Vector Fields for Path Following on Lie Groups with Application in Robot Control

Felipe Bartelt; Luciano C. A. Pimenta; Vinicius M. Gon\c{c}alves; Weijia Yao

REVIEW 3 major objections 4 minor 2 cited by

Vector Fields for Path Following on Lie Groups with Application in Robot Control

T0 review · 3 major / 4 minor · reviewed 2026-08-02 · deepseek-v4-flash

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

Pith's one-line read This paper proves that a single explicit vector field on a matrix Lie group — for SE(3), a six-number twist field — drives a fully actuated robot onto a prescribed pose curve and keeps it moving along the curve, with a real manipulator expe

desk verdict A genuinely useful Lie-group generalization of guiding vector fields, with a real SE(3) twist-level algorithm; the main theorem mostly holds, but the escapability argument and the 'almost all' claim need work before trusting the abstract. read the letter →

arxiv 2602.21450 v2 pith:R3BNBD6E submitted 2026-02-25 cs.RO cs.SYeess.SY

Felipe Bartelt , Luciano C. A. Pimenta , Weijia Yao , Vinicius M. Gonçalves This is my paper · ORCID

classification cs.ROcs.SYeess.SY

keywords vectorfieldsLiegroupspathfollowingSE(3)robotcontrolposetrackingguidancenavigationandtwist

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

The paper aims to show that path following on a connected matrix Lie group can be solved by one explicit vector field, without splitting position and orientation into separate controllers. For the six-dimensional pose group SE(3), the control input is the mechanical twist, so the same six numbers can be sent directly to an omnidirectional drone or a manipulator end-effector. The argument identifies three properties of the distance-to-curve function — left-invariance, chainability, and local linearity — that are sufficient for the field to converge to the curve from almost every initial state and then traverse it with non-zero velocity. A real seven-axis manipulator experiment is presented as evidence that the construction works in real time.

What carries the argument

The central machinery is a pair of Lie-group operators. The Ξ operator extracts the "velocity vector" of a curve on the group — the basis coefficients of the Lie-algebra element that right-translates the tangent — and the L operator plays the role of the gradient of a scalar function on the group. Together they let the authors write the time derivative of the distance-to-curve as ˙D = L[D](H)ξ and define the normal component as the negative transpose of L[D](H). Three distance properties do the load-bearing work: left-invariance yields orthogonality of the normal and tangent components; chainability excludes local minima off the curve; local linearity keeps the normal component from vanishin

What would settle it

Simulate the closed-loop system on a curve whose parametrization has a stationary point (ξ_d(s_0)=0 for some s_0). The paper asserts every curve can be reparametrized to be proper, so the field should still traverse with non-zero velocity; if the state stalls at that point or the Lyapunov derivative stops being negative, the almost-everywhere convergence claim needs a regularity assumption.

Watch

Extended reading notes

Core claim

The central claim is Theorem III.17: if the element-to-element distance is left-invariant, chainable, and locally linear, and the curve parametrization is proper (its twist is never zero), then the closed-loop system ˙H = S(ξ)H with ξ = k_N ξ_N + k_T ξ_T converges either to the target curve C or to a singular set P consisting of non-unique nearest points and non-differentiability points; P is escapable in finite time by arbitrarily small control inputs; and whenever the state converges to C, the tangent term k_T ξ_T never vanishes, so the curve is traversed. The proof works because a left-invariant distance makes the normal and tangent components orthogonal, turning the distance into a Lyapu

Load-bearing premise

The proof needs the target curve to have no stationary tangent and the distance to stay differentiable along the whole motion; on SE(3) that means the robot is never exactly a half-turn away from its nearest curve point, and the experiments never verify this condition.

Editorial extensions

If this is right

On any connected exponential matrix Lie group — SO(n), SE(n), the Heisenberg group — the same construction yields a path-following vector field with control inputs in the dimension of the Lie algebra.
On SE(3) the control input is the six-dimensional twist, so the field can be applied directly to omnidirectional drones and manipulator end-effectors without projecting a higher-dimensional embedding controller.
Almost-global convergence follows: the only failure set is P, which is escapable in finite time by arbitrarily small perturbations, and initial conditions closer to the curve than D_min,C converge to C without ever entering P.
In Euclidean space the framework reduces to the earlier vector-field method and admits any distance with the three properties, so ℓ^p norms with p ≥ 2 are allowed rather than only the Euclidean norm.
Once on the curve, the non-zero tangent term guarantees traversal in the direction fixed by the parametrization, not merely convergence to the curve.

Reading between the lines

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

The paper claims that non-proper parametrizations can always be made proper by reparametrization, but a curve that genuinely pauses (zero tangent at a point) cannot be made proper by any reparametrization. A charitable reading is that the theorem silently assumes regular curves; this should be stated as an explicit premise, and the proof extended or the claim restricted accordingly.
The SE(3) distance is non-differentiable at relative rotations of exactly π, so the positive separation D_min,C = √2π is a real planning constraint. Path designers should keep the curve and its neighbourhood inside the set of relative poses with rotation angle strictly less than π, which is not spelled out in the paper.
The measured 8.9 ms per iteration is dominated by the brute-force nearest-point search. A hierarchical or GPU-parallel search, which the paper notes is possible, would be a natural engineering extension and could move the field to kilohertz update rates.

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.

Desk Editor's Note

read the letter

Short version: this paper extends the Euclidean guiding-vector-field construction to connected matrix Lie groups and lands on a practical SE(3) specialization where the control is the 6-D twist. That is a real advance over Rezende et al. (Euclidean only) and Yao et al. (embedding-space controls). The architecture is clean: left-invariant, chainable, locally linear distances; orthogonality; Lyapunov derivative Ḋ = −k_N‖ξ_N‖². Proposition III.19 correctly proves the log-norm EE-distance satisfies the axioms, and the Euclidean instantiation recovers Rezende et al. The paper is worth reading.

Soft spots, in order of size. First, Theorem III.17(ii) — escapability — is not actually proven. The proof postulates "a non-zero minimum decrease δ" at every step, but no uniform lower bound is established; D(H) can vary over P and the perturbation can be arbitrarily small. Without that, the conclusion that P can be escaped in finite time is an assertion. Second, the abstract's "almost all initial conditions" overstates the theorem: the proven dichotomy is convergence to C or P, and P is not shown to be measure zero or non-attracting. That said, the reader's circularity concern doesn't land: the theorem is a constructive proof, not a fit, and the gains are free design choices. Third, the claim that any non-proper parametrization can be reparametrized to be proper is false for curves with stationary points; regularity is assumed. For the SE(3) log-norm, the separation condition D_min,C = √2π means the relative pose to the nearest curve point must never be a π-rotation, and the experimental curve (§IV.B) is never checked against that. Fourth, the experiment is one 150 s run, no baselines, no error bars, and it has a Jacobian pseudoinverse layer between the theory and the robot. The 8.9 ms cost is useful, but 99.5% in nearest-point search means the headline number is mostly about the brute-force search, not the vector field.

None of this sinks the central idea. The Lie-group framework and the twist-level SE(3) algorithm are new and likely useful for omnidirectional UAVs and manipulators. The gaps are fixable: a proper proof of escapability (or a restriction to D(0) < D_min,P), verification/assumption of the separation condition, and at least one repeat with a baseline. This deserves a serious referee; I'd send it out, expecting major revision. The paper is honest enough to flag its own exponential-group restriction and simplified distance search.

Referee Report

3 major / 4 minor

Summary. The paper develops a vector-field path-following method for fully-actuated systems on connected matrix Lie groups. It defines an Element-to-Curve distance, normal and tangent vector-field components, and identifies sufficient conditions (left-invariance, chainability, local linearity) under which a dichotomy theorem is claimed: trajectories converge either to the desired curve C or to a singular set P, with P claimed to be escapable. The construction is instantiated on exponential Lie groups using the log-Frobenius distance, specialized to SE(3) with an efficient nearest-point algorithm, and validated in real experiments on a Kinova Gen3 manipulator. An open-source implementation is provided.

Significance. The framework is a meaningful extension of Euclidean vector-field guidance to Lie groups, and the identification of the three abstract distance properties is elegant. The orthogonality proof via left-invariance and the log-distance chainability result are nontrivial and useful. The paper also ships an open-source implementation and reports a real-robot experiment, which are concrete strengths. If the gaps identified below are repaired, the contribution would be solid for the journal.

major comments (3)

[Theorem III.17 and abstract] The abstract's 'almost all initial conditions' is not supported. Theorem III.17(i) proves only convergence to C or P. Statement (ii) is about an external perturbation policy ('there exists a policy of choosing arbitrarily small ξ every time H∈P'), not about the closed-loop system (7)-(8); the vector field (8) itself may converge to and remain at P. The proof asserts 'a non-zero minimum decrease δ' with no uniform lower bound; D(H) may vary over P and the perturbations are allowed to be arbitrarily small. Also D_min,P ≜ min_{H∈P} D(H) need not be attained. To support almost-global convergence one must prove P is measure zero and non-attracting under (8), or weaken the abstract/claim.
[Lemma III.5 / Theorem III.17(i)] The first-order optimality condition in Lemma III.5 is asserted for 'the optimal parameter s*' but D is minimized over the closed interval [0,1]. If s*=0 or s*=1, the derivative d/ds bD(H,H_d(s))|_{s=s*} need not vanish. Lemma III.5 is then used in Lemma III.6 to eliminate the ds*/dt term in (11) and in Proposition III.10 to prove orthogonality, so the key inequality ˙D = −k_N‖ξ_N‖² in (17) is not established for points whose nearest curve point is an endpoint. The paper should either exclude boundary minima (e.g., extend the curve or prove s* is interior on the relevant region) or handle them as part of P.
[Section III, before Def. III.4] The claim that any non-proper parametrization of a non-self-intersecting curve can be transformed into a proper one by reparametrization is false. A reparametrization cannot remove a stationary point: e.g., H_d(s)=(s^2,0) on [0,1] is bijective but not proper, and no smooth reparametrization yields ξ_d≠0 at s=0. Thus regularity of the curve is an additional assumption, not a consequence of non-self-intersection. This matters because the paper uses the claim to motivate why parametric curves avoid null-tangent issues.

minor comments (4)

[Definition III.14] The sentence 'bD(A,Φ(σ,A,B)) ≈ o(σ)' should read 'O(σ)' (or 'linear in σ'); as written it contradicts the required positive limit in the same definition.
[Section III.F] The logarithm log is defined by 'any matrix Y' for Z not in R^{n×n}_+, which is not a function. A fixed selection should be specified to make bD well-defined and the chainability proof unambiguous.
[General] The arXiv listing title ('Vector Fields for Path Following on Lie Groups with Application in Robot Control') differs from the internal title ('Constructive Vector Fields for Path Following in Fully-Actuated Systems on Matrix Lie Groups'). Ensure consistency in the final version.
[Section IV.B/IV.E] The experimental curve is designed in joint space and mapped to SE(3), but the assumptions of Theorem III.17 (properness of the discretized curve and that the trajectory never enters P) are not verified. A short statement confirming these checks would strengthen the experimental claim.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: Theorem III.17 is derived from stated definitions and verified distance properties; self-citations are contextual lineage, not load-bearing.

full rationale

The central derivation is self-contained. Theorem III.17 is proved from Definitions III.2–III.14 and auxiliary lemmas (III.5–III.16), and Proposition III.19 independently verifies that the log-norm distance on exponential Lie groups is left-invariant, chainable, and locally linear. No parameter is fitted to experimental data: the gains k_N and k_T are free design choices, and the experimental section reports measured execution times and tracking errors rather than predictions derived from tuned quantities. The citations to Rezende et al. (2022) and Goncalves et al. (2010) are used as background and as a Euclidean special case, but the Lie-group proofs do not invoke those cited results as black boxes; Section III-G re-derives the Euclidean case from the new framework. Thus no step reduces an output to its own input by construction. Correctness caveats exist but are not circularity: (i) Theorem III.17(ii) asserts 'there is a non-zero minimum decrease δ that can be obtained at all steps' without proof, so finite-time escapability of P and the abstract's 'almost all initial conditions' are not fully established; (ii) Definition III.4's statement that 'any non-proper parametrization can be transformed into a proper one through reparametrization' is false for curves with stationary points; (iii) D_min,C > 0 is not verified for the experimental SE(3) curve. These are proof gaps or assumption violations under correctness risk, not equivalences between inputs and claimed results.

Assumptions & free parameters 3 free parameters · 6 assumptions · 0 invented entities

No new physical entities. The Ξ and L operators are re-descriptions of the standard left-trivialized derivative/velocity (the paper acknowledges they are 'not fully novel'); chainability and local linearity are abstract sufficient conditions, not entities. The free parameters are implementation-level; the theorem itself is parameter-free. The main unsupported premises are the regularity/separation of the target curve and the uniform-δ premise inside the escapability proof.

free parameters (3)

Gains k_N, k_T = k_N(D) = 0.1 tanh(0.75√D); k_T(D) = 0.03(1 − tanh(0.75√D))
Hand-chosen in §IV.D for the experiment. Theorem III.17 holds for any positive continuous k_N vanishing on C and positive k_T, so these do not support the theoretical claim; they only affect transient behavior.
Finite-difference step ϵ for L_V[bD] = 0.001
Numerical derivative used in the SE(3) implementation (§IV.A). The proofs assume exact derivatives, so the implementation approximates the theory.
Jacobian damping ε = 1e-4
Regularization in the joint-velocity pseudoinverse for the manipulator experiment (§IV, step 4); outside the Lie-group theory.

assumptions (6)

standard math Lie groups are parallelizable via right-invariant vector fields; every smooth curve G(s) admits a right-trivialized velocity A(s) ∈ g (Lemma II.1, II.3).
Standard fact cited to Grigorian 2024 / Lee 2012.
standard math Matrix exponential/log branch structure: for exponential Lie groups exp is surjective, and a log with eigenvalues in the strip Im ∈ [−π, π] satisfies log(exp(r log Z)) = r log Z for r ∈ [0,1] (Lemma III.18).
Standard analytic facts, but they restrict the construction to exponential Lie groups (acknowledged) and require the asserted branch properties for Z with negative eigenvalues.
domain assumption Control system Ḣ = S(ξ)H — first-order, fully actuated at the twist level (eq. 7, §II.D).
The whole robot-facing theory assumes a pose-level single integrator; real manipulators need the extra Jacobian pseudoinverse layer used in the experiment.
domain assumption Target curve C is non-self-intersecting and admits a proper parametrization (ξ_d(s) ≠ 0) (Definition III.4 and preceding paragraph).
The paper claims any non-proper parametrization can be repaired by reparametrization; false for non-regular curves, so regularity must be assumed.
domain assumption EE-distance satisfies Definition III.2 (bounded directional limits where non-differentiable) with a positive separation D_min,C between C and the singular set P.
Needed for Lemma III.16 and for P-escapability; for the SE(3) log-norm it requires nearest-point relative poses never to be π-rotations.
ad hoc to paper Theorem III.17(ii) postulates a uniform minimum decrease δ of D obtainable by arbitrarily small perturbations at every H ∈ P.
The escape-policy proof sketches this but does not establish that a uniform δ exists over the whole singular set.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Vector Fields for Path Following on Lie Groups with Application in Robot Control." pith.science (2026). https://pith.science/paper/R3BNBD6E

@misc{pith2026260221450,
  author       = {Pith},
  title        = {Pith review of: Vector Fields for Path Following on Lie Groups with Application in Robot Control},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/R3BNBD6E}},
  note         = {Machine review of arXiv:2602.21450}
}

read the original abstract

Many robotic systems allow independent control of position and orientation (pose), including omnidirectional aerial vehicles, underwater robots, and manipulator end-effectors. In many applications, these systems must follow a continuous sequence of poses, leading to either trajectory-tracking or path following formulations. Compared to trajectory-tracking, path following offers important practical advantages. In particular, we focus on the problem of path following on Lie groups. Considering the robots as rigid bodies moving in the 3D space, this path-following problem can be posed as a problem of designing guiding vector fields on the matrix Lie group SE(3). In this paper, we develop a general vector-field framework for path following on connected matrix Lie groups, of which SE(3) is a prominent special case. The proposed vector field guarantees convergence to a desired parametric curve from almost all initial conditions while ensuring continuous motion along the path. Furthermore, another interesting feature is that, as opposed to previous works, the control input is "minimal" in terms of representation and closer to the engineering application (e.g., the body twist in the case SE(3)). After establishing the general case, the framework is then specialized to SE(3), of special interest in robotics, yielding an efficient algorithm suitable for real-time robotic control. Experiments with a robotic manipulator tracking complex pose paths demonstrate the effectiveness of the approach. An open-source implementation is also provided.

Figures

Figures reproduced from arXiv: 2602.21450 by the authors.

**Figure 2.** Example showing the vector field and the components [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗

**Figure 3.** Depiction of the proof in Proposition III.13. [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗

Figures from the paper (3 more)

**Figure 4.** Figure 4: Depiction of Statement (ii) in the proof of Theo [PITH_FULL_IMAGE:figures/full_fig_p010_4.png]

**Figure 5.** Figure 5: The solid cyan line depicts the target curve, with target orientation frames shown by RGB axes. Intermediary [PITH_FULL_IMAGE:figures/full_fig_p014_5.png]

**Figure 6.** Figure 6: Value of EC-distance D, position error in centimeters and orientation error in degrees along time. was limited to parametric curve representations in Euclidean space. To achieve this broader applicability, we examined key properties of distance functions that maintain …

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 2 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

Efficient Computation of Distance Functions for Navigation Vector Fields in Lie Groups
cs.RO 2026-06 conditional novelty 6.0 of 10

A method reduces distance computation to G-polynomial curves on Lie groups to a small number of polynomial root findings, cutting computation time for SE(3) navigation vector fields while preserving accuracy.
Constrained Optimization on Matrix Lie Groups via Interior-Point Method
math.OC 2026-03 conditional novelty 5.0 of 10

A metric-free interior-point optimizer on matrix Lie groups with Lie-algebra Newton steps is claimed to converge quadratically and outperform Riemannian IPMs on SO(7) and SL(7) benchmarks.

Reference graph

Works this paper leans on

2 extracted references · cited by 2 Pith papers

[2020]

A. H. D. Nunes, G. V . Raffo, and L. C. A. Pimenta. Integrated vector field and backstepping control for quadcopters. In 2023 IEEE International Conference on Robotics and Au- tomation (ICRA), pages 1256–1262. IEEE,

2023
[2021]

W. Yao, B. Lin, B. D. O. Anderson, and M. Cao. Topological analysis of vector-field guided path following on manifolds. IEEE Transactions on Automatic Control, 68(3):1353–1368, 2023

2023

Pith tools

Reviewed August 2, 2026 · model on record in the stance chip above.

[1] [2020]

A. H. D. Nunes, G. V . Raffo, and L. C. A. Pimenta. Integrated vector field and backstepping control for quadcopters. In 2023 IEEE International Conference on Robotics and Au- tomation (ICRA), pages 1256–1262. IEEE,

2023

[2] [2021]

W. Yao, B. Lin, B. D. O. Anderson, and M. Cao. Topological analysis of vector-field guided path following on manifolds. IEEE Transactions on Automatic Control, 68(3):1353–1368, 2023

2023