Grasping by parallel shape matching

Fabio Ramos; Fahira Afzal Maken; Tin Lai; Wenzheng Zhang

REVIEW 5 major objections 6 minor 37 references

Grasping by parallel shape matching

T0 review · 5 major / 6 minor · reviewed 2026-08-11 · deepseek-v4-flash

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

Pith's one-line read Grasp planning reduces to rigid point-cloud shape matching, and a parallel annealed ICP optimizer finds stable power grasps in under a second, with 87.3% success on real hardware.

desk verdict A genuinely training-free grasp planner with a plausible new cost combination, but the evidence is too thin and the stability proxy too heuristic to back the 'robust optimal' claim as stated. read the letter →

arxiv 2412.08346 v1 pith:NC5IMOIG submitted 2024-12-11 cs.RO

Wenzheng Zhang , Fahira Afzal Maken , Tin Lai , Fabio Ramos This is my paper · ORCID

classification cs.RO

keywords graspsynthesispointcloudmatchingAnnealedSteinICPpowercollisioncheckingsigneddistancefieldGPUparallelizationtraining-freegrasping

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

This paper tries to establish that grasp synthesis does not need joint-angle optimization, learned priors, or force-closure reasoning: it can be solved as rigid shape matching between the gripper's inner-surface point cloud and the object's point cloud. The authors show that a parallel, GPU-based optimizer running many initial poses with Annealed Stein ICP finds power grasps that generalize across grippers and to unseen objects with noisy single-view point clouds. In real experiments with a three-fingered gripper, the method lifts and holds objects in 87.3% of trials, averaging 0.926 seconds per grasp, and it outperforms a learned baseline in simulation. If this holds, a purely geometric, training-free grasp planner can serve as a fast and portable alternative to data-driven approaches for power grasps.

What carries the argument

The central object is the rigid shape-matching cost $L = L_{\mathrm{ct}} + L_{\mathrm{com}}$, where $L_{\mathrm{ct}}$ is the mean squared distance between paired points on the gripper's inner surface and the object cloud, and $L_{\mathrm{com}}$ pulls the gripper's tool center point toward the object's center of mass. The optimizer is Annealed Stein ICP, a particle-based stochastic-gradient variant of ICP in which transformed gripper clouds are spread over the object by a Stein variational repulsion term and annealed exploration, then refined by SGD-ICP. Collision handling replaces the cost with a point-to-surface distance between penetrating object points and the gripper's signed distance field.

What would settle it

Take a set of 50 objects and record, for every grasp pose the planner returns, whether the pose would pass a force-closure test computed from the contact points and surface normals. If a substantial share (say, over 20%) of the chosen poses are not force-closure yet the objects are still lifted reliably, the geometric cost function alone explains stability; if those same poses slip or drop the object, then the matching-plus-center-of-mass proxy is insufficient. A simpler version: run the same planner on low-friction or flexible objects, where geometric matching and center-of-mass alignment look identical to ordinary rigid objects, and compare the success rate to the reported 87.3%.

Watch

Extended reading notes

Core claim

The paper claims that grasping can be recast as an optimization over rigid poses: slide the gripper's inner-surface point cloud onto the object's point cloud until contact, while keeping the gripper's tool center point close to the object's center of mass. The resulting pose is selected from many parallel starting guesses, each propagated with Annealed Stein ICP to avoid local minima, with a signed-distance-field collision check steering colliding samples out of the object. The authors' experiments report 87.3% average success across eleven objects with a three-fingered gripper using noisy single-view point clouds and 0.926 seconds average computation, and the same formulation works for two other grippers in simulation without retraining.

Load-bearing premise

The load-bearing premise is that keeping the gripper's inner-surface cloud close to the object's surface while aligning the tool center point with the object's center of mass is enough to guarantee a stable power grasp, because no friction, force-closure, or contact-stability model enters the cost.

Editorial extensions

If this is right

The same planner can be applied to any gripper by swapping in its point cloud and preshapes, with no retraining, as demonstrated for three gripper designs.
Because the cost uses only geometry, the method works from partial, noisy single-view point clouds, so it can run on a wrist camera without object models.
The planner can also serve as a post-processor that refines candidate poses from learned grasp detectors, raising their success rate while cutting computation.
Optimization time around one second on a laptop GPU brings the method close to real-time operation for power grasps.
Objects that require precision fingertip grasps or have flexible parts remain failure cases, a boundary the authors acknowledge.

Reading between the lines

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

Beyond the paper's claims: because no friction or contact-force model enters the cost, adding a contact-normals or force-closure check as a post-filter could remove many remaining failures without changing the optimizer.
The TCP-to-CoM term is a compact stability heuristic that could be transferred to trajectory or in-hand manipulation planners, not just grasp pose search.
The sensitivity to starting poses suggests a hybrid pipeline: a cheap learned detector proposing a handful of seeds would cut the number of parallel initializations and reduce computation time further, a direction the paper itself hints at.
The rigid-body treatment could extend to vacuum or magnetic grippers, where the contact-surface cloud is the cup or pad, though the center-of-mass alignment weight would need re-tuning.

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.

Desk Editor's Note

read the letter

Read this if you care about training-free analytic grasp planning. The core idea is genuinely simple: treat grasp synthesis as rigid point-cloud matching between a fixed gripper preshape and the object cloud, optimized with parallel annealed Stein ICP plus a TCP-to-CoM term, with SDF collision checks. It runs in about a second on a laptop GPU, and they back it with 55 real-world lifts at 87.3% success.

What's actually new is the combination: rigid, fixed-preshape gripper matching with AS-ICP, the TCP-to-CoM cost term, SDF-based collision handling, and GPU parallelization for grasp pose generation. That specific package isn't in the ISF/MDISF or data-driven baselines they cite. The real-robot results with partial, noisy views are a positive signal, and they're honest about failures on flexible and precision-grasp objects.

The soft spots are real but not fatal. Evaluation is thin to support the headline: 55 real trials total, no error bars, no significance tests, and baseline protocols aren't matched—AnyGrasp got full point clouds while the proposed method used partial views, and trial counts differ. There's also a sign inconsistency between Eq. (12) and Algorithm 1 for the collision condition: Eq. (12) says dist < 0, Algorithm 1 uses > 0, and the text says positive distance means collision. Algorithm 1 matches the text; Eq. (12) is likely a typo, but it's confusing.

The bigger scientific gap is that the cost function is a geometric proxy, not a force-closure guarantee. Lct is average nearest-neighbour distance from inner surface to object; low values mean geometric proximity, not two-sided frictional contact. Lcom assumes you know the CoM, which for unknown objects you don't have from geometry alone—they never specify a density model. Their own failures on flexible and precision-only objects show the proxy's limits. So the 87.3% is empirical support for a heuristic, not evidence of 'robust optimal' grasps. The correct framing would be to present this as a fast, training-free heuristic that works well on rigid, power-graspable objects, and to add statistics, code, and standardized benchmarks before making stronger claims.

No circularity issue exists; success is measured externally by lifting in simulation and on the physical robot. The self-citations to Stein ICP and SGD-ICP are legitimate algorithm components, not result-derivation.

This deserves peer review, though I'd expect major revisions around evaluation rigor and claim framing. If code and data accompany the revision and they tighten the claims, it becomes a useful reference for the grasping community. My honest verdict: promising heuristic, not yet a proven robust optimal solution.

Referee Report

5 major / 6 minor

Summary. The paper proposes a training-free analytic grasp planner that formulates grasp synthesis as rigid shape matching between the gripper's inner-surface point cloud and the object's point cloud. The pose is optimized with a GPU-parallelized annealed Stein ICP, using a cost function that combines a contact matching error with a distance between the gripper's tool center point and the object's center of mass, and a signed-distance-field collision check. Experiments in Isaac Gym compare the method against AnyGrasp with the Franka Hand and against SplitPSO with the Barrett Hand, and a physical experiment with a Kinova KG3 gripper reports an average success rate of 87.3% over 55 trials with a computation time of 0.926 s. The paper claims robustness to noisy partial point clouds and generalization across gripper types without training.

Significance. If the empirical results hold under tighter statistical scrutiny, the method is a useful contribution: it is gripper-agnostic, avoids expensive finger-joint optimization, and demonstrates a training-free alternative to data-driven grasp planners on unseen objects. The paper also gives a detailed algorithmic recipe, a parameter table, and real-robot experiments, which makes the approach reproducible in principle. The main significance is limited by the lack of statistical rigor in the headline real-robot claim, by protocol differences in the baseline comparisons, and by the paper's reliance on a geometric proxy for grasp stability without a force-closure or contact-models analysis. These issues do not invalidate the approach as a heuristic, but they do not support the paper's 'robust optimal grasp poses' claim as stated.

major comments (5)

[Section IV-B, Eqs. (13)-(15)] The loss L = Lct + Lcom is a geometric proximity score rather than a grasp-stability metric. Lct is an average nearest-neighbor distance between the gripper cloud and the object cloud, and Lcom aligns the TCP to a point-cloud centroid. Nothing in this cost function measures contact forces, friction cones, or force closure, so a low L value does not by itself imply a stable grasp. The manuscript's own Section VI-B reports failures on the flexible Hand, the false-joint Helping Tool, and the flat Toy, which are exactly the regimes where the geometric proxy is insufficient. The abstract's 'robust optimal grasp poses' is therefore stronger than the evidence; either add a wrench/force-closure evaluation or revise the claim to a geometric heuristic with empirical success on selected objects.
[Section VI-B, Table IV] The headline 87.3% success rate is based on only 55 physical trials, five per object. The standard error of a binary mean at this success rate is about 4.5 percentage points, and object-level estimates such as 3/5 or 4/5 have wide confidence intervals. The paper reports no error bars, confidence intervals, or significance tests, and there is no baseline evaluated on the same real setup. The simulation comparisons also use different protocols: AnyGrasp receives a full point cloud, SplitPSO receives a full point cloud with twenty trials per object while the proposed method uses fifty trials, and the object sets are not identical across comparisons. The reported margins are therefore difficult to interpret without matched protocols and additional trials.
[Section VI-A and Section V] The simulation and real experiments use six manually added initializations directly above the object, with the text stating that the sampled initializations alone 'do not provide sufficient coverage.' This is a prior that the object is graspable from above, and it is load-bearing for an ICP-based method that is sensitive to local minima. The ablation in Section V varies the number of initializations and the sampling scheme but does not test the method without these manual additions, so the contribution of this manual prior to the reported success rates is unknown. The statement in Section V that 'we do not provide any prior knowledge of the objects' is also inaccurate.
[Algorithm 1, line 1, and Eq. (15)] The center of mass is set to the center of mass of the target point cloud. For unknown objects and single-view partial point clouds, the centroid of the observed points is not the physical center of mass, and the paper gives no density model or estimation procedure for the true CoM. Since Lcom constitutes half of the optimized cost, the final grasp pose depends on this unstated assumption. The paper should either specify an estimation procedure, for example uniform density over a completed mesh, or report sensitivity of the success rate to CoM error.
[Eq. (12), Section IV-C, Algorithm 1 lines 8-10] The collision condition is internally inconsistent. Equation (12) uses the constraint dist(C, SDF(T(G))) < 0, while Section IV-C states that a collision corresponds to a positive distance value from the SDF, and Algorithm 1 applies the collision gradient when dist(C, SDF(...)) > 0. This ambiguity makes the actual collision behavior irreproducible. Since the collision check is central to filtering infeasible poses, the sign convention and the corresponding test must be stated unambiguously.

minor comments (6)

[Section III-C, Eqs. (7) and (3)] The quaternion parameters are updated with the SVGD and SGD rules, but the paper does not state that the quaternion is renormalized to unit length after each update; this should be specified for reproducibility.
[Section IV-B, Eqs. (16)-(17)] The displayed gradients are missing the factor of 2 that arises from differentiating the squared distances; if this factor is absorbed into the learning rate, it should be stated explicitly.
[Tables II and III] The table captions say the success rates are averages over 50 simulations, but the text for SplitPSO says twenty grasps for each object are simulated; the trial counts and object sets should be reconciled and made identical across the compared methods.
[Table IV and Section VI-B] The reported 0.926 s computation time excludes SDF generation, but it is not stated whether perception and collision checking are included; the paper should report the full pipeline time from raw point cloud to grasp pose.
[Section V and Figure 4] The statement that the success rate 'approaches 100%' with more initializations is not accompanied by the actual asymptotic values or error bars; quantitative values and confidence intervals should be added.
[Table I] The entry 'Learning rate and cost weights for SGD' is set to 1, but the algorithm also has an SVGD learning rate, an annealing schedule, and separate cost weights in Eq. (13); the interaction between these parameters should be clarified.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: grasp success is measured externally and the self-cited ICP components are algorithmic building blocks, not fitted predictions.

full rationale

The derivation chain is: (1) define a rigid shape-matching objective L = Lct + Lcom (Eqs. 13-15) with SDF collision constraints; (2) optimize it with parallel AS-ICP; (3) evaluate by physically lifting and holding objects in Isaac Gym and on a real KG3 arm. Success is defined externally ('lift and hold the object for 5 seconds', Section VI-A), not as a function of the optimized loss, so the 87.3% real-world and 82-89% simulation success rates are independent measurements rather than fitted outputs. The self-citations to Stein ICP [3] and SGD-ICP [20,21] supply the gradient-based optimization machinery (Section III); they do not assert that shape matching produces stable grasps and no uniqueness or optimality theorem from the authors is cited to force the choice of objective. The paper itself reports failure modes (flexible Hand, false-joint Helping Tool, precision-only Toy lying flat, large occlusions, Section VI-B), which shows the evaluation is externally falsifiable. These limitations concern the adequacy of the geometric stability proxy, not circularity. No equation is defined in terms of the target result, and no fitted parameter is renamed as a prediction.

Assumptions & free parameters 11 free parameters · 7 assumptions · 0 invented entities

The central claim rests on standard optimization machinery (ICP, SVGD), a domain assumption that geometric matching implies grasp stability, and several hand-chosen parameters including initializations, preshapes, annealing schedule, and cost weights. No new physical entities are introduced.

free parameters (11)

Initialization count and spread = Approximately 100 Fibonacci-sampled poses per preshape plus 6 manually added poses above the object for Franka…
Success rate climbs with the number of initializations, and the six manual poses are added because Fibonacci sampling alone does not cover the top of the object (Section V).
Gaussian mixture prior means in ablation = 4 means
The ablation shows that heuristically choosing four Gaussian means gives higher success with fewer initializations, indicating the headline results depend on initialization choice.
Mini-batch size schedule = min(k, 2kmax/3) per Table I
This schedule is selected to make SGD-ICP and AS-ICP converge within the fixed iteration budget.
Learning rate and cost weights = 1 for both Lct and Lcom
The relative weight between matching error and TCP-to-CoM distance is not optimized; equal weighting is an ad hoc choice that affects the final poses.
Annealing schedule parameters = p=2, C=5, gamma = (mod(k,kmax/5)/(kmax/5))^2
Controls the exploration-exploitation balance in AS-ICP; chosen empirically without a sensitivity study.
Convergence threshold = 0.02% relative error
Stops the optimizer and therefore influences both success rate and computation time.
SVGD and SGD iteration counts = 15 SVGD iterations, 25 SGD iterations
The split between exploration and refinement is fixed by hand; no ablation is reported.
Voxel sizes for point cloud decimation = 0.005 for gripper and object, 0.025 for Barrett gripper
Downsampling choices affect matching fidelity, collision checking, and runtime.
Preshape set size = 10 Franka preshapes, 2 Barrett preshapes
The fixed finger configurations are hand-selected; the method's performance depends on them, and the authors note that better preshapes would require data-driven selection.
SDF offset epsilon between preshape SDFs = Unspecified
Introduced in Section IV-D to separate gripper preshape SDFs; the exact value is not reported.
RBF kernel bandwidth and von Mises prior concentration = Unspecified
Required for the Stein update in Eqs. (8) and (9); inherited from Stein ICP [3] but not stated in this paper.

assumptions (7)

standard math ICP and Stein variational gradient descent provide a well-behaved optimizer for the non-convex matching problem.
Used implicitly throughout Section III and Algorithm 1; relies on prior convergence properties of SGD-ICP and SVGD that are not re-proven here.
domain assumption The distance between the gripper inner surface point cloud and the object point cloud is a valid proxy for contact and grasp quality.
Appears in Eq. (14) and Section IV-B; no friction, force-closure, or contact-stability model is used, yet the optimizer selects low-cost poses as grasps.
domain assumption The object's center of mass can be approximated from its point cloud centroid.
Algorithm 1 line 1 sets CoM from the target object point cloud; in real partial-view experiments this can be biased by occlusion.
domain assumption Fixed preshapes with no finger-joint optimization are sufficient for power grasps on the tested objects.
The method never changes finger angles during optimization; the paper chooses 10 Franka and 2 Barrett preshapes and explicitly leaves precision grasps for future work.
domain assumption The signed distance field of the gripper, with inverse-transformed object points, gives correct collision information.
Used in Eq. (12) and Algorithm 1 line 8; the sign convention is not fully defined and appears inconsistent between the constraint and the pseudocode.
ad hoc to paper A small set of manually selected objects and the lift-and-hold-5-seconds criterion adequately measures grasp success.
Evaluation design in Section VI; success is defined by simulator and robot lift, not by the optimized cost, and only 55 real trials are reported.
ad hoc to paper Manually added initializations above the object are needed for reliable success.
Section V states that six additional initializations directly above the object are manually added; this is auxiliary knowledge not derived from the shape-matching formulation.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Grasping by parallel shape matching." pith.science (2026). https://pith.science/paper/NC5IMOIG

@misc{pith2026241208346,
  author       = {Pith},
  title        = {Pith review of: Grasping by parallel shape matching},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/NC5IMOIG}},
  note         = {Machine review of arXiv:2412.08346}
}

read the original abstract

Grasping is essential in robotic manipulation, yet challenging due to object and gripper diversity and real-world complexities. Traditional analytic approaches often have long optimization times, while data-driven methods struggle with unseen objects. This paper formulates the problem as a rigid shape matching between gripper and object, which optimizes with Annealed Stein Iterative Closest Point (AS-ICP) and leverages GPU-based parallelization. By incorporating the gripper's tool center point and the object's center of mass into the cost function and using a signed distance field of the gripper for collision checking, our method achieves robust grasps with low computational time. Experiments with the Kinova KG3 gripper show an 87.3% success rate and 0.926 s computation time across various objects and settings, highlighting its potential for real-world applications.

Figures

Figures reproduced from arXiv: 2412.08346 by the authors.

**Figure 2.** (a) Initializations sampled from a mixture of Gaussian [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗

**Figure 3.** Preshapes used for simulation in this paper. On the [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗

Figures from the paper (4 more)

**Figure 5.** Figure 5: Best grasp poses with Franka Hand for 50 trials. On [PITH_FULL_IMAGE:figures/full_fig_p006_5.png]

**Figure 6.** Figure 6: Best grasp poses with Barrett Hand for 50 trials. [PITH_FULL_IMAGE:figures/full_fig_p006_6.png]

**Figure 7.** Figure 7: Picture on the left illustrates fifty grasp poses gener [PITH_FULL_IMAGE:figures/full_fig_p007_7.png]

**Figure 8.** Figure 8: (1) Hand Helping Tool (2)-(6) Grasp poses for Hand [PITH_FULL_IMAGE:figures/full_fig_p007_8.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

37 extracted references · 36 canonical work pages

[1]

Matching Step: pairs the transformed source point cloud,S′ ={s′ i}N i=1, with the reference point cloud R ={ri}M i=1 on the basis of a distance metric, where si and ri∈R3 are N and M points in 3D space. The commonly used point-to-point distance metric finds the pair with the nearest neighbour as follows: ˆri = argmin rj∈R ∥s′ i–rj∥ (1) where ˆri is the cl...

work page
[2]

The updated equation for the point-to-point distance met- ric is defined as follows: θk+1 = argmin θ 1 N N∑ i ∥Rksi +tk–ˆri∥2 (2) where k is the iteration number

Minimization Step: updates θk to minimize a loss function defined by the distance between the paired points in the source and reference point clouds. The updated equation for the point-to-point distance met- ric is defined as follows: θk+1 = argmin θ 1 N N∑ i ∥Rksi +tk–ˆri∥2 (2) where k is the iteration number. Equation (2) can be solved in closed-form us...

work page
[3]

Stein icp for uncertainty estimation in point cloud matching,

F. A. Maken, F. Ramos, and L. Ott, “Stein icp for uncertainty estimation in point cloud matching,” Robotics and Automation Letters, 2022

work page 2022
[4]

A survey on learning-based robotic grasping,

K. Kleeberger, R. Bormann, W. Kraus, et al., “A survey on learning-based robotic grasping,” Current Robot Reports, 2020

work page 2020
[5]

Ising Model on the Fibonacci Sphere

A. Pochinok, A. Molochkov, and M. Chernodub, “Ising model on the fibonacci sphere,” arXiv preprint arXiv:2301.06849 , 2023

work page Pith review arXiv 2023
[6]

Real-time grasp planning for multi-fingered hands by finger splitting,

Y. Fan, T. Tang, H. C. Lin, and M. Tomizuka, “Real-time grasp planning for multi-fingered hands by finger splitting,” IROS, 2018

work page 2018
[7]

Annealed stein variational gradi- ent descent,

F. D’Angelo and V. Fortuin, “Annealed stein variational gradi- ent descent,” in 3rd Symposium on Advances in Approximate Bayesian Inference, 2020

work page 2020
[8]

An overview of 3d object grasp synthesis algorithms,

A. Sahbani, S. El-Khoury, and P. Bidaud, “An overview of 3d object grasp synthesis algorithms,” Robotics and Automation Systems, 2012

work page 2012

Show all 37 references

[9]

Optimization model for planning precision grasps with multi-fingered hands,

Y. Fan, X. Zhu, and M. Tomizuka, “Optimization model for planning precision grasps with multi-fingered hands,” Interna- tional Conference on Intelligent Robots and Systems , 2019

2019
[10]

A geometric approach for grasping unknown objects with multifingered hands,

M. Kiatos, S. Malassiotis, and I. Sarantopoulos, “A geometric approach for grasping unknown objects with multifingered hands,” IEEE Transactions on Robotics, 2021

2021
[11]

Grasp plan- ning for customized grippers by iterative surface fitting,

Y. Fan, H. C. Lin, T. Tang, and M. Tomizuka, “Grasp plan- ning for customized grippers by iterative surface fitting,”Inter- national Conference on Automation Science and Engineering , 2018

2018
[12]

Robotic grasp- ing of unknown objects using novel multilevel convolutional neural networks: From parallel gripper to dexterous hand,

Q. Yu, W. Shang, Z. Zhao, S. Cong, and Z. Li, “Robotic grasp- ing of unknown objects using novel multilevel convolutional neural networks: From parallel gripper to dexterous hand,” Transactions on Automation Science and Engineering, 2021

2021
[13]

Efficient grasp planning and exe- cution with multifingered hands by surface fitting,

Y. Fan and M. Tomizuka, “Efficient grasp planning and exe- cution with multifingered hands by surface fitting,” Robotics and Automation Letters, 2019

2019
[14]

When transformer meets robotic grasping: Exploits context for efficient grasp detection,

S. Wang, Z. Zhou, and Z. Kan, “When transformer meets robotic grasping: Exploits context for efficient grasp detection,” IEEE Robotics and Automation Letters, 2022

2022
[15]

Robot grasp synthesis algorithms: A survey,

K. Shimoga, “Robot grasp synthesis algorithms: A survey,” The International Journal of Robotics Research, 1996

1996
[16]

Anygrasp: Robust and efficient grasp perception in spatial and temporal domains,

H. Fang, C. Wang, H. Fang, et al., “Anygrasp: Robust and efficient grasp perception in spatial and temporal domains,” 2023

2023
[17]

Adagrasp: Learning an adaptive gripper-aware grasping policy,

Z. Xu, B. Qi, S. Agrawal, and S. Song, “Adagrasp: Learning an adaptive gripper-aware grasping policy,” in 2021 IEEE International Conference on Robotics and Automation (ICRA), pp. 4620–4626, IEEE, 2021

2021
[18]

Evaluation and selection of grasp quality criteria for dexterous manipulation,

H. Mnyussiwalla, P. Seguin, P. Vulliez, and J. P. Gazeau, “Evaluation and selection of grasp quality criteria for dexterous manipulation,” Journal of Intelligent and Robotic Systems , 2022

2022
[19]

Grasp quality measures: review and performance,

M. Roa and R. Su´ arez, “Grasp quality measures: review and performance,” Autonomous robots, 2015

2015
[20]

Charac- terisation of grasp quality metrics,

C. Rubert, B. Le´ on, A. Morales, and J. Sancho-Bru, “Charac- terisation of grasp quality metrics,” Journal of Intelligent and Robotic Systems, 2018

2018
[21]

that uses Euler angle representation for rotations. Using the loss function L, at iteration k we have an average gradients of the translation components as ¯g(θ1:3 k ,Sk) = 1 m { ∑ s′ i,ˆri∈Pairs (s′ i− ˆri) ∂tk ∂θ1:3 k + ((RkTCP +tk)–CoM ) ∂tk ∂θ1:3 k } , (16) and for the rot...

1900
[22]

A method for registration of 3-d shapes,

P. J. Besl and N. D. McKay, “A method for registration of 3-d shapes,” IEEE Transactions on Pattern Analysis and Machine Intelligence, 1992

1992
[23]

Speeding up iterative closest point using stochastic gradient descent,

F. A. Maken, F. Ramos, and L. Ott, “Speeding up iterative closest point using stochastic gradient descent,” International Conference on Robotics and Automation, 2019

2019
[24]

Bayesian iterative closest point for mobile robot localization,

F. A. Maken, F. Ramos, and L. Ott, “Bayesian iterative closest point for mobile robot localization,” The International Journal of Robotics Research, 2022

2022
[25]

Closed-form solution of absolute orientation using unit quaternions,

B. K. P. Horn, “Closed-form solution of absolute orientation using unit quaternions,” J. Opt. Soc. Am. , 1987

1987
[26]

Least-squares fitting of two 3-d point sets,

K. S. Arun, T. S. Huang, and S. D. Blostein, “Least-squares fitting of two 3-d point sets,” IEEE Transactions on Pattern Analysis and Machine Intelligence , 1987

1987
[27]

A stochastic approximation method,

H. Robbins and S. Monro, “A stochastic approximation method,” The Annals of Mathematical Statistics , 1951

1951
[28]

Stein variational gradient descent: A general purpose bayesian inference algorithm,

Q. Liu and D. Wang, “Stein variational gradient descent: A general purpose bayesian inference algorithm,” inProc. Neural Inf. Process. Syst. , 2016

2016
[29]

Stochastic parti- cle optimization sampling and the non-asymptotic convergence theory,

J. Zhang, R. Zhang, L. Carin, and C. Chen, “Stochastic parti- cle optimization sampling and the non-asymptotic convergence theory,” in International Conference on Artificial Intelligence and Statistics, 2020

2020
[30]

J. B. Kuipers, Quaternions and Rotation Sequences: A Primer with Applications to Orbits, Aerospace and Virtual Reality . 1999

1999
[31]

Bo-icp: Initializa- tion of iterative closest point based on bayesian optimization,

H. Biggie, A. Beathard, and C. Heckman, “Bo-icp: Initializa- tion of iterative closest point based on bayesian optimization,” in IEEE International Conference on Robotics and Automa- tion, 2023

2023
[32]

Isaac gym: High performance gpu- based physics simulation for robot learning

V. Makoviychuk et al. , “Isaac gym: High performance gpu- based physics simulation for robot learning. ”https://neurips. cc/datasets-benchmarks/2021, 2021

2021
[33]

Open3d: A modern library for 3D data processing,

Q.-Y. Zhou, J. Park, and V. Koltun, “Open3d: A modern library for 3D data processing,” arXiv e-prints , vol. abs/1801.09847, 2018

2018 arXiv
[34]

The kit object models database: An object model database for object recognition, localization, and manipulation in service robotics,

A. Kasper, Z. Xue, and R. Dillmann, “The kit object models database: An object model database for object recognition, localization, and manipulation in service robotics,” Interna- tional Journal of Robotics Research, 2012

2012
[35]

Google scanned objects: A high-quality dataset of 3d scanned household items,

L. Downs, A. Francis, N. Koenig, et al. , “Google scanned objects: A high-quality dataset of 3d scanned household items,” in International Conference on Robotics and Automation , pp. 2553–2560, 2022

2022
[36]

anygrasp-sdk

GraspNet, “anygrasp-sdk. ” https://github.com/graspnet/ anygrasp_sdk, 2024. Accessed: 2024-03

2024
[37]

geometric-object-grasper

M. Kiato, “geometric-object-grasper. ” https://github.com/ mkiatos/geometric-object-grasper, 2024. Accessed: 2024- 03

2024

Pith tools

Reviewed August 11, 2026 · model on record in the stance chip above.

[1] [1]

Matching Step: pairs the transformed source point cloud,S′ ={s′ i}N i=1, with the reference point cloud R ={ri}M i=1 on the basis of a distance metric, where si and ri∈R3 are N and M points in 3D space. The commonly used point-to-point distance metric finds the pair with the nearest neighbour as follows: ˆri = argmin rj∈R ∥s′ i–rj∥ (1) where ˆri is the cl...

work page

[2] [2]

The updated equation for the point-to-point distance met- ric is defined as follows: θk+1 = argmin θ 1 N N∑ i ∥Rksi +tk–ˆri∥2 (2) where k is the iteration number

Minimization Step: updates θk to minimize a loss function defined by the distance between the paired points in the source and reference point clouds. The updated equation for the point-to-point distance met- ric is defined as follows: θk+1 = argmin θ 1 N N∑ i ∥Rksi +tk–ˆri∥2 (2) where k is the iteration number. Equation (2) can be solved in closed-form us...

work page

[3] [3]

Stein icp for uncertainty estimation in point cloud matching,

F. A. Maken, F. Ramos, and L. Ott, “Stein icp for uncertainty estimation in point cloud matching,” Robotics and Automation Letters, 2022

work page 2022

[4] [4]

A survey on learning-based robotic grasping,

K. Kleeberger, R. Bormann, W. Kraus, et al., “A survey on learning-based robotic grasping,” Current Robot Reports, 2020

work page 2020

[5] [5]

Ising Model on the Fibonacci Sphere

A. Pochinok, A. Molochkov, and M. Chernodub, “Ising model on the fibonacci sphere,” arXiv preprint arXiv:2301.06849 , 2023

work page Pith review arXiv 2023

[6] [6]

Real-time grasp planning for multi-fingered hands by finger splitting,

Y. Fan, T. Tang, H. C. Lin, and M. Tomizuka, “Real-time grasp planning for multi-fingered hands by finger splitting,” IROS, 2018

work page 2018

[7] [7]

Annealed stein variational gradi- ent descent,

F. D’Angelo and V. Fortuin, “Annealed stein variational gradi- ent descent,” in 3rd Symposium on Advances in Approximate Bayesian Inference, 2020

work page 2020

[8] [8]

An overview of 3d object grasp synthesis algorithms,

A. Sahbani, S. El-Khoury, and P. Bidaud, “An overview of 3d object grasp synthesis algorithms,” Robotics and Automation Systems, 2012

work page 2012