row is a counterexample. Recommend softening to 'lower calibration loss generally coincides with lower landing error in our experiments, with one exception'.
  2. [IV-B, Table I] The reported '±' values are not defined as standard deviations or standard errors, and the independence of the five trials is not described. Please clarify the definition and how the aggregate 'All Targets' column is computed (pooled trials vs. averaged per-condition means).
  3. [III-B.1] The force low-pass filter cutoff is stated as 0.224 rad/s, which is about 0.036 Hz and seems extremely low for a 9 s interaction. If this is correct, it may remove most of the force dynamics relevant to elasticity; if it is a typo or a different unit, please clarify.
  4. [III-B, Table II] The 'damping scale' parameter appears in the parameter set and in Table II but is never defined or justified. Specify its physical meaning, allowed range, and how it interacts with 'elasticity damping.'
  5. [Tables III and IV] The notation 'Sling2Sim2Real−c,−m' is confusing: the caption says the superscripts denote removal, so this row removes both covariance-informed initialization and multi-start. Consider writing '−c−m' or 'without covariance and multi-start' for clarity.

Circularity Check

1 steps flagged

Minor in-sample reporting of the optimized SI loss; central zero-shot landing-error claim is independently measured.

specific steps
  1. fitted input called prediction [Sec. III-B.2 (Eq. 2, Algorithm 1); Sec. V (Table II)]
    "We use a fitness function represented as a composite loss: L = L_geom + L_force + L_work ... return θ* ← arg min_{θ∈H} L(θ) ... The rightmost column reports the total loss L in Eq. (2), computed using the calibrated simulator. ... Table II demonstrates the effectiveness of the proposed composite loss L for calibrating high-fidelity elastic simulation parameters. Lower landing distance errors correspond to lower composite loss."

    θ* is selected by minimizing L on the same single non-destructive episode D_real, so the L values listed in Table II are the minimized training objective, not an independent calibration metric. Reporting that Sling2Sim2Real has the lowest L is therefore a restatement of the optimization criterion, not evidence that the fitted simulator predicts launch behavior. The paper itself concedes in the Table III discussion that refinement can degrade landing error despite lower calibration loss, confirming L is not a transfer predictor. The central landing-error result is measured on real launches not used in fitting, so the circularity is partial.

full rationale

The derivation chain is largely self-contained and externally validated: the paper fits five elastic parameters to one non-destructive pull-release episode by minimizing L (Eqs. 2–5), trains a policy in the calibrated Isaac Sim, and evaluates landing errors on real projectile launches that were not used in fitting. Those real landing errors are the load-bearing evidence for the one-shot transfer claim, and they do not reduce to the calibration objective. The only circular element is in the presentation of Table II: the loss column re-reports the exact objective minimized to select θ*, so 'lowest L' is in-sample by construction and is not an independent demonstration of calibration quality. The paper's own ablation text concedes that lower calibration loss does not always yield better transfer (g=212.5 cm row), corroborating this reading. The only self-citation, [10], supports a general related-work motivation and is not load-bearing. Hence no substantive circularity in the central claim; score 2 for the minor self-referential loss reporting.

Axiom & Free-Parameter Ledger

4 free parameters · 5 axioms · 1 invented entities

The paper's contribution is a fitting procedure; the ledger is dominated by the five fitted elastic parameters and the unreported loss weights. The damping-scale knob is the clearest invented entity. The two domain assumptions above are the ones that, if false, break the 'one-shot' claim.

free parameters (4)
  • θ: five-parameter elastic set (Young's modulus, Poisson's ratio, elasticity damping, dynamic friction, damping scale) = Band-dependent, e.g. [9.29E+5, 0.46, 1.11E-2, 0.80, 0.79] for Band 1 (Table II)
    These are the unknowns the SI pipeline fits to the single interaction episode; the central transfer claim depends on them being correct.
  • Loss weights w_geom, w_force, w_work = Not reported
    Hand-chosen constants in Eqs. (3)-(5) defining the SI objective; without them the loss is under-specified and the relative emphasis of geometry vs. force vs. work is a free choice affecting fitted parameters.
  • Damping scale = 0.71–0.79 (Sling2Sim2Real, Table II)
    An ad hoc multiplicative simulator knob (part of θ) introduced to absorb unmodeled dissipation; no physical counterpart is justified.
  • Algorithm hyperparameters (K_top=5, K_NN=15, δ=0.1) = As stated in Sec. III-B-2
    Hand-chosen; not ablated, so their necessity for the claimed performance is unverified.
axioms (5)
  • domain assumption The five-parameter tetrahedral-cylinder FEM model in Isaac Sim can represent the real rubber bands closely enough for zero-shot transfer after calibration
    Invoked in Sec. IV-A when the sling is modeled as a tetrahedral cylindrical mesh parameterized with five elasticity variables; if the model class cannot capture real hyperelastic/viscoelastic behavior under slingshot strain, even perfect SI fails. The authors partly concede this: 'simulation inaccuracies about elastic objects' (Sec. V).
  • domain assumption Single non-destructive stretch-and-release data are informative for the projectile-loaded launch dynamics
    The identification episode (Sec. III-B-1) contains no projectile and one prescribed pull/release; the deployment task carries a 39 g projectile and variable pull. The paper provides no consistency check that parameters identified on the pull transfer to launch.
  • domain assumption Rubber bands are treated as linearly elastic with damping and constant Poisson ratio within the tested range
    The five-parameter set is linear-elastic FEM; Treloar [1] (cited for entropic elasticity) indicates large-strain nonlinearity, so the linear model is a simplification, not an established reduction.
  • standard math DE/CMA-ES convergence to a low value of the composite loss implies physically correct parameters
    The paper optimizes a black-box composite loss; there is no analysis that the loss is unimodal or that the identified parameters are unique (NMS is used precisely because the landscape is multimodal). This is a standard stochastic-optimization assumption, not verified here.
  • domain assumption Preprocessing (low-pass at 0.224 rad/s, median filter window 7) preserves the dynamics relevant to SI
    The unusually low cutoff (≈0.036 Hz) could attenuate the release transient; the paper does not justify the value or show it does not distort f_real used in Eqs. (4)-(5).
invented entities (1)
  • Damping scale no independent evidence
    purpose: Multiplicative scaling of elastic damping in Isaac Sim used to match dissipated energy during the release phase
    Simulator-specific knob (values 0.71–0.79 fit per band in Table II) with no independent physical measurement; its only handle is the fitted loss and downstream task success, both inside the paper's own loop.

reviewed 2026-07-31 · how reviews work

0 comments
Cite this review

Pith. "Pith review of Sling2Sim2Real: One-Shot Elastic System Identification for Non-Destructive Slingshot Policy Learning." pith.science (2026). https://pith.science/paper/TCEVAWDJ

@misc{pith2026260723268,
  author       = {Pith},
  title        = {Pith review of: Sling2Sim2Real: One-Shot Elastic System Identification for Non-Destructive Slingshot Policy Learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/TCEVAWDJ}},
  note         = {Machine review of arXiv:2607.23268}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Elastic object manipulation (EOM) involves highdimensional, nonlinear, and elastic deformations. The diverse deformation properties of elastic objects substantially expand the relevant state space, requiring extensive exploration to learn accurate manipulation policies for tasks such as slingshot manipulation. While simulation enables large-scale and safe exploration compared to costly and potentially destructive real-world trials (e.g., repeated projectile launches), accurately calibrating elastic behavior between the real world and simulation remains challenging since elastic properties are largely indistinguishable from visual observations alone. To address these challenges, we propose Sling2Sim2Real, a one-shot Real2Sim2Real framework that identifies elastic parameters from a single non-destructive interaction and enables policy learning in simulation. The framework consists of two stages: 1) a multi-start Real2Sim system identification method that exploits parameter covariance to estimate elastic properties, and 2) simulation-based policy learning followed by zero-shot Sim2Real transfer using the calibrated simulator. We evaluate Sling2Sim2Real on a slingshot manipulation task using a Franka Emika Panda arm and elastic bands with diverse physical properties across varying target distances. Experimental results demonstrate that Sling2Sim2Real achieves accurate policy learning and robust generalization while significantly reducing the amount of required real-world interaction.

Figures

Figures reproduced from arXiv: 2607.23268 by Daehyung Park, Geonwoo Kim, Minseok Song, Wonjae Kang.

Figure 1
Figure 1. Figure 1: An overview of the Sling2Sim2Real steps. After [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: An overview of the Sling2Sim2Real framework that performs one-shot Real2Sim system identification (SI) and [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Three elastic bands used in the quantitative evaluation. [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Real-world demonstrations of Sim2Real results. The overlaid traces show the projectile trajectories after launch. [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗

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

39 extracted references · 4 linked inside Pith

  1. [1]

    The physics of rubber elasticity,

    L. G. Treloar, “The physics of rubber elasticity,” 1975

  2. [2]

    Real2sim2real: Self-supervised learning of physical single-step dynamic actions for planar robot casting,

    V . Lim, H. Huang, L. Y . Chen,et al., “Real2sim2real: Self-supervised learning of physical single-step dynamic actions for planar robot casting,” inProc. Int’l Conf. on Robotics and Automation, 2022

  3. [3]

    Differential evolution - a simple and efficient heuristic for global optimization over continuous spaces,

    R. Storn and K. Price, “Differential evolution - a simple and efficient heuristic for global optimization over continuous spaces,”Journal of Global Optimization, 1997

  4. [4]

    The cma evolution strategy: A tutorial

    N. Hansen, “The cma evolution strategy: A tutorial.” arXiv e-prints, arXiv:1604.00772, 2016

  5. [5]

    Isaac sim - robotics simulation and synthetic data genera- tion,

    NVIDIA, “Isaac sim - robotics simulation and synthetic data genera- tion,” 2024

  6. [6]

    Survey on model-based manipulation planning of de- formable objects,

    P. Jiménez, “Survey on model-based manipulation planning of de- formable objects,”Robotics and Computer-Integrated Manufacturing, 2012

  7. [7]

    Manipulating highly de- formable materials using a visual feedback dictionary,

    B. Jia, Z. Hu, J. Pan, and D. Manocha, “Manipulating highly de- formable materials using a visual feedback dictionary,” inProc. Int’l Conf. on Robotics and Automation, 2018

  8. [8]

    Sim-to-real reinforcement learning for deformable object manipulation,

    J. Matas, S. James, and A. J. Davison, “Sim-to-real reinforcement learning for deformable object manipulation,” inProc. Conf. on Robot Learning, 2018

  9. [9]

    3d diffusion policy: Generalizable visuomotor policy learning via simple 3d representations,

    Y . Ze, G. Zhang, K. Zhang,et al., “3d diffusion policy: Generalizable visuomotor policy learning via simple 3d representations,” inProc. Robotics: Science and Systems, 2024

  10. [10]

    Implicit neural-representation learning for elastic deformable-object manipulations,

    M. Song, J. Ha, B. Park, and D. Park, “Implicit neural-representation learning for elastic deformable-object manipulations,” inProc. Robotics: Science and Systems, 2025

  11. [11]

    Softgym: Benchmarking deep reinforcement learning for deformable object manipulation,

    X. Lin, Y . Wang, J. Olkin, and D. Held, “Softgym: Benchmarking deep reinforcement learning for deformable object manipulation,” in Proc. Conf. on Robot Learning, 2021

  12. [12]

    Phystwin: Physics-informed reconstruction and simulation of deformable objects from videos,

    H. Jiang, H.-Y . Hsu, K. Zhang,et al., “Phystwin: Physics-informed reconstruction and simulation of deformable objects from videos,” Proc. Int’l. Conf. on Computer Vision, 2025

  13. [13]

    Identification of deformable linear object dynamics from input-output measurements in 3d space,

    M. Floren, S. Mamedov, J.-P. Noël, and J. Swevers, “Identification of deformable linear object dynamics from input-output measurements in 3d space,”IFAC Symposium on System Identification SYSID, 2024

  14. [14]

    A distributional treatment of real2sim2real for vision-driven deformable linear object manipula- tion,

    G. Kamaras and S. Ramamoorthy, “A distributional treatment of real2sim2real for vision-driven deformable linear object manipula- tion,”arXiv e-prints, 2025

  15. [15]

    Simultaneous tracking and elasticity parameter estimation of deformable objects,

    A. Sengupta, R. Lagneau, A. Krupa, E. Marchand, and M. Mar- chal, “Simultaneous tracking and elasticity parameter estimation of deformable objects,” inProc. Int’l Conf. on Robotics and Automation, 2020

  16. [16]

    Accurate simulation and parameter identification of deformable linear objects using discrete elastic rods in generalized coordinates,

    Q. J. Chen, T. Bretl, and Q.-C. Pham, “Accurate simulation and parameter identification of deformable linear objects using discrete elastic rods in generalized coordinates,” inProc. RSJ Int’l Conf. on Intelligent Robots and Systems, 2025

  17. [17]

    Real- to-sim parameter learning for deformable packages using high-fidelity simulators for robotic manipulation,

    O. M. Manyar, H. Ye, S. Mayya, F. Wang, and S. K. Gupta, “Real- to-sim parameter learning for deformable packages using high-fidelity simulators for robotic manipulation,” inASME International Design Engineering Technical Conferences and Computers and Information in Engineering Conference, 2025

  18. [18]

    A bayesian treatment of real-to-sim for deformable object manipulation,

    R. Antonova, J. Yang, P. Sundaresan,et al., “A bayesian treatment of real-to-sim for deformable object manipulation,”IEEE Robotics and Automation Letters, 2022

  19. [19]

    Diffcloud: Real-to-sim from point clouds with differentiable simulation and rendering of deformable objects,

    P. Sundaresan, R. Antonova, and J. Bohg, “Diffcloud: Real-to-sim from point clouds with differentiable simulation and rendering of deformable objects,” inProc. RSJ Int’l Conf. on Intelligent Robots and Systems, 2022

  20. [20]

    Empm: Embodied mpm for modeling and simulation of deformable objects,

    Y . Chen*, Y . Hu*, L. Sun,et al., “Empm: Embodied mpm for modeling and simulation of deformable objects,”IEEE Robotics and Automation Letters, 2026

  21. [21]

    Real-to-sim deformable object manipulation: Optimizing physics models with residual mappings for robotic surgery,

    X. Liang, F. Liu, Y . Zhang,et al., “Real-to-sim deformable object manipulation: Optimizing physics models with residual mappings for robotic surgery,” inProc. Int’l Conf. on Robotics and Automation, 2024

  22. [22]

    Gendom: Generalizable one- shot deformable object manipulation with parameter-aware policy,

    S. Kuroki, J. Guo, T. Matsushima,et al., “Gendom: Generalizable one- shot deformable object manipulation with parameter-aware policy,” in Proc. Int’l Conf. on Robotics and Automation, 2024

  23. [23]

    Sim-to-real transfer in deep reinforcement learning for robotics: a survey,

    W. Zhao, J. P. Queralta, and T. Westerlund, “Sim-to-real transfer in deep reinforcement learning for robotics: a survey,” inIEEE symposium series on computational intelligence (SSCI), 2020

  24. [24]

    Dartbot: Overhand throwing of deformable objects with tactile sensing and reinforcement learning,

    S. Aslam, K. Kumar, P. Zhou,et al., “Dartbot: Overhand throwing of deformable objects with tactile sensing and reinforcement learning,” IEEE Transactions on Automation Science and Engineering, 2025

  25. [25]

    Sam 3: Segment anything with concepts,

    N. Carion, L. Gustafson, Y .-T. Hu,et al., “Sam 3: Segment anything with concepts,” 2025

  26. [26]

    Improving the differen- tial evolution strategy by coupling it with cma-es,

    E. Warchulski, J. Arabas, and R. Biedrzycki, “Improving the differen- tial evolution strategy by coupling it with cma-es,” inProceedings of the Genetic and Evolutionary Computation Conference Companion, 2022

  27. [27]

    Rsl-rl: A learning library for robotics research,

    C. Schwarke, M. Mittal, N. Rudin, D. Hoeller, and M. Hutter, “Rsl-rl: A learning library for robotics research,”arXiv preprint arXiv:2509.10771, 2025

  28. [28]

    gradsim: Differentiable simulation for system identification and visuomotor control,

    K. M. Jatavallabhula, M. Macklin, F. Golemo,et al., “gradsim: Differentiable simulation for system identification and visuomotor control,” inProc. Int’l. Conf. on Learning Representations, 2021

  29. [29]

    Differentiable physics-based system identification for robotic manipulation of elastoplastic materials,

    X. Yang, Z. Ji, and Y .-K. Lai, “Differentiable physics-based system identification for robotic manipulation of elastoplastic materials,”Int’l J. of Robotics Research, 2025

  30. [30]

    Sampling- based system identification with active exploration for legged sim2real learning,

    N. Sobanbabu, G. He, T. He, Y . Yang, and G. Shi, “Sampling- based system identification with active exploration for legged sim2real learning,” inProc. Conf. on Robot Learning, 2025

  31. [31]

    Provably-safe, online system identification,

    B. Zhang, Z. Zhou, and R. Vasudevan, “Provably-safe, online system identification,”arXiv preprint arXiv:2504.21486, 2025

  32. [32]

    Tunenet: One- shot residual tuning for system identification and sim-to-real robot task transfer,

    A. Allevato, E. S. Short, M. Pryor, and A. Thomaz, “Tunenet: One- shot residual tuning for system identification and sim-to-real robot task transfer,” inProc. Conf. on Robot Learning, 2020

  33. [33]

    Auto-tuned sim-to-real transfer,

    Y . Du, O. Watkins, T. Darrell, P. Abbeel, and D. Pathak, “Auto-tuned sim-to-real transfer,” inProc. Int’l Conf. on Robotics and Automation, 2021

  34. [34]

    Sam 3d: 3dfy anything in images,

    S. D. Team, X. Chen, F.-J. Chu,et al., “Sam 3d: 3dfy anything in images,” 2025

  35. [35]

    Isaac lab: A gpu-accelerated simulation framework for multi-modal robot learning,

    M. Mittal, P. Roth, J. Tigue,et al., “Isaac lab: A gpu-accelerated simulation framework for multi-modal robot learning,”arXiv preprint arXiv:2511.04831, 2025

  36. [36]

    SciPy 1.0: Fundamen- tal Algorithms for Scientific Computing in Python,

    P. Virtanen, R. Gommers, T. E. Oliphant,et al., “SciPy 1.0: Fundamen- tal Algorithms for Scientific Computing in Python,”Nature Methods, 2020

  37. [37]

    CMA-ES/pycma on Github,

    N. Hansen, Y . Akimoto, and P. Baudis, “CMA-ES/pycma on Github,” 2019

  38. [38]

    Sim-to-real transfer for visual reinforcement learning of deformable object manipulation for robot-assisted surgery,

    P. M. Scheikl, E. Tagliabue, B. Gyenes,et al., “Sim-to-real transfer for visual reinforcement learning of deformable object manipulation for robot-assisted surgery,”IEEE Robotics and Automation Letters, 2022

  39. [39]

    Evaluation of the young’s modulus of rubber-like materials bonded to rigid surfaces with respect to poisson’s ratio,

    D. Koblar, J. Škofic, and M. Boltežar, “Evaluation of the young’s modulus of rubber-like materials bonded to rigid surfaces with respect to poisson’s ratio,”Strojniski Vestnik, 2014

This paper was first reviewed by deepseek-v4-flash on July 31, 2026.