REVIEW 4 major objections 6 minor 2 cited by
DexFlow: A Unified Approach for Dexterous Hand Pose Retargeting and Interaction
T0 review · 4 major / 6 minor · reviewed 2026-08-16 · deepseek-v4-flash
Pith's one-line read Human hand recordings can be retargeted into physically plausible robot grasps at scale: the DexFlow pipeline reports a 7.5-times higher semantic success rate than the DexRetarget baseline and releases a 292K-frame benchmark.
desk verdict Useful retargeting pipeline and a 292K-frame dataset, but the headline 7.5x SSR improvement is not credible because DexFlow's success rule is maximally lenient and the baselines are imported without a shared evaluation protocol. read the letter →
The pith
A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.
The reading
What carries the argument
The load-bearing mechanism is the contact map and its temporal refinement. After a global search optimizer aligns the robot hand to the human hand in task space, each fingertip's distance to the object surface is classified by two thresholds, and the ambiguous band between thresholds is resolved by inheriting the previous frame's contact state. A cubic-spline trajectory fit over a five-frame window supplies a contact-likelihood score $\sigma(\beta(\ddot{T}-\ddot{T}_{\mathrm{obj}}))$, so a smoothed contact state is accepted only when motion continuity and velocity limits hold; otherwise the raw state is kept. A second-stage energy $E_{\mathrm{total}}=E_{\mathrm{dis}}+w_{\mathrm{pen}}E_{\mathrm{pen}}+w_{\mathrm{align}}E_{\mathrm{align}}+w_{\mathrm{spen}}E_{\mathrm{spen}}+w_{\mathrm{joints}}E_{\mathrm{joints}}$ then refines each finger sequentially, starting with the thumb, so that primary functional fingers are not deformed by self-penetration penalties. The temporal-consistency objective couples consecutive frames through $L_{\mathrm{temp}}=\lambda\sum_t\|q_t-2q_{t-1}+q_{t-2}\|^2_{\Sigma^{-1}}$, which is claimed to yield second-order smooth trajectories.
What would settle it
Take a set of MANO interaction sequences with ground-truth contact signals, such as tactile sensor recordings on a real hand or manually labelled contact frames, and run DexFlow's dual-threshold detector on the same sequences; if the predicted contact maps disagree with the ground-truth contacts on a substantial fraction of frames, the contact refinement and the reported success rates inherit that error rather than correcting it.
Extended reading notes
Core claim
The central claim is that a decoupled, hierarchical optimization converts human hand-object interaction sequences into physically plausible robot-hand grasps more reliably than prior retargeters. The method starts from a global task-space alignment between MANO and robot fingertips, then applies a differential loss with a sliding window to keep trajectories $C^2$-smooth, then detects contact through a dual-threshold rule (a fingertip closer than the lower threshold is in contact, farther than the upper threshold is not, and in between inherits the previous frame's state), then smooths contact states over time with a kinematic-consistency check, and finally optimizes one finger at a time from thumb to pinky under distance, penetration, normal-alignment, self-collision, and joint-regularization energies. On the released 292K-frame benchmark, the paper reports an SSR (semantic success rate, the share of grasp frames that survive simulated gravity) of 40.32%, a 7.5-times improvement over DexRetarget's 5.35%, alongside contact distances near the best generative baselines and a 90% reduction in penetration depth relative to traditional retargeting. The authors also report that the pipeline resolves 68% of contact-state fluctuations seen in conventional retargeting, and that the same human motion can be migrated across hand topologies such as ShadowHand and Allegro while preserving the semantic grasp type.
Load-bearing premise
The pipeline inherits whatever error is already baked into the reconstructed MANO hand poses, object meshes, and contact metadata extracted from source video; the authors note that contact information would be more reliable if taken directly from video rather than reconstructed metadata.
Editorial extensions
If this is right
- Human video or MANO motion data can be converted at scale into ShadowHand and Allegro grasp trajectories, giving robot learning pipelines a data source that does not require teleoperation hardware.
- The released 292K-frame benchmark over 50 YCB objects provides a common testbed for comparing retargeting and grasp-synthesis methods on pose accuracy, contact quality, and trajectory smoothness.
- Cross-hand topology migration means a single recorded human manipulation can be mapped to different robot hands while preserving semantic grasp types such as pinch and wrap.
- Resolving 68% of contact-state fluctuations and cutting penetration depth by 90% relative to conventional retargeting implies the generated sequences are closer to usable robot commands rather than raw human poses.
- An SSR of 40.32% versus 5.35% for DexRetarget indicates that retargeting-based data generation can approach the success levels of optimization-based grasp synthesis.
Reading between the lines
- A natural next step the authors do not take is to use the pipeline's contact maps as supervision to train a direct image-to-contact predictor, which would remove the reconstructed-metadata dependency they flag as a limitation.
- If the sequential thumb-to-pinky finger ordering is a genuine kinematic prior, it should be tested against the reverse ordering or against grasps where the pinky carries most of the load; a failure there would localize where the naturalness benefit actually comes from.
- The 292K-frame dataset invites a scaling test: fine-tune a vision-based manipulation policy on DexFlow output and measure real-robot grasp transfer, which the paper's simulation-only evaluation does not yet establish.
- Because the reported gains come from contact refinement guided by reconstructed geometry, the method's ceiling is probably set by current hand-object reconstruction quality; progress in video-based contact estimation should transfer directly into higher SSR for the same pipeline.
Signed reviews
Editorial analysis
A structured set of objections, weighed in public.
Referee Report
Summary. The paper proposes DexFlow, a three-stage pipeline for retargeting MANO human hand-object motion to robotic hands such as ShadowHand and Allegro. The stages are: global task-space optimization with temporal and differential losses; dual-threshold contact detection with frame-to-frame smoothing; and finger-wise contact refinement. The authors contribute a 292K-frame grasp dataset and evaluate their pipeline with Isaac Gym, reporting a semantic success rate (SSR) of 40.32% versus 5.35% for DexRetarget, along with other quality metrics. The central claims are improved pose accuracy, naturalness, and diversity over prior retargeting and grasp-synthesis methods, summarized as a 7.5-times improvement in semantic success.
Significance. If the reported results hold, DexFlow would be a practically useful data-generation tool for dexterous manipulation, converting human motion sequences into physically plausible robot-hand interaction data at scale, and the released 292K-frame dataset would be a resource for the community. The paper's strengths include a clear modular pipeline, cross-hand topology support, an explicit physics-simulation evaluation, and a public project page with code and media. The significance is clouded, however, because the headline comparison in Table II rests on an evaluation protocol whose leniency is not shown to be shared by the baselines, and because several quantities that determine the optimized outputs are not reported.
major comments (4)
- [§IV.B.1 and Table II] The headline 7.5x SSR improvement is not an apples-to-apples comparison. Section IV.B.1 defines DexFlow's success as contact maintained after 100 simulation steps 'regardless of the gravity being applied in any of the six axis-aligned directions' and 'if any frame in the sequence after contacting the object satisfies the condition, the grasp is considered successful.' The rows for DexRetarget, DexGraspNet, SpringGrasp, FRoGGeR, and BODex are imported from the BODex paper, with no evidence that the 5.35% figure for DexRetarget was computed under the same any-frame/any-gravity rule. Because the central claim in the abstract and Section I is the 7.5x improvement, the authors must re-run all baselines under an identical protocol (or provide shared evaluation code) and report both per-frame and per-sequence success rates, and both single-gravity and any-gravity results. As it stands, the comparison may reflect evaluation leniency rather than grasp quality.
- [§III.B, Eq. (2)] Equation (2) introduces a 'kinematic covariance matrix' Σ ∈ R^{28×28} but never defines how Σ is estimated or what norm is being used; without this, the differential loss term is not reproducible. The surrounding text also claims that the objective 'ensures that the generated motion trajectory satisfies continuity C2 through regularization of the Hessian matrix,' but Eq. (2) is a second finite difference (an acceleration-like term), not a Hessian regularization, and no derivation is given. The authors should either provide a precise definition of Σ and a derivation of the C2 claim, or remove the claim and state the finite-difference interpretation.
- [§III.C and §III.D, Eqs. (4)–(8), (14)] Several load-bearing hyperparameters are not reported. The dual-threshold contact extraction depends on dis_min and dis_max; the contact likelihood in Eq. (7) depends on β1; the state imputation in Eq. (8) depends on v_max; and the total energy in Eq. (14) depends on w_pen, w_align, w_spen, and w_joints. None of these values are given, and there is no sensitivity analysis. Since these parameters directly determine the contact maps and the refined finger poses that feed into the Isaac Gym evaluation, the quantitative results in Tables II and III cannot be reproduced or independently assessed without them.
- [Section V (Discussion and Limitations)] The paper itself states that contact information 'would be more reliable if directly extracted from video data instead of relying on the reconstructed metadata' and that the optimization 'struggles with inconsistencies in metadata quality.' This is a load-bearing premise: the dual-threshold contact maps (Section III.C), the finger-wise energy optimization (Section III.D), and the simulated success evaluation (Section IV.B) all inherit errors from reconstructed MANO hand-object parameters and object meshes. The authors should quantify the accuracy of their reconstructed input data, for example by comparing against a subset of hand-annotated or video-derived contacts, or at minimum report how sensitive the final SSR and penetration metrics are to perturbations in the reconstructed input. Without such evidence, the validation is partly self-referential because the quality of the input data is not independently established.
minor comments (6)
- [§III.B, Eq. (1)] Equation (1) sums from i=0 to N with N=13, implying 14 keypoints, while the text states N=13; please make the index range consistent. Also, the notation v_i_H(θ_t, β_t, r_t) uses θ_t, β_t, and r_t without defining these arguments for the human hand model.
- [§III.C, Eq. (4)] In Eq. (4), C_t is described as a contact state, but the expression ∥C_{t−1}+C_{t+1}∥/2 suggests a vector or a normed quantity; if C_t is binary or categorical, the addition and norm need a precise definition. Please clarify whether the indicator applies to a scalar condition or to a vector norm.
- [Table II] The columns SPD, PD, CD, and FVR are not defined in this manuscript; the text only states they are 'measured based on BODex [5].' Please provide one-sentence definitions or a reference to the exact definitions so readers can interpret the numbers, and report standard deviations or confidence intervals over repeated runs.
- [Eq. (15)] The Chamfer distance in Eq. (15) is written as a one-sided distance with min over p ∈ P_ref and q ∈ P_gen; a symmetric Chamfer distance is the common choice in this literature. Please specify whether the metric is symmetric and, if not, state which direction is used.
- [References] Reference [1] is cited for MANO, but the listed reference is the SMPL paper. Please cite the MANO model paper (Romero et al., 2017) or correct the reference.
- [Table III] The row labels 'retarget (Ours)' and 'Optimization (Ours)' are ambiguous; 'Optimization' likely means the full DexFlow pipeline after contact refinement. Please rename the rows to match the terminology used in Section III, e.g., 'DexFlow-retarget' and 'DexFlow-full'.
Circularity Check
No circularity found: the retargeting pipeline, contact refinement, and simulator-based evaluation are self-contained; the SSR comparison protocol mismatch is an evaluation-fairness concern, not a definitional reduction.
full rationale
The paper's derivation chain is self-contained at the equation level. The retargeting objective (Eq. 1-3) maps MANO task-space vectors to robot joint angles using alignment, temporal smoothness, and a differential loss; the contact refinement stage (Eq. 9-14) minimizes distance, penetration, alignment, self-penetration, and regularization energies; and the reported quality metrics are then evaluated in Isaac Gym, a physics simulator external to the optimization. No fitted parameter is renamed as a prediction, and no result is defined in terms of the quantity it is claimed to explain. The CD and PD metrics are related to the Edis and Epen objectives, but this is a standard objective-evaluation relationship rather than a circular derivation: the generated poses are not constructed from the evaluation numbers, and the comparison rows for baselines are imported transparently from BODex (Section IV.B.1). The baseline SSR comparison may be unfair because DexFlow's success rule ('if any frame in the sequence after contacting the object satisfies the condition, the grasp is considered successful') is a maximal-leniency criterion while DexRetarget's 5.35% is taken from another paper, but that is a correctness/fairness risk, not a circularity in the derivation. The paper's own limitation statement (Section V) that contact information 'would be more reliable if directly extracted from video data instead of relying on the reconstructed metadata' is an input-quality caveat, not a circular step. Self-citations to prior work by the authors appear only in related-work examples ([13], [17]) and are not load-bearing. No uniqueness theorem, ansatz, or empirical pattern is imported from the authors' own prior work to force the central claim.
Assumptions & free parameters
free parameters (14)
- object and hand scale factor s =
10/9
- alignment keypoint count N =
13
- temporal consistency weight alpha (Eq 1) =
not reported
- differential loss weight lambda =
0.1
- kinematic covariance matrix Sigma =
not defined (claimed 28x28)
- dynamic smoothing weight gamma =
0.5
- contact distance lower and upper thresholds dis_min and dis_max =
not reported
- raw contact threshold tau_d =
not reported
- velocity modulation weight alpha_c (called alpha in Eq 4) =
0.6
- average finger velocity vf =
0.8 m/s
- contact confidence threshold tau_c =
0.7
- contact likelihood gain beta1 =
not reported
- max contact-state velocity v_max =
not reported
- energy weights w_pen, w_align, w_spen, w_joints =
not reported
assumptions (7)
- domain assumption Forward kinematics and geometry of MANO, ShadowHand, and Allegro are accurate and differentiable as used in Eq 1 and Eqs 9-13.
- domain assumption Reconstructed MANO hand-object sequences and object meshes are accurate enough for contact extraction and optimization.
- domain assumption Matching 13 task-space keypoints (fingertips and palm roots) transfers human grasp intent to a robot hand.
- domain assumption Temporal smoothness and continuity of joint angles improve physical plausibility and naturalness.
- domain assumption Isaac Gym with PhysX is a faithful proxy for real grasp success.
- ad hoc to paper The sequential success criterion, any post-contact frame passing physics conditions counts as success, is a valid measure of semantic success.
- standard math The global optimizer GN_CRS2_LM and the iterative weighted energy minimization converge to useful solutions.
Cite this review
Pith. "Pith review of DexFlow: A Unified Approach for Dexterous Hand Pose Retargeting and Interaction." pith.science (2026). https://pith.science/paper/HVSKSDOW
@misc{pith2026250501083,
author = {Pith},
title = {Pith review of: DexFlow: A Unified Approach for Dexterous Hand Pose Retargeting and Interaction},
year = {2026},
howpublished = {\url{https://pith.science/paper/HVSKSDOW}},
note = {Machine review of arXiv:2505.01083}
}
read the original abstract
Despite advances in hand-object interaction modeling, generating realistic dexterous manipulation data for robotic hands remains a challenge. Retargeting methods often suffer from low accuracy and fail to account for hand-object interactions, leading to artifacts like interpenetration. Generative methods, lacking human hand priors, produce limited and unnatural poses. We propose a data transformation pipeline that combines human hand and object data from multiple sources for high-precision retargeting. Our approach uses a differential loss constraint to ensure temporal consistency and generates contact maps to refine hand-object interactions. Experiments show our method significantly improves pose accuracy, naturalness, and diversity, providing a robust solution for hand-object interaction modeling.
Figures
Figures from the paper (1 more)
Forward citations
Cited by 2 Pith papers
-
C2Dex: Contact-Consistent Reconstruction and Retargeting for Dexterous Manipulation from Monocular Video
C2Dex converts monocular human videos into executable dexterous robot manipulation trajectories by using stable object-side contacts as a shared representation for reconstruction and retargeting, achieving 57.78% and ...
-
The Latent Color Subspace: Emergent Order in High-Dimensional Chaos
FLUX.1’s VAE latent space contains an interpretable Hue–Saturation–Lightness structure that enables training-free color prediction and control via closed-form latent edits.
Reference graph
Works this paper leans on
-
[1]
Smpl: A skinned multi-person linear model
Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons- Moll, and Michael J Black. Smpl: A skinned multi-person linear model. In Seminal Graphics Papers: Pushing the Boundaries, V olume 2, pages 851–866. 2023
work page 2023
-
[2]
Dexmv: Imitation learning for dexterous manipulation from human videos
Yuzhe Qin, Yueh-Hua Wu, Shaowei Liu, Hanwen Jiang, Ruihan Yang, Yang Fu, and Xiaolong Wang. Dexmv: Imitation learning for dexterous manipulation from human videos. In European Conference on Computer Vision , pages 570–587. Springer, 2022
work page 2022
-
[3]
Dexgraspnet: A large-scale robotic dexterous grasp dataset for general objects based on simulation
Ruicheng Wang, Jialiang Zhang, Jiayi Chen, Yinzhen Xu, Puhao Li, Tengyu Liu, and He Wang. Dexgraspnet: A large-scale robotic dexterous grasp dataset for general objects based on simulation. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 11359–11366. IEEE, 2023
work page 2023
-
[4]
Springgrasp: An optimization pipeline for robust and compliant dexterous pre-grasp synthesis
Sirui Chen, Jeannette Bohg, and C Karen Liu. Springgrasp: An optimization pipeline for robust and compliant dexterous pre-grasp synthesis. arXiv e-prints , pages arXiv–2404, 2024
work page 2024
-
[5]
Bodex: Scalable and efficient robotic dexterous grasp synthesis using bilevel optimization
Jiayi Chen, Yubin Ke, and He Wang. Bodex: Scalable and efficient robotic dexterous grasp synthesis using bilevel optimization. arXiv preprint arXiv:2412.16490, 2024
arXiv 2024
-
[6]
Anyteleop: A general vision-based dexterous robot arm-hand teleoperation system
Yuzhe Qin, Wei Yang, Binghao Huang, Karl Van Wyk, Hao Su, Xiaolong Wang, Yu-Wei Chao, and Dieter Fox. Anyteleop: A general vision-based dexterous robot arm-hand teleoperation system. arXiv preprint arXiv:2307.04577, 2023
arXiv 2023
-
[7]
Dexpilot: Vision-based teleoperation of dexterous robotic hand-arm system
Ankur Handa, Karl Van Wyk, Wei Yang, Jacky Liang, Yu-Wei Chao, Qian Wan, Stan Birchfield, Nathan Ratliff, and Dieter Fox. Dexpilot: Vision-based teleoperation of dexterous robotic hand-arm system. In 2020 IEEE International Conference on Robotics and Automation (ICRA), pages 9164–9170. IEEE, 2020
2020
-
[8]
Task-oriented hand motion retargeting for dexterous manipulation imitation
Dafni Antotsiou, Guillermo Garcia-Hernando, and Tae-Kyun Kim. Task-oriented hand motion retargeting for dexterous manipulation imitation. In Proceedings of the European conference on computer vision (ECCV) workshops , pages 0–0, 2018
work page 2018
Show all 27 references
-
[9]
Vividex: Learning vision-based dexterous manipulation from human videos
Zerui Chen, Shizhe Chen, Etienne Arlaud, Ivan Laptev, and Cordelia Schmid. Vividex: Learning vision-based dexterous manipulation from human videos. arXiv preprint arXiv:2404.15709 , 2024
2024 arXiv
-
[10]
Kinematic motion retargeting for contact-rich anthropomorphic ma- nipulations
Arjun S Lakshmipathy, Jessica K Hodgins, and Nancy S Pollard. Kinematic motion retargeting for contact-rich anthropomorphic ma- nipulations. arXiv preprint arXiv:2402.04820 , 2024
2024 arXiv
-
[11]
Task-oriented optimal grasping by multifingered robot hands
Zexiang Li and S Shankar Sastry. Task-oriented optimal grasping by multifingered robot hands. IEEE Journal on Robotics and Automation , 4(1):32–44, 1988
1988
-
[12]
On computing task-oriented grasps
Sahar El-Khoury, Ravin de Souza, and Aude Billard. On computing task-oriented grasps. Robotics and Autonomous Systems , 66:145–158, 2015
2015
-
[13]
Exploiting kinematic redundancy for robotic grasping of multiple objects
Kunpeng Yao and Aude Billard. Exploiting kinematic redundancy for robotic grasping of multiple objects. IEEE Transactions on Robotics , 39(3):1982–2002, 2023
1982
-
[14]
Grasp planning: How to choose a suitable task wrench space
Ch Borst, Max Fischer, and Gerd Hirzinger. Grasp planning: How to choose a suitable task wrench space. In IEEE International Conference on Robotics and Automation, 2004. Proceedings. ICRA’04. 2004 , volume 1, pages 319–325. IEEE, 2004
2004
-
[15]
Grasp planning to maximize task coverage
Yun Lin and Yu Sun. Grasp planning to maximize task coverage. The International Journal of Robotics Research , 34(9):1195–1210, 2015
2015
-
[16]
Partial closure grasps: Metrics and computation
Heinrich Kruger and A Frank van der Stappen. Partial closure grasps: Metrics and computation. In 2011 IEEE International Conference on Robotics and Automation , pages 5024–5030. IEEE, 2011
2011
-
[17]
Learning partial power grasp with task-specific contact
Miao Li. Learning partial power grasp with task-specific contact. In 2016 IEEE International Conference on Robotics and Biomimetics (ROBIO), pages 337–343. IEEE, 2016
2016
-
[18]
Frogger: Fast robust grasp generation via the min-weight metric
Albert H Li, Preston Culbertson, Joel W Burdick, and Aaron D Ames. Frogger: Fast robust grasp generation via the min-weight metric. In 2023 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 6809–6816. IEEE, 2023
2023
-
[19]
Synthesizing diverse and physically stable grasps with arbitrary hand structures using differentiable force closure estimator
Tengyu Liu, Zeyu Liu, Ziyuan Jiao, Yixin Zhu, and Song-Chun Zhu. Synthesizing diverse and physically stable grasps with arbitrary hand structures using differentiable force closure estimator. IEEE Robotics and Automation Letters , 7(1):470–477, 2021
2021
-
[20]
Grasp’d: Differentiable contact-rich grasp synthesis for multi-fingered hands
Dylan Turpin, Liquan Wang, Eric Heiden, Yun-Chun Chen, Miles Macklin, Stavros Tsogkas, Sven Dickinson, and Animesh Garg. Grasp’d: Differentiable contact-rich grasp synthesis for multi-fingered hands. In European Conference on Computer Vision , pages 201–221. Springer, 2022
2022
-
[21]
Handdgp: Camera-space hand mesh prediction with differentiable global posi- tioning
Eugene Valassakis and Guillermo Garcia-Hernando. Handdgp: Camera-space hand mesh prediction with differentiable global posi- tioning. In European Conference on Computer Vision , pages 479–496. Springer, 2024
2024
-
[22]
Unidexgrasp: Universal robotic dexterous grasping via learning diverse proposal generation and goal-conditioned policy
Yinzhen Xu, Weikang Wan, Jialiang Zhang, Haoran Liu, Zikang Shan, Hao Shen, Ruicheng Wang, Haoran Geng, Yijia Weng, Jiayi Chen, et al. Unidexgrasp: Universal robotic dexterous grasping via learning diverse proposal generation and goal-conditioned policy. In Proceedings of the ...
2023
-
[23]
Dexterous reaching, grasp transfer and planning using electro- static representations
Peter Sandilands, Vladimir Ivan, Taku Komura, and Sethu Vijayaku- mar. Dexterous reaching, grasp transfer and planning using electro- static representations. In 2013 13th IEEE-RAS International Confer- ence on Humanoid Robots (Humanoids) , pages 211–218, 2013
2013
-
[24]
Porta, and Llu ´ıs Ros
Carlos Rosales, Josep M. Porta, and Llu ´ıs Ros. Grasp optimization under specific contact constraints. IEEE Transactions on Robotics , 29(3):746–757, 2013
2013
-
[25]
Arjun Lakshmipathy, Dominik Bauer, Cornelia Bauer, and Nancy S. Pollard. Contact transfer: A direct, user-driven method for human to robot transfer of grasps and manipulations. In 2022 International Conference on Robotics and Automation (ICRA) , pages 6195–6201, 2022
2022
-
[26]
Ulrich Hillenbrand and Maximo A. Roa. Transferring functional grasps through contact warping and local replanning. In 2012 IEEE/RSJ International Conference on Intelligent Robots and Systems , pages 2963–2970, 2012
2012
-
[27]
Isaac gym: High performance gpu-based physics simulation for robot learning
Viktor Makoviychuk, Lukasz Wawrzyniak, Yunrong Guo, Michelle Lu, Kier Storey, Miles Macklin, David Hoeller, Nikita Rudin, Arthur Allshire, Ankur Handa, et al. Isaac gym: High performance gpu-based physics simulation for robot learning. arXiv preprint arXiv:2108.10470, 2021
2021 arXiv
Reviewed August 16, 2026 · model on record in the stance chip above.
Discussion (0). Continue with ORCID to comment.