Pith. sign in

REVIEW 4 major objections 6 minor 53 references

Continuous-Time SO(3) Forecasting with Savitzky--Golay Neural Controlled Differential Equations

T0 review · 4 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read SG-neural CDE forecasts SO(3) rotations more accurately than existing baselines on real-world data.

desk verdict Useful idea, honest results, but the control path is not formally well-posed as written and the evaluation is too thin for the claims. read the letter →

arxiv 2506.06780 v1 pith:S46NDOPS submitted 2025-06-07 cs.CV cs.LG

Lennart Bastian , Mohammad Rashed , Nassir Navab , Tolga Birdal This is my paper
classification cs.CVcs.LG
keywords SO(3)forecastingneuralcontrolleddifferentialequationsSavitzky-Golayfilteringrotationmanifoldcontinuous-timedynamicsobjecttrackingsensorfusionrepresentation
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

Forecasting how an object rotates from noisy, sparse sensor readings is hard because rotations live on the curved manifold SO(3), where common assumptions like constant angular velocity fail. This paper proposes SG-nCDE, a neural controlled differential equation whose integration path is built by Savitzky-Golay regression directly on SO(3), with learned weights that make the path noise-resistant. The model learns a continuous-time latent dynamical system from observed rotations and extrapolates it forward, decoding predictions back to rotation matrices through a 6D representation and orthonormalization. On a motion-capture dataset and a camera-plus-IMU sensor-fusion tablet, the method reports lower mean rotational geodesic error than two geometry-aware baselines (a gated recurrent unit and a spline-driven neural CDE) across all tested scenarios. The point of the method is to replace restrictive motion priors with a learned, manifold-respecting dynamics model that works with irregular, noisy input.

What carries the argument

The load-bearing object is the weighted Savitzky-Golay control path on SO(3). A second-order polynomial $p(t; \rho) = \rho_0 + \rho_1 t + \frac{1}{2} \rho_2 t^2$ in the Lie algebra $\mathfrak{so}(3)$ is fit to noisy rotations inside a sliding window by minimizing $\sum_m \|\mathrm{Log}(\tilde{x}_{k+m} \tilde{x}_k^{-1}) - p(t_{k+m} - t_k; \rho)\|^2$, which reduces to the linear system $\rho_k = (A^\top W A)^{-1} A^\top W b$ with learnable weights $W$. Exponentiating this polynomial and applying it to the anchor rotation gives $\varphi(t) = \mathrm{Exp}(p(t - t_k; \rho_k)) \tilde{x}_k$, a smooth, manifold-respecting signal used as the control path in the neural CDE $z_t = z_{t_0} + \int_{t_0}^t f_\theta(z_s)\,dX_s$. The network $f_\theta$ and the weights $W$ are trained end-to-end by minimizing the Frobenius norm between predicted and ground-truth rotation matrices, with predictions recovered via the 6D representation and orthonormalization. This machinery replaces explicit motion priors (constant velocity, minimal torque) with a learned latent dynamical system, while staying cheap enough to differentiate through during training.

What would settle it

Train the same SG-nCDE architecture on real rotational trajectories (or on simulations whose torque and noise distributions deliberately diverge from the training set) and compare RGE to the reported numbers; if performance degrades to the level of the spline-CDE or GRU baselines when dynamics fall outside the simulated family, the claim that the learned model generalizes across real-world scenarios would be falsified. More directly, recording a real object under violent, non-smooth motion such as an impact and checking whether the method's forecasts remain below the baselines would test robustness under dynamics the simulator never covered.

Watch

Extended reading notes

Core claim

The paper's central claim is that a neural controlled differential equation can accurately forecast SO(3) trajectories from noisy pose observations if its control path is constructed by a (learned) Savitzky-Golay filter operating on the rotation manifold, rather than by cubic or Hermite splines that ignore the manifold and amplify noise. Concretely, the control path is $\varphi(t) = \mathrm{Exp}(p(t - t_k; \rho_k)) \tilde{x}_k$, where $p$ is a second-order polynomial in the Lie algebra and $\rho_k$ minimizes a weighted least-squares fit of log-differences inside a sliding window; the weights are learnable. Integrating the hidden state against this path yields a continuous-time latent dynamics that is then projected to a 6D rotation representation and orthonormalized back to SO(3). The authors validate the claim on two real-world datasets, reporting mean rotational geodesic errors of 2.32/2.30/2.18 degrees across three motion scenarios (versus 3.04/2.90/2.83 for the GRU baseline and 6.49/5.43/7.82 for the spline-based CDE) and 8.11 degrees on an irregularly sampled camera-IMU fusion scenario (versus 11.39 and 75.48).

Load-bearing premise

The method's learned filter weights and neural dynamics are trained only on simulated rigid-body trajectories (free rotation, linear control, configuration-dependent torque, damped motion) and then applied unchanged to real sensor data; the paper does not analyze how much the simulation-to-real distribution shift affects the reported gains.

Editorial extensions

If this is right

  • On the motion-capture benchmark, SG-nCDE reports mean RGE of 2.32/2.30/2.18 degrees across unconstrained, multi-object, and static camera-motion scenarios, outperforming the SO(3)-GRU baseline (3.04/2.90/2.83) and the SO(3)-nCDE baseline (6.49/5.43/7.82).
  • On the irregularly sampled camera-IMU sensor-fusion tablet, SG-nCDE reports 8.11 degrees mean RGE versus 11.39 for SO(3)-GRU and 75.48 for SO(3)-nCDE, indicating resilience to multi-rate noisy inputs.
  • Because the control path is built from a closed-form weighted least-squares solution, the method can be differentiated through during training, unlike spline-based extrapolation that requires expensive manifold optimization.
  • The learned latent dynamics avoid explicit motion assumptions such as constant angular velocity or conserved angular momentum, so the method is intended to cover more complex force-driven rotations than earlier approaches.
  • The same 6D projection and orthonormalization used for SO(3) predictions means the model can be plugged into downstream 6D tracking pipelines that already use these rotation representations.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • We infer that the learned Savitzky-Golay weights could be interpreted as a data-driven noise-resistance profile: inspecting which window positions receive high weight might reveal whether the model learns to down-weight outliers or to emphasize recent motion, a testable hypothesis on synthetic noise.
  • We infer that the continuous-time formulation naturally supports forecasting at arbitrary time horizons and asynchronous sensor rates, so an online variant in which the control path is updated incrementally as new frames arrive is a plausible extension.
  • We infer that the same weighted SG control path could be applied to SE(3) by treating translations and rotations separately or working in the Lie algebra of twists, giving a direct route toward the full 6D pose forecasting the paper lists as future work.
  • We infer that if the learned weights transfer across datasets, the method could serve as a drop-in preprocessor that turns noisy rotation streams into smooth CDE control paths, independent of the downstream dynamics model.
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

4 major / 6 minor

Summary. The paper proposes SG-nCDE, a continuous-time model for forecasting rotations in SO(3). It constructs a control path by fitting a degree-two polynomial in the Lie algebra so(3) to a sliding window of noisy rotation measurements, with learnable Savitzky-Golay weights, and uses this path as the driving signal of a neural controlled differential equation. The model is trained on simulated rigid-body trajectories and evaluated without adaptation on the Oxford Motion Dataset and on a camera-IMU sensor-fusion tablet setup. The reported results show lower mean rotational geodesic error for SG-nCDE than SO(3)-GRU and SO(3)-nCDE in both settings.

Significance. The geometric construction is a reasonable and potentially useful contribution: it avoids Euclidean interpolation on rotation matrices, respects SO(3) structure, and the learnable weighting in Lemma 8 is a novel twist on Savitzky-Golay filtering. The idea of using a manifold-adapted, locally polynomial control path for neural CDEs is worth pursuing, and the two real-data experiments speak to practical applicability. The main weaknesses are formal and empirical: the control path is not defined as a continuous or causal global function, and the evaluation is too thin to establish the headline claim of consistent superiority.

major comments (4)
  1. [Sec. 4.2, Defs. 6-7, Eqs. (5)-(6)] Definition 6 defines phi(t) only as a local expression anchored at t_k, and Definition 7 computes rho_k from a window centered at t_k. The manuscript never states how these local polynomials are stitched into a global control path. Under the natural left-continuous interpretation on [t_k, t_{k+1}), phi jumps at every observation time because Exp(p(t_{k+1}-t_k; rho_k)) x_tilde_k is not generally equal to x_tilde_{k+1}. The neural CDE in Eq. (3) requires a continuous or at least bounded-variation control path, so the integral and the Dormand-Prince solver described in Sec. 5 are not well-defined for the path as written. Moreover, at the forecast boundary t_N, Eq. (6) requires observations t_{N+1},...,t_{N+n} that are not available at inference; no causal or one-sided variant is specified. Since Tables 1 and 2 report results for the method as implemented, the implementation must have used an undocumented interpolant or boundary handling. This gap must be closed before the empirical results can be attributed to the proposed SG-nCDE.
  2. [Sec. 4.3, paragraph 'Learning Rotational Kinematics'] The sentence 'we use the 9D rotational derivatives Rdot = R omega from Sec. 4.2 as integration paths for the latent state in Eq. (3)' is internally inconsistent. Equation (3) is integrated with respect to the control path X, not with respect to Rdot; Rdot = R omega is a 3x3 (skew-symmetric) object, not a 9D representation; and Sec. 4.2 does not define Rdot. The authors should state explicitly whether the control path X_t is the 9D coordinate vector of phi(t), or whether some other construction is intended. Without this, the exact input to the CDE integrator is ambiguous.
  3. [Sec. 5.1, Tables 1-2] The headline claim of consistent improvement rests on mean RGE differences of about 0.6-0.7 degrees on OMD and 3.3 degrees on the tablet sequence, with reported standard deviations of order 1 degree on OMD and 1.6 degrees for SG-nCDE on the tablet. The paper reports no ablations of the window size n, polynomial order p, or learnable weights W, no repeated-seed variance, and no significance tests. The tablet experiment appears to be a single sequence, so the comparison could be within run-to-run variability. Additional experiments or at least variance estimates are needed to support 'consistently outperforms.'
  4. [Appendix A.1 and Sec. 5.1] The models are trained exclusively on simulated rigid-body trajectories (free rotation, linear control, configuration-dependent torque, damped motion) and then applied unchanged to real Vicon and camera-IMU data. The manuscript does not analyze distribution shift in torque profiles, noise levels, sampling irregularity, or sensor errors between simulation and real data. Since the simulation is the only training signal, the real-world transfer claim is under-supported; a comparison of the simulation noise/torque ranges with the real recordings, or a sensitivity analysis over those ranges, would make the generalization claim more credible.
minor comments (6)
  1. [Abstract / Introduction] The name 'Savitzy' appears in the abstract and introduction; it should be 'Savitzky'.
  2. [Sec. 5.1] The heading 'Emperical Evaluation' should be 'Empirical Evaluation'.
  3. [Definition 5] In Eq. (4), p(t;rho) is written as an element of so(3) while rho is said to be in R^9; clarify that rho_0, rho_1, rho_2 are vectors that are mapped through the hat operator, or use the vee convention consistently.
  4. [Definition 7 / Lemma 8] The learnable weight matrix W is not specified. If W is unconstrained, the weighted least-squares solution (A^T W A)^{-1} in Eq. (8) may not be positive definite or invertible; state the parametrization and any constraints on W.
  5. [Appendix B] 'Hermit cubic' should be 'Hermite cubic' and 'Kroeneker' should be 'Kronecker'.
  6. [Sec. 5 / Appendix B] The numerical values of the SG window half-size n and polynomial order p used in the experiments are not reported; reporting them is important for reproducibility.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the empirical forecast comparison is evaluated on held-out real-world data, and no equation reduces the prediction to its input by construction.

full rationale

The paper's central claim is an empirical accuracy comparison on real-world datasets (Tables 1 and 2). The learnable parameters — the SG filter weights W in Lemma 8 and the neural vector field f_theta in Definition 4 — are optimized on simulated rigid-body trajectories via the supervised forecasting loss in Eq. (10), while the reported RGE numbers are measured on held-out real Vicon (OMD) and camera-IMU (tablet) data. No equation in the paper defines the forecast y_hat_k in terms of the ground-truth future rotations x_{k+1..k+m} at inference time, and no fitted parameter is renamed as a prediction. The SG control path in Definition 6/7 is a preprocessing function of the observed input rotations (and, when weighted, a learned but input-dependent filter), and the extrapolation is produced by integrating a trained vector field forward; the output is therefore not equivalent to the input by construction. The one substantive concern in the provided skeptical analysis — that the centered Savitzky-Golay window in Definition 7 includes future points at the forecast boundary t_N — is a potential causality and well-posedness gap in the method description, not a circular derivation: it would make the control path ill-defined as written or require an undocumented one-sided variant, but it does not exhibit an equation of the paper making the prediction equal to its own input. Self-citations in the related-work and background sections are present but not load-bearing; the cited SG filter [19] and Neural CDE [21] are external, parameter-free references with stated assumptions, and the claimed contribution is the new combination and its empirical evaluation. For these reasons the derivation chain is self-contained against external benchmarks and no circular step meets the evidentiary bar.

Assumptions & free parameters 3 free parameters · 3 assumptions · 0 invented entities

The method rests on design choices (fixed polynomial order, unreported window size, learned weighting) and on the assumption that simulation-trained models transfer to real sensors. No new physical entities are introduced; the learned weights and networks are ordinary trainable parameters.

free parameters (3)
  • SG window half-size n = not reported
    Controls the smoothing window (2n+1) in Eq. (6); the value is not stated in the paper or appendix, and no sensitivity analysis is provided.
  • Learnable SG weights W = learned end-to-end
    Lemma (8) introduces a weight matrix W optimized by gradient descent on the forecasting objective, making the control path task-specific rather than a fixed robust filter.
  • Polynomial order p = 2
    Definition 5 fixes a second-order Lie algebra polynomial; no comparison with p=1 or p=3 is reported.
assumptions (3)
  • ad hoc to paper A second-order Lie algebra polynomial, mapped back via Exp, is an adequate control path for the latent dynamics over the forecast horizon.
    Definition 5 and Eq. (5); the order is fixed without ablation or theoretical justification.
  • domain assumption The dynamics f_theta and filter weights W trained on simulated rigid-body motions generalize to real-world sensor data without adaptation.
    Section 5 states models are trained in simulation and applied to real data; no distribution shift analysis is included.
  • standard math The Frobenius norm in the learning objective is a suitable training proxy for the geodesic error used at evaluation.
    Definition 9 versus Eq. (11); the Frobenius norm of the rotation difference is monotonic in the angle, so this is a mild assumption.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Continuous-Time SO(3) Forecasting with Savitzky--Golay Neural Controlled Differential Equations." pith.science (2026). https://pith.science/paper/S46NDOPS

@misc{pith2026250606780,
  author       = {Pith},
  title        = {Pith review of: Continuous-Time SO(3) Forecasting with Savitzky--Golay Neural Controlled Differential Equations},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/S46NDOPS}},
  note         = {Machine review of arXiv:2506.06780}
}
abstract

Tracking and forecasting the rotation of objects is fundamental in computer vision and robotics, yet SO(3) extrapolation remains challenging as (1) sensor observations can be noisy and sparse, (2) motion patterns can be governed by complex dynamics, and (3) application settings can demand long-term forecasting. This work proposes modeling continuous-time rotational object dynamics on $SO(3)$ using Neural Controlled Differential Equations guided by Savitzky-Golay paths. Unlike existing methods that rely on simplified motion assumptions, our method learns a general latent dynamical system of the underlying object trajectory while respecting the geometric structure of rotations. Experimental results on real-world data demonstrate compelling forecasting capabilities compared to existing approaches.

Figures

Figures reproduced from arXiv: 2506.06780 by the authors.

Figure 1
Figure 1. We propose Savitzy Golay neural Controlled Differ [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Qualitative comparison of the forecasting results produced by our method (top row) and [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Evaluating the robustness of the proposed extrapolation [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Comparison of interpolation methods on a particularly complex and noisy trajectory from the [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

53 extracted references · 49 canonical work pages

  1. [1]

    Ceres Solver, 2023

    Sameer Agarwal, Keir Mierle, and The Ceres Solver Team. Ceres Solver, 2023. 2

    work page 2023

  2. [2]

    Borisov and I.S

    A. Borisov and I.S. Mamaev.Rigid Body Dynamics. De Gruyter, 2018. 1

    work page 2018

  3. [3]

    Continuous 3d scan-matching with a spinning 2d laser

    Michael Bosse and Robert Zlot. Continuous 3d scan-matching with a spinning 2d laser. InInt. Conf. Robot. Automation., pages 4312–4319. IEEE, 2009. 1

    work page 2009

  4. [4]

    Deep regression on manifolds: a 3d rotation case study

    Romain Br´egier. Deep regression on manifolds: a 3d rotation case study. InInt. Conf. 3D Vision, pages 166–174. IEEE,

    work page

  5. [5]

    Cam- era pose filtering with local regression geodesics on the rie- mannian manifold of dual quaternions

    Benjamin Busam, Tolga Birdal, and Nassir Navab. Cam- era pose filtering with local regression geodesics on the rie- mannian manifold of dual quaternions. InICCVW, pages 2436–2445, 2017. 2

    work page 2017

  6. [6]

    Se3-nets: Learning rigid body motion using deep neural networks

    Arunkumar Byravan and Dieter Fox. Se3-nets: Learning rigid body motion using deep neural networks. InInt. Conf. Robot. Automation., pages 173–180. IEEE, 2017. 2

    work page 2017

  7. [7]

    Projective manifold gra- dient layer for deep rotation regression

    Jiayi Chen, Yingda Yin, Tolga Birdal, Baoquan Chen, Leonidas J Guibas, and He Wang. Projective manifold gra- dient layer for deep rotation regression. InCVPR, pages 6646–6655, 2022. 2

    work page 2022

  8. [8]

    Ricky T.Q. Chen. torchdiffeq, 2018. 3, 1

    work page 2018

Show all 53 references
  1. [9]

    Chen, Yulia Rubanova, Jesse Bettencourt, and David K Duvenaud

    Ricky T.Q. Chen, Yulia Rubanova, Jesse Bettencourt, and David K Duvenaud. Neural ordinary differential equations. 31, 2018. 2

    2018

  2. [10]

    Gru-ode-bayes: Continuous modeling of sporadically-observed time series

    Edward De Brouwer, Jaak Simm, Adam Arany, and Yves Moreau. Gru-ode-bayes: Continuous modeling of sporadically-observed time series. 32, 2019. 2

    2019

  3. [11]

    Large scale interactive motion forecasting for autonomous driving: The waymo open motion dataset

    Scott Ettinger, Shuyang Cheng, Benjamin Caine, Chenxi Liu, Hang Zhao, Sabeek Pradhan, Yuning Chai, Ben Sapp, Charles R Qi, Yin Zhou, et al. Large scale interactive motion forecasting for autonomous driving: The waymo open motion dataset. InCVPR, pages 9710–9719, 2021. 1

    2021

  4. [12]

    Benchmarks and challenges in pose estimation for egocentric hand interactions with objects

    Zicong Fan, Takehiko Ohkawa, Linlin Yang, Nie Lin, Zhishan Zhou, Shihao Zhou, Jiajun Liang, Zhong Gao, Xuanyang Zhang, Xue Zhang, et al. Benchmarks and challenges in pose estimation for egocentric hand interactions with objects. In ECCV, pages 428–448. Springer, 2025. 1, 2

    2025

  5. [13]

    Auto- matic generation and detection of highly reliable fiducial markers under occlusion.Pattern Recognition, 47(6):2280– 2292, 2014

    Sergio Garrido-Jurado, Rafael Mu˜noz-Salinas, Francisco Jos´e Madrid-Cuevas, and Manuel Jes ´us Mar´ın-Jim´enez. Auto- matic generation and detection of highly reliable fiducial markers under occlusion.Pattern Recognition, 47(6):2280– 2292, 2014. 4

    2014

  6. [14]

    Learning with 3d rotations, a hitchhiker’s guide to so (3)

    Andreas Ren´e Geist, Jonas Frey, Mikel Zhobro, Anna Levina, and Georg Martius. Learning with 3d rotations, a hitchhiker’s guide to so (3). InInt. Conf. Mach. Lear., 2024. 2, 3, 1

    2024

  7. [15]

    Consistent parame- ter estimation for systems of ordinary differential equations: bypassing numerical integration via smoothing.Bernoulli, pages 1061–1098, 2012

    Shota Gugushvili and Chris AJ Klaassen. Consistent parame- ter estimation for systems of ordinary differential equations: bypassing numerical integration via smoothing.Bernoulli, pages 1061–1098, 2012. 1

    2012

  8. [16]

    Survey of higher order rigid body motion interpolation methods for keyframe animation and continuous-time trajectory estima- tion

    Adrian Haarbach, Tolga Birdal, and Slobodan Ilic. Survey of higher order rigid body motion interpolation methods for keyframe animation and continuous-time trajectory estima- tion. InInt. Conf. 3D Vision. IEEE, 2018. 2, 4

    2018

  9. [17]

    Metrics for 3d rotations: Comparison and analysis.Journal of Mathematical Imaging and Vision, 35: 155–164, 2009

    Du Q Huynh. Metrics for 3d rotations: Comparison and analysis.Journal of Mathematical Imaging and Vision, 35: 155–164, 2009. 3

    2009

  10. [18]

    Constrained 3d rotation smooth- ing via global manifold regression for video stabilization

    Chao Jia and Brian L Evans. Constrained 3d rotation smooth- ing via global manifold regression for video stabilization. IEEE Trans. on Sig. Proc., 62(13):3293–3304, 2014. 2

    2014

  11. [19]

    Geometric savitzky-golay filtering of noisy rotations on so (3) with si- multaneous angular velocity and acceleration estimation

    Maarten Jongeneel and Alessandro Saccon. Geometric savitzky-golay filtering of noisy rotations on so (3) with si- multaneous angular velocity and acceleration estimation. In Int. Conf. Intell. Robot. and Syst., pages 2962–2968. IEEE,

  12. [20]

    The oxford multimotion dataset: Multiple se (3) motions with ground truth.Robo

    Kevin Michael Judd and Jonathan D Gammell. The oxford multimotion dataset: Multiple se (3) motions with ground truth.Robo. Auto. Let., 4(2):800–807, 2019. 4

    2019

  13. [21]

    Neural controlled differential equations for irregular time series

    Patrick Kidger, James Morrill, James Foster, and Terry Lyons. Neural controlled differential equations for irregular time series. 33:6696–6707, 2020. 2, 3, 1

    2020

  14. [22]

    A general construction scheme for unit quaternion curves with simple high order derivatives

    Myoung-Jun Kim, Myung-Soo Kim, and Sung Yong Shin. A general construction scheme for unit quaternion curves with simple high order derivatives. InComputer graphics and interactive techniques, pages 369–376, 1995. 2

    1995

  15. [23]

    Adam: A method for stochastic opti- mization.arXiv preprint arXiv:1412.6980, 2014

    Diederik P Kingma. Adam: A method for stochastic opti- mization.arXiv preprint arXiv:1412.6980, 2014. 3

    2014 arXiv

  16. [24]

    3d-rcnn: Instance- level 3d object reconstruction via render-and-compare

    Abhijit Kundu, Yin Li, and James M Rehg. 3d-rcnn: Instance- level 3d object reconstruction via render-and-compare. In CVPR, 2018. 2

    2018

  17. [25]

    Computationally efficient rigid-body gaussian process for motion dynamics.Robo

    Muriel Lang and Sandra Hirche. Computationally efficient rigid-body gaussian process for motion dynamics.Robo. Auto. Let., 2, 2017. 2

    2017

  18. [26]

    Gaussian process kernels for rotations and 6d rigid body motions

    Muriel Lang, Oliver Dunkley, and Sandra Hirche. Gaussian process kernels for rotations and 6d rigid body motions. In Int. Conf. Robot. Automation., pages 5165–5170. IEEE, 2014. 2

    2014

  19. [27]

    Global formulations of lagrangian and hamiltonian dynamics on manifolds.Springer, 13:31, 2017

    Taeyoung Lee, Melvin Leok, and N Harris McClamroch. Global formulations of lagrangian and hamiltonian dynamics on manifolds.Springer, 13:31, 2017. 2

    2017

  20. [28]

    An analysis of svd for deep rotation estimation

    Jake Levinson, Carlos Esteves, Kefan Chen, Noah Snavely, Angjoo Kanazawa, Afshin Rostamizadeh, and Ameesh Maka- dia. An analysis of svd for deep rotation estimation. 33: 22554–22565, 2020. 2

    2020

  21. [29]

    Incorporating con- trol inputs in the estimation of continuous mobile robot trajectories and continuum robot shapes.arXiv preprint arXiv:2408.01333, 2024

    Sven Lilge and Timothy D Barfoot. Incorporating con- trol inputs in the estimation of continuous mobile robot trajectories and continuum robot shapes.arXiv preprint arXiv:2408.01333, 2024. 2

    2024 arXiv

  22. [30]

    Spline fusion: A continuous-time representation for visual- inertial fusion with application to rolling shutter cameras

    Steven Lovegrove, Alonso Patron-Perez, and Gabe Sibley. Spline fusion: A continuous-time representation for visual- inertial fusion with application to rolling shutter cameras. In BMVC, page 8, 2013. 2

    2013

  23. [31]

    Dynamic grasp and trajectory planning for moving objects.Autonomous Robots, 43:1241–1256, 2019

    Naresh Marturi, Marek Kopicki, Alireza Rastegarpanah, Vi- jaykumar Rajasekaran, Maxime Adjigble, Rustam Stolkin, Aleˇs Leonardis, and Yasemin Bekiroglu. Dynamic grasp and trajectory planning for moving objects.Autonomous Robots, 43:1241–1256, 2019. 1

    2019

  24. [32]

    Learn- ing to predict 3d rotational dynamics from images of a rigid body with unknown mass distribution.Aerospace, 10(11): 921, 2023

    Justice J Mason, Christine Allen-Blanchette, Nicholas Zol- man, Elizabeth Davison, and Naomi Ehrich Leonard. Learn- ing to predict 3d rotational dynamics from images of a rigid body with unknown mass distribution.Aerospace, 10(11): 921, 2023. 2, 1

    2023

  25. [33]

    On the choice of interpolation scheme for neural cdes.Trans

    James Morrill, Patrick Kidger, Lingyi Yang, and Terry Lyons. On the choice of interpolation scheme for neural cdes.Trans. Mach. Lear. Resea., 2022(9), 2022. 3, 1, 2

    2022

  26. [34]

    Orb-slam: A versatile and accurate monocular slam system.IEEE transactions on robotics, 31(5):1147–1163,

    Raul Mur-Artal, Jose Maria Martinez Montiel, and Juan D Tardos. Orb-slam: A versatile and accurate monocular slam system.IEEE transactions on robotics, 31(5):1147–1163,

  27. [35]

    Noise modeling and analysis of an imu-based attitude sensor: improvement of performance by filtering and sensor fusion

    K Nirmal, AG Sreejith, Joice Mathew, Mayuresh Sarpotdar, Ambily Suresh, Ajin Prakash, Margarita Safonova, and Jayant Murthy. Noise modeling and analysis of an imu-based attitude sensor: improvement of performance by filtering and sensor fusion. InAdvances in optical and mechan...

  28. [36]

    Kinematic state estimation and motion planning for stochastic nonholonomic systems using the ex- ponential map.Robotica, 26(4):419–434, 2008

    Wooram Park, Yan Liu, Yu Zhou, Matthew Moses, and Gre- gory S Chirikjian. Kinematic state estimation and motion planning for stochastic nonholonomic systems using the ex- ponential map.Robotica, 26(4):419–434, 2008. 1

    2008

  29. [37]

    Pytorch: An imperative style, high-performance deep learning library

    Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. 32,

  30. [38]

    Nicholas Greene, David Rosen, Jonathan Kelly, and Nicholas Roy

    Valentin Peretroukhin, Matthew Giamou, W. Nicholas Greene, David Rosen, Jonathan Kelly, and Nicholas Roy. A smooth representation of belief over so(3) for deep rotation learning with uncertainty.Robotics: Science and Systems, 2020. 2

    2020

  31. [39]

    Practical pose trajectory splines with explicit reg- ularization

    Mikael Persson, Gustav H¨ager, Hannes Ovr´en, and Per-Erik Forss´en. Practical pose trajectory splines with explicit reg- ularization. InInt. Conf. 3D Vision, pages 156–165. IEEE,

  32. [40]

    Latent ordinary differential equations for irregularly-sampled time series

    Yulia Rubanova, Ricky TQ Chen, and David K Duvenaud. Latent ordinary differential equations for irregularly-sampled time series. 32, 2019. 2

    2019

  33. [41]

    Why and how savitzky–golay filters should be replaced.ACS Measure- ment Science Au, 2(2):185–196, 2022

    Michael Schmid, David Rath, and Ulrike Diebold. Why and how savitzky–golay filters should be replaced.ACS Measure- ment Science Au, 2(2):185–196, 2022. 3

    2022

  34. [42]

    Uniform random rotations

    Ken Shoemake. Uniform random rotations. InGraphics Gems III (IBM Version), pages 124–132. Elsevier, 1992. 1

    1992

  35. [43]

    Efficient deriva- tive computation for cumulative b-splines on lie groups

    Christiane Sommer, Vladyslav Usenko, David Schubert, Nikolaus Demmel, and Daniel Cremers. Efficient deriva- tive computation for cumulative b-splines on lie groups. In CVPR, 2020. 2, 4

    2020

  36. [44]

    Sophus: C++ implementation of lie groups using eigen

    Hauke Strasdat. Sophus: C++ implementation of lie groups using eigen. https : / / github . com / strasdat / Sophus, 2011. 2

    2011

  37. [45]

    Continuous-time state estimation methods in robotics: A survey.arXiv preprint arXiv:2411.03951, 2024

    William Talbot, Julian Nubert, Turcan Tuna, Cesar Cadena, Frederike D¨umbgen, Jesus Tordesillas, Timothy D Barfoot, and Marco Hutter. Continuous-time state estimation methods in robotics: A survey.arXiv preprint arXiv:2411.03951, 2024. 2

    2024 arXiv

  38. [46]

    A white-noise-on-jerk motion prior for continuous-time trajectory estimation on se (3).Robo

    Tim Yuqing Tang, David Juny Yoon, and Timothy D Bar- foot. A white-noise-on-jerk motion prior for continuous-time trajectory estimation on se (3).Robo. Auto. Let., 4, 2019. 2

    2019

  39. [47]

    Tangent space backpropagation for 3d transformation groups

    Zachary Teed and Jia Deng. Tangent space backpropagation for 3d transformation groups. InIEEE/CVF conference on computer vision and pattern recognition, pages 10338–10347,

  40. [48]

    The double sphere camera model

    Vladyslav Usenko, Nikolaus Demmel, and Daniel Cremers. The double sphere camera model. In2018 International Conference on 3D Vision (3DV), pages 552–560. IEEE, 2018. 2

    2018

  41. [49]

    A data-driven motion prior for continuous- time trajectory estimation on se(3).Robo

    Jeremy N Wong, David J Yoon, Angela P Schoellig, and Tim- othy D Barfoot. A data-driven motion prior for continuous- time trajectory estimation on se(3).Robo. Auto. Let., 5, 2020. 2

    2020

  42. [50]

    Grasparl: Dy- namic grasping via adversarial reinforcement learning.arXiv preprint arXiv:2203.02119, 2022

    Tianhao Wu, Fangwei Zhong, Yiran Geng, Hongchen Wang, Yongjian Zhu, Yizhou Wang, and Hao Dong. Grasparl: Dy- namic grasping via adversarial reinforcement learning.arXiv preprint arXiv:2203.02119, 2022. 1

    2022 arXiv

  43. [51]

    Quater- nion equivariant capsule networks for 3d point clouds

    Yongheng Zhao, Tolga Birdal, Jan Eric Lenssen, Emanuele Menegatti, Leonidas Guibas, and Federico Tombari. Quater- nion equivariant capsule networks for 3d point clouds. In ECCV, pages 1–19. Springer, 2020. 2

    2020

  44. [52]

    On the continuity of rotation representations in neural networks

    Yi Zhou, Connelly Barnes, Jingwan Lu, Jimei Yang, and Hao Li. On the continuity of rotation representations in neural networks. InCVPR, pages 5745–5753, 2019. 2, 3, 1 Continuous-Time SO(3) Forecasting with Savitzky–Golay Neural Controlled Differential Equations Supplementary M...

    2019

  45. [53]

    Following [21], the method encodes an initial value z0, which is then integrated forward in time usingtorchdiffeq[ 8] and Dormand-Prince 4/5 with respect to the constructed spline

    via backward differences. Following [21], the method encodes an initial value z0, which is then integrated forward in time usingtorchdiffeq[ 8] and Dormand-Prince 4/5 with respect to the constructed spline. The latent representation is then decoded into the 6D rotation represe...

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.