{"id":"122d8161-586d-4b2d-9e3f-55160fee1526","arxiv_id":"2501.17963","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A differentiable simulator for soft growing robots with a new wrinkling-based bending stiffness model, fitted and validated against real robot trajectories.","lead":"This paper builds a fast, differentiable simulator for vine-like soft growing robots, and adds a new bending model based on wrinkling of inflated tubes. The simulator can be tuned automatically from video of real robots, and the bending model matches measured shapes better than previous linear models.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The stiffness model is never tested as a physics prediction: εcrit is refit for the robot trials, and Eq. (2) uses a global cantilever angle in calibration but a joint angle in simulation, so Fig. 6 may only show curve fitting.","rationale":"Agreeing with the reader that Eq. (2) is a strong simplification, I think the sharper version of the problem is that the model is never confronted with a true held-out test of its physical content. The Fig. 3(a) moment curves are best fits to the same data; Fig. 6 uses a refit εcrit. Consequently, the comparison against linear/MLP is a comparison of function classes after per-model parameter fitting, and the reported advantage may be due to the saturating shape of Eq. (1) rather than to the geometric derivation of Eq. (2). The open-source implementation is a real strength, as is the use of differentiable optimization, but it does not by itself validate the physical bridge. A no-refit prediction test using the independently measured εcrit would settle whether the model transfers from the cantilever calibration to the multi-link simulator. I therefore keep the reader's CONDITIONAL verdict but add the transfer and discretization conditions.","tokens_in":14344,"tokens_out":13768,"duration_ms":154647,"concrete_test":"Use the εcrit measured in Sec. III-A (Fig. 3(b)) directly in the simulator, with no refitting to the robot-video training set, and evaluate MSE on the Fig. 6 test environment; also repeat with dsegment halved. If the non-refit model is not as accurate as the refit model, or if results change markedly with dsegment, then the central claim is supported only as parameter fitting, not as a validated physics-grounded stiffness model.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central validation (Fig. 6, Fig. 7) compares three stiffness models after their parameters are optimized on robot-video training data. For the proposed model the key parameter εcrit is therefore refit on the target task, not taken from the cantilever measurement in Sec. III-A; the measurement only checks Eq. (1)-(2) against the same data used to fit εcrit (Fig. 3(a)), which is interpolation, not prediction. This matters because the kinematic bridge Eq. (2) assumes the wrinkle angle γ0 is set by the overlap of two rigid circular cross-sections, with no length scale or loading configuration. In the calibration setup (Fig. 3(c)) the measured θ is a global deflection angle of a cantilever under a point load, where local curvature varies along the beam; in the simulator (Sec. III-B) θ is the relative pin-joint angle of adjacent rigid links. Nothing in Eq. (2) shows these two situations produce the same γ0(θ) relation, and nothing normalizes by the segment length dsegment that defines the discretization. If Eq. (2) is applied at different dsegment values, the same physical curvature gives different joint angles and hence different predicted moments, so the model is not a material law. The improved MSE in Fig. 6 may therefore reflect a flexible one-parameter curve fit to the training environment rather than validated first-principles physics.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a differentiable simulator for vine robots, building on the prior Jitosho et al. impulse-velocity formulation and extending it with variable-length links, contact complementarity, and a differentiable quadratic programming layer. Its main contribution is a closed-form nonlinear stiffness model that connects the Comer-Levy wrinkling moment (Eq. 1) to a bending angle via a wrinkling criterion and a geometric bridge (Eq. 2). The authors experimentally fit the wrinkling criterion to cantilever moment data for an inflated LDPE tube, fit a polynomial for pressure dependence, and integrate the model into the simulator. They then optimize simulator parameters (mass, inertia, growth rate, damping, and stiffness) against real-robot video data and compare the proposed stiffness model against linear and MLP baselines on a held-out environment.","tokens_in":14602,"tokens_out":4288,"duration_ms":44855,"significance":"If the stiffness model is genuinely predictive, the work would provide a valuable physics-based component for sim-to-real transfer, planning, and control of vine robots, and the open-source differentiable simulator with batched rollouts is a useful engineering contribution. The paper also demonstrates a practical pipeline for fitting simulator parameters through differentiable optimization. However, the validation as presented does not yet establish the main claim that the proposed model 'captures non-constant curvature bending' better than baselines in a physics-predictive sense, because the stiffness parameter is refit on the target task and the moment-curve comparison is in-sample.","major_comments":[{"comment":"The kinematic bridge γ0 = cos⁻¹(2 ε_crit / sin(θ/2) − 1) is derived from the overlap of two rigid circular cross-sections and contains no length scale such as the segment length dsegment. In the calibration setup (Fig. 3c), θ is a global cantilever deflection angle under a point load, where local curvature varies along the beam; in the simulator (Sec. III-B), θ is the relative pin-joint angle between adjacent virtual links separated by dsegment. These are different quantities, so the same physical curvature produces different θ at different discretizations. The authors should either express Eq. (2) in terms of local curvature and dsegment, or demonstrate that the fitted ε_crit is invariant to dsegment. Without this, the improved MSE in Fig. 6 could be explained by a one-parameter curve fit rather than by first-principles physics.","section":"§III-A, Eq. (2)"},{"comment":"The 'model prediction' in Fig. 3(a) is obtained by least-squares fitting ε_crit to the same measured moment-angle data (as stated in the text: 'ɛcrit is found via a least-square fitting to the data'). The agreement shown is therefore an in-sample fit, not an independent prediction. In addition, no error bars or repeated-trial statistics are reported for the moment measurements, so the reader cannot assess the scatter or the statistical significance of the mismatch at small angles. The authors should report a hold-out validation (e.g., fit on a subset of pressures or bending angles and predict the rest) and include measurement uncertainty.","section":"§III-A, Fig. 3(a)"},{"comment":"The trajectory comparison is not a test of the independently measured stiffness model. The text explicitly states that the stiffness parameter is derived again on a different vine robot using fitting: 'we derive this parameter again on a different vine robot using fitting to demonstrate the capabilities of our model.' Thus, for the proposed model, ε_crit is optimized on the same robot-video training data that is used for all other parameters, and the held-out environment is only unseen after this optimization. This makes Fig. 6 a comparison of curve-fitting capacity on the target task, not a validation that the physics-based model predicts without task-specific tuning. The authors should either fix ε_crit from the Sec. III-A cantilever measurement in the trajectory evaluation, or compare all models with matched parameter counts and report the fitted ε_crit value. The dataset of 10 demonstrations in 4 environments is also too small to support the strong claim about consistency in Fig. 7 without confidence intervals or per-environment breakdowns.","section":"§IV-C, Figs. 6 and 7"}],"minor_comments":[{"comment":"The notation for the wrinkling criterion is inconsistent: the text uses 'ε_crit', 'ϵcrit', and Fig. 3(b) uses 'ε'. Please standardize the symbol throughout.","section":"§III-A, after Eq. (2)"},{"comment":"The third-degree polynomial fit for ε_crit(P) is described as 'heuristic' but the polynomial coefficients or an equation are not given in the text; since the code is open source this may be recoverable, but stating the polynomial would improve reproducibility and make the model self-contained.","section":"§III-A, Fig. 3(b)"},{"comment":"The notation K(q, ·) is unclear; it would help to define explicitly that K is a stiffness function returning the joint torque as a function of the configuration, and to state the domain of C and the meaning of the damping matrix.","section":"§III-B, Eq. (3)"},{"comment":"The video-processing pipeline is described briefly; please provide the key parameters (e.g., blur kernel size, morphological closure size, skeletonization method) or a reference to a repository file so that the data extraction is reproducible.","section":"§IV-B"},{"comment":"The loss function minimizes positional squared error, but it is not stated whether the rotational degrees of freedom are included in the error or whether the ground-truth link spacing matches the simulator's fixed dsegment. This should be clarified.","section":"§IV-C, optimization"}],"recommendation":"major_revision","confidential_remarks":"The core engineering contribution—a batched differentiable vine-robot simulator with an open-source implementation—is solid and likely useful to the community. The main risk is overclaiming the physics grounding of the stiffness model: the moment-curve comparison is in-sample, and the trajectory comparison refits the key parameter on the target task. I would recommend major revision rather than rejection because the issues are addressable: report an independent validation of the stiffness model (e.g., hold-out pressures or bending angles, error bars), test discretization invariance of Eq. (2), and either fix ε_crit from the independent measurement or frame the trajectory results as a model-class comparison rather than a physics-prediction validation."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's my read. This is a solid engineering contribution: a batched, differentiable simulator for vine robots built on Jitosho et al., with a new closed-form stiffness model based on Comer-Levy wrinkling, released open source. The new piece is Eq. (2), a geometric bridge between the wrinkling angle γ0 and the joint bending angle θ, which gives a nonlinear moment–angle curve without needing a constant curvature assumption. The simulator's batching and variable link lengths are useful, and the performance scaling plots look credible.\n\nThe comparison against linear and MLP stiffness models on real-robot video is a reasonable way to show that the nonlinear form is expressive. But the validation is not what the abstract implies. εcrit is fitted to the same cantilever moment curves shown in Fig. 3(a), so that figure is an in-sample fit, not a prediction. Then, in Sec. IV-C, εcrit is re-fit on the robot-video training data for the simulator. So Fig. 6 is a curve-fitting comparison, not a test of the physics. The paper is transparent about this, which I appreciate, but it undercuts the 'physics-grounded' framing.\n\nThe bigger conceptual issue is that Eq. (2) contains no length scale. The same physical curvature will give different joint angles at different discretization lengths, so the model is not a material law. The calibration measures a global cantilever deflection angle, while the simulator uses relative pin-joint angles; the paper never shows these are equivalent. That needs to be addressed. Lesser issues: no error bars or repeated trials on the moment data, and the robot dataset is small (10 demonstrations, 4 environments) and not released.\n\nStill, the open-source code and the new closed-form relation make this worth engaging. I'd send it to review, and I'd ask the authors to clarify the discretization issue and to run a true predictive test—fix εcrit from the cantilever data, use it in simulation, and see if the MSE improvement holds. As is, it's a useful tool with a promising model, but the evidence for the physics is slim.","headline":"A useful batched differentiable simulator for vine robots with a new wrinkling-based stiffness model, but the central validation is curve fitting rather than physics prediction.","tokens_in":15179,"tokens_out":4225,"would_cite":true,"duration_ms":41564,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"New stiffness model makes vine robot simulation match reality","keywords":["soft growing robots","vine robots","differentiable simulation","inflated beam wrinkling","nonlinear stiffness model","sim-to-real transfer","gradient-based parameter fitting","closed-form bending model"],"falsifier":"Measure the moment-angle curve for the same tube material at two different radii and pressures, fit $\\epsilon_{\\mathrm{crit}}$ to one curve, and check whether the same value (or the same non-dimensional scaling) predicts the other curve; if the predicted wrinkling-onset angle $2\\sin^{-1}(\\epsilon_{\\mathrm{crit}})$ and the saturation level do not track the measured curves, the kinematic bridge is not capturing the physics. A direct simulator-level test would compare predicted shapes to skeletonized video in an environment with a single hard obstacle across several approach angles and pressures, looking specifically at the curvature just behind the contact point.","tokens_in":14084,"feed_emoji":"🤖","tokens_out":6009,"duration_ms":56177,"temperature":0.7,"pith_summary":"This paper tries to establish that a closed-form, first-principles stiffness model for thin-walled inflated tubes, built from the physics of surface wrinkling, can be embedded in a differentiable simulator for soft growing (vine) robots and that the resulting simulator predicts real robot shapes more accurately than the linear or learned stiffness models used in prior work. Vine robots are long pressurized tubes that extend by eversion, and their bending is governed by localized loss of tension on the compressed side of the tube, not by constant or linearly elastic stiffness. The authors derive a bending-angle-dependent moment law from a wrinkling criterion, fit its single parameter to one tube's measured moment data, and then use the simulator's gradients to fit the remaining dynamics parameters to real video demonstrations. On an environment not used for fitting, the proposed model produces lower and more consistent mean-squared error than the alternatives. If correct, this gives planners and control designers a fast, gradient-bearing simulator whose shapes match reality.","feed_headline":"Wrinkling model makes vine robot simulation match reality","feed_subtitle":"A differentiable simulator with the closed-form nonlinear bending model beats linear and learned stiffness on unseen test paths.","key_machinery":"The load-bearing object is the kinematic bridge between wrinkling angle and bending angle, Eq. (2), together with the experimentally fitted wrinkling criterion $\\epsilon_{\\mathrm{crit}}$. The bridge treats the onset of wrinkling as the point where the overlap of two rigid circular cross-sections implies a length change $\\Delta l$ exceeding a threshold $\\Delta l_{\\mathrm{crit}}$ at some height on the tube surface; geometry then converts that critical height into $\\gamma_0$. This converts the classical moment formula $M = \\pi P R^3 [\\sin 2\\gamma_0 + 2\\pi - 2\\gamma_0] / [4(\\sin\\gamma_0 + (\\pi-\\gamma_0)\\cos\\gamma_0)]$ into a function of bending angle only, with $\\epsilon_{\\mathrm{crit}}$ fit to experimental moment data and interpolated by a polynomial for use in simulation. The differentiable simulator wraps this stiffness law in a maximal-coordinate rigid-body formulation where feasible velocities are obtained from a differentiable quadratic program, so gradients flow from a shape-matching loss back to physical parameters.","core_discovery":"The central claim is that vine robot bending can be captured by a nonlinear moment law obtained by inserting a geometric wrinkling-onset condition into the classical wrinkled-beam moment formula. The formula relates the restoring moment $M$ to the wrinkling angle $\\gamma_0$, and the paper's bridge equation $\\gamma_0 = \\cos^{-1}(2\\epsilon_{\\mathrm{crit}}/\\sin(\\theta/2)-1)$ connects $\\gamma_0$ to the joint bending angle $\\theta$ through a single dimensionless wrinkling criterion $\\epsilon_{\\mathrm{crit}} = \\Delta l_{\\mathrm{crit}}/2R$. Substituting the bridge into the moment formula yields a closed-form $M(\\theta)$ that rises steeply at small angles and saturates near the fully wrinkled moment, matching the measured behavior of inflated LDPE tubes. Integrated into a differentiable rigid-body simulator with variable-length growth and contact complementarity, this stiffness model reproduces non-constant curvature bending when a growing vine contacts obstacles, and parameter fitting by gradient descent on real video data produces lower and more consistent mean-squared error on a held-out environment than both a linear stiffness model and a learned multilayer-perceptron stiffness model.","pith_inferences":["If $\\epsilon_{\\mathrm{crit}}$ could be predicted from material properties, tube geometry, and pressure rather than fitted per tube, the model would generalize across robot designs without new bending experiments; the paper leaves this derivation open.","The wrinkling-bridge idea could transfer to other inflated structures, such as inflatable booms or soft actuators, wherever bending is dominated by tension loss rather than material elasticity.","The differentiable coupling between observed motion and stiffness parameters suggests an inverse-problem route: infer operating pressure or material state from video of a deployed vine robot.","Replacing the differentiable quadratic-program solver with a learned surrogate or a GPU-accelerated solver, as the paper notes is future work, could remove the current CPU bottleneck and make the simulator useful for reinforcement learning."],"forward_implications":["The differentiable simulator can be placed inside gradient-based optimization loops for planning, control, and parameter identification, since gradients flow back through both the quadratic program and the nonlinear stiffness law.","Batched parallel rollouts make high-throughput simulation practical, so many launch angles or environment variations can be evaluated simultaneously.","The closed-form stiffness model reduces sim-to-real shape error compared with constant-moment and linear-stiffness baselines on unseen environments.","The single wrinkling criterion $\\epsilon_{\\mathrm{crit}}$ can be re-fit by gradient descent on real robot video, avoiding hand-tuned stiffness parameters.","The same framework can be extended to other vine robot actuation modes, such as pre-formed welds, pinches, and tensioning cables."],"supporting_citations":[{"why":"Supplies the classical inflated-beam wrinkling moment formula that the proposed model starts from.","marker":"[26]"},{"why":"Supplies the maximal-coordinate rigid-body dynamics, Lagrange multiplier constraints, and the linear stiffness baseline that this work extends and replaces.","marker":"[25]"},{"why":"Documents the obstacle-interaction bending behavior, non-constant curvature, that the proposed model is validated against.","marker":"[23]"},{"why":"Provides evidence that neither constant nor linear stiffness fully predicts inflated-beam bending, motivating the nonlinear model.","marker":"[27]"},{"why":"Supplies the differentiable convex optimization layers used to pass gradients through the quadratic-program feasibility solve.","marker":"[28]"},{"why":"Supplies the conic solver used to solve the quadratic program efficiently and differentiably.","marker":"[29]"},{"why":"Shows the constant-moment simplification of the wrinkling model that the paper argues overestimates stiffness at small angles.","marker":"[34]"}],"fun_headline_variants":["Nonlinear wrinkling model sharpens vine robot simulations","Differentiable sim nails vine robot bending via wrinkling law","Closed-form wrinkling model boosts vine robot sim fidelity","Physics-grounded sim matches vine robot shapes with wrinkling","Vine robot sim learns real bending from a wrinkling equation"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole model leans on the assumption that wrinkling begins and grows according to a single scalar length-change threshold $\\epsilon_{\\mathrm{crit}}$ fitted to the moment data of one tube, with the shell treated as inextensible and free of shear, so an error in that fitted threshold propagates directly into every simulated bend.","fun_headline_variants_meta":{"raw":{"variants":["Nonlinear wrinkling model sharpens vine robot simulations","Differentiable sim nails vine robot bending via wrinkling law","Closed-form wrinkling model boosts vine robot sim fidelity","Physics-grounded sim matches vine robot shapes with wrinkling","Vine robot sim learns real bending from a wrinkling equation"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000182,"raw_usage":{"total_tokens":1333,"prompt_tokens":991,"completion_tokens":342,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":607,"completion_tokens_details":{"reasoning_tokens":262}},"tokens_in":607,"tokens_out":342,"duration_ms":3758,"temperature":1.0,"reasoning_tokens":262,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T04:27:16.336459+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Measure the moment-angle curve for the same tube material at two different radii and pressures, fit $\\epsilon_{\\mathrm{crit}}$ to one curve, and check whether the same value (or the same non-dimensional scaling) predicts the other curve; if the predicted wrinkling-onset angle $2\\sin^{-1}(\\epsilon_{\\mathrm{crit}})$ and the saturation level do not track the measured curves, the kinematic bridge is not capturing the physics. A direct simulator-level test would compare predicted shapes to skeletonized video in an environment with a single hard obstacle across several approach angles and pressures, looking specifically at the curvature just behind the contact point.","supporting_citations":[{"cited_title":"Deflections of an inflated circular-cylindrical cantilever beam","cited_arxiv_id":null,"evidence_quote":"Supplies the classical inflated-beam wrinkling moment formula that the proposed model starts from."},{"cited_title":"A dy- namics s imulator f or soft growing robots","cited_arxiv_id":null,"evidence_quote":"Supplies the maximal-coordinate rigid-body dynamics, Lagrange multiplier constraints, and the linear stiffness baseline that this work extends and replaces."},{"cited_title":"Robust navigation o f a so ft growing robot by exploiting contact with the environment","cited_arxiv_id":null,"evidence_quote":"Documents the obstacle-interaction bending behavior, non-constant curvature, that the proposed model is validated against."},{"cited_title":"Anisotropic Stiffness and Programmable Actuation for Soft Robots Enabled by an Inflated Rotational Joint","cited_arxiv_id":"2410.13003","evidence_quote":"Provides evidence that neither constant nor linear stiffness fully predicts inflated-beam bending, motivating the nonlinear model."},{"cited_title":"Differentiable con vex optimization layers","cited_arxiv_id":null,"evidence_quote":"Supplies the differentiable convex optimization layers used to pass gradients through the quadratic-program feasibility solve."},{"cited_title":"Operator splitting for a homogeneous embedding of the linear complementarity problem","cited_arxiv_id":null,"evidence_quote":"Supplies the conic solver used to solve the quadratic program efficiently and differentiably."},{"cited_title":"Character- izing environmental i nteractions for soft growing robots","cited_arxiv_id":null,"evidence_quote":"Shows the constant-moment simplification of the wrinkling model that the paper argues overestimates stiffness at small angles."}],"review_version":1}