arxiv: 2604.28161 · v1 · submitted 2026-04-30 · 💻 cs.RO

Recognition: unknown

RopeDreamer: A Kinematic Recurrent State Space Model for Dynamics of Flexible Deformable Linear Objects

Tim Missal , Lucas Domingues , Berk Guler , Simon Manschitz , Jan Peters , Paula Dornhofer Paro Costa

Authors on Pith no claims yet

Pith reviewed 2026-05-07 06:22 UTC · model grok-4.3

classification 💻 cs.RO

keywords deformable linear objectsrope dynamicsrecurrent state space modelquaternionic kinematicsrobotic manipulationlatent dynamicspick-and-placetopological consistency

0 comments

The pith

A recurrent state space model with quaternionic kinematics predicts long-term rope dynamics more accurately by enforcing physical link lengths.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper develops a dynamics model for flexible linear objects such as ropes that robots must manipulate. Prior data-driven approaches often allow non-physical behaviors like stretching or improper tangling because they treat positions independently. The new method represents the object as a chain of relative rotations using quaternions inside a recurrent state space model, which automatically keeps segment lengths constant. A dual-decoder structure separates the task of reconstructing the current state from predicting future states, pushing the latent representation to reflect actual deformation physics. On simulated pick-and-place sequences that include crossings, the model lowers open-loop prediction error by 40.52 percent over 50 steps while running 31 percent faster and preserving topology better than the baseline.

Core claim

Encoding the DLO as a sequence of relative rotations (quaternions) rather than independent Cartesian positions inherently constrains the model to a physically valid manifold that preserves link-length constancy. The dual-decoder architecture decouples state reconstruction from future-state prediction, forcing the latent space to capture the underlying physics of deformation. Evaluated on large-scale simulated pick-and-place trajectories with self-intersections, this yields a 40.52 percent reduction in open-loop prediction error over 50-step horizons and a 31.17 percent reduction in inference time compared to the state-of-the-art baseline, with improved topological consistency at multiple cro

What carries the argument

The quaternionic kinematic chain representation inside a recurrent state space model with dual decoder, which encodes the object through sequences of relative rotations to enforce physical manifold constraints during latent dynamics forecasting.

If this is right

The model achieves a 40.52 percent reduction in open-loop prediction error over 50-step horizons compared to the baseline.
Inference time is reduced by 31.17 percent while maintaining higher topological consistency during multiple crossings.
The approach serves as a compositional primitive suitable for long-horizon manipulation planning tasks.
By construction the quaternion encoding prevents link stretching and non-physical deformations that affect prior methods.

Where Pith is reading between the lines

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

The same kinematic encoding could be tested on other deformable objects such as cloth or soft tubes if the relative-rotation manifold generalizes beyond linear chains.
Embedding the model inside a planner might allow robots to generate collision-free cable-routing sequences that remain valid over dozens of steps.
If sim-to-real transfer holds, the method could lower the amount of real-robot data needed to train reliable dynamics predictors for contact-rich tasks.

Load-bearing premise

The simulated pick-and-place trajectories with self-intersections sufficiently capture the distribution of real-world contact forces, friction, and material properties so that the learned latent dynamics transfer without retraining.

What would settle it

Collecting real-world data from physical rope manipulation experiments and measuring whether the model's open-loop prediction error over 50 steps remains substantially lower than the baseline without any retraining on that data.

Figures

Figures reproduced from arXiv: 2604.28161 by Berk Guler, Jan Peters, Lucas Domingues, Paula Dornhofer Paro Costa, Simon Manschitz, Tim Missal.

**Figure 1.** Figure 1: After observing an initial state s0 from simulation, RopeDreamer predicts the following evolution of the Deformable Linear Object given a sequence of actions u0:T . Abstract— The robotic manipulation of Deformable Linear Objects (DLOs) is a fundamental challenge due to the highdimensional, non-linear dynamics of flexible structures and the complexity of maintaining topological integrity during contact-ric… view at source ↗

**Figure 2.** Figure 2: Problem formulation of DLO shape prediction task. a) The DLO view at source ↗

**Figure 3.** Figure 3: Architecture of the World Model for DLO manipulation. The current DLO state and manipulation actions are first fed into their respective encoders view at source ↗

**Figure 4.** Figure 4: Comparison of RopeDreamer variants (Ours) and baseline variants view at source ↗

**Figure 5.** Figure 5: Comparison of mean single-step inference time over different model view at source ↗

**Figure 7.** Figure 7: Example of a loop and prediction at t = 20 of our small model (red) and GA-Net S (blue) compared against the ground truth (black). Our model effectively identifies the loop and accurately reconstructs the rope. GA-Net fails to maintain DLO topology and predicts physically unrealistic link lengths. truth codes at each step, we measure the model’s preservation of the DLO’s structural identity. A match in Gau… view at source ↗

read the original abstract

The robotic manipulation of Deformable Linear Objects (DLOs) is a fundamental challenge due to the high-dimensional, non-linear dynamics of flexible structures and the complexity of maintaining topological integrity during contact-rich tasks. While recent data-driven methods have utilized Recurrent and Graph Neural Networks for dynamics modeling, they often struggle with self-intersections and non-physical deformations, such as tangling and link stretching. In this paper, we propose a latent dynamics framework that combines a Recurrent State Space Model with a Quaternionic Kinematic Chain representation to enable robust, long-term forecasting of DLO states. By encoding the DLO as a sequence of relative rotations (quaternions) rather than independent Cartesian positions, we inherently constrain the model to a physically valid manifold that preserves link-length constancy. Furthermore, we introduce a dual-decoder architecture that decouples state reconstruction from future-state prediction, forcing the latent space to capture the underlying physics of deformation. We evaluate our approach on a large-scale simulated dataset of complex pick-and-place trajectories involving self-intersections. Our results demonstrate that the proposed model achieves a 40.52% reduction in open-loop prediction error over 50-step horizons compared to the state-of-the-art baseline, while reducing inference time by 31.17%. Our model further maintains superior topological consistency in scenarios with multiple crossings, proving its efficacy as a compositional primitive for long-horizon manipulation planning.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

RopeDreamer cuts simulated DLO prediction error 40% with RSSM plus quaternion kinematics and dual decoders, but the gains rest entirely on simulation with no real-robot checks.

read the letter

The key takeaway is that this paper builds a recurrent state space model for deformable linear objects by encoding them as a chain of relative quaternions and adding a dual-decoder split that separates reconstruction from prediction. On their simulated pick-and-place trajectories with self-intersections, the model reports a 40.52% drop in open-loop error over 50 steps and 31% faster inference than the prior baseline, while keeping better topological consistency across crossings. The quaternion representation enforces link-length constancy by construction, which directly tackles the stretching and tangling problems that plague position-based models. The dual decoder appears to push the latent space toward actual dynamics rather than simple copying of the current state. These choices are a clean way to inject kinematic priors into an RSSM without heavy penalty terms. The main limitation is that all results come from simulation. No physical robot trials, real sensor data, or domain-randomization tests are shown, so it remains unclear whether the error reductions survive real friction, contact forces, or material variation. The abstract gives no error bars, ablation numbers, or detailed baseline descriptions, which makes it hard to judge how much of the gain is architectural versus tuning. Free parameters such as latent dimension and segment count are listed, suggesting room for sensitivity analysis that is not yet visible. This work is aimed at researchers who need long-horizon dynamics models for cable routing, fabric handling, or surgical tasks. Anyone already using state-space models or kinematic constraints in robot learning will find the specific architecture worth examining. I would send it to peer review. The technical framing is coherent and the simulation gains are large enough to justify expert feedback, even if real-world validation is still needed.

Referee Report

3 major / 2 minor

Summary. The paper proposes RopeDreamer, a latent dynamics framework that integrates a Recurrent State Space Model (RSSM) with a quaternionic kinematic chain representation for modeling the dynamics of Deformable Linear Objects (DLOs). The DLO is encoded as a sequence of relative rotations (quaternions) to enforce link-length constancy and avoid non-physical deformations such as stretching or tangling. A dual-decoder architecture separates state reconstruction from future prediction in the latent space. The model is evaluated exclusively on a large-scale simulated dataset of pick-and-place trajectories involving self-intersections, where it claims a 40.52% reduction in open-loop 50-step prediction error and a 31.17% reduction in inference time relative to a state-of-the-art baseline, along with improved topological consistency.

Significance. If the reported gains hold under more rigorous scrutiny, the work could meaningfully advance data-driven dynamics modeling for robotic DLO manipulation by embedding kinematic constraints directly into an RSSM, addressing common failure modes like self-intersections that plague prior RNN- and GNN-based approaches. The dual-decoder design and quaternion encoding are well-motivated choices that promote physical validity in the latent space. The quantitative improvements on simulation data are a positive signal for long-horizon forecasting. However, the exclusive use of simulation without real-world validation or transfer studies limits the immediate significance for practical manipulation planning.

major comments (3)

Evaluation section: The headline claims of a 40.52% reduction in open-loop prediction error and 31.17% faster inference are presented without error bars, standard deviations across trials, the number of evaluation trajectories, or random seeds. This absence makes it impossible to judge whether the gains are statistically reliable or sensitive to particular test conditions.
Methods section: The state-of-the-art baseline is referenced only generically without a specific citation, architectural description, training procedure, or hyperparameter details. This prevents verification that the comparison is fair and obscures which elements (quaternion encoding, dual decoder, or RSSM structure) drive the reported improvements.
Introduction and Conclusion: The manuscript asserts utility as a 'compositional primitive for long-horizon manipulation planning,' yet all results are confined to simulation with no real-robot experiments, sim-to-real transfer tests, or domain-randomization ablations. Because the central performance numbers rest on the assumption that simulated contact forces, friction, and material properties are representative of real DLOs, this gap is load-bearing for the broader claims.

minor comments (2)

Abstract: The phrase 'state-of-the-art baseline' should include a specific reference to the prior work being compared against.
Notation and figures: The description of the kinematic chain and dual-decoder would benefit from an explicit diagram or additional equations clarifying how relative quaternions are integrated into the RSSM latent state.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the constructive and detailed feedback. The comments highlight important areas for improving the rigor and clarity of the manuscript. We address each major comment point-by-point below and will incorporate revisions to strengthen the paper.

read point-by-point responses

Referee: Evaluation section: The headline claims of a 40.52% reduction in open-loop prediction error and 31.17% faster inference are presented without error bars, standard deviations across trials, the number of evaluation trajectories, or random seeds. This absence makes it impossible to judge whether the gains are statistically reliable or sensitive to particular test conditions.

Authors: We agree that statistical details are essential for assessing the reliability of the reported gains. In the revised manuscript, we will add error bars to all relevant plots, report standard deviations computed across 5 independent random seeds, explicitly state the number of evaluation trajectories (500 held-out test trajectories), and include a table summarizing mean performance with variability measures. This will allow readers to better evaluate the consistency of the 40.52% error reduction and 31.17% inference speedup. revision: yes
Referee: Methods section: The state-of-the-art baseline is referenced only generically without a specific citation, architectural description, training procedure, or hyperparameter details. This prevents verification that the comparison is fair and obscures which elements (quaternion encoding, dual decoder, or RSSM structure) drive the reported improvements.

Authors: We apologize for the insufficient detail on the baseline. In the revised Methods section, we will provide the full citation to the baseline method, a complete architectural description, the training procedure (including loss functions and optimization settings), and the specific hyperparameters used in our re-implementation. This will make the comparison fully transparent and help isolate the contributions of the quaternionic kinematic chain and dual-decoder components. revision: yes
Referee: Introduction and Conclusion: The manuscript asserts utility as a 'compositional primitive for long-horizon manipulation planning,' yet all results are confined to simulation with no real-robot experiments, sim-to-real transfer tests, or domain-randomization ablations. Because the central performance numbers rest on the assumption that simulated contact forces, friction, and material properties are representative of real DLOs, this gap is load-bearing for the broader claims.

Authors: We acknowledge that the current evaluation is limited to simulation and that this constrains the strength of claims about real-world manipulation planning. The work prioritizes rigorous validation of the kinematic constraints and dual-decoder design on complex simulated trajectories with self-intersections. In the revised manuscript, we will qualify the language in the Introduction and Conclusion to present the model as a promising foundation rather than a proven primitive, add an explicit Limitations section discussing simulation assumptions, and outline future directions for sim-to-real transfer and real-robot experiments. We will not add new real-world results at this stage. revision: partial

Circularity Check

0 steps flagged

No significant circularity; modeling choices and empirical results are independent

full rationale

The paper defines a kinematic RSSM using relative quaternions to enforce link-length constancy and a dual-decoder to separate reconstruction from prediction. These are explicit architectural decisions that impose manifold constraints by construction, but the paper does not claim any performance metric or physical prediction as following mathematically from those choices alone. The headline 40.52% error reduction and 31.17% inference speedup are reported as direct experimental outcomes on a held-out simulated dataset; they are not obtained by refitting parameters to the same quantities or by renaming inputs. No self-citation chain, uniqueness theorem, or ansatz smuggling is present in the derivation. The approach therefore remains a standard application of RSSM techniques with added kinematic priors, evaluated externally rather than tautologically.

Axiom & Free-Parameter Ledger

2 free parameters · 2 axioms · 0 invented entities

The central claim rests on standard quaternion algebra for rotation representation and the assumption that separating reconstruction and prediction decoders forces the latent space to encode physical deformation rules. No new physical constants or entities are introduced; the main free parameters are the usual neural-network weights and latent dimensionality chosen during training.

free parameters (2)

latent dimensionality
Hyperparameter controlling the size of the hidden state in the recurrent model; its value is selected to fit the training data.
number of kinematic segments
Discretization level of the DLO into rigid links; chosen as a modeling decision before training.

axioms (2)

standard math Quaternion multiplication preserves unit length and therefore inter-link distances
Invoked to guarantee that the kinematic chain representation cannot produce stretching.
domain assumption Decoupling reconstruction and prediction decoders forces the latent variables to capture underlying physics
Stated as the motivation for the dual-decoder design.

pith-pipeline@v0.9.0 · 5571 in / 1392 out tokens · 121170 ms · 2026-05-07T06:22:45.910187+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

35 extracted references · 17 canonical work pages · 2 internal anchors

[1]

Planning and Control for Cable-routing with Dual-arm Robot,

G. A. Waltersson, R. Laezza, and Y . Karayiannidis, “Planning and Control for Cable-routing with Dual-arm Robot,” in2022 International Conference on Robotics and Automation (ICRA), (Philadelphia, PA, USA), pp. 1046–1052, IEEE, May 2022

2022
[2]

Cable Routing and Assembly using Tactile-driven Motion Primitives,

A. Wilson, H. Jiang, W. Lian, and W. Yuan, “Cable Routing and Assembly using Tactile-driven Motion Primitives,” 2023

2023
[3]

Hierarchical Planning for Rope Manipulation using Knot Theory and a Learned Inverse Model,

M. Sudry, T. Jurgenson, A. Tamar, and E. Karpas, “Hierarchical Planning for Rope Manipulation using Knot Theory and a Learned Inverse Model,” inProceedings of The 7th Conference on Robot Learning, vol. 229 ofProceedings of Machine Learning Research, pp. 1596–1609, PMLR, 2023

2023
[4]

Untangling Dense Knots by Learning Task-Relevant Keypoints,

J. Grannen, P. Sundaresan, B. Thananjeyan, J. Ichnowski, A. Bal- akrishna, V . Viswanath, M. Laskey, J. Gonzalez, and K. Goldberg, “Untangling Dense Knots by Learning Task-Relevant Keypoints,” in Proceedings of the 2020 Conference on Robot Learning, pp. 782–800, PMLR, Oct. 2021. ISSN: 2640-3498

2020
[5]

Towards Assistive Teleoperation for Knot Untangling,

B. Guler, K. Pompetzki, S. Manschitz, and J. Peters, “Towards Assistive Teleoperation for Knot Untangling,” inGerman Robotics Conference (GRC), Mar. 2025

2025
[6]

Learning Graph Dynamics With Interaction Effects Propagation for Deformable Linear Objects Shape Control,

F. Gu, H. Sang, Y . Zhou, J. Ma, R. Jiang, Z. Wang, and B. He, “Learning Graph Dynamics With Interaction Effects Propagation for Deformable Linear Objects Shape Control,”IEEE Transactions on Automation Science and Engineering, vol. 22, pp. 10881–10892, 2025

2025
[7]

Graph Neural Networks Exponen- tially Lose Expressive Power for Node Classification,

K. Oono and T. Suzuki, “Graph Neural Networks Exponen- tially Lose Expressive Power for Node Classification,” Jan. 2021. arXiv:1905.10947 [cs]

work page arXiv 2021
[8]

How does over-squashing affect the power of GNNs?,

F. D. Giovanni, T. K. Rusch, M. M. Bronstein, A. Deac, M. Lackenby, S. Mishra, and P. Veli ˇckovi´c, “How does over-squashing affect the power of GNNs?,” Feb. 2024. arXiv:2306.03589 [cs]

work page arXiv 2024
[9]

and Yahav, E

U. Alon and E. Yahav, “On the Bottleneck of Graph Neural Networks and its Practical Implications,” Mar. 2021. arXiv:2006.05205 [cs]

work page arXiv 2021
[10]

Dream to Control: Learning Behaviors by Latent Imagination

D. Hafner, T. Lillicrap, J. Ba, and M. Norouzi, “Dream to Control: Learning Behaviors by Latent Imagination,” Mar. 2020. arXiv:1912.01603 [cs]

work page internal anchor Pith review arXiv 2020
[11]

Learning Predictive Representations for Deformable Objects Using Contrastive Estima- tion,

W. Yan, A. Vangipuram, P. Abbeel, and L. Pinto, “Learning Predictive Representations for Deformable Objects Using Contrastive Estima- tion,” Mar. 2020. arXiv:2003.05436 [cs]

work page arXiv 2020
[12]

De- formable Linear Object Prediction Using Locally Linear Latent Dy- namics,

W. Zhang, K. Schmeckpeper, P. Chaudhari, and K. Daniilidis, “De- formable Linear Object Prediction Using Locally Linear Latent Dy- namics,” Mar. 2021. arXiv:2103.14184 [cs]

work page arXiv 2021
[13]

Sample- Efficient Learning of Deformable Linear Object Manipulation in the Real World Through Self-Supervision,

R. Lee, M. Hamaya, T. Murooka, Y . Ijiri, and P. Corke, “Sample- Efficient Learning of Deformable Linear Object Manipulation in the Real World Through Self-Supervision,”IEEE Robotics and Automa- tion Letters, vol. 7, pp. 573–580, Jan. 2022

2022
[14]

Self-Supervised Learning of State Estimation for Manipulating Deformable Linear Objects,

M. Yan, Y . Zhu, N. Jin, and J. Bohg, “Self-Supervised Learning of State Estimation for Manipulating Deformable Linear Objects,” Oct
[15]

arXiv:1911.06283 [cs]

work page arXiv 1911
[16]

LSTM-GCN Hy- brid Architecture for Model Predictive Control of Deformable Linear Objects,

Z. Yue, X. Zhang, Y . Wang, S. Jiang, and J. Zhao, “LSTM-GCN Hy- brid Architecture for Model Predictive Control of Deformable Linear Objects,” in2025 IEEE International Conference on Mechatronics and Automation (ICMA), pp. 303–309, Aug. 2025. ISSN: 2152-744X

2025
[17]

A Hybrid Force-Position Strategy for Shape Control of Deformable Linear Objects With Graph Attention Networks,

Y . Yu, H. Yang, J. Tan, and X. Wang, “A Hybrid Force-Position Strategy for Shape Control of Deformable Linear Objects With Graph Attention Networks,” Aug. 2025. arXiv:2508.07319 [cs]

work page arXiv 2025
[18]

Shape Control of Deformable Linear Objects with Offline and Online Learning of Local Linear Deformation Models,

M. Yu, H. Zhong, and X. Li, “Shape Control of Deformable Linear Objects with Offline and Online Learning of Local Linear Deformation Models,” Feb. 2022. arXiv:2109.11091 [cs]

work page arXiv 2022
[19]

Learning to Propagate Inter- action Effects for Modeling Deformable Linear Objects Dynamics,

Y . Yang, J. A. Stork, and T. Stoyanov, “Learning to Propagate Inter- action Effects for Modeling Deformable Linear Objects Dynamics,” in2021 IEEE International Conference on Robotics and Automation (ICRA), pp. 1950–1957, May 2021. ISSN: 2577-087X

1950
[20]

DeformNet: Latent Space Modeling and Dynamics Prediction for Deformable Object Manipulation,

C. Li, Z. Ai, T. Wu, X. Li, W. Ding, and H. Xu, “DeformNet: Latent Space Modeling and Dynamics Prediction for Deformable Object Manipulation,” Feb. 2024. arXiv:2402.07648 [cs]

work page arXiv 2024
[21]

TrackDLO: Tracking Deformable Linear Objects Under Occlusion With Motion Coherence,

J. Xiang, H. Dinkel, H. Zhao, N. Gao, B. Coltin, T. Smith, and T. Bretl, “TrackDLO: Tracking Deformable Linear Objects Under Occlusion With Motion Coherence,”IEEE Robotics and Automation Letters, vol. 8, pp. 6179–6186, Oct. 2023

2023
[22]

mBEST: Realtime Deformable Linear Object Detection Through Minimal Bending Energy Skeleton Pixel Traversals,

A. Choi, D. Tong, B. Park, D. Terzopoulos, J. Joo, and M. K. Jawed, “mBEST: Realtime Deformable Linear Object Detection Through Minimal Bending Energy Skeleton Pixel Traversals,”IEEE Robotics and Automation Letters, vol. 8, pp. 4863–4870, Aug. 2023. arXiv:2302.09444 [cs]

work page arXiv 2023
[23]

RT-DLO: Real-Time Deformable Linear Objects Instance Segmentation,

A. Caporali, K. Galassi, B. L. ˇZagar, R. Zanella, G. Palli, and A. C. Knoll, “RT-DLO: Real-Time Deformable Linear Objects Instance Segmentation,”IEEE Transactions on Industrial Informatics, vol. 19, pp. 11333–11342, Nov. 2023

2023
[24]

Limitations of Normalization in Attention Mechanism,

T. Mudarisov, M. Burtsev, T. Petrova, and R. State, “Limitations of Normalization in Attention Mechanism,” Oct. 2025. arXiv:2508.17821 [cs]

work page arXiv 2025
[25]

Mamba-360: Survey of State Space Models as Transformer Alternative for Long Sequence Modelling: Methods, Applications, and Challenges,

B. N. Patro and V . S. Agneeswaran, “Mamba-360: Survey of State Space Models as Transformer Alternative for Long Sequence Modelling: Methods, Applications, and Challenges,” Apr. 2024. arXiv:2404.16112 [cs]

work page arXiv 2024
[26]

Global Model Learning for Large Deformation Control of Elastic Deformable Linear Objects: An Efficient and Adaptive Approach,

M. Yu, K. Lv, H. Zhong, S. Song, and X. Li, “Global Model Learning for Large Deformation Control of Elastic Deformable Linear Objects: An Efficient and Adaptive Approach,”IEEE Transactions on Robotics, vol. 39, pp. 417–436, Feb. 2023

2023
[27]

Robust Deformation Model Approximation for Robotic Cable Manipulation,

S. Jin, C. Wang, and M. Tomizuka, “Robust Deformation Model Approximation for Robotic Cable Manipulation,” in2019 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pp. 6586–6593, Nov. 2019. ISSN: 2153-0866

2019
[28]

A Survey on Robotic Manipulation of Deformable Objects: Recent Advances, Open Challenges and New Frontiers,

F. Gu, Y . Zhou, Z. Wang, S. Jiang, and B. He, “A Survey on Robotic Manipulation of Deformable Objects: Recent Advances, Open Challenges and New Frontiers,” Dec. 2023. arXiv:2312.10419 [cs]

work page arXiv 2023
[29]

Learning Latent Dynamics for Planning from Pixels,

D. Hafner, T. Lillicrap, I. Fischer, R. Villegas, D. Ha, H. Lee, and J. Davidson, “Learning Latent Dynamics for Planning from Pixels,” inProceedings of the 36th International Conference on Machine Learning, pp. 2555–2565, PMLR, May 2019. ISSN: 2640-3498

2019
[30]

Reducing the Dimensionality of Data with Neural Networks,

G. E. Hinton and R. R. Salakhutdinov, “Reducing the Dimensionality of Data with Neural Networks,”Science, vol. 313, pp. 504–507, July 2006

2006
[31]

Auto-Encoding Variational Bayes

D. P. Kingma and M. Welling, “Auto-Encoding Variational Bayes,” Dec. 2022. arXiv:1312.6114 [stat]

work page internal anchor Pith review arXiv 2022
[32]

Interaction networks for learning about objects, relations and physics.arXiv:1612.00222,

P. W. Battaglia, R. Pascanu, M. Lai, D. Rezende, and K. Kavukcuoglu, “Interaction Networks for Learning about Objects, Relations and Physics,” Dec. 2016. arXiv:1612.00222 [cs]

work page arXiv 2016
[33]

Propagation Networks for Model-Based Control Under Partial Obser- vation,

Y . Li, J. Wu, J.-Y . Zhu, J. B. Tenenbaum, A. Torralba, and R. Tedrake, “Propagation Networks for Model-Based Control Under Partial Obser- vation,” Apr. 2019. arXiv:1809.11169 [cs]

work page arXiv 2019
[34]

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, pp. 5026–5033, Oct. 2012. ISSN: 2153-0866

2012
[35]

The Unknotting Problem,

L. H. Kauffman, “The Unknotting Problem,” inOpen Problems in Mathematics(J. Nash, John Forbes and M. T. Rassias, eds.), pp. 303– 345, Cham: Springer International Publishing, 2016

2016