Pith. sign in

REVIEW 4 major objections 7 minor 14 references

REVISE: Robust Probabilistic Motion Planning in a Gaussian Random Field

T0 review · 4 major / 7 minor · reviewed 2026-08-12 · deepseek-v4-flash

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

Pith's one-line read Rewiring belief roadmaps provably reaches more goal distributions.

desk verdict A genuinely useful robust edge controller and rewiring scheme for belief roadmaps, but the coverage theorem is not yet proved and the safety claim overreaches. read the letter →

arxiv 2411.13369 v1 pith:UZ5PZ2S2 submitted 2024-11-20 cs.RO cs.SYeess.SY

Alex Rose , Naman Aggarwal , Christopher Jewison , Jonathan P. How This is my paper · ORCID
classification cs.ROcs.SYeess.SY
keywords beliefroadmapcovariancesteeringGaussianrandomfieldchanceconstraintssigmapointsedgerewiringRRT*motionplanningunderuncertainty
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

REVISE tackles a practical problem: a robot whose linear dynamics are perturbed by a spatially correlated random field, such as wind over a map, and which must steer an initial Gaussian belief to a goal Gaussian belief while respecting chance constraints. The paper argues that combining a robust sigma-point edge controller with an edge-rewiring step in belief roadmap construction yields a roadmap with provably at least as much coverage as a non-rewired roadmap, and with much better plan accuracy and cost in practice. If true, this gives multi-query planners a way to precompute a library of robust stochastic maneuvers that can be reused for many goals, even in nonlinear disturbance fields.

What carries the argument

Two mechanisms carry the argument. The first is the robust edge controller: for each of $4n$ $\sigma$ points placed on the $\sqrt{n}$ covariance contour of the initial distribution, it linearizes the Gaussian random field around both the nominal mean trajectory and the $\sigma$-point trajectory, forms a worst-case second-moment matrix $S^{(i)}$, and solves a semidefinite program (Problem IV.1) that minimizes the largest eigenvalue of a matrix $\Sigma_f$ dominating all these contributions. The second is the rewiring step (Algorithm 2), which checks whether steering through a newly added node lowers the covariance of neighboring nodes and recursively propagates the lower cost to descendants, matching CC-RRT*; the coverage proof rests on Lemmas VII.1 and VII.2, which establish monotonicity of the edge objective in the initial covariance and equality of node means across the two algorithms.

What would settle it

Run the single-query experiment with the high-variance box (wind variance 6 vs 0.2 m²/s²) and compare the predicted final covariance envelope from Problem IV.1 with empirical Monte Carlo covariance over 100 seeds; if the empirical $\lambda_{\max}$ of the final state covariance exceeds the predicted $\lambda_{\max}(\Sigma_f)$ plus the chance-constraint slack, the robustness envelope is not conservative and the central claim fails.

Watch

Extended reading notes

Core claim

The central discovery is that edge rewiring, borrowed from RRT* and applied to belief roadmaps whose edges are covariance-steering controllers, preserves or improves the set of goal distributions reachable from the initial belief. The proof (Theorem V.1) works for the convex objective of Problem III.1 and, under two mild conditions, for the robust $\sigma$-point objective of Problem IV.1: both algorithms produce the same node means under the same random seed, but the rewired tree has node covariances that are no larger, so any goal reachable in the non-rewired tree is reachable in the rewired tree. Complementing this, the robust edge controller samples $4n$ $\sigma$ points on the $\sqrt{n}$ covariance contour and minimizes the worst-case contribution to the final second moment, which makes planned trajectories track the actual closed-loop distribution much more closely: the experiments report a 10x reduction in median Wasserstein distance in multi-query planning and a 2.5x reduction in the largest eigenvalue of the planned goal covariance in single-query planning.

Load-bearing premise

The method assumes that the worst-case second moment computed from a fixed set of 4n sigma points and two linearizations of the disturbance field is a conservative envelope for the true reachable distribution; if the real disturbance variation is larger than this envelope, the accuracy gains and safe constraint satisfaction do not transfer.

Editorial extensions

If this is right

  • Any goal distribution reachable from a non-rewired belief roadmap is reachable from the rewired roadmap, so existing plans remain valid after switching to REVISE.
  • Rewiring propagates lower-cost covariances to descendant nodes, reducing the largest eigenvalue of the planned covariance at the goal in single-query planning.
  • The robust sigma-point edge controller can be used as a drop-in edge policy for other belief roadmap constructions, trading some conservatism for substantially higher plan accuracy.
  • Because the coverage theorem does not require an admissible cost for RRT*, the rewiring technique applies to objectives like Problem III.1 that are not designed for optimal rewiring.

Reading between the lines

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

  • This suggests that the largest accuracy gains will appear in environments with strong spatial gradients in the disturbance field, since that is precisely where a single linearization diverges from the sigma-point envelope; in nearly uniform fields the robust objective adds conservatism without much benefit.
  • A natural testable extension is to enforce the chance constraints using the robust envelope itself, which the paper defers to future work; this would trade a modest computational increase for better constraint satisfaction in high-variance regions.
  • The monotonicity conditions in Lemma VII.1 may transfer to other state-dependent disturbance models beyond Gaussian random fields, as long as the edge controller's output covariance is monotone in the input covariance.
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 / 7 minor

Summary. The paper proposes REVISE, a belief-roadmap algorithm for covariance steering of linear systems disturbed by a state-dependent Gaussian random field. The edge controller (Problem IV.1) approximates the initial distribution by 4n sigma points and minimizes the maximum final-state second-moment contribution via an SDP. The roadmap is built by a rewiring algorithm (Algorithm 2) that the authors claim provably improves coverage relative to no rewiring (Theorem V.1). Experiments on a 6-DoF triple-integrator quadrotor in a 2D wind field report a 10x median Wasserstein accuracy improvement in multi-query planning and a 2.5x reduction in planned final covariance in single-query planning.

Significance. If the theoretical claims were fully established, the rewiring result would be a useful extension of RRT* ideas to belief space, and the sigma-point robust objective addresses a real need: disturbance statistics that depend on the realized state. The SDP formulation is coherent, and the claim that the semidefinite relaxation for Sigma_f is lossless is correct as stated. The experiments are reproducible in structure and show consistent gains. However, the paper's central theorem currently rests on unproved lemmas and a pseudocode bug, and the robust guarantee for state constraints is not delivered. The contribution is therefore promising but not yet at the level claimed.

major comments (4)
  1. [§V, Algorithm 2, lines 6-10, 14, 23] The second call to the edge controller at line 10 ignores the status flag. If the optimization returns failure, the node is still added to V at line 19 with the edge e_{k->q} from line 10, even though that edge need not be feasible. The same problem occurs at lines 14 and 23, where covariances are compared without checking whether the corresponding steering maneuver was feasible. Consequently, V can contain nodes that are not reachable from I, contradicting the invariant used in Theorem V.1. The pseudocode must check status at each Pi call and handle infeasibility explicitly, for example by falling back to the edge computed from the Vsample parent or by rejecting the candidate.
  2. [§VII-A, Lemma VII.2 and Theorem V.1] The proof sketch of Theorem V.1 depends entirely on Lemma VII.1 and Lemma VII.2, but neither lemma is proved. Lemma VII.2's assertion that Algorithm 2 produces identical node means with componentwise smaller covariances is nontrivial and is not guaranteed by the pseudocode as written, given the status-handling issue above. The monotonicity conditions (1)-(2) in Theorem V.1 are not verified for the experimental wind-field model, and no argument is given that they hold for the counterclockwise linear wind field used in Section VI. Without these proofs or verifications, the central claim that rewiring provably improves coverage is unsupported.
  3. [§IV, Problem IV.1, Eq. (9)] The chance constraints in Problem IV.1 are those of Problem III.1 and use the nominal covariance S = A_bar Sigma_I A_bar^T + G_bar Sigma_W G_bar^T. They are not evaluated at the sigma points and do not use S(i). Thus the robust sigma-point method changes only the objective, not the constraint satisfaction. The abstract's statement that the controller 'safely steer[s]... satisfying state constraints' is therefore not supported; the conclusion explicitly defers robust constraints to future work. The authors should either robustify the constraints with the sigma-point covariances or substantially soften the claims.
  4. [§IV, Eqs. (10)-(11)] The sigma-point approximation is not shown to be conservative. The 4n points are chosen on the sqrt(n) covariance contour, but no moment-matching condition (e.g., sum_i w_i (x0^(i)-x0)(x0^(i)-x0)^T = Sigma_I) is imposed, and the two linearizations per sigma point are heuristic. As a result, S(i) is not proven to cover the true reachable second moment under the state-dependent field. The claim that the worst-case second moment is an upper bound on the largest eigenvalue of the final distribution is conditional on this unproved representation. Please provide an error bound or clearly label the robust objective as a heuristic with empirical validation only.
minor comments (7)
  1. [§V, Algorithm 2, line 6] The third output of the Pi call is blank; it should be the edge e_{k->q,spl} from the sampled parent.
  2. [§V, Algorithm 2, line 8] There is a missing closing parenthesis in 'Vsample <- Vsample union {(mu_q, Sigma_q}'.
  3. [§IV, sigma-point construction] The construction of the 4n sigma points should be specified precisely, for example using the columns of Sigma_I^{1/2} and their negatives, so that the dependence of S(i) on Sigma_I is unambiguous.
  4. [§IV, Eq. (13)] The quantities X^(i) are used before being defined; please define the closed-loop mean trajectory for each sigma point.
  5. [§VI-A, Figure 2] The caption states that trials with W2 > 1 are not shown; this removes exactly the heavy tail that distinguishes the methods in Table I (max W2 up to 308.5). Consider plotting the full distribution or using a log scale.
  6. [§VI-B, Table II] It is unclear whether the robust-ablation trial that failed to reach the goal is included in the reported statistics; please state this explicitly.
  7. [§VI, wind-field sampling] The text says the wind field is sampled over a 121 m^2 space, but x, y in [0,10] is a 100 m^2 domain; presumably 121 sample points at 1 m spacing are meant.

Circularity Check

1 steps flagged · score 2.0 of 10

No prediction reduces to a fit; one load-bearing self-citation to [2] supplies the funnel-composition premise, while the coverage theorem itself is left unsupported by an unproved monotonicity lemma.

  1. self citation load bearing [Section III (Preliminaries), paragraph before Problem III.1]
    "In our prior work [2], we demonstrated that when steering backward from a goal distribution G, roadmap coverage is maximized by maximizing the spectral radius of the initial distribution I and composing funnels such that for two funnels I1 → G1 and I2 → G2, in order to have the tail of the first funnel end inside the mouth of the second, µG1 = µI2 and λmin(I2) ≥ λmax(G1). Accordingly, for forward steering, coverage is maximized by minimizing the spectral radius of the goal distribution G and composing funnels as in our prior work [2]."

    The objective of Problem III.1 (minimizing the goal spectral radius to maximize roadmap coverage) is justified solely by the authors' own prior work [2] (Aggarwal and How, 2024), not by a derivation in this paper. The same self-cited work supplies the funnel-composition rule (λmin(I2) ≥ λmax(G1)) that is later used to define what it means for a goal distribution to be reachable in Theorem V.1. No independent proof, external benchmark, or machine-checked certificate for this rule is provided here, so the coverage-maximizing premise is imported from the authors themselves. This is a partial circularity: the robust sigma-point edge controller and the experimental measurements are independent of this premise, but the roadmap-coverage formulation inherits an unverified self-cited result.

full rationale

The robust sigma-point edge controller (Eq. (10), Problem IV.1) is a new construction, and the experimental numbers (Wasserstein distances, MSE, planned covariance eigenvalues) are measurements rather than fitted parameters renamed as predictions, so no experimental claim reduces to a fit. Problem III.1's definition of ΣG as the minimal scalar-identity covariance satisfying the funnel constraint is a construction, not a prediction. The central concern is not circularity but unsupported proof: Theorem V.1's proof sketch invokes Lemma VII.2, which asserts that Algorithm 1 and Algorithm 2 produce the same node means with componentwise no-larger covariances, but no proof of Lemma VII.2 is given, and the conditions (1)-(2) of Theorem V.1 are not verified for the experimental wind-field model. Moreover, Algorithm 2 (lines 9-20) does not check the status of the line-10 edge computation before adding the node to V, so a node could be added without a feasible incoming edge, breaking the reachability comparison. These are omitted-proof and correctness gaps, not circular reductions, and I flag them explicitly per the reviewing rule. The only concrete circularity is the load-bearing self-citation to [2] for the funnel-composition and coverage-maximization premise, which warrants a score of 2 rather than 0.

Assumptions & free parameters 2 free parameters · 5 assumptions · 0 invented entities

No new physical entities are introduced. The free parameters are heuristic design choices rather than fitted constants. The key axioms are the Gaussian-field linearization and the unproved conditions behind the coverage theorem.

free parameters (2)
  • Sigma point count and placement = 4n points on sqrt(n) covariance contour
    Design choice approximating the initial Gaussian; not tuned to experimental data, but the robustness envelope depends on it.
  • Number of linearizations per sigma point = 2 (around nominal and sigma-point nominal trajectory)
    Heuristic choice in Section IV; not justified by an error bound.
assumptions (5)
  • domain assumption The Gaussian random field can be discretized around the nominal trajectory so that W and Sigma_W are constant and the disturbance is approximated as Gaussian N(W, Sigma_W).
    Section III, after Eq. (7); inherited from [1]; breaks if the field varies rapidly along the trajectory.
  • ad hoc to paper The initial distribution is adequately represented by 4n sigma points with two linearizations, so the worst-case second moment S(i) is a conservative envelope.
    Section IV, Eqs. (10)-(11); no formal error bound is given.
  • ad hoc to paper Lemma VII.1 monotonicity conditions (1)-(2): the eigenvalues of S(i) are monotonically increasing in the eigenvalues of the initial covariance, and the sum over sigma points of (W(i) minus W) is proportional to a constant vector.
    Appendix VII-A; asserted without proof and needed for the robust case of Theorem V.1.
  • domain assumption Roadmap coverage is maximized by minimizing the spectral radius of the goal distribution, with funnel composition requiring lambda_min of the next mouth to bound lambda_max of the previous tail.
    Section III and Problem III.1, based on self-cited prior work [2].
  • standard math Chance constraints can be reformulated exactly as convex constraints using the Gaussian CDF and the Schur complement.
    Section III, after Eq. (7); valid under the Gaussian approximation of the state distribution.

how reviews work

0 comments
Cite this review

Pith. "Pith review of REVISE: Robust Probabilistic Motion Planning in a Gaussian Random Field." pith.science (2026). https://pith.science/paper/UZ5PZ2S2

@misc{pith2026241113369,
  author       = {Pith},
  title        = {Pith review of: REVISE: Robust Probabilistic Motion Planning in a Gaussian Random Field},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/UZ5PZ2S2}},
  note         = {Machine review of arXiv:2411.13369}
}
read the original abstract

This paper presents Robust samplE-based coVarIance StEering (REVISE), a multi-query algorithm that generates robust belief roadmaps for dynamic systems navigating through spatially dependent disturbances modeled as a Gaussian random field. Our proposed method develops a novel robust sample-based covariance steering edge controller to safely steer a robot between state distributions, satisfying state constraints along the trajectory. Our proposed approach also incorporates an edge rewiring step into the belief roadmap construction process, which provably improves the coverage of the belief roadmap. When compared to state-of-the-art methods, REVISE improves median plan accuracy (as measured by Wasserstein distance between the actual and planned final state distribution) by 10x in multi-query planning and reduces median plan cost (as measured by the largest eigenvalue of the planned state covariance at the goal) by 2.5x in single-query planning for a 6DoF system. We will release our code at https://acl.mit.edu/REVISE/.

Figures

Figures reproduced from arXiv: 2411.13369 by the authors.

Figure 1
Figure 1. (a-b) Baseline covariance steering steers between Gaussian [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Distribution of Wasserstein distance between the planned and actual [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Trajectories and final state distribution for the single query [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

14 extracted references · 13 canonical work pages

  1. [1]

    Chance-constrained covariance steering in a gaussian random field via successive convex programming,

    J. Ridderhof and P. Tsiotras, “Chance-constrained covariance steering in a gaussian random field via successive convex programming,” J. of Guidance, Control, and Dynamics , vol. 45, no. 4, pp. 599–610, 2022

    work page 2022

  2. [2]

    SDP Synthesis of Maximum Coverage Trees for Probabilistic Planning under Control Constraints

    N. Aggarwal and J. P. How, “SDP synthesis of maximum cover- age trees for probabilistic planning under control constraints,” arXiv preprint arXiv:2403.14605, 2024

    work page Pith review arXiv 2024

  3. [3]

    Prob- abilistic roadmaps for path planning in high-dimensional configuration spaces,

    L. E. Kavraki, P. Svestka, J.-C. Latombe, and M. H. Overmars, “Prob- abilistic roadmaps for path planning in high-dimensional configuration spaces,” IEEE Trans. on Robotics and Automation , vol. 12, no. 4, pp. 566–580, 1996

    work page 1996

  4. [4]

    RRT-connect: An efficient approach to single-query path planning,

    J. J. Kuffner and S. M. LaValle, “RRT-connect: An efficient approach to single-query path planning,” in 2000 IEEE Int. Conf. on Robotics and Automation (ICRA) , vol. 2. IEEE, 2000, pp. 995–1001

    work page 2000

  5. [5]

    Sampling-based algorithms for optimal motion planning,

    S. Karaman and E. Frazzoli, “Sampling-based algorithms for optimal motion planning,” The Int. J. of Robotics Research , vol. 30, no. 7, pp. 846–894, 2011

    work page 2011

  6. [6]

    Chance constrained RRT for prob- abilistic robustness to environmental uncertainty,

    B. Luders, M. Kothari, and J. How, “Chance constrained RRT for prob- abilistic robustness to environmental uncertainty,” in AIAA Guidance, Navigation, and Control Conf. , 2010, p. 8160

    work page 2010

  7. [7]

    Robust sampling-based motion planning with asymptotic optimality guarantees,

    B. D. Luders, S. Karaman, and J. P. How, “Robust sampling-based motion planning with asymptotic optimality guarantees,” in AIAA Guidance, Navigation, and Control (GNC) Conf. , 2013, p. 5097

    work page 2013

  8. [8]

    CS-BRM: A probabilistic roadmap for consistent belief space planning with reachability guarantees,

    D. Zheng, J. Ridderhof, Z. Zhang, P. Tsiotras, and A.-A. Agha- Mohammadi, “CS-BRM: A probabilistic roadmap for consistent belief space planning with reachability guarantees,”IEEE Trans. on Robotics, 2024

    work page 2024

Show all 14 references
  1. [9]

    LQR- trees: Feedback motion planning via sums-of-squares verification,

    R. Tedrake, I. R. Manchester, M. Tobenkin, and J. W. Roberts, “LQR- trees: Feedback motion planning via sums-of-squares verification,”The Int. J. of Robotics Research , vol. 29, no. 8, pp. 1038–1052, 2010

    2010

  2. [10]

    Funnel libraries for real-time robust feedback motion planning,

    A. Majumdar and R. Tedrake, “Funnel libraries for real-time robust feedback motion planning,” The Int. J. of Robotics Research , vol. 36, no. 8, pp. 947–982, 2017

    2017

  3. [11]

    FIRM: Sampling-based feedback motion-planning under motion uncertainty and imperfect measurements,

    A.-A. Agha-Mohammadi, S. Chakravorty, and N. M. Amato, “FIRM: Sampling-based feedback motion-planning under motion uncertainty and imperfect measurements,” The Int. J. of Robotics Research, vol. 33, no. 2, pp. 268–304, 2014

    2014

  4. [12]

    Optimal covariance control for stochastic systems under chance constraints,

    K. Okamoto, M. Goldshtein, and P. Tsiotras, “Optimal covariance control for stochastic systems under chance constraints,” IEEE Control Systems Letters, vol. 2, no. 2, pp. 266–271, 2018

    2018

  5. [13]

    Optimal covariance control for discrete-time stochastic linear systems subject to constraints,

    E. Bakolas, “Optimal covariance control for discrete-time stochastic linear systems subject to constraints,” in 2016 IEEE 55th Conf. on Decision and Control (CDC) . IEEE, 2016, pp. 1153–1158

    2016

  6. [14]

    Design of affine controllers via convex optimization,

    J. Skaf and S. P. Boyd, “Design of affine controllers via convex optimization,” IEEE Trans. on Automatic Control , vol. 55, no. 11, pp. 2476–2487, 2010

    2010

Pith tools

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