Pith. sign in

REVIEW 4 major objections 5 minor 20 references

BoxTwin establishes that the full dynamics of elastoplastic articulated objects — nonlinear elasticity, plastic yielding, and damage accumulation — can be learned from videos and used to predict post-contact deformation over long horizons.

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 · deepseek-v4-flash

2026-08-01 18:54 UTC pith:BGSQ44TF

load-bearing objection A clean forward model for elastoplastic hinges in a paper that overclaims 'learning from videos' because the identification algorithm is missing. the 4 major comments →

arxiv 2607.17132 v1 pith:BGSQ44TF submitted 2026-07-19 cs.RO

BoxTwin: Learning Elastoplastic Articulated Object Dynamics from Videos

Heng Zhang , Gehan Zheng , Kaifeng Zhang , Jay Song , Shivansh Patel , Sonny Hu , Yunzhu Li , Changxi Zheng
show 1 more author
Peter Yichen Chen
This is my paper
classification cs.RO
keywords elastoplastic articulated objectsdigital twinvideo-driven reconstructionplastic deformationdamage evolutionhinge joint modelrobotic manipulationsystem identification
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.

BoxTwin is a digital-twin pipeline that learns the behavior of elastoplastic articulated objects — objects like cardboard boxes and sheet-metal assemblies that bend, permanently crease, and degrade with repeated use — directly from videos. The paper argues that by modeling each crease as a hinge with a nonlinear elastic torque, a plastic rest-angle that shifts past a yield threshold, and a damage variable that accumulates with plastic slip, all three effects can be identified and simulated in one framework. The reported experiments show that the resulting twin tracks joint-angle trajectories and reproduces post-contact plastic behavior over long horizons for both hand folding and dual-arm manipulation. If this holds, robots could anticipate how such objects will settle after a manipulation and plan around progressive wear without requiring physical trials.

Core claim

The paper's central claim is that the full mechanical life of an elastoplastic articulated object can be captured by a physics-aware constitutive model per crease, learned from video: each crease is an ideal hinge whose elastic torque is a smooth nonlinear function of the angle gap; when the gap exceeds a yield threshold, the rest angle plastically shifts; and a scalar damage variable, grown monotonically by accumulated plastic slip, attenuates both elastic and plastic responses. On top of this, a video-driven reconstruction pipeline supplies the object's geometry and identifies the per-hinge coefficients. In a manual folding test and in dual-arm manipulation replays, the identified BoxTwin

What carries the argument

The load-bearing object is the per-crease constitutive model: a kinematic tree whose edges are ideal hinges with fixed axes, each carrying (i) a nonlinear elastic torque f(θe) as a function of the gap between current and rest angle, (ii) a plastic flow rule that moves the rest angle when the gap exceeds a yield threshold, and (iii) a damage variable d_e ∈ [0,1] that multiplies both elastic torque and plastic flow and evolves monotonically with total accumulated plastic slip. This single scalar-per-hinge parameterization is what makes identification tractable and keeps the dynamics expressible in standard Lagrange form.

Load-bearing premise

The model assumes each crease is an ideal hinge with a fixed, known rotation axis and that all bending across the crease collapses into one scalar angle, so the representation fails for creases that bend over a wide zone or whose effective hinge migrates as damage accumulates.

What would settle it

Fold the same crease repeatedly past its yield point while tracking markers on both panels; if the inferred rotation axis drifts or out-of-plane deformation appears, the fixed-axis single-scalar-hinge representation cannot support the claimed prediction accuracy.

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

If this is right

  • A robot equipped with such a twin can predict the final configuration of a folded object, including where it has permanently creased, before the motion is executed.
  • Manipulation sequences for new objects can be developed and debugged in simulation using the identified parameters, reducing the number of physical trials.
  • Because damage accumulates in the model, the twin can flag when a crease is approaching mechanical failure and trigger a change in strategy.
  • The same identified model can be reused for control, making post-contact states part of the loop rather than a post-hoc measurement.
  • Long-horizon tracking of joint trajectories becomes possible despite hysteresis and wear, which is the key requirement for adaptive manipulation.

Where Pith is reading between the lines

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

  • If the vision-based identification is accurate, damage estimates could be extracted from video in settings where no force or torque sensors are available, enabling predictive maintenance of packaging and sheet-metal products.
  • A natural stress test is large-amplitude or highly localized folds: the fixed-axis hinge assumption implies that the model's prediction error should grow as the effective bending region widens; measuring this error would define the model's operational envelope.
  • The same constitutive structure could be ported to continuum elastoplastic plates by replacing a sparse set of hinges with a dense network of plastic hinges, effectively turning the method into a discrete shell model.
  • A testable extension is to use the identified yield threshold and damage variable as a control objective — planning trajectories that keep folds below yield to extend object lifetime.

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. BoxTwin proposes a digital-twin framework for elastoplastic articulated objects (EAOs) that is claimed to learn full dynamics from videos. The forward model in §II represents an EAO as a kinematic tree of rigid links connected by ideal hinge joints, with a nonlinear elastic crease torque f_θe(Δe), plastic evolution of the rest angle via a yield condition and flow rule, and a scalar damage variable d_e that attenuates elastic torque and plastic flow, driven by accumulated plastic slip. The evaluation in §III reports two experiments: a manual single-crease folding test with marker-based joint-angle tracking, and a Trossen Aloha dual-arm manipulation test in which recorded follower-arm trajectories are replayed in MuJoCo using 'experimentally identified parameters.' The paper claims accurately tracked joint trajectories and long-horizon post-contact plastic behavior, and concludes that BoxTwin is the first interactive digital-twin framework capturing full EAO dynamics. However, the manuscript provides no identification algorithm, no quantitative error metrics, and no evidence that the reported agreements are predictive rather than fitted.

Significance. If fully realized, a system that reconstructs EAO geometry and identifies elastoplastic-damage constitutive parameters from raw video would be a valuable contribution to robotic manipulation and digital twins. The constitutive model in §II is internally consistent and plausibly expressive: the yield threshold, plastic rest-angle evolution, and damage attenuation cover the three coupled effects advertised in the introduction. The strengths of the paper are its clear forward-model formulation and the use of a physics engine (MuJoCo) for the replay experiments. Nevertheless, the central claim is 'learning from videos,' and the manuscript contains no description of the inference procedure that maps video frames to the model parameters: no loss function, optimization loop, data-association step, or evaluation protocol. The reported results are qualitative and, in the prediction test, potentially circular because the replayed simulation uses parameters identified from the same recordings. As written, the contribution is a forward constitutive model plus a sketch of an application, not a demonstrated learning or identification method.

major comments (4)
  1. [§II, §III-B, Abstract] The central claim that BoxTwin 'learns the full dynamics of EAOs from videos' is not supported by any described algorithm. §II presents only the forward constitutive model. The parameters to be identified include the kinematic tree G, hinge axes a_e, initial rest angles q_e^0, and coefficient sets θ_e, ψ_e, ξ_e, Δy_e. The paper nowhere specifies how these are estimated from video; there is no optimization objective, no gradient or sampling procedure, no data-association step, and no description of the 'video-driven reconstruction' shown in Fig. 1. §III-B refers to 'experimentally identified parameters' but never states the identification method. Without this, the reader cannot determine whether the reported agreement is attributable to BoxTwin or to an unspecified calibration process. This is a load-bearing omission for the paper's main claim.
  2. [§III-B, Prediction Test] The prediction test is circular as described. Recorded follower-arm trajectories are replayed in MuJoCo with 'experimentally identified parameters,' and agreement with the same physical measurements is cited as validation. If those parameters were identified from the same videos or sensor logs — which the text implies but never rules out — then the comparison measures training-set fit, not predictive generalization. No train/test split, cross-validation, or held-out trial is reported. The claim of 'long-horizon' prediction also lacks a definition of the horizon and a comparison against a baseline or a model with no plastic/damage terms.
  3. [§III-A, Fig. 2] The folding-test evaluation is entirely qualitative. The text states that the model 'accurately tracks the joint-angle trajectories' and 'faithfully reproducing' the behavior, but the paper reports no error metrics such as mean absolute error, root-mean-square error, or correlation between simulated and measured joint angles. Figure 2(b) appears to show only a single trajectory overlay. Without quantitative error measures and a baseline comparison (e.g., a purely elastic model, or a model without damage), the accuracy claim is unsubstantiated.
  4. [§II-A] The model assumes each EAO crease is an ideal hinge joint with a fixed, known rotation axis a_e and a fixed kinematic tree G, and that all deformation across the crease collapses into a single scalar joint angle q_e. Real folded creases bend over a finite zone, and their effective hinge axis may migrate during large plastic folds or after damage. The paper gives no procedure for determining a_e or G from video and no evidence that this representation is valid for the tested objects. This assumption directly limits the claimed 'full dynamics' and 'long-horizon' accuracy, since the model cannot represent geometric changes that violate the fixed-hinge idealization.
minor comments (5)
  1. [Abstract and §II] There are minor language issues, e.g., 'identifies a physics aware constitutive model' should be 'physics-aware'; a comma splice in the pipeline sentence. These do not affect substance.
  2. [§III-A] The manual folding test relies on color markers to derive joint angles, but the paper does not explain how marker positions are converted to joint angles, nor how the 'same actuation' is transferred to simulation. This transfer is essential for the comparison.
  3. [§III] The opening sentence says 'Both experiments were implemented using the MuJoCo simulator.' The first experiment is described as a real manual deformation; the intended meaning is presumably that the simulation was implemented in MuJoCo, but the wording is ambiguous.
  4. [§III-B, Fig. 3] Figure 3 shows qualitative side-by-side video frames, but no numerical comparison of joint angles or object configurations for the multi-joint box. Please provide a quantitative plot similar to Fig. 2(b) for the prediction test.
  5. [References] Some citations are to arXiv preprints, which is acceptable, but please ensure the most relevant recent work on video-based physics identification (e.g., PAC-NeRF, PhysDreamer, PhysTwin) is discussed beyond a listing, especially to clarify the claimed novelty over existing video-to-simulation pipelines.

Circularity Check

0 steps flagged

No demonstrable circularity: forward model has no fitted-prediction identity; missing identification algorithm is a correctness gap, not a circularity.

full rationale

The paper's derivations in §II are a forward dynamical model only: hinge kinematics, a nonlinear elastic torque f_θ(Δ), a plastic flow rule g_ψ(Δ), and a damage-rate law h_ξ(P) are all introduced as free per-hinge constitutive maps. No equation is defined in terms of the quantity that is later claimed to be predicted, and no load-bearing argument reduces to a self-citation; the self-citations in the references are motivational context, not the justification for the constitutive model. The principal weakness is that the 'video-driven reconstruction' and the identification of θ_e, ψ_e, ξ_e, Δ_y^e are never described—there is no loss function, optimizer, or data-association procedure—so the central claim of learning from videos is unsubstantiated and unreproducible. However, an omitted inference algorithm is a correctness and reproducibility problem, not a circularity: the paper's text does not exhibit a specific equation or fitting step in which an identified parameter is renamed as a prediction. The §III-B replay of recorded control signals with 'experimentally identified parameters' would be an in-sample fit check if those parameters were fit to the same trials, but the paper never states that identification was performed on the same data, so the constructional identity required for a circularity finding cannot be established without speculation. The evaluation is therefore weak evidence, but not demonstrably circular.

Axiom & Free-Parameter Ledger

5 free parameters · 7 axioms · 2 invented entities

The central simulation depends on five per-hinge parameter sets and two internal variables; none are specified, measured independently, or validated. The paper therefore contributes a model skeleton whose predictive content is carried entirely by unstated fitting.

free parameters (5)
  • θe (elastic torque coefficients) = not reported
    Per-edge coefficients of f_θe in §II-B; no identification method or fitted values given.
  • ψe (plastic flow coefficients) = not reported
    Per-edge coefficients of g_ψe in §II-C; flow rule is parameterized but never fit or specified.
  • ξe (damage-rate coefficients) = not reported
    Per-edge coefficients of h_ξe in §II-D; damage law is chosen as monotone function of P_e but no form or fit reported.
  • Δy_e (yield threshold) = not reported
    Material-specific plastic yield threshold introduced in §II-C; no measurement protocol.
  • q0_e (nominal rest angle) = not reported
    Initial rest angles per hinge; presumably inferred from video but inference is not described.
axioms (7)
  • domain assumption EAO can be represented as rigid links connected by ideal hinge joints with known axes a_e.
    §II-A states this; if creases have finite bending region or migrating axes, model cannot represent them.
  • ad hoc to paper Elastic torque is a smooth, differentiable function of scalar gap Δe only.
    §II-B; no specific form or empirical justification beyond smoothness.
  • domain assumption Plastic flow activates when |Δe| exceeds Δy_e and evolves via ˙αe = g_ψe(Δe).
    §II-C; standard plasticity framework but specifics not validated.
  • ad hoc to paper Damage evolution depends only on accumulated plastic slip P_e via monotone h_ξe, and attenuates torque/flow multiplicatively as (1−d_e).
    §II-D; no independent damage measure or validation.
  • domain assumption Replaying recorded follower-arm trajectories in MuJoCo with the 'same' parameters reproduces the physical contact sequence.
    §III-B; ignores unmodeled slip, compliance, sensing error, and parameter uncertainty.
  • domain assumption Marker-derived joint angles from a monocular view are accurate enough for validation.
    §III-A; no camera calibration or tracking error analysis.
  • standard math Joint-space Lagrange dynamics M(q)q¨ + C(q,q˙)q˙ + g(q) = τ describe the articulated body.
    §II-A; standard rigid-body mechanics.
invented entities (2)
  • α_e (internal plastic rotation variable) no independent evidence
    purpose: Updates rest angle q̄_e = q0_e + α_e; records permanent set.
    Internal state never observed directly; only inferred via joint angle evolution.
  • d_e (scalar damage variable) no independent evidence
    purpose: Attenuates elastic torque and plastic flow via (1−d_e), modeling stiffness loss and wear.
    No direct measurement of damage; no comparison to force-stiffness or microstructural measurements.

pith-pipeline@v1.3.0-alltime-deepseek · 4734 in / 14046 out tokens · 123623 ms · 2026-08-01T18:54:49.314382+00:00 · methodology

0 comments
read the original abstract

Digital twins enable robots to anticipate and adapt to physical interactions, but existing models struggle with elastoplastic articulated objects (EAOs) that exhibit nonlinear elasticity, plastic yielding, and damage accumulation. We present BoxTwin, an interactive digital twin framework that learns the full dynamics of EAOs from videos. Our pipeline reconstructs the scene, identifies a physics aware constitutive model for each EAO. Experiments on manual folding and dual arm manipulation of EAOs show that BoxTwin accurately tracks joint trajectories and reproduces post contact plastic behavior over long horizons. By integrating video driven reconstruction with elastoplastic damage modeling, BoxTwin advances digital twins toward predictive, adaptive control of deformable articulated objects in unstructured environments.

Figures

Figures reproduced from arXiv: 2607.17132 by Changxi Zheng, Gehan Zheng, Heng Zhang, Jay Song, Kaifeng Zhang, Peter Yichen Chen, Shivansh Patel, Sonny Hu, Yunzhu Li.

Figure 1
Figure 1. Figure 1: Our workflow first reconstructs the full scene from [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Figure (a) depicts the experimental setup. Figure (b) [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Figure (a) shows the frames captured from the [PITH_FULL_IMAGE:figures/full_fig_p003_3.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

20 extracted references · 3 linked inside Pith

  1. [1]

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

    H. Jiang, H.-Y . Hsu, K. Zhang, H.-N. Yu, S. Wang, and Y . Li, “Phys- twin: Physics-informed reconstruction and simulation of deformable objects from videos,”arXiv preprint arXiv:2503.17973, 2025

    Pith/arXiv arXiv 2025

  2. [2]

    Digital twins for collaborative robots: A case study in human-robot interaction,

    A. A. Malik and A. Brem, “Digital twins for collaborative robots: A case study in human-robot interaction,”Robotics and Computer- Integrated Manufacturing, vol. 68, p. 102092, 2021

    2021

  3. [3]

    Robotwin: Dual-arm robot benchmark with generative digital twins,

    Y . Mu, T. Chen, Z. Chen, S. Peng, Z. Lan, Z. Gao, Z. Liang, Q. Yu, Y . Zou, M. Xuet al., “Robotwin: Dual-arm robot benchmark with generative digital twins,” inProceedings of the Computer Vision and Pattern Recognition Conference, 2025, pp. 27 649–27 660

    2025

  4. [4]

    Pie-nerf: Physics-based interactive elastodynamics with nerf,

    Y . Feng, Y . Shang, X. Li, T. Shao, C. Jiang, and Y . Yang, “Pie-nerf: Physics-based interactive elastodynamics with nerf,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recog- nition, 2024, pp. 4450–4461

    2024

  5. [5]

    Vr-gs: A physical dynamics-aware interactive gaussian splatting system in virtual reality,

    Y . Jiang, C. Yu, T. Xie, X. Li, Y . Feng, H. Wang, M. Li, H. Lau, F. Gao, Y . Yanget al., “Vr-gs: A physical dynamics-aware interactive gaussian splatting system in virtual reality,” inACM SIGGRAPH 2024 Conference Papers, 2024, pp. 1–1

    2024

  6. [6]

    Phys- gaussian: Physics-integrated 3d gaussians for generative dynamics,

    T. Xie, Z. Zong, Y . Qiu, X. Li, Y . Feng, Y . Yang, and C. Jiang, “Phys- gaussian: Physics-integrated 3d gaussians for generative dynamics,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024, pp. 4389–4398

    2024

  7. [7]

    Virtual elastic objects,

    H.-y. Chen, E. Tretschk, T. Stuyck, P. Kadlecek, L. Kavan, E. V ouga, and C. Lassner, “Virtual elastic objects,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022, pp. 15 827–15 837

    2022

  8. [8]

    Pac-nerf: Physics augmented continuum neural radiance fields for geometry-agnostic system identification,

    X. Li, Y .-L. Qiao, P. Y . Chen, K. M. Jatavallabhula, M. Lin, C. Jiang, and C. Gan, “Pac-nerf: Physics augmented continuum neural radiance fields for geometry-agnostic system identification,”arXiv preprint arXiv:2303.05512, 2023

    Pith/arXiv arXiv 2023

  9. [9]

    Neuphysics: Editable neural ge- ometry and physics from monocular videos,

    Y .-L. Qiao, A. Gao, and M. Lin, “Neuphysics: Editable neural ge- ometry and physics from monocular videos,”Advances in Neural Information Processing Systems, vol. 35, pp. 12 841–12 854, 2022

    2022

  10. [10]

    Physdreamer: Physics-based interaction with 3d objects via video generation,

    T. Zhang, H.-X. Yu, R. Wu, B. Y . Feng, C. Zheng, N. Snavely, J. Wu, and W. T. Freeman, “Physdreamer: Physics-based interaction with 3d objects via video generation,” inEuropean Conference on Computer Vision. Springer, 2024, pp. 388–406

    2024

  11. [11]

    Reconstruction and simula- tion of elastic objects with spring-mass 3d gaussians,

    L. Zhong, H.-X. Yu, J. Wu, and Y . Li, “Reconstruction and simula- tion of elastic objects with spring-mass 3d gaussians,” inEuropean Conference on Computer Vision. Springer, 2024, pp. 407–423

    2024

  12. [12]

    Time-stepping for three-dimensional rigid body dynamics,

    M. Anitescu, F. A. Potra, and D. E. Stewart, “Time-stepping for three-dimensional rigid body dynamics,”Computer methods in applied mechanics and engineering, vol. 177, no. 3-4, pp. 183–197, 1999

    1999

  13. [13]

    Affine body dynamics: Fast, stable & intersection-free simulation of stiff materials,

    L. Lan, D. M. Kaufman, M. Li, C. Jiang, and Y . Yang, “Affine body dynamics: Fast, stable & intersection-free simulation of stiff materials,”arXiv preprint arXiv:2201.10022, 2022

    Pith/arXiv arXiv 2022

  14. [14]

    Stramigioli,Modeling and IPC control of interactive mechanical systems—A coordinate-free approach

    S. Stramigioli,Modeling and IPC control of interactive mechanical systems—A coordinate-free approach. Springer, 2001

    2001

  15. [15]

    Simplicits: Mesh-free, geometry-agnostic elastic simulation,

    V . Modi, N. Sharp, O. Perel, S. Sueda, and D. I. Levin, “Simplicits: Mesh-free, geometry-agnostic elastic simulation,”ACM Transactions on Graphics (TOG), vol. 43, no. 4, pp. 1–11, 2024

    2024

  16. [16]

    Vid2sim: Realistic and interactive simulation from video for urban navigation,

    Z. Xie, Z. Liu, Z. Peng, W. Wu, and B. Zhou, “Vid2sim: Realistic and interactive simulation from video for urban navigation,” inProceedings of the Computer Vision and Pattern Recognition Conference, 2025, pp. 1581–1591

    2025

  17. [17]

    Incremental potential contact: intersection-and inversion-free, large-deformation dynamics

    M. Li, Z. Ferguson, T. Schneider, T. R. Langlois, D. Zorin, D. Panozzo, C. Jiang, and D. M. Kaufman, “Incremental potential contact: intersection-and inversion-free, large-deformation dynamics.”ACM Trans. Graph., vol. 39, no. 4, p. 49, 2020

    2020

  18. [18]

    A constitutive model for paperboard including wrinkle prediction and post-wrinkle behavior ap- plied to deep drawing,

    E. Linvill, M. Wallmeier, and S. ¨Ostlund, “A constitutive model for paperboard including wrinkle prediction and post-wrinkle behavior ap- plied to deep drawing,”International Journal of Solids and Structures, vol. 117, pp. 143–158, 2017

    2017

  19. [19]

    Creasing and folding,

    D. W. Coffin and M. Nyg ˚ards, “Creasing and folding,” inTransactions of the 16th fundamental research symposium, 2017, pp. 69–136

    2017

  20. [20]

    Mujoco: A physics engine for model-based control,

    E. Todorov, T. Erez, and Y . Tassa, “Mujoco: A physics engine for model-based control,” in2012 IEEE/RSJ International Conference on Intelligent Robots and Systems. IEEE, 2012, pp. 5026–5033

    2012