Pith. sign in

REVIEW 3 major objections 5 minor 69 references

Learning Smooth SE(3) Trajectories under Left-Invariant Riemannian Metrics

T0 review · 3 major / 5 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 This paper claims that a learned polynomial in body twist can approximate higher-order optimal SE(3) trajectories in about a millisecond.

desk verdict A useful fast SE(3) trajectory generator, but the 'closely approximate' claim outruns the numbers: energy and EL residuals are orders of magnitude off BVP-optimal, and the enforced boundary conditions are ordinary twist derivatives, not covariant ones. read the letter →

arxiv 2608.01562 v1 pith:IAAPYWPP submitted 2026-08-03 cs.RO

Yuwei Wu , Vijay Kumar This is my paper · ORCID
classification cs.RO
keywords SE(3)trajectorygenerationleft-invariantRiemannianmetricpolynomialtwistparameterizationcoefficientcompletionneurallearningEuler-Lagrangeresidualreal-timereplanningquadrotormotionprimitives
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

Optimal rigid-body trajectory generation on SE(3) is normally a boundary value problem: minimizing higher-order covariant-derivative energy under general left-invariant metrics requires expensive numerical solvers. The paper claims a neural network can produce near-optimal smooth trajectories in about a millisecond by parameterizing the body twist as a high-degree polynomial. Boundary conditions on twist and its derivatives are enforced exactly through coefficient completion, while the terminal pose and optimality are enforced through training losses. If the claim holds, real-time replanning, motion primitive libraries, and robot deployments like quadrotor flight become computationally feasible without abandoning geometric consistency.

What carries the argument

The central mechanism is coefficient completion on a polynomial parameterization of the body twist. For a degree-$n$ polynomial $\xi(t)$, boundary conditions at $t=0$ and $t=T$ on twist and its derivatives determine the lowest $2(s-1)$ coefficient vectors as the solution to a linear system; the remaining $n_f = n+1-2(s-1)$ coefficient vectors and the duration are network outputs. A Cholesky-based embedding of the left-invariant metric is fed into FiLM or low-rank modulation layers so the same network adapts to different metric matrices, and the Euler-Lagrange residual (Eq. 10) is evaluated at collocation points as the optimality loss.

What would settle it

Run a high-accuracy collocation solver on the Euler-Lagrange equations with covariant-derivative boundary conditions for a coupled metric such as W_CB and compare the resulting twist profiles with the network output; if the mismatch in covariant derivatives or terminal pose exceeds the reported tolerances, the polynomial surrogate is falsified.

Watch

Extended reading notes

Core claim

On its own terms, the paper claims that the optima of higher-order Riemannian variational problems on SE(3), which normally require expensive boundary value solvers, can be closely approximated by finite-dimensional polynomial curves in the body twist. The body twist is $\xi(t)=\sum_{j=0}^{n} c_j t^j$ with $n$ up to 12, the lower-order coefficients are fixed by solving a linear system that imposes initial and terminal twist boundary conditions, and a metric-conditioned network predicts the remaining coefficients and duration. Training minimizes a weighted sum of a metric smoothness cost, an Euler-Lagrange residual, dynamic-feasibility penalties, and a terminal pose loss. The paper reports that the resulting trajectories match numerical optimization closely, have boundary twist errors of zero by construction, keep terminal pose errors around 0.1, and run at about 1 ms inference.

Load-bearing premise

The load-bearing premise is that a degree-12 polynomial in the body twist, with boundary conditions placed on ordinary time derivatives of twist, faithfully represents the true variational solution whose boundary conditions involve covariant derivatives of the pose; the paper asserts this approximation without an error bound.

Editorial extensions

If this is right

  • A single forward pass generates a smooth, dynamically feasible SE(3) trajectory in about 1 ms, making online replanning during flight practical.
  • Boundary conditions on twist and its derivatives are satisfied exactly by construction, so non-rest-to-rest transitions with nonzero velocity and acceleration are handled without extra constraints.
  • Because the network is conditioned on the metric tensor, one model covers isotropic, anisotropic, inertia-based, and rotated-displaced frame metrics rather than requiring a separate solver per metric.
  • Multiple waypoints can be traversed by concatenating segments whose terminal twist state seeds the next segment, yielding C^1 twist continuity and a motion primitive library.
  • The same generator can be fine-tuned for underactuated platforms by adding robot-specific feasibility layers and anchor regularization.

Reading between the lines

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

  • Editorial extension: if the polynomial surrogate and soft terminal pose hold at the reported accuracy, the framework could plausibly be extended to time-varying or configuration-dependent metrics, but that is beyond the paper's constant-metric assumption.
  • Editorial extension: because terminal pose is only softly enforced, safety-critical tasks may need a downstream correction or refinement step; the paper does not address that failure mode.
  • Editorial extension: the Euler-Lagrange residual is used as a regularization rather than a certified optimality bound, so a natural test is to compare residual values against a tight discretization of the exact Euler-Lagrange equations for coupled metrics.
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 proposes a learning-based framework for generating smooth SE(3) trajectories under general left-invariant Riemannian metrics. Body-twist trajectories are parameterized as high-degree polynomials; a neural network predicts a subset of polynomial coefficients and the trajectory duration, while the remaining coefficients are determined analytically by solving a linear system that enforces boundary conditions on the twist and its ordinary time derivatives. Training losses include the Riemannian energy, the Euler–Lagrange residual, a time-regularization term, a soft terminal-pose penalty, and dynamic-feasibility penalties. The framework is evaluated in numerical benchmarks against a collocation BVP solver and a per-instance PINN, in ablations of loss components and architectural variants, and in applications to waypoint traversal, replanning, and quadrotor trajectory generation. The central claim is that the generated trajectories closely approximate numerical optimization solutions while achieving millisecond-level inference times.

Significance. If the central claim were substantiated, the framework would be a practically valuable real-time alternative to BVP solvers for geometrically consistent trajectory generation on SE(3), with exact enforcement of twist-derivative boundary conditions and generalization across metric families. The analytic coefficient-completion module, the metric-conditioned network design, and the systematic treatment of several left-invariant metric families are solid engineering contributions, and the empirical study is broad in scope, including ablations and hardware-oriented applications. Even as a feasible smooth-trajectory generator with soft terminal-pose satisfaction, the system has clear practical utility. However, the quantitative evidence does not currently support the 'closely approximate solutions from numerical optimization' claim, and the gap between the covariant-derivative boundary conditions of the variational problem and the ordinary-derivative twist boundary conditions actually enforced is not resolved. The paper therefore needs substantial revision before its main claim can be accepted.

major comments (3)
  1. [Section V.C / VI.A, Eqs. (12) and (46)] The trajectories are trained to satisfy ordinary time-derivative boundary conditions on the body twist, whereas the variational problem in Eqs. (1) and (12) imposes boundary conditions on covariant derivatives of the pose. For s=3, the terminal boundary value is (γ(T), V(T), ∇_V V(T)); even for the isotropic metric, ∇_V V = (ωdot, vdot + ω×v) in Eq. (30), and for coupled metrics additional connection terms appear in Eqs. (31)–(32). Enforcing ξ(T)=ξ_T and ξdot(T)=ξdot_T does not enforce the covariant-acceleration boundary condition unless ω×v and the connection terms vanish. The paper acknowledges this as an approximation in Section IV.B and V.C, but it provides no error bound and no comparison of the true covariant-derivative boundary residual. Because the abstract and introduction claim approximation to the variational problem, this issue is load-bearing and must be addressed either by reformulating the claim or by supplying a quantitative analysis of the surrogate error.
  2. [Table III, W0 row and other metric rows] The reported energy and Euler–Lagrange residual for the proposed method are orders of magnitude larger than those of BVP-Opt when the latter converges. For W0, the energy is 79.245 versus 1.236 for BVP-Opt, and the EL residual is 317.540 versus 0.222. The paper's statement that 'the proposed method achieves high-quality trajectories across all metrics' and the abstract's 'closely approximate solutions from numerical optimization' are not supported by these numbers. The comparison is also confounded because BVP-Opt has large boundary errors (pose 0.795, vel/acc 1.159 on W0) and a low success rate, so it is not a validated optimality reference. Still, the burden is on the authors to show that the large gap is due to the baseline's failure rather than to the learned trajectories being far from optimal; a fairer BVP baseline with boundary-error reporting and warm-starting, or an independent optimality certificate, is needed.
  3. [Section VI.C, Eq. (50), and Problem 2 constraint (42d)] The introduction and contribution list state that boundary conditions are enforced exactly via coefficient completion, but only the twist and its derivatives are enforced exactly; the terminal pose is enforced only through the soft loss in Eq. (50). The reported pose errors in Tables I–III are 0.022–0.113, which may be acceptable for control applications but contradict the exact-constraint language and the hard equality constraint (42d) in Problem 2. The authors do note in the introduction that 'exact satisfaction of end states in SE(3) is not guaranteed,' but the framing, the 'Constraint Satisfaction by Construction' heading, and the table column 'Boundary Error' obscure this distinction. The manuscript should consistently separate exactly enforced twist-derivative conditions from softly enforced pose conditions and discuss the implications for tasks requiring precise endpoint positioning.
minor comments (5)
  1. [Section VII.B / VIII.B] The introduction claims 'real quadrotor hardware experiments,' but Section VIII.B describes Crazyflie 2.1 with RotorPy simulations and only mentions real-world flight tests in passing, without presenting tracking plots, error statistics, or experimental details. Either add the hardware data or soften the claim to simulation-only validation.
  2. [Eq. (30) and related formulas] The notation in Eq. (30) is unclear: the term 'RT d(2)' should be written as R^T \ddot d or explicitly defined. Similar notational inconsistencies appear in other covariant-derivative expressions, making them hard to verify.
  3. [Eq. (22)] The displayed formula for ∇^W_X Y omits the definition of the X(Y^i \hat L_i) term and mixes coordinate and coordinate-free notation. A self-contained derivation or a reference to the standard left-invariant connection formula would improve reproducibility.
  4. [Section VII.E] The comparison with BVP-Opt fixes the duration to the value produced by the proposed method, and the solver is limited to 100 iterations. This is a reasonable first step, but the text should discuss how a warm-started or more tightly converged BVP solver would change the comparison, and it should report the distribution of BVP-Opt boundary errors in the success cases.
  5. [Section VI.D / VII.B] The loss weights α_t, α_e, the scaling factors λ_{d,i}, and the augmented Lagrangian multiplier schedules are not reported. Without these values, the loss balance and the reported trade-offs are difficult to reproduce, especially the claim that the time regularization is 'intentionally small.'

Circularity Check

1 steps flagged · score 4.0 of 10

Partial evidence-level circularity: the 'Energy' and 'E.L.' smoothness metrics used to claim near-optimality are the training-loss terms of Eq. (48), so the favorable quality evidence is the fit itself; the independent BVP-Opt comparison instead runs 1-3 orders of magnitude better on those same metrics, making the main risk correctness rather than circular derivation.

  1. fitted input called prediction [Section VI.D (Problem 3, Eq. 48) and Section VII.D / Table III]
    "Trajectory quality: Produces smooth, near-optimal trajectories that minimize cost functionals under arbitrary state conditions and left-invariant Riemannian metrics. ... Performance is evaluated in terms of trajectory smoothness (energy and Euler-Lagrange residual), boundary accuracy (velocity/acceleration and pose errors), trajectory duration, and dynamic feasibility. ... The overall loss can be written as Ĵ=J_m(ξ^(s−1)) + α_e 1/M Σ_{k=1}^M ∥E(ξ[2s−1])(k/M T))∥² + α_t T(τ)."

    The metrics used to support near-optimality are, by construction, the training objective: Section VII.D evaluates 'smoothness (energy and Euler-Lagrange residual)' and Tables I-III report 'Energy'/'E.L.', while Problem 3/Eq. (48) minimizes exactly J_m + α_e Σ∥E(ξ[2s−1])∥² + α_t T. The network weights were fitted by minimizing these quantities, so low values on the held-out split report the fit itself; citing them as evidence that outputs are 'smooth, near-optimal' is the fitted input renamed as a prediction of optimality. Partial: completion (Eq. 46), metric conditioning, and the BVP-Opt benchmark are independent of the loss. But the only non-circular check (Table III) shows Ours's Energy 1-2 orders higher (W0: 79.245 vs 1.236) and E.L.

full rationale

The core derivation is not circular: the network is trained on physics-derived losses (Eqs. 47-48), evaluated on a held-out split, and never trained on the outputs of the BVP-Opt optimizer; coefficient completion (Eq. 46), metric conditioning (Eqs. 52-54), and the boundary linear system are all independent of the target solutions. The one genuine circular step sits at the evidence level: the smoothness metrics reported as trajectory quality ('Energy', 'E.L.') are exactly the first two terms of the training loss (Eq. 48), so low values on the test split are the fit itself, not an independent certification of near-optimality. The step is partial because the BVP-Opt baseline is external, but that external check does not support the abstract's 'closely approximate' claim: where BVP-Opt converges its Energy is one to two orders of magnitude lower (e.g., W0: 1.236 vs 79.245) and its E.L. residual three orders lower (0.222 vs 317.540), while its large boundary errors (pose 0.53-0.92) make the comparison inconclusive; hence the unsupported claim is a correctness risk, not a circularity. The paper also flags its own validity gap: Section IV.B states that solutions of Eq. (10) 'are not polynomials in local coordinates' and that boundary constraints are enforced on body-twist (ordinary) derivatives rather than the covariant-derivative boundary conditions of Eq. (12), and Section IX concedes the trajectories 'approximate the solution of the variational optimization problem and do not guarantee exact optimality,' with no error bound supplied; these are missing-support issues weighed here. Self-citations [3] and [60] (Kumar as co-author) are used for background and motivation only - the general Euler-Lagrange equation is cited externally [8] and the metric matrices are rederived in Eqs. (17)-(21) - so they are not load-bearing. No uniqueness theorem is imported from the authors, no ansatz arrives solely via a citation, and no known result is renamed. Score 4 reflects one partial circularity in the evaluation evidence; the central derivation retains independent content.

Assumptions & free parameters 6 free parameters · 5 assumptions · 0 invented entities

The central claim rests on a handful of free loss weights and design choices that are not fully reported, on standard geometric identities from the cited literature, and on an unproven surrogate assumption: a degree-12 polynomial in the body twist with time-derivative boundary conditions approximates the true covariant-derivative variational problem. No new physical entities are introduced.

free parameters (6)
  • Time regularization weight α_t and EL residual weight α_e = Not specified (α_t=50 in Fig. 3 example)
    These weights determine the balance between duration, energy, and optimality residual in the training loss; without them the exact objective is not reproducible.
  • Polynomial degree n = 12
    Chosen as a compromise between expressiveness and numerical stability; directly limits approximation accuracy.
  • Dynamic limit scaling factors λ_{d,i} = Not specified
    Adjust the permitted magnitudes of the twist derivatives in Eq. (47c); values are not reported.
  • Condition-number filter threshold = Not specified
    Samples above a spectral condition-number threshold are removed from the training set; the threshold value is not given.
  • Augmented Lagrangian multipliers update rate and penalty parameters = Not specified
    The update rule λ_i ← Π(λ_i + ηλ φ(c_i)) and ρ_i are stated symbolically but numerical values and choices are not reported.
  • T_max and ε = T_max=10 s, ε=0.01
    Bound the duration parameterization; chosen by hand for the experiments.
assumptions (5)
  • standard math Euler-Lagrange equation for higher-order variational problems (Eq. 10) is the correct necessary condition for the functional in Eq. (9).
    Restated from Camarinha et al. [8]; used as the optimality signal in the loss.
  • standard math The Levi-Civita connection formulas for left-invariant metrics on SE(3) (Eqs. 22-29) are correct.
    Derived from Do Carmo [59] and Zefran et al. [60]; the learning framework depends on these formulas.
  • ad hoc to paper A finite-degree polynomial in the body twist with boundary conditions on ordinary twist derivatives is an adequate surrogate for a higher-order Riemannian variational curve.
    Central modeling assumption in Section VI.A; no error bound or comparison to covariant-derivative boundary conditions is provided.
  • domain assumption The metric W is constant along the trajectory.
    Acknowledged in Limitations; the framework does not handle configuration-dependent metrics.
  • domain assumption The discrete time-ordered product in Eq. (44) accurately reconstructs SE(3) poses from the polynomial twist at the chosen sampling rate.
    The pose loss and terminal pose accuracy depend on this integration being accurate at N collocation points.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Learning Smooth SE(3) Trajectories under Left-Invariant Riemannian Metrics." pith.science (2026). https://pith.science/paper/IAAPYWPP

@misc{pith2026260801562,
  author       = {Pith},
  title        = {Pith review of: Learning Smooth SE(3) Trajectories under Left-Invariant Riemannian Metrics},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/IAAPYWPP}},
  note         = {Machine review of arXiv:2608.01562}
}
read the original abstract

Optimal trajectory generation for rigid-body motions on Lie groups can be formulated as a variational problem that minimizes energy functionals defined by Riemannian metrics. While closed-form solutions exist for special cases such as product metrics and rest-to-rest boundary conditions, solving the general problem with arbitrary boundary states and coupled rotational-translational metrics often requires computationally expensive numerical boundary value solvers. These limitations restrict the use of geometrically consistent trajectory generation in real-time robotic planning and control. This paper presents a learning-based framework for approximating higher-order smooth trajectories on SE(3) under general left-invariant Riemannian metrics. The method parameterizes body-twist trajectories using high-order polynomials and relies on a neural network to learn a subset of the polynomial coefficients and the trajectory duration. The remaining coefficients are analytically determined to enforce the boundary conditions. The training of the network is guided by losses derived from Euler-Lagrange optimality conditions, metric-weighted smoothness objectives, and feasibility constraints. The metric-conditioned framework enables generalization across diverse metric structures and motion conditions. Extensive numerical experiments demonstrate that the proposed approach generates smooth trajectories that closely approximate solutions from numerical optimization while achieving millisecond-level inference times. We demonstrate two practical applications of the proposed framework: real-time generation of diverse motion primitives with waypoint traversal, and refinement for quadrotor flight under dynamic conditions. These results suggest that learning-based motions with geometric structure can provide an efficient alternative to conventional optimization-based methods for trajectory generation on SE(3).

Figures

Figures reproduced from arXiv: 2608.01562 by the authors.

Figure 1
Figure 1. SE(3) trajectories with different boundary conditions and metrics generated by the proposed framework. (a) Generation of single motion primitives (first row) and trajectories traversing multiple waypoints with higher-order boundary conditions (second row). (b) Fine-tuning adapted for quadrotor applications, evaluated over six different experiments with randomly generated waypoints. To address these challenges, sever… view at source ↗
Figure 2
Figure 2. Proposed framework. (A) The learning module predicts the coefficients [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Time-optimal trajectory generation with trade-off between duration [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (6 more)
Figure 4
Figure 4. Figure 4: Trajectories optimized using numerical solvers under five distinct Riemannian metrics with different functionals: (a) minimum distance, (b) minimum [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]
Figure 5
Figure 5. Figure 5: Evaluation of different metrics difficulty in condition number (left) [PITH_FULL_IMAGE:figures/full_fig_p010_5.png]
Figure 6
Figure 6. Figure 6: Two representative metric classes (W0 and WCB) with trajectories from all motion phases generated by the proposed framework. For each case, we show the R3 trajectory with body axes (left) and the corresponding SO(3) trajectory on the unit sphere [63], where color encod…
Figure 7
Figure 7. Figure 7: Distribution of evaluation metrics. Histograms show the distribution [PITH_FULL_IMAGE:figures/full_fig_p012_7.png]
Figure 8
Figure 8. Figure 8: Trajectories between fixed random initial and terminal poses under [PITH_FULL_IMAGE:figures/full_fig_p013_8.png]
Figure 10
Figure 10. Figure 10: Replanning under non-rest motions. The left shows trajectories [PITH_FULL_IMAGE:figures/full_fig_p014_10.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

69 extracted references · 58 canonical work pages

  1. [1]

    Reduction of euler lagrange problems for constrained variational problems and relation with optimal control problems,

    A. Bloch and P. Croach, “Reduction of euler lagrange problems for constrained variational problems and relation with optimal control problems,” inProceedings of 1994 33rd IEEE Conference on Decision and Control, vol. 3, 1994, pp. 2584–2590 vol.3

    work page 1994

  2. [2]

    The dynamic interpolation problem: on riemannian manifolds, lie groups, and symmetric spaces,

    P. Crouch and F. S. Leite, “The dynamic interpolation problem: on riemannian manifolds, lie groups, and symmetric spaces,”Journal of Dynamical and control systems, vol. 1, no. 2, pp. 177–202, 1995

    work page 1995

  3. [3]

    On the generation of smooth three- dimensional rigid body motions,

    M. Zefran, V . Kumar, and C. Croke, “On the generation of smooth three- dimensional rigid body motions,”IEEE Transactions on Robotics and Automation, vol. 14, no. 4, pp. 576–589, 1998

    work page 1998

  4. [4]

    Kinematic controllability and decoupled tra- jectory planning for underactuated mechanical systems,

    F. Bullo and K. Lynch, “Kinematic controllability and decoupled tra- jectory planning for underactuated mechanical systems,” inProceedings 2001 ICRA. IEEE International Conference on Robotics and Automation (Cat. No.01CH37164), vol. 4, 2001, pp. 3300–3307 vol.4

    work page 2001

  5. [5]

    An SVD-based projection method for inter- polation onSE(3),

    C. Belta and V . Kumar, “An SVD-based projection method for inter- polation onSE(3),”IEEE Transactions on Robotics and Automation, vol. 18, no. 3, pp. 334–345, 2002

    work page 2002

  6. [6]

    Optimal control on rieman- nian manifolds by interpolation,

    R. Giambo, F. Giannoni, and P. Piccione, “Optimal control on rieman- nian manifolds by interpolation,”Mathematics of Control, Signals and Systems, vol. 16, no. 4, pp. 278–296, 2004

    work page 2004

  7. [7]

    Bézier curves and c2 interpolation in riemannian manifolds,

    T. Popiel and L. Noakes, “Bézier curves and c2 interpolation in riemannian manifolds,”Journal of Approximation Theory, vol. 148, no. 2, pp. 111–127, 2007. [Online]. Available: https://www. sciencedirect.com/science/article/pii/S0021904507000469

    work page 2007

  8. [8]

    Splines of class c k on non-euclidean spaces,

    M. Camarinha, F. S. Leite, and P. E. Crouch, “Splines of class c k on non-euclidean spaces,”IMA Journal of Mathematical Control and Information, vol. 12, no. 4, pp. 399–410, 12 1995. [Online]. Available: https://doi.org/10.1093/imamci/12.4.399

    work page doi:10.1093/imamci/12.4.399 1995

Show all 69 references
  1. [9]

    Bi-jacobi fields and riemannian cubics for left- invariantSO(3),

    L. Noakes and T. Ratiu, “Bi-jacobi fields and riemannian cubics for left- invariantSO(3),”Communications in Mathematical Sciences, vol. 14, pp. 55–68, 01 2016

    2016

  2. [10]

    The de casteljau algorithm on lie groups and spheres,

    P. Crouch, G. Kun, and F. Silva Leite, “The de casteljau algorithm on lie groups and spheres,”Journal of Dynamical and Control Systems, vol. 5, pp. 397–429, 07 1999

    1999

  3. [11]

    A numerical algorithm for C2-splines on symmetric spaces,

    G. Bogfjellmo, K. Modin, and O. Verdier, “A numerical algorithm for C2-splines on symmetric spaces,”SIAM J. Numer. Analysis, vol. 56, no. 4, pp. 2623–2647, 2018. [Online]. Available: https: //doi.org/10.1137/17M1123353

    2018 doi

  4. [12]

    Trajectory optimization on manifolds with applications toSO(3)andR3×S2

    M. Watterson, S. Liu, K. Sun, T. Smith, and V . Kumar, “Trajectory optimization on manifolds with applications toSO(3)andR3×S2.” inRobotics: Science and Systems, vol. 20, 2018

    2018

  5. [13]

    Flatness and defect of non-linear systems: introductory theory and examples,

    M. Fliess, J. Lévine, P. Martin, and P. Rouchon, “Flatness and defect of non-linear systems: introductory theory and examples,”International Journal of Control, vol. 61, no. 6, pp. 1327–1361, 1995. [Online]. Available: https://doi.org/10.1080/00207179508921959

    1995 doi

  6. [14]

    Minimum snap trajectory generation and control for quadrotors,

    D. Mellinger and V . Kumar, “Minimum snap trajectory generation and control for quadrotors,” in2011 IEEE International Conference on Robotics and Automation, 2011, pp. 2520–2525

    2011

  7. [15]

    A computationally ef- ficient motion primitive for quadrocopter trajectory generation,

    M. W. Mueller, M. Hehn, and R. D’Andrea, “A computationally ef- ficient motion primitive for quadrocopter trajectory generation,”IEEE Transactions on Robotics, vol. 31, no. 6, pp. 1294–1310, 2015

    2015

  8. [16]

    Search-based motion planning for quadrotors using linear quadratic minimum time control,

    S. Liu, N. Atanasov, K. Mohta, and V . Kumar, “Search-based motion planning for quadrotors using linear quadratic minimum time control,” in2017 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2017, pp. 2872–2879

    2017

  9. [17]

    Real-time trajectory generation for quadro- copters,

    M. Hehn and R. D’Andrea, “Real-time trajectory generation for quadro- copters,”IEEE Transactions on Robotics, vol. 31, no. 4, pp. 877–892, 2015

    2015

  10. [18]

    Deep learning for opti- mization of trajectories for quadrotors,

    Y . Wu, X. Sun, I. Spasojevic, and V . Kumar, “Deep learning for opti- mization of trajectories for quadrotors,”IEEE Robotics and Automation Letters, vol. 9, no. 3, pp. 2479–2486, 2024

    2024

  11. [19]

    Fast uav trajectory optimization using bilevel optimization with analytical gradients,

    W. Sun, G. Tang, and K. Hauser, “Fast uav trajectory optimization using bilevel optimization with analytical gradients,” in2020 American Control Conference (ACC), 2020, pp. 82–87

    2020

  12. [20]

    Planning dynamically feasible trajectories for quadrotors using safe flight corridors in 3-d complex environments,

    S. Liu, M. Watterson, K. Mohta, K. Sun, S. Bhattacharya, C. J. Taylor, and V . Kumar, “Planning dynamically feasible trajectories for quadrotors using safe flight corridors in 3-d complex environments,”IEEE Robotics and Automation Letters, vol. 2, no. 3, pp. 1688–1695, 2017

    2017

  13. [21]

    FASTER: Fast and safe trajectory planner for navigation in unknown environments,

    J. Tordesillas and J. P. How, “FASTER: Fast and safe trajectory planner for navigation in unknown environments,”IEEE Transactions on Robotics, 2021

    2021

  14. [22]

    Time-optimal path tracking for robots: A convex optimization approach,

    D. Verscheure, B. Demeulenaere, J. Swevers, J. De Schutter, and M. Diehl, “Time-optimal path tracking for robots: A convex optimization approach,”IEEE Transactions on Automatic Control, vol. 54, no. 10, pp. 2318–2327, 2009. 16

    2009

  15. [23]

    Fitting smooth paths to speherical data,

    P. E. Jupp and J. T. Kent, “Fitting smooth paths to speherical data,”Journal of the Royal Statistical Society. Series C (Applied Statistics), vol. 36, no. 1, pp. 34–46, 1987. [Online]. Available: http://www.jstor.org/stable/2347843

    1987

  16. [24]

    Rotation interpolation based on the geometric structure of unit quaternions,

    D. Han, X. Fang, and Q. Wei, “Rotation interpolation based on the geometric structure of unit quaternions,” in2008 IEEE International Conference on Industrial Technology, 2008, pp. 1–6

    2008

  17. [25]

    Trajectory optimization on manifolds: A theoretically-guaranteed embedded sequential convex programming approach,

    R. Bonalli, A. Bylard, A. Cauligi, T. Lew, and M. Pavone, “Trajectory optimization on manifolds: A theoretically-guaranteed embedded sequential convex programming approach,” inRobotics: Science and Systems, Freiburg im Breisgau, Germany, June 2019. [Online]. Available: https:/...

    2019 arXiv

  18. [26]

    Geometrically constrained tra- jectory optimization for multicopters,

    Z. Wang, X. Zhou, C. Xu, and F. Gao, “Geometrically constrained tra- jectory optimization for multicopters,”IEEE Transactions on Robotics, vol. 38, no. 5, pp. 3259–3278, 2022

    2022

  19. [27]

    Control of quadrotors using the hopf fibration onSO(3),

    M. Watterson and V . Kumar, “Control of quadrotors using the hopf fibration onSO(3),” inRobotics Research, N. M. Amato, G. Hager, S. Thomas, and M. Torres-Torriti, Eds. Cham: Springer International Publishing, 2020, pp. 199–215

    2020

  20. [28]

    Trajectory optimization on manifolds with applications to quadrotor systems,

    M. Watterson, S. Liu, K. Sun, T. Smith, and V . Kumar, “Trajectory optimization on manifolds with applications to quadrotor systems,” The International Journal of Robotics Research, vol. 39, no. 2- 3, pp. 303–320, 2020. [Online]. Available: https://doi.org/10.1177/ 0278364919891775

    2020

  21. [29]

    Riemannian direct trajectory optimization of rigid bodies on matrix lie groups,

    S. Teng, T.-Y . Lin, W. A. Clark, R. Vasudevan, and M. Ghaffari, “Riemannian direct trajectory optimization of rigid bodies on matrix lie groups,”arXiv preprint arXiv:2505.02323, 2025

    2025 arXiv

  22. [30]

    Reach- ing the limit in autonomous racing: Optimal control versus reinforcement learning,

    Y . Song, A. Romero, M. Müller, V . Koltun, and D. Scaramuzza, “Reach- ing the limit in autonomous racing: Optimal control versus reinforcement learning,”Science Robotics, vol. 8, no. 82, p. eadg1462, 2023

    2023

  23. [31]

    Champion-level drone racing using deep reinforcement learning,

    E. Kaufmann, L. Bauersfeld, A. Loquercio, M. Müller, V . Koltun, and D. Scaramuzza, “Champion-level drone racing using deep reinforcement learning,”Nature, vol. 620, no. 7976, pp. 982–987, 2023

    2023

  24. [32]

    Learning high-speed flight in the wild,

    A. Loquercio, E. Kaufmann, R. Ranftl, M. Müller, V . Koltun, and D. Scaramuzza, “Learning high-speed flight in the wild,”Science Robotics, vol. 6, no. 59, p. eabg5810, 2021

    2021

  25. [33]

    Learning trajectories for real- time optimal control of quadrotors,

    G. Tang, W. Sun, and K. Hauser, “Learning trajectories for real- time optimal control of quadrotors,” in2018 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2018, pp. 3620– 3625

    2018

  26. [34]

    Real-time generation of time- optimal quadrotor trajectories with semi-supervised seq2seq learning,

    G. Ryou, E. Tal, and S. Karaman, “Real-time generation of time- optimal quadrotor trajectories with semi-supervised seq2seq learning,” in Proceedings of The 6th Conference on Robot Learning, ser. Proceedings of Machine Learning Research, K. Liu, D. Kulic, and J. Ichnowski, Eds...

    2023

  27. [35]

    Accelerating trajectory generation for quadrotors using transformers,

    S. Tankasala and M. Pryor, “Accelerating trajectory generation for quadrotors using transformers,” inProceedings of The 5th Annual Learning for Dynamics and Control Conference, ser. Proceedings of Machine Learning Research, N. Matni, M. Morari, and G. J. Pappas, Eds., vol. 211...

    2023

  28. [36]

    Multi-fidelity reinforcement learning for time-optimal quadrotor re-planning,

    G. Ryou, G. Wang, and S. Karaman, “Multi-fidelity reinforcement learning for time-optimal quadrotor re-planning,” 2024. [Online]. Available: https://arxiv.org/abs/2403.08152

    2024 arXiv

  29. [37]

    Sequence modeling for time-optimal quadrotor trajectory optimization with sampling-based robustness analysis,

    K. Mao, H. Yu, R. Zhang, I. Spasojevic, S. Gao, and V . Kumar, “Sequence modeling for time-optimal quadrotor trajectory optimization with sampling-based robustness analysis,” inConference on Robot Learning. PMLR, 2025, pp. 1531–1540

    2025

  30. [38]

    Dynamically feasible trajectory gen- eration with optimization-embedded networks for autonomous flight,

    Z. Han, L. Xu, L. Pei, and F. Gao, “Dynamically feasible trajectory gen- eration with optimization-embedded networks for autonomous flight,” IEEE Robotics and Automation Letters, 2025

    2025

  31. [39]

    Optnet: Differentiable optimization as a layer in neural networks,

    B. Amos and J. Z. Kolter, “Optnet: Differentiable optimization as a layer in neural networks,” inProceedings of the 34th International Conference on Machine Learning - Volume 70, ser. ICML’17. JMLR.org, 2017, p. 136–145

    2017

  32. [40]

    Differentiable convex optimization layers,

    A. Agrawal, B. Amos, S. Barratt, S. Boyd, S. Diamond, and J. Z. Kolter, “Differentiable convex optimization layers,”Advances in neural information processing systems, vol. 32, 2019

    2019

  33. [41]

    Dc3: A learning method for optimization with hard constraints,

    P. Donti, D. Rolnick, and J. Z. Kolter, “Dc3: A learning method for optimization with hard constraints,” inInternational Conference on Learning Representations, 2021

    2021

  34. [42]

    Learning differ- entiable solvers for systems with hard constraints,

    G. Négiar, M. W. Mahoney, and A. Krishnapriyan, “Learning differ- entiable solvers for systems with hard constraints,” inThe Eleventh International Conference on Learning Representations, 2023

    2023

  35. [43]

    Hard-constrained neural networks with universal approximation guarantees,

    Y . Min, A. Sonar, and N. Azizan, “Hard-constrained neural networks with universal approximation guarantees,” 2024. [Online]. Available: https://arxiv.org/abs/2410.10807

    2024

  36. [44]

    Physics-informed neural networks: A deep learning framework for solving forward and inverse problems involving nonlinear partial differential equations,

    M. Raissi, P. Perdikaris, and G. Karniadakis, “Physics-informed neural networks: A deep learning framework for solving forward and inverse problems involving nonlinear partial differential equations,” Journal of Computational Physics, vol. 378, pp. 686–707, 2019. [Online]. Ava...

    2019

  37. [45]

    Physics-informed neural networks with hard constraints for inverse design,

    L. Lu, R. Pestourie, W. Yao, Z. Wang, F. Verdugo, and S. G. Johnson, “Physics-informed neural networks with hard constraints for inverse design,”SIAM Journal on Scientific Computing, vol. 43, no. 6, pp. B1105–B1132, 2021. [Online]. Available: https://doi.org/10.1137/ 21M1397908

    2021

  38. [46]

    Enhanced physics-informed neural networks with augmented lagrangian relaxation method (al- pinns),

    H. Son, S. W. Cho, and H. J. Hwang, “Enhanced physics-informed neural networks with augmented lagrangian relaxation method (al- pinns),”Neurocomputing, vol. 548, p. 126424, 2023. [Online]. Available: https://www.sciencedirect.com/science/article/pii/S0925231223005477

    2023

  39. [47]

    Learning to sequence and blend robot skills via differentiable optimization,

    N. Jaquier, Y . Zhou, J. Starke, and T. Asfour, “Learning to sequence and blend robot skills via differentiable optimization,”IEEE Robotics and Automation Letters, vol. 7, no. 3, pp. 8431–8438, 2022

    2022

  40. [48]

    Dif- ferentiable mpc for end-to-end planning and control,

    B. Amos, I. D. J. Rodriguez, J. Sacks, B. Boots, and J. Z. Kolter, “Dif- ferentiable mpc for end-to-end planning and control,” inProceedings of the 32nd International Conference on Neural Information Processing Systems, ser. NIPS’18. Red Hook, NY , USA: Curran Associates Inc.,...

    2018

  41. [49]

    Rayen: Imposition of hard convex constraints on neural networks,

    J. Tordesillas, J. P. How, and M. Hutter, “Rayen: Imposition of hard convex constraints on neural networks,” 2023. [Online]. Available: https://arxiv.org/abs/2307.08336

    2023 arXiv

  42. [50]

    Hamiltonian-based Neural ODE Networks on the SE(3) Manifold For Dynamics Learning and Control,

    T. Duong and N. Atanasov, “Hamiltonian-based Neural ODE Networks on the SE(3) Manifold For Dynamics Learning and Control,” inPro- ceedings of Robotics: Science and Systems, Virtual, July 2021

    2021

  43. [51]

    Lietransformer: equivariant self-attention for lie groups,

    M. J. Hutchinson, C. Le Lan, S. Zaidi, E. Dupont, Y . W. Teh, and H. Kim, “Lietransformer: equivariant self-attention for lie groups,” in International Conference on Machine Learning. PMLR, 2021, pp. 4533–4543

    2021

  44. [52]

    Learn- ing so (3) equivariant representations with spherical cnns,

    C. Esteves, C. Allen-Blanchette, A. Makadia, and K. Daniilidis, “Learn- ing so (3) equivariant representations with spherical cnns,” inProceed- ings of the European Conference on Computer Vision (ECCV), 2018, pp. 52–68

    2018

  45. [53]

    Learning riemannian manifolds for geodesic motion skills,

    H. Beik-Mohammadi, S. Hauberg, G. Arvanitidis, G. Neumann, and L. Rozo, “Learning riemannian manifolds for geodesic motion skills,” inRobotics: Science and System XVII, July 12 - July 16, 2021 –Held Virtually–, Ed.: D. A. Shell, 2021

    2021

  46. [54]

    Fast and robust shortest paths on manifolds learned from data,

    G. Arvanitidis, S. Hauberg, P. Hennig, and M. Schober, “Fast and robust shortest paths on manifolds learned from data,” inThe 22nd Interna- tional Conference on Artificial Intelligence and Statistics. PMLR, 2019, pp. 1506–1515

    2019

  47. [55]

    Meshdmp: Motion planning on discrete manifolds using dynamic movement primitives,

    M. D. Vedove, F. J. Abu-Dakka, L. Palopoli, D. Fontanelli, and M. Saveriano, “Meshdmp: Motion planning on discrete manifolds using dynamic movement primitives,” in2025 IEEE International Conference on Robotics and Automation (ICRA), 2025, pp. 895–901

    2025

  48. [56]

    Equivariant motion manifold primitives,

    B. Lee, Y . Lee, S. Kim, M. Son, and F. C. Park, “Equivariant motion manifold primitives,” inProceedings of The 7th Conference on Robot Learning, ser. Proceedings of Machine Learning Research, J. Tan, M. Toussaint, and K. Darvish, Eds., vol. 229. PMLR, 06–09 Nov 2023, pp. 1199...

    2023

  49. [57]

    Mmp++: Motion manifold primitives with parametric curve models,

    Y . Lee, “Mmp++: Motion manifold primitives with parametric curve models,”IEEE Transactions on Robotics, vol. 40, pp. 3950–3963, 2024

    2024

  50. [58]

    Fast kinodynamic planning on the constraint manifold with deep neural networks,

    P. Kicki, P. Liu, D. Tateo, H. Bou-Ammar, K. Walas, P. Skrzypczy ´nski, and J. Peters, “Fast kinodynamic planning on the constraint manifold with deep neural networks,”IEEE Transactions on Robotics, vol. 40, pp. 277–297, 2024

    2024

  51. [59]

    M. P. Do Carmo and J. Flaherty Francis,Riemannian geometry. Springer, 1992, vol. 2

    1992

  52. [60]

    Metrics and connections for rigid-body kinematics,

    M. Žefran, V . Kumar, and C. Croke, “Metrics and connections for rigid-body kinematics,”The International Journal of Robotics Research, vol. 18, no. 2, pp. 242–1–242–16, 1999. [Online]. Available: https://doi.org/10.1177/027836499901800208

    1999 doi

  53. [61]

    Kinematic dexterity of robotic mechanisms,

    F. C. Park and R. W. Brockett, “Kinematic dexterity of robotic mechanisms,”The International Journal of Robotics Research, vol. 13, no. 1, pp. 1–15, 1994. [Online]. Available: https://doi.org/10.1177/ 027836499401300101

    1994

  54. [62]

    Learning with 3D rotations, a hitchhiker’s guide to SO(3),

    A. R. Geist, J. Frey, M. Zhobro, A. Levina, and G. Martius, “Learning with 3D rotations, a hitchhiker’s guide to SO(3),” inProceedings of the 41st International Conference on Machine Learning, ser. Proceedings of Machine Learning Research, vol. 235. PMLR, 21–27 Jul 2024, pp. 1...

    2024

  55. [63]

    Embedding-based interpolation on the special orthogonal group,

    E. S. Gawlik and M. Leok, “Embedding-based interpolation on the special orthogonal group,”SIAM Journal on Scientific Computing, vol. 40, no. 2, pp. A721–A746, 2018. [Online]. Available: https: //doi.org/10.1137/17M1129416

    2018 doi

  56. [64]

    Polynomial trajectory planning for aggressive quadrotor flight in dense indoor environments,

    C. Richter, A. Bry, and N. Roy, “Polynomial trajectory planning for aggressive quadrotor flight in dense indoor environments,” inRobotics Research: The 16th International Symposium ISRR. Springer, 2016, pp. 649–666

    2016

  57. [65]

    SciPy 1.0: Fundamental algorithms for scientific computing in python,

    P. Virtanen, R. Gommers, T. E. Oliphant, M. Haberland, T. Reddy, D. Cournapeau, E. Burovski, P. Peterson, W. Weckesser, J. Bright, S. J. van der Walt, M. Brett, J. Wilson, K. J. Millman, N. Mayorov, A. R. J. Nelson, E. Jones, R. Kern, E. Larson, C. Carey, ˙Ibrahim Polat, Y . F...

    2020

  58. [66]

    Dispersion- minimizing motion primitives for search-based motion planning,

    L. Jarin-Lipschitz, J. Paulos, R. Bjorkman, and V . Kumar, “Dispersion- minimizing motion primitives for search-based motion planning,” in 2021 IEEE International Conference on Robotics and Automation (ICRA), 2021, pp. 12 625–12 631

    2021

  59. [67]

    Primitive-swarm: An ultra-lightweight and scalable planner for large- scale aerial swarms,

    J. Hou, X. Zhou, N. Pan, A. Li, Y . Guan, C. Xu, Z. Gan, and F. Gao, “Primitive-swarm: An ultra-lightweight and scalable planner for large- scale aerial swarms,”IEEE Transactions on Robotics, vol. 41, pp. 3629– 3648, 2025

    2025

  60. [68]

    RotorPy: A python-based multirotor simulator with aerodynamics for education and research,

    S. Folk, J. Paulos, and V . Kumar, “RotorPy: A python-based multirotor simulator with aerodynamics for education and research,”arXiv preprint arXiv:2306.04485, 2023

    2023 arXiv

  61. [69]

    Mighty: Hermite spline- based efficient trajectory planning,

    K. Kondo, Y . Wu, V . Kumar, and J. P. How, “Mighty: Hermite spline- based efficient trajectory planning,”IEEE Robotics and Automation Letters, vol. 11, no. 6, pp. 6664–6671, 2026

    2026

Pith tools

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