Pith. sign in

REVIEW 3 major objections 6 minor 17 references

RIT* replaces Euclidean informed search with a Riemannian one, claiming up to 63.5% lower final costs in a 14-DOF bimanual task and preserving asymptotic optimality.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review

2026-08-05 00:12 UTC pith:AR5EJSRZ

load-bearing objection A genuinely useful Riemannian-informed planner whose high-D gains are convincing, but the AO proof leaves a real gap between the whitened sampling in the code and the uniform-on-I_R assumption in the theorem; worth refereeing with a required fix. the 3 major comments →

arxiv 2608.00822 v1 pith:AR5EJSRZ submitted 2026-08-01 cs.RO

RIT*: Riemannian Informed Trees for Cost-Adaptive Optimal Motion Planning

Muhayy Ud Din , Ahmed Nadar , Jan Rosell , Irfan Hussain This is my paper
classification cs.RO
keywords motion planningsampling-based planningRiemannian metricinformed samplingasymptotic optimalitybatch informed treesanisotropic configuration spacesonline metric learning
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

This paper argues that the Euclidean primitives used by batch-informed sampling planners—the informed set, nearest-neighbour search, and edge cost—are the reason they underperform in anisotropic configuration spaces, and that replacing them with Riemannian counterparts yields tighter informed sets and lower-cost paths. It introduces RIT*, which samples from a Riemannian informed set, connects neighbours under a metric-aligned ellipsoid, and screens edge costs with a cascade that rejects most candidates cheaply. It also introduces CARM, which learns an obstacle-proximity cost field online from collision feedback and applies it as a conformal scaling, so the planner adapts without prior obstacle knowledge. The paper claims asymptotic optimality is preserved and reports median final-cost gains of 9% over BIT* in a 6-DOF task and 24.8-63.5% in a 14-DOF bimanual task, with gains growing with anisotropy and dimension.

Core claim

RIT* claims that an asymptotically optimal informed planner can operate natively under a spatially varying Riemannian metric without losing its guarantees, and that doing so pays off when costs are anisotropic. The central object is the Riemannian informed set I_R = {x : d_R(xs,x) + d_R(x,xg) ≤ c_best}, which contains exactly the configurations that can still improve the current solution under the metric. The paper constructs this set approximately by a whitening transform using the mean metric along the start-goal segment, performs nearest-neighbour search in Riemannian balls, and evaluates edges by Gauss-Legendre quadrature screened by a three-level cascade. It proves probabilistic complet

What carries the argument

The Riemannian informed set I_R, defined by the geodesic distance under a positive-definite metric tensor field G(x), is the object that replaces the Euclidean prolate hyperspheroid. Because geodesic distance has no closed form, the set is sampled approximately via a whitening transform L = chol(Gbar), where Gbar is the mean metric along the start-goal segment; this maps I_R to an ellipsoid so the direct ellipsoid-sampling method of Informed RRT* applies. The same metric defines the nearest-neighbour connection radius and the edge-cost integral, and CARM refines G by a learned scalar conformal factor s(x) = 1 + alpha * KDE(collision points).

Load-bearing premise

The proof assumes samples are drawn uniformly from the exact Riemannian informed set, but the implementation draws from a whitened ellipsoid built from the mean metric along the start-goal segment; if that ellipsoid excludes part of the true informed set, the optimality proof and the claimed cost gains may not carry over to unbenchmarked environments.

What would settle it

Construct a 3-D environment with a sharply spatially varying metric, for example G(x)=diag(1,3,6)(1+beta||x-x_c||) with large beta, compute the true Riemannian informed set I_R by fine numerical geodesic evaluation, then compare it pointwise and by volume with the whitened ellipsoid L^{-T}E used by RIT*. If a positive-measure subset of I_R lies outside the whitened ellipsoid, then the implementation is not sampling uniformly on I_R, contradicting the premise of Theorem 3; a planner that then fails to find lower-cost paths in that environment would settle the matter.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

Share X Bluesky LinkedIn Reddit HN

If this is right

  • If the central claim holds, replacing Euclidean informed sets with Riemannian ones is a generally applicable upgrade for batch-informed planners under anisotropic costs, with gains increasing with metric condition number and dimension.
  • Planners using joint-space inertia as the metric should expect lower-cost solutions in manipulation tasks; real-robot runs show trajectories that preferentially use low-inertia joints.
  • CARM's learned cost field can substitute for hand-designed obstacle-aware metrics: with no metric prior it recovers roughly two-thirds of the oracle gap from collision feedback alone and stays within about 1% of oracle final costs.
  • The cascading edge evaluation keeps the added Riemannian computation cheap, with roughly 5% of candidate edges reaching full quadrature and collision checking, so the gains do not come at prohibitive compute.
  • The CARM-refined metric preserves the asymptotic optimality guarantee, so users can adapt the metric online without sacrificing convergence.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • We infer that CARM is a portable component: since it is a conformal scaling of whatever edge-cost metric the planner uses, a Euclidean BIT* equipped with CARM is a natural next test. Our expectation is that CARM would narrow the gap but not eliminate it, because in high dimensions the metric-aligned informed set, not just the learned cost field, drives most of the gain.
  • A direct stress-test extension would be to sweep beta in G(x)=diag(1,3,6)(1+beta||x-x_c||) and measure both the whitening volume mismatch and median cost. The paper reports mismatch growing to 6.1% at beta=2; measuring whether final-cost gains degrade at the same rate would quantify how much of the 24.8% advantage survives beyond the tested anisotropy.
  • We infer that the whitening approximation is the main robustness boundary: a sampling scheme that rejects points outside the true Riemannian informed set, or that uses local metric frames instead of a single global mean, would harden the method against environments with sharp metric gradients.
  • We infer that CARM's almost-sure termination claim could be tested in a maze-like environment where repeated collision probing keeps injecting new collision points; that would probe whether the convergence of the KDE field is as fast as the analysis assumes.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 6 minor

Summary. The paper proposes RIT*, a batch-informed sampling-based motion planner that replaces the Euclidean primitives of BIT* with Riemannian counterparts: a Riemannian informed set, Riemannian nearest-neighbour search, Riemannian edge costs with a cascading quadrature scheme, and an online collision-adaptive conformal metric refinement (CARM). The authors claim that RIT* preserves the asymptotic optimality guarantees of BIT* (Theorem 3, Corollary 4) and demonstrate empirically that it produces lower-cost solutions in anisotropic, high-dimensional settings, with reported gains of up to 24.8--63.5% over Euclidean baselines in a 14-DOF bimanual task, while remaining competitive in isotropic 2-D/3-D benchmarks. A real-robot UR10e execution is also reported. The central theoretical claim is that replacing Euclidean informed sets, neighbour selection, and edge costs with their Riemannian analogues is both asymptotically optimal and practically beneficial. The strongest support is the experimental evaluation, which is broad and uses a common cost objective across planners. The main weakness is a gap between the theoretical assumptions (uniform sampling on the exact Riemannian informed set with exact Riemannian volume) and the implementation (whitening transform based on a mean metric, with an approximate volume term).

Significance. If the AO claim holds, the paper makes a useful contribution: it extends informed, anytime sampling-based planning from Euclidean to anisotropic cost geometries, where existing planners degrade, and it provides a practical online metric-learning mechanism (CARM) that works without prior obstacle knowledge. The experimental scope (2-D to 14-D, 50 trials, ablations, sensitivity sweep, real robot) is a genuine strength, and the authors are transparent about limitations and about the approximation involved in whitening. The paper ships code and videos, which aids reproducibility. The provided proof sketches are not machine-checked and the central AO theorem is not fully established as stated for the implemented algorithm; this is the main barrier to acceptance.

major comments (3)
  1. [Sec. IV-A, Eq. (4), Theorem 3] The implementation samples from the whitened ellipsoid L^{-T}E, where L=chol(Gbar) and Gbar is the mean metric along the start-goal segment, and evaluates the connection-radius volume as mu_R(I_R) ~ det(Gbar)^{1/2} mu(I_E) in Eq. (6). Theorems 2 and 3 assume uniform sampling on the exact I_R and exact mu_R(I_R). The paper itself states this transformation 'approximates' the informed set and reports only median relative volume mismatch (<1% on benchmarks, 6.1% in a controlled sweep). Volume mismatch is neither necessary nor sufficient: the whitened ellipsoid could have the correct volume yet be shifted so as to exclude a positive-measure portion of I_R containing the optimal-path corridor, or it could contain I_R while underestimating mu_R(I_R), shrinking r_R^n below the Karaman-Frazzoli threshold. Either failure breaks the proof of Theorem 3. A support-containment bound, a total-variatio
  2. [Sec. IV-F, proof of Theorem 3] The proof asserts that a Riemannian ball of radius r_R^n contains a Euclidean ball of radius r_R^n/sqrt(lambda_max), so the edge set is a superset of BIT*'s at that radius and 'inherits its neighbour-graph properties.' This does not establish the Karaman-Frazzoli condition for the actual sampling distribution, which is uniform with respect to the Riemannian volume on I_R, not uniform in Euclidean coordinates. The statement 'gamma_R >= lambda_min^{1/2} gamma_KF' compares the gamma_R defined in Eq. (6) (based on mu_R(I_R)) with an undefined gamma_KF, and no derivation shows that the resulting r_R^n exceeds the required threshold for a sampling density that is only bounded between lambda_min^{d/2} and lambda_max^{d/2}. Even aside from the whitening gap, the central AO derivation is therefore incomplete as written.
  3. [Corollary 4 (Sec. IV-F)] The CARM guarantee assumes that after a finite t0 the metric is frozen at G_infinity and Theorems 2 and 3 apply. However, the tree at t0 contains vertices and edges generated under earlier (changing) metrics; these are not i.i.d. samples from the final I_R under G_infinity. The proof does not show that this finite biased prefix is asymptotically negligible, nor that the pruning rule restores the i.i.d. condition. A formal argument (e.g., a coupling argument or a demonstration that the prefix has zero asymptotic effect) is needed to support the claimed AO guarantee under CARM.
minor comments (6)
  1. [Sec. V-A, Table II] The reported percentage improvements are computed relative to RIT*, not the baseline: e.g., 24.8% = (10.7092-8.5778)/8.5778. State this convention explicitly; readers will otherwise interpret 'improvement over BIT*' as (baseline - new)/baseline.
  2. [Theorem 2 proof] The set I_R^+ = {x : d_R(xs,x)+d_R(x,xg) <= c*+epsilon} is said to be 'contained in I_R'; this is true only while the current c_best >= c*+epsilon, which fails as c_best approaches c*. The proof can be repaired by using the fact that waypoints on the optimal path have d_R-sum exactly c*, but the current text is inaccurate.
  3. [Remark 1 / Lemma 1] The statement 'the tree-accumulated sum of c_R converges to d_R under densification' is asserted without proof. It is plausible via polygonal approximation of geodesics, but it is load-bearing for the rewiring/pruning asymptotic equivalence and should be argued explicitly.
  4. [Sec. IV-C] The deflation argument deserves a small clarification: hat(c)_1 and hat(c)_2 are not themselves integrand samples but are formed from such samples; the bound eta_e^{-1} hat(c) <= sqrt(lambda_min)||Delta|| relies on hat(c) being bounded above by sqrt(lambda_max)||Delta||, which is true for both L1 and L2 estimates. State this explicitly.
  5. [Sec. V-A] The statement that 'the cascading evaluation of Sec. IV-C is applied to all planners' means the baselines are not the published BIT*, AIT*, etc., but modified versions. This may be a fair timing ablation, but it should be stated that the baselines are re-implemented with RIT*'s cascade, so the comparison is to modified baselines.
  6. [Sec. IV-A / IV-B] The volume mismatch is reported for the Euclidean volume of the whitened ellipsoid vs. the Euclidean volume of I_R, while Eq. (6) uses the Riemannian volume mu_R(I_R) approximated by det(Gbar)^{1/2} mu(I_E). Clarify which quantity is used in the experiments and in the theorem; the two are not interchangeable.

Circularity Check

0 steps flagged

No significant circularity: RIT*'s claims are evaluated on held-out benchmarks; the Riemannian primitives and CARM are not defined in terms of the reported cost gains.

full rationale

The paper's derivation chain does not reduce to its own inputs. The Riemannian informed set I_R in Eq. (4) is defined from geodesic distances under the given metric, and the whitening transform of Sec. IV-A is explicitly labeled an approximation: 'This transformation approximates the Riemannian informed set, with accuracy governed by the smoothness of G' and the volume mismatch is reported empirically. Thus the implementation gap between uniform sampling on I_R assumed in Theorem 3 and the whitened-ellipsoid sampler is a stated approximation/omitted containment proof, not a hidden equivalence. Similarly, the connection radius in Eq. (6) uses an approximate volume det(Gbar)^{1/2} mu(I_E) that is disclosed as exact only for constant G. CARM's learned field is not used to score the reported results: 'All path costs... are evaluated a posteriori under the fixed base metric G... never under G_CARM', so the experimental comparisons cannot be circular. Theorem 3 invokes the Karaman-Frazzoli conditions and builds on the external, prior algorithms of [2] and [4]; there is no load-bearing self-citation. The main caveat—that the proof assumes exact Riemannian quantities while the code uses mean-metric approximations—is a correctness/robustness gap, not a circularity, and is explicitly flagged by the paper's own discussion of approximation error. No fitted parameter is renamed as a prediction, and no known result is merely re-labeled. Score 0.

Axiom & Free-Parameter Ledger

7 free parameters · 4 axioms · 0 invented entities

The method introduces no new physical entities. The free parameters are hand-chosen algorithmic hyperparameters (CARM and numerical integration). The central theoretical results rely on standard AO background, a smoothness/clearance domain assumption, and two unproved convergence premises: the densification equivalence for edge costs and the CARM freezing behavior.

free parameters (7)
  • CARM kernel bandwidth sigma = 0.1
    Chosen by hand; controls spatial resolution of the learned cost field. Default in sensitivity sweep (Sec. V-E).
  • CARM inflation strength alpha = 10
    Chosen by hand; controls how aggressively traversal cost increases near obstacles (Sec. V-A).
  • CARM update interval Kc = 15
    Chosen by hand; how often the metric is updated from the collision buffer (Sec. V-A).
  • CARM freezing threshold epsilon0 = 1e-3
    Used to declare metric frozen for the AO analysis (Corollary 4); hand-chosen.
  • Gauss-Legendre quadrature order n = 10 (configurable)
    Numerical integration order for edge costs; default in Remark 1, chosen as a computational accuracy vs cost trade-off.
  • Whitening segment count N_seg = 10
    Number of uniform points along start-goal segment to compute mean metric Gbar (Sec. IV-A); chosen by hand.
  • Metric cache resolution RES = 32 in 2-D/3-D, 3 in 6-D/14-D
    Grid resolution for metric lookups; hand-chosen trade-off between memory and accuracy (Sec. V-A).
axioms (4)
  • domain assumption Assumption 1: G satisfies lambda_min I_d <= G(x) <= lambda_max I_d on C_free with 0 < lambda_min <= lambda_max < inf and ||grad G||_inf < inf; optimal path has strong delta-clearance
    Needed for bi-Lipschitz bound (13) and the AO proofs in Sec. IV-F. Assumed to hold for the environments tested.
  • standard math Karaman-Frazzoli conditions for asymptotic optimality are applicable to the Riemannian metric setting
    Theorem 3 proof verifies the KF conditions using (13); the conditions themselves are prior results taken as given.
  • domain assumption The tree-accumulated sum of c_R converges to d_R under densification (Remark 1)
    Assumed without proof in Remark 1; standard for straight-line edge costs on smooth manifolds but not established in the paper.
  • ad hoc to paper CARM updates terminate at a finite t0 when the KDE change drops below epsilon0 (Corollary 4)
    The paper argues per-update change is O(B/N_t) -> 0, but does not formally prove this holds in all newly explored regions; needed for applying Theorems 2-3 to the CARM-modified planner.

reviewed 2026-08-05 · how reviews work

0 comments
Cite this review

Pith. "Pith review of RIT*: Riemannian Informed Trees for Cost-Adaptive Optimal Motion Planning." pith.science (2026). https://pith.science/paper/AR5EJSRZ

@misc{pith2026260800822,
  author       = {Pith},
  title        = {Pith review of: RIT*: Riemannian Informed Trees for Cost-Adaptive Optimal Motion Planning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/AR5EJSRZ}},
  note         = {Machine review of arXiv:2608.00822}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

We present Riemannian Informed Trees (RIT*), a planning framework that replaces Euclidean primitives in batch-informed search with their Riemannian counterparts. RIT* constructs a tighter, cost-consistent informed set, performs a nearest-neighbour search under an anisotropic distance metric, and evaluates edge costs efficiently via a cascading scheme. We further introduce a Collision-Adaptive Metric Refinement (CARM), which learns an obstacle-proximity cost field online from collision feedback, reducing the reliance on prior metric design in practical settings. Experiments across environments from 2-D to 14-D show that RIT* is competitive in low-dimensional and spatially constant-metric settings and produces substantially lower-cost solutions when the metric varies spatially in high-dimensional configuration spaces. Performance gains scale with anisotropy and dimension, reaching up to 13.0% improvement in median initial cost over BIT* in the 3-D anisotropic benchmark, up to 9.0% in median final cost over BIT* in 6-DOF manipulation, and 24.8-63.5% in a 14-DOF bimanual planning problem, where Euclidean-informed baselines degrade. Videos and code can be found here: https://muhayyuddin.github.io/ritstar/

Figures

Figures reproduced from arXiv: 2608.00822 by Ahmed Nadar, Irfan Hussain, Jan Rosell, Muhayy Ud Din.

Figure 1
Figure 1. Figure 1: Overview of RIT*. (a) BIT* samples from the Eu￾clidean informed set IE, ignoring anisotropic cost structure. (b) RIT* uses a cost-aligned Riemannian informed set; CARM (red shading) learns a cost field from collision feedback to guide paths away from obstacles. Collision-Adaptive Metric Refinement (CARM), which learns a collision-induced conformal cost field while planning from collision feedback and adapt… view at source ↗
Figure 2
Figure 2. Figure 2: Riemannian NN selection and cascading edge evalua [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Tree rewiring and pruning in RIT*. (a) Rewiring [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: CARM metric evolution: as collision points accumu [PITH_FULL_IMAGE:figures/full_fig_p005_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Benchmark environments from 2-D to 14-D, covering [PITH_FULL_IMAGE:figures/full_fig_p006_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Success rate (top) and Riemannian path cost (bottom) vs. time across all benchmarks; median over trials with shaded [PITH_FULL_IMAGE:figures/full_fig_p007_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: UR10e real-robot execution of RIT*-planned trajecto [PITH_FULL_IMAGE:figures/full_fig_p007_7.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

17 extracted references · 14 canonical work pages

  1. [1]

    S. M. LaValle,Planning Algorithms. Cambridge Univ. Press, 2006

    2006

  2. [2]

    Sampling-based algorithms for optimal motion planning,

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

    work page 2011

  3. [3]

    Informed sampling for asymptotically optimal path planning,

    J. D. Gammell, T. D. Barfoot, and S. S. Srinivasa, “Informed sampling for asymptotically optimal path planning,”IEEE Trans. Robot., vol. 34, no. 4, pp. 966–984, 2018

    work page 2018

  4. [4]

    Batch Informed Trees (BIT*): Informed asymptotically optimal anytime search,

    ——, “Batch Informed Trees (BIT*): Informed asymptotically optimal anytime search,”Int. J. Robot. Res., vol. 39, no. 5, pp. 543–567, 2020

    work page 2020

  5. [5]

    Adaptively Informed Trees (AIT*) and Effort Informed Trees (EIT*): Asymmetric bidirectional sampling-based path planning,

    M. P. Strub and J. D. Gammell, “Adaptively Informed Trees (AIT*) and Effort Informed Trees (EIT*): Asymmetric bidirectional sampling-based path planning,”Int. J. Robot. Res., vol. 41, no. 4, pp. 390–417, 2022

    work page 2022

  6. [6]

    APT*: Asymptotically optimal motion planning via adaptively prolated ellipticalr-nearest neighbors,

    L. Zhang, S. Wang, K. Cai, Z. Bing, F. Wu, C. Wang, S. Haddadin, and A. Knoll, “APT*: Asymptotically optimal motion planning via adaptively prolated ellipticalr-nearest neighbors,”IEEE Robot. Autom. Lett., vol. 10, no. 10, pp. 10 242–10 249, 2025

    work page 2025

  7. [7]

    Flexible Informed Trees (FIT*): Adaptive batch-size approach in informed sampling-based path planning,

    L. Zhang, Z. Bing, K. Chen, L. Chen, K. Cai, Y . Zhang, F. Wu, P. Krumbholz, Z. Yuan, S. Haddadin, and A. Knoll, “Flexible Informed Trees (FIT*): Adaptive batch-size approach in informed sampling-based path planning,” inProc. IEEE/RSJ Int. Conf. Intell. Robot. Syst. (IROS), 2024, pp. 3146–3152

    work page 2024

  8. [8]

    CHOMP: Gradient optimization techniques for efficient motion planning,

    N. Ratliff, M. Zucker, J. A. Bagnell, and S. Srinivasa, “CHOMP: Gradient optimization techniques for efficient motion planning,” inProc. IEEE Int. Conf. Robot. Autom. (ICRA), 2009, pp. 489–494

    work page 2009

  9. [9]

    STOMP: Stochastic trajectory optimization for motion planning,

    M. Kalakrishnan, S. Chitta, E. Theodorou, P. Pastor, and S. Schaal, “STOMP: Stochastic trajectory optimization for motion planning,” in Proc. IEEE Int. Conf. Robot. Autom. (ICRA), 2011, pp. 4569–4574

    work page 2011

  10. [10]

    Tangent bundle RRT: A randomized algorithm for constrained motion planning,

    B. Kim, T. T. Um, C. Suh, and F. C. Park, “Tangent bundle RRT: A randomized algorithm for constrained motion planning,”Robotica, vol. 34, no. 1, pp. 202–225, 2016

    work page 2016

  11. [11]

    Geometry-aware sampling-based motion plan- ning on Riemannian manifolds,

    P. T. Kyaw and J. Kelly, “Geometry-aware sampling-based motion plan- ning on Riemannian manifolds,” 2026, arXiv preprint arXiv:2602.00992

    Pith/arXiv arXiv 2026

  12. [12]

    Bi-AM-RRT*: A fast and efficient sampling-based motion planning algorithm in dynamic environments,

    Y . Zhang, H. Wang, M. Yin, J. Wang, and C. Hua, “Bi-AM-RRT*: A fast and efficient sampling-based motion planning algorithm in dynamic environments,”IEEE Trans. Intell. Veh., vol. 9, no. 1, pp. 1282–1293, 2024

    work page 2024

  13. [13]

    Sampling-based path planning on configuration-space costmaps,

    L. Jaillet, J. Cortés, and T. Siméon, “Sampling-based path planning on configuration-space costmaps,”IEEE Trans. Robot., vol. 26, no. 4, pp. 635–646, 2010

    work page 2010

  14. [14]

    On the design of region-avoiding metrics for collision-safe motion generation on Rieman- nian manifolds,

    H. Klein, N. Jaquier, A. Meixner, and T. Asfour, “On the design of region-avoiding metrics for collision-safe motion generation on Rieman- nian manifolds,” inProc. IEEE/RSJ Int. Conf. Intell. Robot. Syst. (IROS), 2023, pp. 2346–2353

    work page 2023

  15. [15]

    RMPflow: A computational graph for automatic motion policy generation,

    C.-A. Cheng, M. Mukadam, J. Issac, S. Birchfield, D. Fox, B. Boots, and N. Ratliff, “RMPflow: A computational graph for automatic motion policy generation,” inProc. Worksh. Algorithmic Found. Robot. (WAFR), ser. Springer Proc. Adv. Robot., vol. 14. Springer, 2018, pp. 441–457

    work page 2018

  16. [16]

    Reactive motion generation on learned Riemannian mani- folds,

    H. Beik-Mohammadi, S. Hauberg, G. Arvanitidis, G. Neumann, and L. Rozo, “Reactive motion generation on learned Riemannian mani- folds,”Int. J. Robot. Res., vol. 42, no. 10, pp. 729–754, 2023

    work page 2023

  17. [17]

    PyBullet, a Python module for physics simulation for games, robotics and machine learning,

    E. Coumans and Y . Bai, “PyBullet, a Python module for physics simulation for games, robotics and machine learning,” http://pybullet.org, 2016–2021

    2016

This paper was first reviewed by deepseek-v4-flash on August 5, 2026.