Pith. sign in

REVIEW 3 major objections 3 minor 43 references

ReMoSPLAT shows that a Gaussian Splat scene representation can replace an analytic signed-distance field in reactive collision avoidance for mobile manipulation, matching baseline performance in cluttered reaching tasks.

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 · deepseek-v4-flash

2026-08-03 17:23 UTC pith:XSXK4QHM

load-bearing objection First reactive mobile-manipulation controller that uses Gaussian-Splat distances directly, with a solid synthetic benchmark, but the safety gradient in Eq. (19) is inconsistent with the text as written and needs fixing or a code release. the 3 major comments →

arxiv 2512.09656 v2 pith:XSXK4QHM submitted 2025-12-10 cs.RO

ReMoSPLAT: Reactive Mobile Manipulation Control on a Gaussian Splat

Nicolas Marticorena , Tobias Fischer , Niko Suenderhauf This is my paper
classification cs.RO
keywords reactive controlmobile manipulationGaussian Splattingcollision avoidancequadratic programdepth rasterisationsigned distance fieldwhole-body control
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.

The paper sets out to show that a mobile manipulator can avoid obstacles reactively using only the Gaussian Splat representation of an offline-mapped scene, without global planning or an explicit signed-distance field. It builds a quadratic-program controller that treats each robot link as a set of spheres, queries distances from those spheres to the splatted surfaces, and turns those distances into hard velocity-damper constraints and a soft collision cost. Two distance queries are compared: a geometric sphere-to-ellipsoid computation and a six-camera depth-rasterisation sensor that respects Gaussian opacity. In 1,000 randomised synthetic scenes the splat-based controller matches a ground-truth SDF baseline on success rate and clearance while running at 17–24 Hz; a real-world scan and a noisy-reconstruction study show the rasterisation method is the more robust of the two. A sympathetic reader would take the result as evidence that high-fidelity learned scene representations can carry reactive control, not just planning.

Core claim

ReMoSPLAT's central claim is that collision avoidance for reactive mobile manipulation can be performed directly on a Gaussian Splat map by computing robot-to-surface distances and feeding them into a QP as inequality constraints plus a distance-weighted collision cost. The paper demonstrates that this learned representation performs comparably to a controller with perfect distance information from an analytical SDF: over common successful trials, success rates, average clearance, and collision counts are statistically similar, with the splat-based controller even improving success rate by up to 8% when hard and soft constraints are combined. It further claims that accounting for opacity thr

What carries the argument

The load-bearing object is the distance query on the Gaussian Splat. The robot is covered by 77 spheres; distances are computed either by solving the point-to-ellipsoid problem as a convex root-finding via bisection, or by mounting six virtual cameras per sphere with 90° fields of view, rasterising median depth from 2D Gaussian Splats, and backprojecting the closest pixel. These distances become linear inequality constraints through a velocity damper and a weighted collision cost, so the existing QP controller sees only standard Jacobian-distance rows. The median-depth rule is what lets the rasterisation query respect opacity, which is the property that saves it in noisy scenes.

Load-bearing premise

The load-bearing premise is that the six fixed camera orientations per robot sphere see every obstacle that could collide; if an obstacle lies between those frustums, no distance constraint is generated and the collision-free guarantee lapses.

What would settle it

Run the controller with the depth-rasterisation distance query in a scene containing a thin pole placed along a direction outside the six camera axes — e.g., the diagonal between two adjacent 90° frustums — within the influence distance of a sphere. The query returns no obstacle distance, the constraint set is empty for that sphere, and the robot drives into the pole; equivalently, compute the union of the six frustums per sphere and show it does not cover the sphere's full solid angle.

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

If this is right

  • If the claim holds, a mobile manipulator can navigate cluttered, offline-mapped scenes with a reactive QP and no global planner, at control rates above 20 Hz.
  • The Gaussian Splat can substitute for a signed-distance field in optimisation-based reactive control, matching ground-truth performance on success rate and clearance.
  • Combining hard velocity-damper constraints with a soft collision cost yields measurable gains: up to 8% higher success rate and larger average obstacle clearance, at the price of about 4% longer end-effector paths.
  • Opacity-aware depth rasterisation is the robust distance primitive: the geometric sphere-to-ellipsoid query collapses under noisy reconstructions, while rasterisation still completes 4 of 7 tasks with no collisions.
  • The controller is representation-agnostic: any geometry-accurate Gaussian Splatting variant can be swapped in, so advances in splat reconstruction directly improve control.

Where Pith is reading between the lines

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

  • The six fixed camera directions per sphere are a discretisation; a diagonal obstacle can fall between frustums and produce no constraint. An analytic angular-coverage check or adaptive view selection would close that gap, and the real-world 6/7 success may partly reflect scenes whose surfaces align with the camera axes.
  • The opacity-weighted median depth behaves like a probabilistic occupancy filter; the same mechanism could weight constraint confidence by transmittance, making the controller degrade gracefully when reconstruction quality drops.
  • Since the distance query is the only splat-dependent piece, hooking the controller to incremental or dynamic Gaussian Splatting would lift the static fully-mapped assumption, a step the paper names as future work.
  • The collision cost is built from the already-computed constraint matrix and distances, so similar QP-based controllers can add safety costs without extra scene queries.

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 / 3 minor

Summary. The manuscript introduces ReMoSPLAT, a reactive quadratic-programming (QP) controller for a mobile manipulator that uses a Gaussian Splatting (GS) scene representation for collision avoidance. The robot's body is approximated by 77 spheres. Two distance-query methods are developed: a sphere-to-ellipsoid geometric computation and a depth-rasterisation method that places six virtual cameras on each sphere and renders depth maps from the GS. These distances enter the QP as velocity-damper hard constraints (Eq. (21)) and an active collision-avoidance cost (Eq. (23)). The method is evaluated on 1000 synthetic scenes against the RMMI controller with a ground-truth SDF, on one real-world office scan, and on a noisy-reconstruction ablation. The paper reports success rates of 77.6-85.6% versus 77.8-84.6% for the SDF baseline, zero collisions, and control rates of 17-24 Hz, concluding comparable performance to a perfect-distance baseline.

Significance. If correct, the work demonstrates that a GS representation can support reactive collision avoidance for mobile manipulation without global planning or an explicit SDF, at control rates sufficient for real-time operation. The experimental design has notable strengths: the synthetic benchmark is large (1000 scenes), the baseline uses ground-truth analytical distances, and the noisy-reconstruction ablation isolates the opacity-handling advantage of the rasterisation approach. The real-world scan adds credibility. However, the central safety-constraint equations are not self-consistent as written: the depth-rasterisation gradient in Eq. (19) is neither unit-normalised nor expressed in the world frame, and the sign of the velocity-damper constraint in Eq. (21) is inconsistent with the gradient defined in Eq. (14). The six-view sampling in Eq. (16) is not a blind-spot issue (the six 90° frustums form a cube map covering all directions), but the gradient inconsistency is load-bearing because every reported collision-free result depends on it. Because no code is provided, the reader cannot determine whether the evaluated controller matches the described one.

major comments (3)
  1. [§IV-C.2 and §IV-D, Eq. (19)-(21)] The gradient in Eq. (19) is not a unit vector and is expressed in the virtual camera frame. Its Euclidean norm is sqrt(((u-cx)/f)^2 + ((v-cy)/f)^2 + 1), which ranges from 1 to about sqrt(3) for a 90° FOV. Section IV-D nevertheless states that every method yields 'a corresponding unit vector ∇dj' in the world frame, and Eq. (20) multiplies it by the world-frame Jacobian J_v. The paper omits the necessary normalisation and the rotation R_c from camera to world frame. If the implementation follows the text literally, the left-hand side of Eq. (21) is not the time derivative of the measured distance; it is a scaled, frame-mismatched quantity. Since the safety constraints and the active-collision cost (Eq. (23)) are built from this quantity, the collision-free claim rests on an unspecified implementation detail. Please provide the exact transformation or correct the equations.
  2. [§IV-D, Eq. (20)-(21)] The sign of the velocity-damper constraint is inconsistent with the gradient definition. For the sphere-to-ellipsoid method, Eq. (14) defines ∇d_j as the gradient of d_j with respect to the sphere-centre position, which points away from the obstacle (from p_e to p_j). With this choice, \dot d_j in Eq. (20) is the rate of increase of clearance, and a safety constraint should limit -\dot d_j, not \dot d_j. As written, Eq. (21) restricts only how fast the robot moves away from an obstacle; the approach speed is unconstrained. If instead ∇d_j is intended to point toward the obstacle as stated in Section IV-D, then Eq. (14) has the wrong sign. This is a load-bearing inconsistency because both the hard constraints and the cost in Eq. (23) use J_d.
  3. [§V-B.4, Table I] The metrics marked '*' in Table I are averaged only over trials in which all methods succeeded. This selection bias can make a method that fails on hard scenes appear comparable to the baseline by removing its failures from the clearance, gracefulness, and path-length averages. Success rate is reported over all trials, but the other metrics are not. In addition, only QP solving time is reported with a standard deviation; no confidence intervals or statistical tests are given for success rate, distance, or collision outcomes. Please report the unselected metrics (or per-method success sets) and include variability estimates.
minor comments (3)
  1. [§IV-C.2, Eq. (15)] In the definition of the virtual sensor set D, the tuple (T_Dj, r_j, K) does not include p_j, although the text says a sensor is placed on each sphere and Eq. (18) uses the sphere-centre distance. Clarify how T_Dj relates to the sphere position p_j.
  2. [§V-B.4] The reported control frequencies (17.2 Hz and 24.2 Hz) are far below what the sub-millisecond QP solving times in Table I would allow; the bottleneck is the distance query. This is worth stating explicitly, since one of the paper's contributions is the efficiency trade-off.
  3. [§IV-E] The 'dynamic gain of [38]' is not defined; please give the value or the relevant equation from [38] so that the cost magnitude is reproducible.

Circularity Check

0 steps flagged

No significant circularity: GS-based distances feed a QP whose claims are evaluated against an independent analytical SDF baseline.

full rationale

The paper's derivation chain is self-contained in the relevant sense. The controller QP is inherited from an external formulation [1] and from the same group's RMMI [38], but the novel component — obtaining robot-obstacle distances from a Gaussian Splat — is computed by two concrete algorithms (bisection-based sphere-to-ellipsoid distance, and median-depth rasterisation from six virtual cameras), neither of which is fit to the reported success metrics. The central comparison is against a baseline that uses an analytical signed distance function, i.e. ground-truth geometry, so the claim of comparable performance is not forced by construction. The reused 'dynamic gain of [38]' is a parameter transfer, not a fitted prediction. The paper also explicitly acknowledges its main limitation (the environment must be fully mapped), which is a scope condition, not a circular step. The reviewer-flagged issue of the Eq. (19) gradient not being a unit vector and not being explicitly rotated into the world frame is a correctness/soundness concern about the implementation as written, not a circularity: it does not make any output equal to an input by definition. No equation in the paper reduces to another by construction, and no fitted parameter is renamed as a prediction. Therefore the circularity score is 0.

Axiom & Free-Parameter Ledger

4 free parameters · 6 axioms · 1 invented entities

The central method depends on hand-chosen controller parameters (d_i, d_s, active-cost gain), a manually specified robot-sphere model, a fixed six-view sensor geometry, and the domain assumption that a trained Gaussian Splat is a metric and geometric proxy for the scene. These are not fitted to the reported success metrics, but they are load-bearing and mostly unspecified, limiting reproducibility.

free parameters (4)
  • influence distance d_i and stopping distance d_s = not reported (inherited from RMMI [38])
    Used in Eq. (21) to scale velocity-damper constraints and in Eq. (24) to weight the collision cost; hand-chosen values directly affect how early reactive avoidance begins.
  • 77-sphere robot approximation = not provided
    Section V-A.2; manually placed sphere radii and link offsets define all distance queries; different approximations change the collision constraints and thus the results.
  • virtual camera intrinsics K and six-view orientation set C_R = 90° FOV; Rx(0), Rx(±π/2), Ry(π/2), Ry(π), Ry(3π/2)
    Eqs. (15)–(16); hand-chosen sampling determines which directions are checked for obstacles and therefore the blind-spot coverage of the rasterisation method.
  • active collision cost gain = not reported (dynamic gain of [38])
    Section IV-E; scalar gain in the collision cost set by prior work; tuning would change avoidance aggressiveness versus tracking performance.
axioms (6)
  • domain assumption The trained Gaussian Splat / 2DGS reconstruction is a geometrically accurate metric-scale proxy for the scene's occupied surfaces.
    Used throughout; Section V-A.1 (training from RGB-D point clouds) and Section V-C rely on this. Section V-D shows performance degrades when the assumption fails, confirming it is load-bearing.
  • domain assumption Median-depth rasterisation (Eqs. 7–8) returns a meaningful surface distance per pixel, and opacity-weighted transmittance separates real surfaces from floaters.
    Section III-C and IV-C.2; the depth-rasterisation distance sensor depends on this to filter low-opacity ellipsoids.
  • domain assumption The six fixed virtual-camera directions C_R cover all relevant obstacle directions for each sphere at every control step.
    Eq. (16); if an obstacle lies outside all six frustums, it generates no constraint, so collision avoidance is blind in that direction.
  • domain assumption The base/arm kinematic QP controller from [1]/[38] remains stable when the distance constraints and collision cost are appended.
    Sections III-A and IV-D/E; the method inherits rather than re-derives the whole-body controller, including manipulability and base-orientation costs.
  • domain assumption The environment is static and fully mapped before execution.
    Stated as a limitation in Section VI; incremental updates are left to future work.
  • standard math The closest point on an ellipsoid can be obtained robustly via the convex root-finding/bisection method of [37].
    Used in Section IV-C.1 to compute sphere-to-ellipsoid distances; accepted numerical algorithm.
invented entities (1)
  • Virtual camera distance sensor (six virtual cameras per robot sphere) no independent evidence
    purpose: Rasterises six depth images per sphere to estimate obstacle distance and gradient for the QP constraints and cost.
    A design construct; its coverage and FOV are hand-chosen, and it is validated only indirectly through control success, not by direct comparison to measured distances.

pith-pipeline@v1.3.0-alltime-deepseek · 11083 in / 13785 out tokens · 138992 ms · 2026-08-03T17:23:02.553978+00:00 · methodology

0 comments
read the original abstract

Reactive control can gracefully coordinate the motion of the base and the arm of a mobile manipulator. However, incorporating an accurate representation of the environment to avoid obstacles without involving costly planning remains a challenge. In this work, we present ReMoSPLAT, a reactive controller based on a quadratic program formulation for mobile manipulation that leverages a Gaussian Splat representation for collision avoidance. By integrating additional constraints and costs into the optimisation formulation, a mobile manipulator platform can reach its intended end effector pose while avoiding obstacles, even in cluttered scenes. We investigate the trade-offs of two methods for efficiently calculating robot-obstacle distances, comparing a purely geometric approach with a rasterisation-based approach. Our simulation experiments on both synthetic and real-world scans demonstrate the feasibility of the proposed method, achieving performance comparable to controllers that rely on perfect ground-truth information. We further validate the approach on a real robot platform more details: https://remosplat.github.io

Figures

Figures reproduced from arXiv: 2512.09656 by Nicolas Marticorena, Niko Suenderhauf, Tobias Fischer.

Figure 1
Figure 1. Figure 1: Illustration of our reactive control approach for a mobile manipulator [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Depiction of the median depth rasterisation process. In blue, we → [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: High-level diagram of our approach, showing the same robot state and how it models the constraints depending on the chosen distance method. [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figure 2
Figure 2. Figure 2: The depth rasterisation process returns a set of 6 depth maps Z j c [u, v] ∈ R + for each j-th sensor, where each value encodes the distance along the camera’s optical (z-axis) direction. Each depth map is backprojected into a 3D point cloud expressed in the c-th camera frame: cPj =    cp(u, v) = Z j c (u, v) K−1   u v 1   [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 4
Figure 4. Figure 4: Resulting 2DGS Ellipsoids on a sample of the bookshelf and table [PITH_FULL_IMAGE:figures/full_fig_p006_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Resulting Gaussian Splats of the printer scene, including the resulting motions showing in red and black the trajectories of the end effector and the [PITH_FULL_IMAGE:figures/full_fig_p007_5.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

43 extracted references · 9 linked inside Pith

  1. [1]

    A Holistic Approach to Reactive Mobile Manipulation,

    J. Haviland, N. Sunderhauf, and P. Corke, “A Holistic Approach to Reactive Mobile Manipulation,”IEEE Robotics and Automation Letters, vol. 7, no. 2, pp. 3122–3129, 2022

    2022

  2. [2]

    Mobile Manipulation in Unknown Environments with Differential Inverse Kinematics Control,

    A. Heins, M. Jakob, and A. P. Schoellig, “Mobile Manipulation in Unknown Environments with Differential Inverse Kinematics Control,” Conference on Robots and Vision, pp. 64–71, 2021

    2021

  3. [3]

    Local reactive control for mobile manipulators with whole-body safety in complex environments,

    C. Zhenget al., “Local reactive control for mobile manipulators with whole-body safety in complex environments,”IEEE Robotics and Automation Letters, vol. 10, no. 5, pp. 4556–4563, 2025

    2025

  4. [4]

    A review of the challenges in mobile manipulation: systems design and robocup challenges,

    M. Sereinig, W. Werth, and L.-M. Faller, “A review of the challenges in mobile manipulation: systems design and robocup challenges,” Elektrotechnik und Informationstechnik, vol. 137, pp. 297–308, 10 2020

    2020

  5. [5]

    Decentralized planning and control for UA V–UGV cooperative teams,

    B. Arbanaset al., “Decentralized planning and control for UA V–UGV cooperative teams,”Autonomous Robots, vol. 42, no. 8, pp. 1601–1618, 2018

    2018

  6. [6]

    Rgb-only reconstruction of tabletop scenes for collision-free manipulator control,

    Z. Tanget al., “Rgb-only reconstruction of tabletop scenes for collision-free manipulator control,” inIEEE International Conference on Robotics and Automation, 2023, pp. 1778–1785

    2023

  7. [7]

    Closing the loop: Real-time perception and control for robust collision avoidance with occluded obstacles,

    A. Tulbure and O. Khatib, “Closing the loop: Real-time perception and control for robust collision avoidance with occluded obstacles,” in IEEE/RSJ International Conference on Intelligent Robots and Systems, 2020, pp. 5700–5707

    2020

  8. [8]

    3D Gaussian Splatting for Real-Time Radiance Field Rendering,

    B. Kerblet al., “3D Gaussian Splatting for Real-Time Radiance Field Rendering,”ACM Transactions on Graphics, vol. 42, no. 4, pp. 1–14, 2023

    2023

  9. [9]

    3D Gaussian Splatting in Robotics: A Survey,

    S. Zhuet al., “3D Gaussian Splatting in Robotics: A Survey,”arXiv preprint arXiv:2410.12262, 2024

    Pith/arXiv arXiv 2024

  10. [10]

    Splat-nav: Safe real-time robot navigation in gaussian splatting maps,

    T. Chenet al., “Splat-nav: Safe real-time robot navigation in gaussian splatting maps,”IEEE Transactions on Robotics, vol. 41, pp. 2765– 2784, 2025

    2025

  11. [11]

    FOCI: Trajectory Optimization on Gaussian Splats,

    M. G. Andreuet al., “FOCI: Trajectory Optimization on Gaussian Splats,”arXiv preprint arXiv:2505.08510, 2025

    Pith/arXiv arXiv 2025

  12. [12]

    SAFER-Splat: A Control Barrier Function for Safe Navigation with Online Gaussian Splatting Maps,

    T. Chenet al., “SAFER-Splat: A Control Barrier Function for Safe Navigation with Online Gaussian Splatting Maps,”arXiv preprint arXiv:2409.09868, 2024

    Pith/arXiv arXiv 2024

  13. [13]

    Motion planning with sequential convex opti- mization and convex collision checking,

    J. Schulmanet al., “Motion planning with sequential convex opti- mization and convex collision checking,”The International Journal of Robotics Research, vol. 33, no. 9, pp. 1251–1270, 8 2014

    2014

  14. [14]

    CHOMP: Covariant Hamiltonian optimization for motion planning,

    M. Zuckeret al., “CHOMP: Covariant Hamiltonian optimization for motion planning,”International Journal of Robotics Research, vol. 32, no. 9-10, pp. 1164–1193, 2013

    2013

  15. [15]

    Differentiable Collision Detection for a Set of Convex Primitives,

    K. Tracy, T. A. Howell, and Z. Manchester, “Differentiable Collision Detection for a Set of Convex Primitives,” inIEEE International Conference on Robotics and Automation, 2023, pp. 3663–3670

    2023

  16. [16]

    OctoMap: An efficient probabilistic 3D mapping framework based on octrees,

    A. Hornunget al., “OctoMap: An efficient probabilistic 3D mapping framework based on octrees,”Autonomous Robots, vol. 34, no. 3, pp. 189–206, 2013

    2013

  17. [17]

    V oxfield: Non-Projective Signed Distance Fields for Online Planning and 3D Reconstruction,

    Y . Panet al., “V oxfield: Non-Projective Signed Distance Fields for Online Planning and 3D Reconstruction,” inIEEE International Conference on Intelligent Robots and Systems, 2022, pp. 5331–5338

    2022

  18. [18]

    Perceptive model predictive control for continuous mobile manipulation,

    J. Pankert and M. Hutter, “Perceptive model predictive control for continuous mobile manipulation,”IEEE Robotics and Automation Letters, vol. 5, no. 4, pp. 6177–6184, 2020

    2020

  19. [19]

    NeRF: Representing Scenes as Neural Radiance Fields for View Synthesis,

    B. Mildenhallet al., “NeRF: Representing Scenes as Neural Radiance Fields for View Synthesis,”https://dl.acm.org/doi/10.1145/3503250, 2020

    doi:10.1145/3503250 2020

  20. [20]

    Splatam: Splat, track & map 3d gaussians for dense rgb-d slam,

    N. Keethaet al., “Splatam: Splat, track & map 3d gaussians for dense rgb-d slam,” inIEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024, pp. 21 357–21 366

    2024

  21. [21]

    Gaussian Splatting SLAM,

    H. Matsukiet al., “Gaussian Splatting SLAM,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024, pp. 18 039–18 048

    2024

  22. [22]

    Splat-slam: Globally optimized rgb-only slam with 3d gaussians,

    E. Sandstr ¨omet al., “Splat-slam: Globally optimized rgb-only slam with 3d gaussians,”arXiv preprint arXiv:2405.16544, 2024

    Pith/arXiv arXiv 2024

  23. [23]

    GraspSplats: Efficient Manipulation with 3D Feature Splatting,

    M. Jiet al., “GraspSplats: Efficient Manipulation with 3D Feature Splatting,”arXiv preprint arXiv:2409.02084, 9 2024

    Pith/arXiv arXiv 2024

  24. [24]

    Gaussiangrasper: 3d language gaussian splatting for open-vocabulary robotic grasping,

    Y . Zhenget al., “Gaussiangrasper: 3d language gaussian splatting for open-vocabulary robotic grasping,”IEEE Robotics and Automation Letters, vol. 9, no. 9, pp. 7827–7834, 2024

    2024

  25. [25]

    Physgaussian: Physics-integrated 3d gaussians for generative dynamics,

    T. Xieet al., “Physgaussian: Physics-integrated 3d gaussians for generative dynamics,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024, pp. 4389–4398

    2024

  26. [26]

    A Purely-Reactive Manipulability- Maximising Motion Controller,

    J. Haviland and P. Corke, “A Purely-Reactive Manipulability- Maximising Motion Controller,”arXiv preprint arXiv:2002.11901, 2020

    Pith/arXiv arXiv 2002

  27. [27]

    Manipulability of Robotic Mechanisms,

    T. Yoshikawa, “Manipulability of Robotic Mechanisms,”The Interna- tional Journal of Robotics Research, vol. 4, no. 2, pp. 3–9, 1985

    1985

  28. [28]

    Structure-from-Motion Revisited,

    J. L. Sch ¨onberger and J.-M. Frahm, “Structure-from-Motion Revisited,” inIEEE/CVF Conference on Computer Vision and Pattern Recognition, 2016, pp. 4104–4113

    2016

  29. [29]

    2D Gaussian Splatting for Geometrically Accurate Radiance Fields,

    B. Huanget al., “2D Gaussian Splatting for Geometrically Accurate Radiance Fields,”SIGGRAPH 2024 Conference Papers, vol. 1, no. 1, pp. 1–13, 2024

    2024

  30. [30]

    RaDe-GS: Rasterizing Depth in Gaussian Splatting,

    B. Zhanget al., “RaDe-GS: Rasterizing Depth in Gaussian Splatting,” arXiv preprint arXiv:2406.01467, 2024

    Pith/arXiv arXiv 2024

  31. [31]

    SuGaR: Surface-Aligned Gaussian Splat- ting for Efficient 3D Mesh Reconstruction and High-Quality Mesh Rendering,

    A. Gu ´edon and V . Lepetit, “SuGaR: Surface-Aligned Gaussian Splat- ting for Efficient 3D Mesh Reconstruction and High-Quality Mesh Rendering,” inIEEE/CVF Conference on Computer Vision and Pattern Recognition, 6 2024, pp. 5354–5363

    2024

  32. [32]

    GSFusion: Online RGB-D Mapping Where Gaussian Splatting Meets TSDF Fusion,

    J. Wei and S. Leutenegger, “GSFusion: Online RGB-D Mapping Where Gaussian Splatting Meets TSDF Fusion,”IEEE Robotics and Automation Letters, vol. 9, no. 12, pp. 11 865–11 872, 12 2024

    2024

  33. [33]

    GSDF: 3DGS Meets SDF for Improved Rendering and Reconstruction,

    M. Yuet al., “GSDF: 3DGS Meets SDF for Improved Rendering and Reconstruction,”arXiv preprint arXiv:2403.16964, 2024

    Pith/arXiv arXiv 2024

  34. [34]

    CuRobo: Parallelized Collision-Free Robot Motion Generation,

    B. Sundaralingamet al., “CuRobo: Parallelized Collision-Free Robot Motion Generation,”IEEE International Conference on Robotics and Automation, pp. 8112–8119, 2023

    2023

  35. [35]

    Foam: A tool for spherical approximation of robot geometry,

    S. Coumaret al., “Foam: A tool for spherical approximation of robot geometry,”arXiv preprint arXiv:2503.13704, 2025

    Pith/arXiv arXiv 2025

  36. [36]

    Point-to-ellipse and point-to- ellipsoid distance equation analysis,

    A. Y . Uteshev and M. V . Goncharova, “Point-to-ellipse and point-to- ellipsoid distance equation analysis,”Journal of Computational and Applied Mathematics, vol. 328, pp. 232–251, 2018

    2018

  37. [37]

    Distance from a point to an ellipse, an ellipsoid, or a hyperellipsoid,

    D. Eberly, “Distance from a point to an ellipse, an ellipsoid, or a hyperellipsoid,” 2013, accessed: 2025-02-19. [Online]. Available: https://www.geometrictools.com/Documentation/ DistancePointEllipseEllipsoid.pdf

    2013

  38. [38]

    RMMI: Reactive Mobile Manipulation using an Implicit Neural Map,

    N. Marticorenaet al., “RMMI: Reactive Mobile Manipulation using an Implicit Neural Map,” inIEEE/RSJ International Conference on Intelligent Robots and Systems, 2025

    2025

  39. [39]

    Not your grandmother’s toolbox - the Robotics Toolbox reinvented for Python,

    P. Corke and J. Haviland, “Not your grandmother’s toolbox - the Robotics Toolbox reinvented for Python,”IEEE International Confer- ence on Robotics and Automation, pp. 11 357–11 363, 2021

    2021

  40. [40]

    NeRFCapture,

    J. Abou-Chakra, “NeRFCapture,” 3 2023. [Online]. Available: https://github.com/jc211/NeRFCapture

    2023

  41. [41]

    Physically embodied gaussian splatting: A realtime correctable world model for robotics,

    J. Abou-Chakraet al., “Physically embodied gaussian splatting: A realtime correctable world model for robotics,” inConference on Robot Learning, 2024

    2024

  42. [42]

    Dynamic 3d gaussians: Tracking by persistent dynamic view synthesis,

    J. Luitenet al., “Dynamic 3d gaussians: Tracking by persistent dynamic view synthesis,” in3DV, 2024

    2024

  43. [43]

    LangSplat: 3D Language Gaussian Splatting,

    M. Qinet al., “LangSplat: 3D Language Gaussian Splatting,” in IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024, pp. 20 051–20 060

    2024