Pith. sign in

REVIEW 3 major objections 5 minor 31 references

CLAP: Clustering to Localize Across n Possibilities, A Simple, Robust Geometric Approach in the Presence of Symmetries

T0 review · 3 major / 5 minor · reviewed 2026-08-15 · deepseek-v4-flash

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

Pith's one-line read The paper claims that clustering many candidate poses generated from pairs of landmarks localizes a robot robustly, because true candidates coincide while false ones scatter.

desk verdict Genuinely simple and effective localization for humanoid soccer, with a real insight about clustering, but the robustness headline is under-tested against coherent false landmarks; deserves a careful referee. read the letter →

arxiv 2509.08495 v1 pith:VROXULIZ submitted 2025-09-10 cs.RO

Gabriel I. Fernandez , Ruochen Hou , Alex Xu , Colin Togashi , Dennis W. Hong This is my paper · ORCID
classification cs.RO
keywords robotlocalizationlandmark-basedposeclusteringK-meansfalsepositiverobustnesshumanoidroboticsgeometricestimationsymmetryambiguity
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

CLAP claims that a robot can estimate its global pose by generating one candidate pose from every observed pair of landmarks and letting the dense cluster of mutually consistent candidates be the answer. Correct pairings all agree on the pose, while wrong pairings and false detections spread out, so K-means clustering plus the previous pose for disambiguation gives an accurate estimate even when many detected landmarks are false. This matters for autonomous humanoid soccer, where motion blur, changing light, bystanders, and neighboring fields corrupt vision and an unstable pose estimate can make a robot miss goals or score on its own. The paper reports accuracy comparable to existing landmark filters and markedly better robustness under 20-120% false landmarks, and the method ran fast enough on competition hardware to help win the 2024 RoboCup adult-sized humanoid title.

What carries the argument

The engine is Eq. (1), the closed-form pose-from-pair calculation: for two observed landmark positions in the robot frame matched to two landmarks in the a priori world map, the robot orientation is the difference of their bearing angles, and the robot position follows by rotating the first observed landmark into the world frame. This map turns every pair of observations into a hypothesis. The identity that carries the argument is that correct hypotheses coincide, in the noiseless case exactly and in practice in a tight cluster, while incorrect hypotheses, including the mirror solution produced by pairs of identical landmarks, scatter across the field because they must still be consistent with the global map's geometry. K-means then finds the two dominant clusters produced by the twofold field symmetry, iterative outlier removal refines the centroids, and the previous estimate selects the correct one.

What would settle it

Take a map with an additional set of landmarks arranged so a wrong pairing reproduces the same pairwise distances as a true pairing, such as a second identical goal area, then run CLAP repeatedly over trajectories that keep both configurations in view; if the wrong cluster is as dense as the true cluster and persists across several frames, the method should occasionally lock onto the mirrored or repeated pose. Counting those divergence episodes would directly test the assumption that incorrect estimates spread apart.

Watch

Extended reading notes

Core claim

The paper's central claim is that global localization on a known field can be solved by generating a large set of candidate poses from every observed pair of landmarks and then clustering those candidates. For a pair of observed landmarks and a candidate matching pair from the a priori map, the closed-form formula of Eq. (1) gives one robot pose; every correct pairing returns essentially the same pose, while incorrect pairings, mirror ambiguities, and false detections generate poses scattered across the field. The paper argues that the true pose is therefore the dense cluster, and that two clusters appear because a soccer field is symmetric under a 180-degree rotation; the previous pose selects which cluster is the robot's. In real-field comparisons, CLAP's mean absolute position error stayed between roughly 0.18 and 0.22 m, comparable to the two landmark-matching baselines, and under artificially inflated false-landmark ratios of 20-120% CLAP produced far fewer velocity jumps and far less divergence from motion-capture ground truth.

Load-bearing premise

The entire method relies on false or wrongly matched landmarks failing to form their own persistent dense cluster of pose estimates; if repeated structures on or near the field generated a stable wrong cluster, CLAP could lock onto a pose that is not the robot's.

Editorial extensions

If this is right

  • CLAP's accuracy is comparable to the two baselines: position mean absolute error roughly 0.18-0.22 m and orientation error roughly 2.9-3.1 degrees across the three real-field trajectories.
  • Under 20-120% injected false landmarks, CLAP stays close to ground truth with 22-25 velocity jumps, while aMCL has 72-310 jumps and ILM without RANSAC diverges for more than half of the trajectory once the false-to-true ratio exceeds 40%.
  • Because each pose estimate is a simple closed-form calculation from one landmark pair, CLAP runs at 100 Hz on competition hardware while observing up to seven landmarks, and faster in simulation.
  • The pairing formula implies incorrect estimates grow quadratically with false detections, yet they do not cohere into a dense cluster, so the method's robustness degrades gradually rather than suddenly as false landmarks are added.

Reading between the lines

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

  • Editorial inference: The two-centroid structure is a direct consequence of the field's twofold rotational symmetry; on maps with higher-order or broken symmetries, the same pipeline should generalize to n centroids with the previous pose or a map-consistency check selecting among them, a testable extension the paper's final section points toward.
  • Editorial inference: CLAP's robustness depends on the fraction of unambiguous landmark pairs rather than simply the total landmark count; one testable prediction is that inserting false landmarks that form repeated pair distances with true landmarks should hurt more than isolated false landmarks.
  • Editorial inference: Because correct estimates coincide exactly only in the noise-free limit, position noise will smear the true cluster; the method's accuracy should therefore scale with landmark range accuracy, and weighting each estimate by the inverse landmark depth error could improve the centroid without changing the clustering.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. The paper proposes CLAP, a vision-based global localization method for humanoid soccer robots. From every pair of detected field landmarks, CLAP generates pose hypotheses by matching the observed pair against all same-distance pairs in a known map, using the two-point geometric relation in Eq. (1). It then clusters the resulting hypotheses in pose space: a fast local cluster near the previous estimate and a slower global K-means with iterative outlier removal provide the final pose, while a particle filter and an InEKF fuse IMU data between camera updates. The paper reports simulation and real-world accuracy comparable to aMCL and ILM without RANSAC, and reports better robustness under 20-120% randomly appended false landmarks. The method was used during the authors' RoboCup 2024 adult-sized humanoid soccer win.

Significance. If the central claim holds, CLAP is a valuable simple geometric alternative to Monte Carlo and ICP-based localization: it avoids iterative optimization, runs at high rate, and withstands high fractions of random false detections. The paper gives a clear two-point geometric derivation, explicit combinatorics of hypothesis counts, and real-hardware comparison with two baseline methods, and the competitive deployment is a strong practical validation. However, the generality of the 'incorrect estimates spread out' premise is not established for coherent false structures such as neighboring fields, and the strongest natural baseline (ILM with RANSAC) is not evaluated. These gaps limit the strength of the advertised robustness claim.

major comments (3)
  1. [Section II-E and Section III-B]
  2. [Section III-A, Table I, and Figures 8-11]
  3. [Section III-B and Figures 8-11]
minor comments (5)
  1. [Algorithm 1 and Section II-E]
  2. [Section II-D, Eq. (3)]
  3. [Section II-E]
  4. [Section III-A]
  5. [Section III-B and Figure 5b]

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the derivation is standard two-point geometry plus unsupervised clustering, and comparisons are external benchmarks.

full rationale

I walked the derivation chain. Equation (1) is a standard geometric identity: orientation is recovered from the angle difference between a world-frame landmark pair and a body-frame landmark pair, and position is then recovered by rotating one measured landmark into the world frame. Equation (2) counts unique and symmetric landmark pairs, and Eq. (3) counts false-positive pairings; both are combinatorics. The clustering step (Section II-E) is an algorithmic choice, not a definition of the output in terms of itself, and the thresholds δt and δc are tuning parameters, not quantities fitted to force Table I. The robustness experiments are compared against Vicon motion-capture ground truth, which is external to CLAP, and the reported accuracy is comparable to, not derived from, the baselines aMCL and ILM. The self-citations ([1], [11], [28]) are used for context, for the ILM comparison baseline, and for the competition record; none of them supplies the load-bearing derivation of CLAP. The premise that incorrect estimates 'spread out' is an empirical assumption and is not asserted by construction; the paper's Future Work explicitly admits that false positives can introduce false symmetries with more than one possibility, which is a limitation of the robustness claim rather than a circular step. No equation, parameter, or cited result was found to reduce to the paper's own output by definition.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The geometric core has no fitted constants; the algorithm depends on two distance thresholds and an outlier-pruning fraction whose values are not stated. The central robustness behavior rests on an empirical cluster-separability assumption rather than on free parameters, plus planar map and two-fold symmetry assumptions.

free parameters (3)
  • delta_t
    Local clustering distance threshold in Algorithm 1; not reported numerically, so its tuning is unknown.
  • delta_c
    Agreement threshold between local and global cluster in Algorithm 1; not reported numerically.
  • K-means outlier removal fraction
    Section II-E says a percentage of outliers is removed iteratively, but the percentage and number of iterations are unspecified.
assumptions (4)
  • domain assumption The robot pose is planar and the field has a known a priori map of labeled landmarks (corners, T-junctions, goal posts, crosses); camera observations are projected onto the field plane.
    Section II-C and II-D define the problem relative to a known global frame and use planar projections for Eq. (1). If the map or planar projection is wrong, every pair estimate inherits that error.
  • domain assumption Incorrect pose estimates spread out in position and orientation while correct estimates form a dense cluster, so clustering can separate them.
    Section II-E states that incorrect estimates 'spread out' and only consistent estimates 'densely cluster'; this is the central statistical premise of CLAP.
  • domain assumption The soccer field has two-fold symmetry, so K-means with two centroids and selection of the centroid nearest the previous pose resolves the ambiguity.
    Section II-E: 'we can consistently identify 2 centroids due to the two-fold symmetry of soccer fields.' On asymmetric maps or with persistent false symmetries this step can fail.
  • domain assumption The previous state estimate is close enough to the true pose that choosing the nearest centroid is safe.
    Algorithm 1 and Section II-E rely on r_{t-1} to select among the two mirrored clusters; a lost or wrong previous pose could pull selection to the wrong cluster.

how reviews work

0 comments
Cite this review

Pith. "Pith review of CLAP: Clustering to Localize Across n Possibilities, A Simple, Robust Geometric Approach in the Presence of Symmetries." pith.science (2026). https://pith.science/paper/VROXULIZ

@misc{pith2026250908495,
  author       = {Pith},
  title        = {Pith review of: CLAP: Clustering to Localize Across n Possibilities, A Simple, Robust Geometric Approach in the Presence of Symmetries},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/VROXULIZ}},
  note         = {Machine review of arXiv:2509.08495}
}
abstract

In this paper, we present our localization method called CLAP, Clustering to Localize Across $n$ Possibilities, which helped us win the RoboCup 2024 adult-sized autonomous humanoid soccer competition. Competition rules limited our sensor suite to stereo vision and an inertial sensor, similar to humans. In addition, our robot had to deal with varying lighting conditions, dynamic feature occlusions, noise from high-impact stepping, and mistaken features from bystanders and neighboring fields. Therefore, we needed an accurate, and most importantly robust localization algorithm that would be the foundation for our path-planning and game-strategy algorithms. CLAP achieves these requirements by clustering estimated states of our robot from pairs of field features to localize its global position and orientation. Correct state estimates naturally cluster together, while incorrect estimates spread apart, making CLAP resilient to noise and incorrect inputs. CLAP is paired with a particle filter and an extended Kalman filter to improve consistency and smoothness. Tests of CLAP with other landmark-based localization methods showed similar accuracy. However, tests with increased false positive feature detection showed that CLAP outperformed other methods in terms of robustness with very little divergence and velocity jumps. Our localization performed well in competition, allowing our robot to shoot faraway goals and narrowly defend our goal.

Figures

Figures reproduced from arXiv: 2509.08495 by the authors.

Figure 1
Figure 1. (a) An image from the Robocup championship match with our robot, [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Flowchart of the full localization framework in [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. In this figure the red player are matched with T and L pairs that share the same relative distance between markers. Note that in order to keep the spatial relationship in the local frame that it cannot be simply mirrored and only one solution exists. The green dotted line always touches the L and the blue dotted line the T. Also notice that the red robots are far from one another. Conversely, the blue robots in the … view at source ↗
Figures from the paper (8 more)
Figure 4
Figure 4. Figure 4: This will be resolved in Section II-E. Noise: As the noise in the relative position of the land￾marks increases, matching becomes far more difficult due to field features becoming indistinguishable between other similar pairs. It is unknown if the distance between land…
Figure 4
Figure 4. Figure 4: Depicts all estimated states, assuming we observe all the landmarks [PITH_FULL_IMAGE:figures/full_fig_p004_4.png]
Figure 6
Figure 6. Figure 6: Real localized state at the center of the [PITH_FULL_IMAGE:figures/full_fig_p006_6.png]
Figure 7
Figure 7. Figure 7: Three trajectories recorded on a real soccer field. Only a single lap [PITH_FULL_IMAGE:figures/full_fig_p006_7.png]
Figure 8
Figure 8. Figure 8: Comparing the localization results of trajectory 3 in full from CLAP, aMCL, and ILM without RANSAC when subjected to false landmarks. As [PITH_FULL_IMAGE:figures/full_fig_p007_8.png]
Figure 9
Figure 9. Figure 9: Closeup results of Fig. 8 with 120% more false landmarks than [PITH_FULL_IMAGE:figures/full_fig_p007_9.png]
Figure 10
Figure 10. Figure 10: Number of velocity jumps in a similar but longer path to trajectory [PITH_FULL_IMAGE:figures/full_fig_p007_10.png]
Figure 11
Figure 11. Figure 11: Percentage of the total path that has diverged from the motion [PITH_FULL_IMAGE:figures/full_fig_p008_11.png]

Review history (2 revisions) →

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

31 extracted references · 26 canonical work pages

  1. [1]

    Model predictive control with visibility graphs for humanoid path planning and tracking against adversarial opponents,

    R. Hou, G. Fernandez, M. Zhu, and D. Hong, “Model predictive control with visibility graphs for humanoid path planning and tracking against adversarial opponents,” in2025 IEEE International Conference on Robotics and Automation. IEEE Robotics and Automation Society,

    work page

  2. [2]

    Vision-based odometric localization for humanoids using a kinematic ekf,

    G. Oriolo, A. Paolillo, L. Rosa, and M. Vendittelli, “Vision-based odometric localization for humanoids using a kinematic ekf,” in 2012 12th IEEE-RAS International Conference on Humanoid Robots (Humanoids 2012). IEEE, 2012, pp. 153–158

    work page 2012

  3. [3]

    Posenet: A convolutional network for real-time 6-dof camera relocalization,

    A. Kendall, M. Grimes, and R. Cipolla, “Posenet: A convolutional network for real-time 6-dof camera relocalization,” inProceedings of the IEEE international conference on computer vision, 2015, pp. 2938–2946

    2015

  4. [4]

    Enhancing adultsize humanoid localization accuracy: A vision-based amcl leveraging object detection model and hungarian algorithm,

    J. Y . Kim, M. S. Ahn, and J. Han, “Enhancing adultsize humanoid localization accuracy: A vision-based amcl leveraging object detection model and hungarian algorithm,” in2023 IEEE-RAS 22nd Interna- tional Conference on Humanoid Robots (Humanoids). IEEE, 2023, pp. 1–8

    work page 2023

  5. [5]

    Thrun, W

    S. Thrun, W. Burgard, and D. Fox,Probabilistic Robotics. MIT Press, 2005

    2005

  6. [6]

    Robust monte carlo localization for humanoid soccer robot,

    W. Hong, C. Zhou, and Y . Tian, “Robust monte carlo localization for humanoid soccer robot,” in2009 IEEE/ASME International Confer- ence on Advanced Intelligent Mechatronics. IEEE, 2009, pp. 934– 939

    work page 2009

  7. [7]

    Humanoid robot localization in complex indoor environments,

    A. Hornung, K. M. Wurm, and M. Bennewitz, “Humanoid robot localization in complex indoor environments,” in2010 IEEE/RSJ International Conference on Intelligent Robots and Systems. IEEE, 2010, pp. 1690–1695

    work page 2010

  8. [8]

    Vision-based monte- carlo localization for humanoid soccer robots,

    A. C. Almeida, A. H. Costa, and R. A. Bianchi, “Vision-based monte- carlo localization for humanoid soccer robots,” in2017 Latin American robotics symposium (LARS) and 2017 Brazilian symposium on robotics (SBR). IEEE, 2017, pp. 1–6

    work page 2017

Show all 31 references
  1. [9]

    Vision-based monte carlo localization for robocup humanoid kid-size league,

    I. Nagi, W. Adiprawita, and K. Mutijarsa, “Vision-based monte carlo localization for robocup humanoid kid-size league,” in2014 13th International Conference on Control Automation Robotics & Vision (ICARCV). IEEE, 2014, pp. 1433–1438

    2014

  2. [10]

    Method for registration of 3-D shapes,

    P. J. Besl and N. D. McKay, “Method for registration of 3-D shapes,” inSensor Fusion IV: Control Paradigms and Data Structures, P. S. Schenker, Ed., vol. 1611, International Society for Optics and Photonics. SPIE, 1992, pp. 586 – 606. [Online]. Available: https://doi.org/10.11...

    1992 doi

  3. [11]

    Fast and robust localization for humanoid soccer robot via iterative landmark matching,

    R. Hou, G. Fernandez, and D. Hong, “Fast and robust localization for humanoid soccer robot via iterative landmark matching,” 2025. [Online]. Available: https://arxiv.org/abs/2503.11020

    2025 arXiv

  4. [12]

    Some fundamental concepts and synthesis procedures for pattern recognition preprocessors,

    G. Ball, “Some fundamental concepts and synthesis procedures for pattern recognition preprocessors,” inProc. Int. Conf. Microwaves, Circuit Theory, and Information Teory, 1964, pp. 281–297

    1964

  5. [13]

    Algorithms for clustering data,

    A. Jain, “Algorithms for clustering data,” 1988

    1988

  6. [14]

    Dbscan: Past, present and future,

    K. Khan, S. U. Rehman, K. Aziz, S. Fong, and S. Sarasvady, “Dbscan: Past, present and future,” inThe fifth international conference on the applications of digital information and web technologies (ICADIWT 2014). IEEE, 2014, pp. 232–238

    2014

  7. [15]

    Maximum likelihood from incomplete data via the em algorithm,

    A. P. Dempster, N. M. Laird, and D. B. Rubin, “Maximum likelihood from incomplete data via the em algorithm,”Journal of the royal statistical society: series B (methodological), vol. 39, no. 1, pp. 1– 22, 1977

    1977

  8. [16]

    The estimation of the gradient of a density function, with applications in pattern recognition,

    K. Fukunaga and L. Hostetler, “The estimation of the gradient of a density function, with applications in pattern recognition,”IEEE Transactions on information theory, vol. 21, no. 1, pp. 32–40, 1975

    1975

  9. [17]

    Mean shift, mode seeking, and clustering,

    Y . Cheng, “Mean shift, mode seeking, and clustering,”IEEE transac- tions on pattern analysis and machine intelligence, vol. 17, no. 8, pp. 790–799, 1995

    1995

  10. [18]

    Some methods for classification and analysis of multi- variate observations,

    J. Macqueen, “Some methods for classification and analysis of multi- variate observations,” inProceedings of 5-th Berkeley Symposium on Mathematical Statistics and Probability/University of California Press, 1967

    1967

  11. [19]

    Cluster analysis of multivariate data: efficiency versus interpretability of classifications,

    E. W. Forgy, “Cluster analysis of multivariate data: efficiency versus interpretability of classifications,”biometrics, vol. 21, pp. 768–769, 1965

    1965

  12. [20]

    Ik-means-+: An iterative clustering algorithm based on an enhanced version of the k-means,

    H. Ismkhan, “Ik-means-+: An iterative clustering algorithm based on an enhanced version of the k-means,”Pattern Recognition, vol. 79, pp. 402–413, 2018

    2018

  13. [21]

    An iterative improved k-means clustering,

    M. Harale and U. L. Kulkarni, “An iterative improved k-means clustering,” inInternational Conference on Advances in Computer Engineering. Citeseer, 2011, pp. 25–28

    2011

  14. [22]

    Contact- aided invariant extended kalman filtering for robot state estimation,

    R. Hartley, M. Ghaffari, R. M. Eustice, and J. W. Grizzle, “Contact- aided invariant extended kalman filtering for robot state estimation,”

  15. [23]

    A particle filtering framework for joint video tracking and pose estimation,

    C. Chen and D. Schonfeld, “A particle filtering framework for joint video tracking and pose estimation,”IEEE Transactions on Image Processing, vol. 19, no. 6, pp. 1625–1634, 2010

    2010

  16. [24]

    2d particle filter accelerator for mobile robot indoor localization and pose estimation,

    O. Tariq and D. Han, “2d particle filter accelerator for mobile robot indoor localization and pose estimation,”IEEE Access, 2024

    2024

  17. [25]

    Ultralytics YOLO,

    G. Jocher, A. Chaurasia, and J. Qiu, “Ultralytics YOLO,” inGithub,

  18. [26]

    Real-time object recognition for football field landmark detection based on deep neural networks,

    M. R. Dwijayanto, S. Kurniawan, and B. Sugandi, “Real-time object recognition for football field landmark detection based on deep neural networks,” in2019 2nd International Conference on Applied Engi- neering (ICAE). IEEE, 2019, pp. 1–5

    2019

  19. [27]

    Development and real-time optimization-based control of a full-sized humanoid for dynamic walking and running,

    M. S. Ahn, “Development and real-time optimization-based control of a full-sized humanoid for dynamic walking and running,” inPhD Thesis. University of California, Los Angeles, 2023

    2023

  20. [28]

    Robocup 2024 adult-sized humanoid cham- pions guide for hardware, vision, and strategy,

    G. I. Fernandezet al., “Robocup 2024 adult-sized humanoid cham- pions guide for hardware, vision, and strategy,” inRoboCup 2024: Robot World Cup XXVII. Cham: Springer Nature Switzerland, 2025, pp. 502–514

    2024

  21. [2019]

    Available: https://arxiv.org/abs/1904.09251

    [Online]. Available: https://arxiv.org/abs/1904.09251

    1904 arXiv

  22. [2023]

    Available: https://github.com/ultralytics/ultralytics

    [Online]. Available: https://github.com/ultralytics/ultralytics

  23. [2025]

    Available: https://arxiv.org/abs/2504.02184

    [Online]. Available: https://arxiv.org/abs/2504.02184

    arXiv

Pith tools

Reviewed August 15, 2026 · model on record in the stance chip above.