Pith. sign in

REVIEW 4 major objections 5 minor 24 references

Self-Curriculum Model-based Reinforcement Learning for Shape Control of Deformable Linear Objects

T0 review · 4 major / 5 minor · reviewed 2026-08-02 · deepseek-v4-flash

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

Pith's one-line read A two-stage policy-servo controller brings real cables to under 10 mm shape error, zero-shot.

desk verdict Solid two-stage recipe for DLO shape control, but the zero-shot sim2real headline rests on 30 hand-picked, single-run real trials and one advertised generalization test that never appears. read the letter →

arxiv 2602.21816 v2 pith:4EU4335I submitted 2026-02-25 cs.RO

Zhaowei Liang , Song Wang , Zhao Jin , Shirui Wu , Dan Wu This is my paper
classification cs.RO
keywords deformablelinearobjectsshapecontrolmodel-basedreinforcementlearningself-curriculumgoalgenerationvisualservoingsim-to-realtransferdual-armmanipulationgoal-conditionedRL
verification ladder T0 review T1 audit T2 compute T3 formal

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 show that precise shape control of deformable linear objects under large, opposite-curvature deformations can be solved by splitting the task at a 30 mm error threshold: a reinforcement-learning policy trained only in simulation handles the coarse large-deformation phase, and an online Jacobian visual servo handles fine convergence. The authors argue this decomposition sidesteps RL's sample inefficiency and poor final precision while avoiding the sim-to-real fragility of end-to-end learned control. A self-curriculum goal generator that filters goals by imagined difficulty and spreads them by diversity is what makes the policy learn efficiently. If correct, the method would give a practical recipe for zero-shot sim-to-real DLO manipulation across materials, sizes, and stiffnesses.

What carries the argument

The central mechanism is the two-stage task decomposition with a thresholded switch: RL for large deformation, Jacobian-based visual servoing for small deformation. Supporting it are (i) an ensemble of bi-directional LSTM dynamics models trained on simulator data, whose short rollouts augment the replay buffer and train a soft actor-critic policy, and (ii) a self-curriculum goal generator that uses imagined rollouts through the ensemble to label candidate goals as too easy, too hard, or intermediate, then selects a diverse, uncertainty-weighted subset via farthest-point sampling. The Jacobian is continuously updated during the RL stage so the servo stage starts with a good estimate.

What would settle it

Run the simulation-trained policy on a real DLO with an arbitrary target shape not drawn from a pre-collected achievable set and record whether the shape error ever drops below 30 mm; if it does not, the zero-shot transfer claim fails. Alternatively, measure the sim-to-real dynamics gap directly by comparing predicted and observed keypoint displacements under identical actions—a large gap at the switching boundary would predict failure.

Watch

Extended reading notes

Core claim

The central claim is that a policy trained purely in a MuJoCo simulator, without any real-world data, can drive electric wire, USB cable, and braided cotton into a small-deformation basin (error below 30 mm) from which a model-free Jacobian visual servo converges to under 10 mm RMSE. The paper reports 30/30 successful real-world cases and superior success rates and precision over MPC, pure visual servo, and RL-only baselines. The load-bearing insight is that the RL policy need only be approximately right—coarse directional guidance—while the servo stage, which generalizes naturally, supplies the final precision.

Load-bearing premise

The MuJoCo model of the cable—40 articulated capsules with fixed bending stiffness and damping—is faithful enough that the simulation-only policy brings every real DLO into the 30 mm switching basin, and the real targets are pre-selected to be achievable.

Editorial extensions

If this is right

  • DLO shape control can be made sample-efficient: a single simulator-trained policy handles diverse initial and target shapes including opposite-curvature cases.
  • The 30 mm switching threshold converts a long-horizon RL problem into a coarse reaching problem, making the policy trainable and robust to the sim-to-real gap.
  • The same policy transfers zero-shot to unseen real DLOs of different materials and sizes, with final errors near 2 mm on average in real-world tests.
  • Compared with MPC, pure visual servo, and RL-only, the two-stage method achieves higher success rates and precision, while MPC degrades near straight configurations and visual servo falls into local minima.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Because the policy is only required to reach a coarse threshold, the sim-to-real gap is absorbed by the servo stage; this suggests a design principle—deliberately relax the learned policy's accuracy target to improve transfer robustness—that may extend to other deformable-object tasks.
  • The self-curriculum goal generator, which evaluates goals in imagination using the learned dynamics model, could be lifted to other goal-conditioned manipulation problems where a cheap simulator model exists.
  • The paper's real-world targets are pre-collected and guaranteed achievable; a natural stress test is to run the method on arbitrary user-specified targets, where the 30/30 claim may not hold.
  • The approach assumes primarily elastic DLOs; extending to strongly plastic or extremely soft DLOs would likely require retraining or a different dynamics model.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. This paper addresses 2D shape control of deformable linear objects (DLOs) with a two-stage framework. In the large-deformation stage, a model-based reinforcement learning (MBRL) policy is trained in MuJoCo using an ensemble of Bi-LSTM dynamics models and a self-curriculum goal generator that selects intermediate-difficulty, diverse goals via imagined evaluation and weighted farthest-point sampling. In the small-deformation stage, an online Jacobian-based visual servo refines the shape. The paper reports simulation comparisons against MPC, visual serving, and RL-only baselines, with ablations of the curriculum components, and real-world experiments on three DLOs (electric wire, USB cable, braided cotton) claiming zero-shot sim-to-real transfer with 30/30 success. The central claim is that the two-stage decomposition jointly solves sample efficiency and precision for large-deformation DLO shape control, including opposite-curvature cases.

Significance. If the results hold, the paper makes a useful contribution to DLO manipulation: it shows a principled way to combine coarse RL guidance with fine model-free servoing, and the self-curriculum mechanism is a reasonable extension of existing goal-generation ideas. The paper's strengths include evaluation on a fixed external test set in simulation, ablation of the curriculum components, comparison with three reasonable baselines, and real-world validation on multiple DLO materials. The reported 10x-stiffness generalization test is a good idea, though its reporting needs clarification. However, the real-world evidence is currently too selective to support the unqualified 'strong sim2real transfer' and 'diverse initial and target shapes' claims, because the target shapes are pre-selected achievable ones, no repeated trials are reported, and the stage-transition error is not quantified. These gaps should be addressed before publication.

major comments (4)
  1. [Sec. VI.B / Table II] The real-world target shapes are 'pre-collected and guaranteed to be achievable [6]' (Sec. VI.B). The 30/30 success rate is therefore demonstrated only on a hand-picked set of feasible targets, not on arbitrary user-specified shapes. The abstract and conclusion claim 'strong sim2real transfer' and 'diverse initial and target shapes'; these claims are stronger than the evidence. Please either soften the claims or provide a broader/randomized target set with a justification of representativeness.
  2. [Sec. VI.B / Table II] Each real-world case is run only once; the reported success rates (15/15, 12/15, 6/15, etc.) are single-run estimates with no confidence intervals or repeated trials. With 15 trials, 15/15 success does not statistically distinguish a high success probability from a substantially lower one. Please report repeated trials (at least several per case) or per-case variance, and/or provide a statistical characterization such as confidence intervals or bootstrap estimates.
  3. [Sec. VI.A / Fig. 9 / Table II] The paper never reports the real-world shape error at the moment of stage transition (when e falls below the 30 mm threshold). The load-bearing claim is that the simulation-trained RL policy brings the real DLO into the visual-servo basin. Without the transition-error distribution, one cannot verify this: the visual servo itself may be robust enough to succeed from larger errors, which would weaken the zero-shot transfer claim. Please report the transition errors (e.g., histogram or per-case values) and, ideally, the error reduction attributable to the RL stage.
  4. [Sec. V.C / Table I] The text states 'To assess generalization, we test with the DLO's bending stiffness increased by 10 times' but Table I does not separately report results for the nominal vs. 10x-stiffness conditions. It is unclear whether the table aggregates both conditions or whether the 10x results are omitted. If aggregated, the reader cannot evaluate stiffness generalization; if omitted, the experiment is not presented. Please report these results separately or clarify how they are included.
minor comments (5)
  1. [Sec. V.B / Fig. 5] Learning curves are described as averaged over three random seeds, but no error bars or variance shading are shown. Adding variance bands would help assess the statistical significance of the ablation and baseline comparisons.
  2. [Sec. IV.B / Eq. (3)] The Bi-LSTM dynamics model is described only briefly. Please specify the input sequence ordering (e.g., along the DLO arc length) and how the ensemble models are initialized/trained, to improve reproducibility.
  3. [Sec. IV.C / Eq. (4)] The criterion for 'intermediate difficulty' (ϵ_RL < ebar_min < ϵ_upper) is introduced as an assumption. It is reasonable, but please cite or discuss evidence that this interval maximizes learning efficiency, since it is central to the self-curriculum design.
  4. [Sec. III / Sec. IV.A] The stage-switch threshold ϵ and the RL success threshold ϵ_RL are defined, but there is no mention of hysteresis. If the error oscillates around the threshold, could the system switch repeatedly between stages? Please clarify or add a mechanism.
  5. [Sec. VI.B / Fig. 9] The caption of Fig. 9 lists (a) initial, (b) large-deformation stage, (c) transition point, (d) final shape, but the text refers to 'six cases' without clearly mapping them to panels. Labeling the cases or referencing them explicitly would improve readability.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: central sim-to-real claim rests on fixed external test sets and independent baselines; the self-curriculum only shapes training.

full rationale

The paper's derivation chain is self-contained on the point that matters: the reported success rates and errors are measured on fixed held-out test sets, not on the goals generated by the self-curriculum. In Sec. V.B the authors evaluate on 'a separate fixed test set of 50 random target shapes for each condition,' and Tables I and II report environment-measured outcomes; the curriculum's imagined evaluations (Eq. 4 and Algorithm 1) use the learned ensemble model and current policy only to select training goals, so any feedback loop is confined to training and cannot by construction determine test success. The real-world generalization claim is narrower than the abstract suggests: Sec. VI.B states target shapes are 'pre-collected and guaranteed to be achievable [6],' and each of the 30 cases appears to be run once, with no confidence intervals or transition-error measurements; however, this is an evidence/validity limitation, not circularity. The small-deformation stage uses an external model-free Jacobian visual servo [4,16], and the MBRL component builds on MBPO [22] rather than on the paper's own results. The only author self-citation, Ref. [3], supports a general related-work remark on physics-based planning difficulty and is not load-bearing for the central two-stage claim. No equation or fitted parameter is reused as its own prediction, so there is no demonstrated circularity.

Assumptions & free parameters 7 free parameters · 4 assumptions · 0 invented entities

The paper introduces no new physical entities. Its central claims depend on several hand-set thresholds (epsilon, epsilon_RL, epsilon_upper, lambda, alpha) and on the fidelity of the MuJoCo DLO model to real cables. The self-curriculum difficulty criterion is an ad hoc modeling assumption, but it is ablated in the experiments. The most important unquantified input is the sim-to-real dynamics gap and visual tracking accuracy.

free parameters (7)
  • epsilon (stage-switch threshold) = 30 mm
    Decides when the RL policy hands control to the visual servo. Hand-selected; affects the division of labor and final precision.
  • epsilon_RL (RL success threshold) = 20 mm
    Used to define RL success and the intermediate-difficulty band. Hand-set; controls curriculum difficulty and reward density.
  • epsilon_upper (curriculum upper bound) = 30 mm
    Upper bound of the intermediate-difficulty goal set in Eq. (4). Hand-chosen; determines which goals are considered reachable-but-hard.
  • Evaluation success threshold = 10 mm
    The RMSE threshold used for success in Tables I and II. Not derived from application requirements; smaller thresholds would lower success rates.
  • Visual servo gain lambda = 0.05
    Gain in Eq. (2) for the Jacobian-based visual servo. Hand-tuned.
  • Weighted FPS weight alpha = 0.8
    Balances goal diversity versus model uncertainty in Algorithm 1. Hand-tuned.
  • MuJoCo DLO model parameters = stiffness 5e6, damping 0.2, 40 capsules
    Simulator parameters defining the training environment. The zero-shot transfer claim assumes these are representative of the three real DLOs tested.
assumptions (4)
  • domain assumption DLOs considered undergo primarily elastic deformation; the method is not applicable to extremely low stiffness (Sec. III, Assumption 4).
    The sim-to-real transfer claim depends on real DLOs behaving quasielastically like the MuJoCo capsule chain. Plastic USB cable and braided cotton are treated as approximately elastic without a quantified validity check.
  • domain assumption A vision-based algorithm provides real-time, high-accuracy tracking of the 13 keypoints (Sec. III, Assumption 2).
    The state feedback X and the error e come directly from keypoint positions. No quantitative tracking accuracy is reported, so the effect of tracking noise on the 2 mm claims is unknown.
  • ad hoc to paper Optimal learning occurs when goals are slightly above the current success threshold (Sec. IV.C, Eq. 4).
    The intermediate-difficulty band (epsilon_RL, epsilon_upper) is assumed to give the most informative training signal. The ablations support the value of difficulty filtering, but the specific band is not derived.
  • domain assumption Short branched rollouts with the learned elite ensemble models provide a faithful distribution of synthetic states (Sec. IV.B).
    Policy training mixes 99% synthetic data with 1% real transitions. If the model rollouts drift from true dynamics, the policy may not reflect the real DLO; elite selection and short rollouts mitigate but do not eliminate this risk.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Self-Curriculum Model-based Reinforcement Learning for Shape Control of Deformable Linear Objects." pith.science (2026). https://pith.science/paper/4EU4335I

@misc{pith2026260221816,
  author       = {Pith},
  title        = {Pith review of: Self-Curriculum Model-based Reinforcement Learning for Shape Control of Deformable Linear Objects},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/4EU4335I}},
  note         = {Machine review of arXiv:2602.21816}
}
read the original abstract

Precise shape control of Deformable Linear Objects (DLOs) is crucial in robotic applications such as industrial and medical fields. However, existing methods face challenges in handling complex large deformation tasks, especially those involving opposite curvatures, and lack efficiency and precision. To address this, we propose a two-stage framework combining Reinforcement Learning (RL) and online visual servoing. In the large-deformation stage, a model-based reinforcement learning approach using an ensemble of dynamics models is introduced to significantly improve sample efficiency. Additionally, we design a self-curriculum goal generation mechanism that dynamically selects intermediate-difficulty goals with high diversity through imagined evaluations, thereby optimizing the policy learning process. In the small-deformation stage, a Jacobian-based visual servo controller is deployed to ensure high-precision convergence. Simulation results show that the proposed method enables efficient policy learning and significantly outperforms mainstream baselines in shape control success rate and precision. Furthermore, the framework effectively transfers the policy trained in simulation to real-world tasks with zero-shot adaptation. It successfully completes all 30 cases with diverse initial and target shapes across DLOs of different sizes and materials. The project website is available at: https://anonymous.4open.science/w/sc-mbrl-dlo-EB48/

Figures

Figures reproduced from arXiv: 2602.21816 by the authors.

Figure 1
Figure 1. Illustration of the task with large deformations [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 3
Figure 3. The architecture of the MBRL framework with [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figure 4
Figure 4. The proposed self-curriculum goal generation method. [PITH_FULL_IMAGE:figures/full_fig_p004_4.png] view at source ↗
Figures from the paper (5 more)
Figure 5
Figure 5. Figure 5: Policy learning curves under two initial conditions: (a) [PITH_FULL_IMAGE:figures/full_fig_p005_5.png]
Figure 6
Figure 6. Figure 6: Self-curriculum goal evolution under two initial [PITH_FULL_IMAGE:figures/full_fig_p005_6.png]
Figure 7
Figure 7. Figure 7: Representative shape control cases in simulation. (a) Six test cases showing the initial (transparent) and final (opaque) [PITH_FULL_IMAGE:figures/full_fig_p006_7.png]
Figure 8
Figure 8. Figure 8: Real-world experiment setup and DLOs used in the [PITH_FULL_IMAGE:figures/full_fig_p007_8.png]
Figure 9
Figure 9. Figure 9: Real-world shape control process of six cases. (a) Initial shape. (b) Shape at the large deformation stage. (c) Transition [PITH_FULL_IMAGE:figures/full_fig_p008_9.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

24 extracted references · 3 linked inside Pith

  1. [6]

    Global model learning for large deformation control of elastic deformable linear objects: An efficient and adaptive approach,

    M. Yu, K. Lv,et al., “Global model learning for large deformation control of elastic deformable linear objects: An efficient and adaptive approach,”IEEE Trans. Robot., vol. 39, no. 1, pp. 417–436, 2022

    2022

  2. [1]

    Learning graph dynamics with interaction effects propagation for deformable linear objects shape control,

    F. Gu, H. Sang,et al., “Learning graph dynamics with interaction effects propagation for deformable linear objects shape control,”IEEE Trans. Autom. Sci. Eng., 2025

    2025

  3. [2]

    Self-supervised learning of reconstructing deformable linear objects under single-frame occluded view,

    S. Wang, G. Shen,et al., “Self-supervised learning of reconstructing deformable linear objects under single-frame occluded view,” inProc. IEEE Int. Conf. Robot. Autom. (ICRA). IEEE, 2025, pp. 16 028– 16 034

    2025

  4. [3]

    An optimization-based motion planner for dual-arm manipulation of the soft deformable linear objects with nonnegligible gravity,

    S. Wu, J. Zhang, and D. Wu, “An optimization-based motion planner for dual-arm manipulation of the soft deformable linear objects with nonnegligible gravity,”Adv. Eng. Inform., vol. 62, p. 102874, 2024. (a) Case1 USB cable Case3 Braided cotton Case2 Electric wire Case4 USB cable Case6 Braided cotton Case5 Electric wire (b) (c) (d) (e) Overstreched Overstr...

    2024

  5. [4]

    Dual-arm robotic manipulation of flexible cables,

    J. Zhu, B. Navarro,et al., “Dual-arm robotic manipulation of flexible cables,” inProc. IEEE/RSJ Int. Conf. Intell. Robot. Syst. (IROS). IEEE, 2018, pp. 479–484

    2018

  6. [5]

    Automatic shape control of deformable wires based on model-free visual servoing,

    R. Lagneau, A. Krupa, and M. Marchal, “Automatic shape control of deformable wires based on model-free visual servoing,”IEEE Robot. Autom. Lett., vol. 5, no. 4, pp. 5252–5259, 2020

    2020

  7. [7]

    Robotic co- manipulation of deformable linear objects for large deformation tasks,

    K. Almaghout, A. Cherubini, and A. Klimchik, “Robotic co- manipulation of deformable linear objects for large deformation tasks,” Robot. Auton. Syst., vol. 175, p. 104652, 2024

    2024

  8. [8]

    Multi actor-critic ddpg for robot action space decomposition: A framework to control large 3d defor- mation of soft linear objects,

    M. Daniel, A. Magassouba,et al., “Multi actor-critic ddpg for robot action space decomposition: A framework to control large 3d defor- mation of soft linear objects,”IEEE Robot. Autom. Lett., vol. 9, no. 2, pp. 1318–1325, 2023

    2023

Show all 24 references
  1. [9]

    Offline goal-conditioned reinforcement learning for shape control of deformable linear objects,

    R. Laezza, M. Shetab-Bushehri,et al., “Offline goal-conditioned reinforcement learning for shape control of deformable linear objects,” arXiv preprint arXiv:2403.10290, 2024

    2024 arXiv

  2. [10]

    Modeling, learning, perception, and control methods for deformable object manipulation,

    H. Yin, A. Varava, and D. Kragic, “Modeling, learning, perception, and control methods for deformable object manipulation,”Sci. Robot., vol. 6, no. 54, p. eabd8803, 2021

    2021

  3. [11]

    Goal-conditioned reinforcement learning: Problems and solutions,

    M. Liu, M. Zhu, and W. Zhang, “Goal-conditioned reinforcement learning: Problems and solutions,”arXiv preprint arXiv:2201.08299, 2022

    2022 arXiv

  4. [12]

    Learning shape control of elasto- plastic deformable linear objects,

    R. Laezza and Y . Karayiannidis, “Learning shape control of elasto- plastic deformable linear objects,” inProc. IEEE Int. Conf. Robot. Autom. (ICRA). IEEE, 2021, pp. 4438–4444

    2021

  5. [13]

    Robotic control of the defor- mation of soft linear objects using deep reinforcement learning,

    M. H. D. Zakaria, M. Aranda,et al., “Robotic control of the defor- mation of soft linear objects using deep reinforcement learning,” in Proc. IEEE Int. Conf. Autom. Sci. Eng. (CASE). IEEE, 2022, pp. 1516–1522

    2022

  6. [14]

    Learning graph dynamics with external contact for deformable linear objects shape control,

    Y . Huang, C. Xia,et al., “Learning graph dynamics with external contact for deformable linear objects shape control,”IEEE Robot. Autom. Lett., vol. 8, no. 6, pp. 3892–3899, 2023

    2023

  7. [15]

    Robocraft: Learning to see, simulate, and shape elasto-plastic objects in 3d with graph networks,

    H. Shi, H. Xu,et al., “Robocraft: Learning to see, simulate, and shape elasto-plastic objects in 3d with graph networks,”Int. J. Robot. Res., vol. 43, no. 4, pp. 533–549, 2024

    2024

  8. [16]

    Robust deformation model approximation for robotic cable manipulation,

    S. Jin, C. Wang, and M. Tomizuka, “Robust deformation model approximation for robotic cable manipulation,” inProc. IEEE/RSJ Int. Conf. Intell. Robot. Syst. (IROS). IEEE, 2019, pp. 6586–6593

    2019

  9. [17]

    Online model learning for shape control of deformable linear objects,

    Y . Yang, J. A. Stork, and T. Stoyanov, “Online model learning for shape control of deformable linear objects,” inProc. IEEE/RSJ Int. Conf. Intell. Robot. Syst. (IROS). IEEE, 2022, pp. 4056–4062

    2022

  10. [18]

    Goal-conditioned on-policy reinforcement learning,

    X. Gong, D. Feng,et al., “Goal-conditioned on-policy reinforcement learning,”Adv. Neural Inf. Process. Syst. (NeurIPS), vol. 37, pp. 45 975–46 001, 2024

    2024

  11. [19]

    Stein variational goal gener- ation for adaptive exploration in multi-goal reinforcement learning,

    N. Castanet, O. Sigaud, and S. Lamprier, “Stein variational goal gener- ation for adaptive exploration in multi-goal reinforcement learning,” in Proc. Int. Conf. Mach. Learn. (ICML). PMLR, 2023, pp. 3714–3731

    2023

  12. [20]

    Variational curriculum reinforcement learning for unsupervised discovery of skills,

    S. Kim, K. Lee, and J. Choi, “Variational curriculum reinforcement learning for unsupervised discovery of skills,” inProc. Int. Conf. Mach. Learn. (ICML). PMLR, 2023, pp. 16 668–16 695

    2023

  13. [21]

    Automatic goal generation for rein- forcement learning agents,

    C. Florensa, D. Held,et al., “Automatic goal generation for rein- forcement learning agents,” inProc. Int. Conf. Mach. Learn. (ICML). PMLR, 2018, pp. 1515–1528

    2018

  14. [22]

    When to trust your model: Model-based policy optimization,

    M. Janner, J. Fu,et al., “When to trust your model: Model-based policy optimization,”Adv. Neural Inf. Process. Syst. (NeurIPS), vol. 32, 2019

    2019

  15. [23]

    Self-supervised learning of state estimation for manipulating deformable linear objects,

    M. Yan, Y . Zhu,et al., “Self-supervised learning of state estimation for manipulating deformable linear objects,”IEEE Robot. Autom. Lett., vol. 5, no. 2, pp. 2372–2379, 2020

    2020

  16. [24]

    Learning particle dynamics for manipulat- ing rigid bodies, deformable objects, and fluids,

    Y . Li, J. Wu,et al., “Learning particle dynamics for manipulat- ing rigid bodies, deformable objects, and fluids,”arXiv preprint arXiv:1810.01566, 2018

    2018 arXiv

Pith tools

Reviewed August 2, 2026 · model on record in the stance chip above.