Pith. sign in

REVIEW 3 major objections 5 minor 47 references

Let the Dynamics Flow: Stable Flow Matching Dynamical Systems

T0 review · 3 major / 5 minor · reviewed 2026-07-12 · grok-4.5

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

Pith's one-line read Flow matching can learn multimodal robot motions that stay stable by design, not just by hope.

desk verdict Coherent soft/hard flow-matching DS with Lyapunov/LaSalle constraints and Lie-group extensions; multimodal gains are real, discrete-time transfer is the honest soft spot. read the letter →

arxiv 2606.03834 v2 pith:2XQSFE6L submitted 2026-06-02 cs.RO

Rodrigo Pérez-Dattari , Francisco Leiva , Andrea Testa , Leonel Rozo , Javier Ruiz del Solar , Noémie Jaquier This is my paper
classification cs.RO
keywords flowmatchingimitationlearningdynamicalsystemsLyapunovstabilitypositiveinvarianceLiegroupsrobotmotiongenerationmultimodalpolicies
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

Robot motion policies learned by modern generative models are expressive and multimodal, but they usually lack formal stability guarantees that keep trajectories safe and goal-driven. This paper shows how to keep the generative power of flow matching while forcing the learned vector field into admissible sets defined by positive invariance and Lyapunov conditions. Two practical variants are given: a soft penalty that encourages the right velocities, and a hard architectural projection that guarantees them, both extended to Lie groups so orientations can be handled correctly. Across handwriting, torus, multimodal, pouring, humanoid insertion, and a 1682-dimensional reaction–diffusion system, the resulting systems match or beat prior stable methods on unimodal data and capture multimodal behaviors that those methods miss, while still converging from out-of-distribution starts. The practical payoff is robot motion that is both richly expressive and provably well-behaved.

What carries the argument

Stable Flow Matching Dynamical Systems (SFMDS): a state-conditioned flow-matching vector field whose auxiliary flow is forced, by LaSalle half-space or ball constraints (soft penalty or hard projection through a latent map), to land in the set of velocities that keep the system positively invariant and/or decrease a latent Lyapunov function.

What would settle it

Train hard SFMDS on a multimodal dataset or the humanoid insertion task, roll out many trajectories from a dense grid of out-of-distribution starts with the same discrete integrator used at deployment, and measure whether a non-negligible fraction fails to reach the goal or escapes the intended workspace.

Watch

Extended reading notes

Core claim

Dynamical systems can be parametrized by flow matching and simultaneously constrained—softly via a hinge penalty or hard via projection—to admissible velocity sets derived from positive invariance and latent Lyapunov decrease, including on Lie groups, so that the learned policies remain scalable and multimodal yet asymptotically stable.

Load-bearing premise

Stability is proved for continuous trajectories obtained by freezing or tracking a single noise sample, so the same guarantees need not hold automatically after discrete Euler integration or when the latent map’s Jacobian is sharp.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. The paper introduces Stable Flow Matching Dynamical Systems (SFMDS), which model robot motion policies as dynamical systems whose velocities are generated by a state-conditioned flow-matching process, while restricting those velocities to admissible sets derived from positive invariance and/or Lyapunov (LaSalle) conditions. Two variants are proposed: soft SFMDS (hinge penalties on the flow vector field) and hard SFMDS (structural projection through an invertible latent map ψ_θ). Both are extended to Lie groups (T², SE(3)) via first-cover restrictions on the Lie algebra. Training combines multi-step imitation losses with the stability constraints (Algorithms 1–2). Experiments cover LASA, a new LASA-T² benchmark, a new multimodal R² set, SE(3) pouring, a real humanoid insertion task, and a 1682-dimensional Gray–Scott system, with comparisons to CLF-DM, Euclideanizing flows, PUMA, LieImFlow, and unconstrained BC. Soft/hard SFMDS match or beat unimodal baselines and substantially improve multimodal coverage while reporting near-zero unsuccessful trajectories under OOD initial conditions.

Significance. If the constructions hold as stated, the work is a meaningful bridge between high-capacity generative imitation (flow matching) and classical stable dynamical-system policies. The multimodal gains (Table II) are particularly valuable: competing latent-Lyapunov methods remain effectively unimodal and fail to cover intersecting modes, whereas SFMDS captures them while retaining asymptotic stability. The Lie-group extension, new T² and multimodal R² benchmarks, real-robot SE(3) insertion, and 1682-D reaction–diffusion experiment strengthen the contribution beyond a pure methods paper. Public code and video support reproducibility. The free parameters (loss weights, margins, architectures) are standard for this literature; the main novelty is the admissible-set formulation of flow matching rather than a parameter-free derivation.

major comments (3)
  1. [Sec. V-B–V-C; App. A; Tables I–III] Sec. V-B (Conditions 1–2) and Sec. V-C: The formal asymptotic-stability argument is developed for continuous-time sampled trajectories obtained by fixing or tracking a single ω, then adapted to discrete forward-Euler maps via admissible balls (48)–(56). The abstract and introduction nevertheless advertise “formal stability guarantees” for robot motion generation without clearly scoping that the deployed controller is always the discrete Euler map (and, on the humanoid, a Cartesian impedance loop). App. A and residual failures in Tables I–III (e.g., 0.044% hard RealNVP on LASA, <0.03% hard NODE on SE(3), 2.3% soft on GSRD) show that discretization and sharp Jacobians of ψ_θ can still escape the admissible set. The manuscript should state a precise discrete-time theorem (what is guaranteed for the implemented map under stated step-size/margin conditions) and quantify the measure of the res
  2. [Abstract; Sec. I; Sec. V-B; Algorithm 1] Sec. V-B and Algorithm 1: Soft SFMDS is only locally asymptotically stable by construction (positive-invariance loss over a closed region plus latent hinge). Hard SFMDS claims global asymptotic stability only when ψ_θ is a diffeomorphism and continuous-time idealization holds. The abstract and contribution list do not distinguish these scopes; readers can reasonably infer global formal guarantees for both variants. Please qualify soft vs. hard claims consistently in the abstract, Sec. I, and conclusions, and report the positively invariant region used for soft SFMDS in each experiment.
  3. [Sec. VII-A3; Table II; Sec. VII-C; Table IV] Sec. VII-A3 and Table II: Hard SFMDS with RealNVP failed to learn the multimodal R² behaviors robustly, so all subsequent hard experiments use only NODE. This architectural sensitivity is load-bearing for the claim that hard structural constraints preserve expressivity: if the invertible map must be a carefully tuned NODE (with extra inner integration and Jacobian cost; Table IV ~13 Hz), the practical “hard” guarantee is narrower than the general construction suggests. Please report RealNVP multimodal numbers (even if poor), discuss when the hard projection becomes too restrictive, and clarify whether global stability is retained under the Taylor Jacobian approximations used on the real robot.
minor comments (5)
  1. [Sec. III-B1; Remark 1] Notation: u_θ is overloaded for the full augmented field and for the learnable ḣ component after Remark 1; a consistent superscript or rename would help.
  2. [Fig. 4; Sec. V-C] Fig. 4 caption and surrounding text: the overlap of X, Ẋ, and U is dense; a short legend mapping dotted/solid and blue/purple sets to equations (40)–(52) would improve readability.
  3. [Algorithm 1; free parameters] Loss weights λ_IL, λ_PI, λ_AS (and γ) and margins m̃_h, m_h are free parameters; a brief sensitivity or default schedule in the appendix would aid reproduction.
  4. [Throughout] Typos/style: “Euclideanzing” vs “Euclideanizing”; “must must be taken” (Sec. VI-B); “assesstrajectory” spacing artifacts in Sec. VII; arXiv header “arXiv:2606.03834v1” is fine for review but check final DOI/venue formatting.
  5. [Sec. I Related work] Related work: briefly position against concurrent stable diffusion/flow policy certificates (e.g., S2Diff is cited but the imitation vs. cost-optimization distinction could be one sentence sharper).

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: admissible sets and constraints are derived from standard Lyapunov/LaSalle conditions, enforced by construction on the flow-matching field, and evaluated empirically against independent accuracy and OOD stability metrics.

full rationale

The paper's core derivation (Secs. III–V) defines flow-matching dynamical systems f_θ via the conditional FM objective (Eqs. 19–24), then independently constructs admissible velocity sets ˙X_A from positive-invariance half-spaces (Eq. 36) or latent Lyapunov decrease conditions (Eqs. 40–41, Conditions 1–2). Soft penalties (Eqs. 34, 44) and hard projections (Eqs. 35, 45–46) are applied to u_θ so that h_1 lands in those sets by τ=1 (via LaSalle, Sec. IV). Training minimizes the sum of imitation losses plus these independent stability terms; success is measured by separate held-out metrics (RMSE/DTWD/FD/Chamfer on demos, % unsuccessful OOD trajectories) against external baselines (CLF-DM, Euc. Flows, PUMA, LieImFlow, BC). Self-citations (e.g., to PUMA [31], Euclideanizing flows [33], stable autonomous FM [36]) supply related methods or baselines but are not load-bearing uniqueness theorems or fitted inputs renamed as predictions. Discrete-time edge cases (App. A) are acknowledged as implementation limitations, not circular reductions. The construction is self-contained and falsifiable outside the training losses.

Assumptions & free parameters 4 free parameters · 6 assumptions · 2 invented entities

The central claim rests on standard Lyapunov/LaSalle theory, the flow-matching training objective, and the modeling choice that stability of sampled ω-conditioned trajectories implies useful stability of the generative policy. Free parameters are loss weights, margins, and network choices; invented entities are the SFMDS constructions themselves rather than new physical objects.

free parameters (4)
  • Loss weights λ_IL, λ_PI, λ_AS (and γ for discrete ball term)
    Trade off imitation accuracy against positive-invariance and asymptotic-stability penalties; chosen for training success, not derived from first principles.
  • Latent/task margins m̃_h, m_h, m_u and λ_h, λ_s (ϵ_s, α)
    Control strictness of half-space/ball constraints and finite-time convergence by τ=1; set by design heuristics from stable autonomous flow matching.
  • Network architectures and latent dimension for ψ_θ / u_θ
    MLP/NODE/RealNVP/Glow/UNet choices and capacities are free modeling decisions that affect expressivity and Jacobian invertibility.
  • Integration step sizes Δt, Δτ and single-step flow approximation
    Discrete Euler and one-step h1 approximation (Eq. 23) are implementation choices that interact with stability guarantees.
assumptions (6)
  • standard math LaSalle’s invariance principle: if H decreases along the flow outside S and is constant on S, trajectories approach S.
    Used in Sec. II-C and IV to force h_τ into admissible velocity sets by τ=1.
  • standard math Lyapunov asymptotic stability: positive-definite V with V̇<0 implies asymptotic stability of the equilibrium.
    Basis for Conditions 1–2 and admissible sets ḊL_A / Ẋ_A in Sec. V-B.
  • domain assumption Conditional flow matching regression yields the same gradients as matching the marginal probability path.
    Inherited from Lipman et al. / Liu et al.; underpins the imitation loss (Sec. II-B, III-B).
  • ad hoc to paper Stability of trajectories obtained by sampling a single ω_0 (or tracking ω) is an adequate certificate for the generative dynamical system used at deployment.
    Sec. V-B explicitly replaces non-differentiable distributional dynamics with sampled ODEs for Lyapunov analysis.
  • domain assumption For hard SFMDS, ψ_θ is a diffeomorphism (invertible Jacobian) so latent half-space/ball constraints map back to task space without losing Cond. 1.
    Required in Sec. V-B hard variant and Lie-group first-cover constructions (Sec. VI).
  • domain assumption Restricting Lie-algebra maps to the first cover (e.g., ∥ξ^r∥<π) preserves bijectivity of log/exp for stability on T² and SE(3).
    Sec. VI-B; standard but load-bearing for hard SFMDS on orientations.
invented entities (2)
  • Stable Flow Matching Dynamical Systems (SFMDS)
    purpose: Name the combined generative dynamical system with soft/hard stability constraints.
    Framework label for the method; not an external physical entity. Independent evidence is empirical performance, not a separate measurement channel.
  • Admissible velocity / u-space half-spaces and discrete balls Ẋ_A, U_A, ḊL_A, Ũ^B_A
    purpose: Encode positive invariance and Lyapunov decrease as geometric constraints on flow-matching velocities.
    Derived from standard stability conditions but packaged as the paper’s constraint geometry (Sec. IV–V).

how reviews work

0 comments
Cite this review

Pith. "Pith review of Let the Dynamics Flow: Stable Flow Matching Dynamical Systems." pith.science (2026). https://pith.science/paper/2XQSFE6L

@misc{pith2026260603834,
  author       = {Pith},
  title        = {Pith review of: Let the Dynamics Flow: Stable Flow Matching Dynamical Systems},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/2XQSFE6L}},
  note         = {Machine review of arXiv:2606.03834}
}
read the original abstract

Flow matching has recently emerged as a powerful approach for imitation learning, enabling scalable, expressive, and multimodal motion policies. However, when modeling these policies as dynamical systems, incorporating formal stability guarantees into these generative models is a prerequisite to ensure safe and generalizable robot behaviors, which remains a significant challenge. This paper introduces Stable Flow Matching Dynamical Systems (SFMDS), a novel framework that bridges the gap between highly expressive generative modeling and formal stability guarantees. SFMDS parametrizes dynamical systems via flow matching while constraining the model to satisfy positive invariance and/or Lyapunov stability conditions. We propose two variants: a soft constraint based on a penalty term, and a hard structural constraint embedded directly into the model architecture. We further extend both formulations to Lie groups to robustly handle orientation trajectories. Experiments on benchmark datasets, in simulation, and on a humanoid robot show that SFMDS learns stable, scalable, and multimodal dynamical systems in low- and high-dimensional state spaces, enabling safe and expressive robot motion generation. SFMDS matches or outperforms state-of-the-art methods on unimodal datasets, while substantially improving performance on multimodal datasets, where competing approaches fail to capture multi-modal behaviors. Accompanying source code and video are available at: https://let-the-dynamics-flow.github.io/SFMDS/.

Figures

Figures reproduced from arXiv: 2606.03834 by the authors.

Figure 1
Figure 1. Robot behavior modeled as a stable flow matching dynamical system [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. (b). This can be expressed as satisfying d [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. 2D illustration of trajectory inference for soft/hard SFMDS. Each [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figures from the paper (11 more)
Figure 4
Figure 4. Figure 4: Overlap between X, X˙ , and U. Dotted lines denote the admissible half-spaces, while solid lines denote the admissible balls, whose depiction illustrates their effect in X. The admissible sets in velocity space are shown in blue, and those in U-space are shown in purpl…
Figure 5
Figure 5. Figure 5: Vector fields of learned soft (left) and hard (right) SFMDS on LASA. [PITH_FULL_IMAGE:figures/full_fig_p011_5.png]
Figure 6
Figure 6. Figure 6: Vector fields of learned soft (left) and hard (right) SFMDS on LASA [PITH_FULL_IMAGE:figures/full_fig_p012_6.png]
Figure 7
Figure 7. Figure 7: Vector fields of learned soft (top) and hard (bottom) SFMDS in the Multimodal [PITH_FULL_IMAGE:figures/full_fig_p013_7.png]
Figure 8
Figure 8. Figure 8: Learned SFMDS on the water-pouring task in [PITH_FULL_IMAGE:figures/full_fig_p013_8.png]
Figure 9
Figure 9. Figure 9: Sequences of robot poses along the hard SFMDS trajectories on the [PITH_FULL_IMAGE:figures/full_fig_p014_9.png]
Figure 10
Figure 10. Figure 10: Ground-truth (GT) and generated trajectories via Soft and Hard SFMDS associated with two different demonstrations ((a) and (b), respectively). [PITH_FULL_IMAGE:figures/full_fig_p015_10.png]
Figure 11
Figure 11. Figure 11: Failure edge case caused by discretizing ˙ [PITH_FULL_IMAGE:figures/full_fig_p017_11.png]
Figure 12
Figure 12. Figure 12: Example of ground-truth trajectory for the [PITH_FULL_IMAGE:figures/full_fig_p019_12.png]
Figure 13
Figure 13. Figure 13: Diagram of the UNet-like architecture used to parameterize [PITH_FULL_IMAGE:figures/full_fig_p019_13.png]
Figure 14
Figure 14. Figure 14: Examples of trajectories generated by Soft and Hard SFMDS when starting from OOD initial states. Out-of-bound pixels are white if below the [PITH_FULL_IMAGE:figures/full_fig_p019_14.png]

Review history (2 revisions) →

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

47 extracted references · 3 linked inside Pith

  1. [1]

    Parametric correspondence and cham- fer matching: Two new techniques for image matching

    Harry G Barrow, Jay M Tenenbaum, Robert C Bolles, and Helen C Wolf. Parametric correspondence and cham- fer matching: Two new techniques for image matching. Technical report, 1977

    1977

  2. [2]

    Learning for adaptive and reactive robot control: a dynamical systems approach

    Aude Billard, Sina Mirrazavi, and Nadia Figueroa. Learning for adaptive and reactive robot control: a dynamical systems approach. Mit Press, 2022

    2022

  3. [3]

    Cambridge university press, 2004

    Stephen Boyd and Lieven Vandenberghe.Convex opti- mization. Cambridge university press, 2004

    2004

  4. [4]

    Riemannian flow matching policy for robot motion learning

    Max Braun, No ´emie Jaquier, Leonel Rozo, and Tamim Asfour. Riemannian flow matching policy for robot motion learning. InIEEE/RSJ Intl. Conf. on Intelligent Robots and Systems (IROS), pages 5144–5151, 2024

    2024

  5. [5]

    Neural ordinary differential equations.Advances in neural information processing systems, 31, 2018

    Ricky TQ Chen, Yulia Rubanova, Jesse Bettencourt, and David K Duvenaud. Neural ordinary differential equations.Advances in neural information processing systems, 31, 2018. 15 (a) (b) Fig. 10. Ground-truth (GT) and generated trajectories via Soft and Hard SFMDS associated with two different demonstrations ((a) and (b), respectively)

    2018

  6. [6]

    Safe and stable control via Lyapunov-guided diffusion models

    Xiaoyuan Cheng, Xiaohang Tang, and Yiming Yang. Safe and stable control via Lyapunov-guided diffusion models. InNeural Information Processing Systems (NeurIPS), 2025

    2025

  7. [7]

    Learn- ing robotic manipulation policies from point clouds with conditional flow matching

    Eugenio Chisari, Nick Heppert, Max Argus, Tim Welschehold, Thomas Brox, and Abhinav Valada. Learn- ing robotic manipulation policies from point clouds with conditional flow matching. InConference on Robot Learning (CoRL), 2024

    2024

  8. [8]

    Fast and robust visuomotor riemannian flow matching policy.IEEE Transactions on Robotics, 41: 5327–5343, 2025

    Haoran Ding, No ´emie Jaquier, Jan Peters, and Leonel Rozo. Fast and robust visuomotor riemannian flow matching policy.IEEE Transactions on Robotics, 41: 5327–5343, 2025

    2025

Show all 47 references
  1. [9]

    Density estimation using real nvp

    Laurent Dinh, Jascha Sohl-Dickstein, and Samy Bengio. Density estimation using real nvp. InInternational Conference on Learning Representations, 2017

    2017

  2. [10]

    Fast and stable learning of dynamical systems based on extreme learning machine.IEEE Transactions on Systems, Man, and Cybernetics: Systems, 49(6):1175–1185, 2017

    Jianghua Duan, Yongsheng Ou, Jianbing Hu, Zhiyang Wang, Shaokun Jin, and Chao Xu. Fast and stable learning of dynamical systems based on extreme learning machine.IEEE Transactions on Systems, Man, and Cybernetics: Systems, 49(6):1175–1185, 2017

    2017

  3. [11]

    Computing discrete Fr´echet distance

    Thomas Eiter and Heikki Mannila. Computing discrete Fr´echet distance. Technical Report CD-TR 94/64, Tech- nische Universitat Wien, 1994

    1994

  4. [12]

    Action- Flow: Equivariant, accurate, and efficient policies with spatially symmetric flow matching

    Niklas Funk, Julen Urain, Joao Carvalho, Vignesh Prasad, Georgia Chalvatzaki, and Jan Peters. Action- Flow: Equivariant, accurate, and efficient policies with spatially symmetric flow matching. 2024

    2024

  5. [13]

    Mohammad Khansari-Zadeh and Aude Billard

    S. Mohammad Khansari-Zadeh and Aude Billard. Learn- ing stable nonlinear dynamical systems with gaussian mixture models.IEEE Transactions on Robotics, 27(5): 943–957, 2011

    2011

  6. [14]

    Mohammad Khansari-Zadeh and Aude Billard

    S. Mohammad Khansari-Zadeh and Aude Billard. Learn- ing control lyapunov function to ensure stability of dy- namical system-based robot reaching motions.Robotics and Autonomous Systems, 62(6):752–765, 2014

    2014

  7. [15]

    Glow: Generative flow with invertible 1x1 convolutions.Advances in neural information processing systems, 31, 2018

    Durk P Kingma and Prafulla Dhariwal. Glow: Generative flow with invertible 1x1 convolutions.Advances in neural information processing systems, 31, 2018

    2018

  8. [16]

    Normalizing flows: An introduction and review of current methods.IEEE Transactions on Pattern Analysis and Machine Intelligence, 43(11):3964–3979, 2020

    Ivan Kobyzev, Simon JD Prince, and Marcus A Brubaker. Normalizing flows: An introduction and review of current methods.IEEE Transactions on Pattern Analysis and Machine Intelligence, 43(11):3964–3979, 2020

    2020

  9. [17]

    Learning stable deep dynamics models.Advances in neural information processing systems, 32, 2019

    J Zico Kolter and Gaurav Manek. Learning stable deep dynamics models.Advances in neural information processing systems, 32, 2019

    2019

  10. [18]

    An invariance principle in the theory of stability

    Joseph P La Salle. An invariance principle in the theory of stability. Technical report, 1966

    1966

  11. [19]

    J. M. Lee.Introduction to smooth manifolds. Springer, 2013

    2013

  12. [20]

    Smooth manifolds

    John M Lee. Smooth manifolds. InIntroduction to smooth manifolds, pages 1–29. Springer, 2003

    2003

  13. [21]

    Mmp++: Motion manifold primitives with parametric curve models.IEEE Transactions on Robotics, 2024

    Yonghyeon Lee. Mmp++: Motion manifold primitives with parametric curve models.IEEE Transactions on Robotics, 2024

    2024

  14. [22]

    Neural learning of vector fields for encoding stable dynamical systems.Neurocomputing, 141:3–14, 2014

    Andre Lemme, Klaus Neumann, Ren ´e Felix Reinhart, and Jochen J Steil. Neural learning of vector fields for encoding stable dynamical systems.Neurocomputing, 141:3–14, 2014

    2014

  15. [23]

    Yaron Lipman, Ricky T. Q. Chen, Heli Ben-Hamu, Maximilian Nickel, and Matthew Le. Flow matching for generative modeling. InIntl. Conf. on Learning Representations (ICLR), 2023

    2023

  16. [24]

    Flow straight and fast: Learning to generate and transfer data with rectified flow

    Xingchao Liu, Chengyue Gong, et al. Flow straight and fast: Learning to generate and transfer data with rectified flow. InIntl. Conf. on Learning Representations (ICLR), 16 2023

    2023

  17. [25]

    Neural contractive dynamical systems

    Hadi Beik Mohammadi, Søren Hauberg, Georgios Ar- vanitidis, Nadia Figueroa, Gerhard Neumann, and Leonel Rozo. Neural contractive dynamical systems. InThe Twelfth International Conference on Learning Represen- tations, 2023

    2023

  18. [26]

    Dynamic time warping.Information retrieval for music and motion, pages 69–84, 2007

    Meinard M ¨uller. Dynamic time warping.Information retrieval for music and motion, pages 69–84, 2007

    2007

  19. [27]

    Springer Science & Business Media, 2013

    Bernt Oksendal.Stochastic differential equations: an introduction with applications. Springer Science & Business Media, 2013

    2013

  20. [28]

    Normalizing flows for probabilistic modeling and inference.Journal of Machine Learning Research, 22(57):1–64, 2021

    George Papamakarios, Eric Nalisnick, Danilo Jimenez Rezende, Shakir Mohamed, and Balaji Lakshmi- narayanan. Normalizing flows for probabilistic modeling and inference.Journal of Machine Learning Research, 22(57):1–64, 2021

    2021

  21. [29]

    Complex patterns in a simple system

    John E Pearson. Complex patterns in a simple system. Science, 261(5118):189–192, 1993

    1993

  22. [30]

    Stable motion primitives via imitation and contrastive learning.IEEE Transactions on Robotics, 39(5):3909–3928, 2023

    Rodrigo P ´erez-Dattari and Jens Kober. Stable motion primitives via imitation and contrastive learning.IEEE Transactions on Robotics, 39(5):3909–3928, 2023

    2023

  23. [31]

    Puma: Deep metric imitation learning for stable motion primitives.Advanced Intelligent Systems, 6(11): 2400144, 2024

    Rodrigo P ´erez-Dattari, Cosimo Della Santina, and Jens Kober. Puma: Deep metric imitation learning for stable motion primitives.Advanced Intelligent Systems, 6(11): 2400144, 2024

    2024

  24. [32]

    Fast diffeomorphic matching to learn globally asymptotically stable nonlinear dynamical systems.Systems & Control Letters, 96:51–59, 2016

    Nicolas Perrin and Philipp Schlehuber-Caissier. Fast diffeomorphic matching to learn globally asymptotically stable nonlinear dynamical systems.Systems & Control Letters, 96:51–59, 2016

    2016

  25. [33]

    Euclideaniz- ing flows: Diffeomorphic reduction for learning stable dynamical systems

    Muhammad Asif Rana, Anqi Li, Dieter Fox, Byron Boots, Fabio Ramos, and Nathan Ratliff. Euclideaniz- ing flows: Diffeomorphic reduction for learning stable dynamical systems. InConference on Learning for Dynamics and Control (L4DC), pages 630–639, 2020

    2020

  26. [34]

    U-net: Convolutional networks for biomedical image segmentation

    Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. InInternational Conference on Medical image computing and computer-assisted intervention, pages 234–241. Springer, 2015

    2015

  27. [35]

    A micro lie theory for state estimation in robotics.arXiv preprint arXiv:1812.01537, 2018

    Joan Sola, Jeremie Deray, and Dinesh Atchuthan. A micro lie theory for state estimation in robotics.arXiv preprint arXiv:1812.01537, 2018

    2018 arXiv

  28. [36]

    Stable autonomous flow matching.arXiv preprint arXiv:2402.05774, 2024

    Christopher Iliffe Sprague, Arne Elofsson, and Hossein Azizpour. Stable autonomous flow matching.arXiv preprint arXiv:2402.05774, 2024

    2024 arXiv

  29. [37]

    Imitationflow: Learning deep stable stochastic dynamic systems by normalizing flows

    Julen Urain, Michele Ginesi, Davide Tateo, and Jan Peters. Imitationflow: Learning deep stable stochastic dynamic systems by normalizing flows. InIEEE/RSJ Intl. Conf. on Intelligent Robots and Systems (IROS), pages 5231–5237, 2020

    2020

  30. [38]

    Learning stable vector fields on lie groups.IEEE Robotics and Automation Letters, 7(4):12569–12576, 2022

    Julen Urain, Davide Tateo, and Jan Peters. Learning stable vector fields on lie groups.IEEE Robotics and Automation Letters, 7(4):12569–12576, 2022

    2022

  31. [39]

    Attention is all you need.Advances in neural information processing systems, 30, 2017

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Łukasz Kaiser, and Illia Polosukhin. Attention is all you need.Advances in neural information processing systems, 30, 2017

    2017

  32. [40]

    Neural geometric fabrics: Efficiently learning high-dimensional policies from demonstration

    Mandy Xie, Ankur Handa, Stephen Tyree, Dieter Fox, Harish Ravichandar, Nathan D Ratliff, and Karl Van Wyk. Neural geometric fabrics: Efficiently learning high-dimensional policies from demonstration. InCon- ference on Robot Learning, pages 1355–1367. PMLR, 2023

    2023

  33. [41]

    Mani- Flow: A general robot manipulation policy via consis- tency flow training.arXiv preprint arXiv:2509.01819, 2025

    Ge Yan, Jiyue Zhu, Yuquan Deng, Shiqi Yang, Ri-Zhao Qiu, Xuxin Cheng, Marius Memmel, Ranjay Krishna, Ankit Goyal, Xiaolong Wang, and Dieter Fox. Mani- Flow: A general robot manipulation policy via consis- tency flow training.arXiv preprint arXiv:2509.01819, 2025

    2025 arXiv

  34. [42]

    Learning riemannian stable dynamical systems via diffeomorphisms

    Jiechao Zhang, Hadi Beik-Mohammadi, and Leonel Rozo. Learning riemannian stable dynamical systems via diffeomorphisms. InConference on Robot Learning (CoRL), 2022

    2022

  35. [43]

    Dif- feomorphic transforms for generalised imitation learning

    Weiming Zhi, Tin Lai, Lionel Ott, and Fabio Ramos. Dif- feomorphic transforms for generalised imitation learning. InL4DC, Proceedings of Machine Learning Research, pages 508–519. PMLR, 2022. APPENDIXA EDGECASES FORHARD-CONSTRAINEDMETHODS As discussed in Secs. V-C and VII-A, a ...

    2022

  36. [44]

    However, the same approach can be employed in the case when ˙XA(xt;θ)is a ball

    Escaping ˙XA:Next, we consider the case when ˙XA(xt;θ)is a half-space. However, the same approach can be employed in the case when ˙XA(xt;θ)is a ball. When hτ ∈ ˙XA(xt;θ), the ReLU-based projection in (29) leaves the vector fieldu θ unconstrained. In such cases, ifu θ induces ...

  37. [45]

    In this case, the latent dynamics may remain stable, while the corresponding deformation induced byJ −1 ψθ changes too abruptly in task space

    Overfitting ofψ θ:In cases where the transformation ψθ varies too sharply, for instance, due to overfitting, the discretized dynamics may violate the stability conditions even when the admissible set is defined as a ball in the latent space. In this case, the latent dynamics m...

  38. [46]

    SinceS 1 is isomorphic to the set of unit complex numbers{e iθ |θ∈R} ⊂C, elements on the torus can be written asx t = (eiθ1,t , eiθ2,t)∈T 2 ⊂C 2 [20, Ch

    TorusT 2:The2-torus is identified as the Lie group T2 =S 1 × S1. SinceS 1 is isomorphic to the set of unit complex numbers{e iθ |θ∈R} ⊂C, elements on the torus can be written asx t = (eiθ1,t , eiθ2,t)∈T 2 ⊂C 2 [20, Ch. 7]. Its Lie algebra ist 2 ∼= R2, whose elements can be wri...

  39. [47]

    patternλ

    Special Euclidean groupSE(3):Each element ofSE(3) is a pose xt = Rt pt 0 1 ∈SE(3), whereR t ∈SO(3)andp t ∈R 3 denote the rotation and translation components, respectively. A Lie algebra element is written as ξt = [ωt]× vt 0⊤ 0 ∈se(3).(70) The exponential map is defined as [35]...

Pith tools

Reviewed July 12, 2026 · model on record in the stance chip above.