Pith. sign in

REVIEW 4 major objections 5 minor 27 references

SAGP: Semantic Affordance-Guided Grasp Planning via Coarse-Zone VLM Reasoning

T0 review · 4 major / 5 minor · reviewed 2026-08-03 · deepseek-v4-flash

Pith's one-line read SAGP claims that zero-shot VLM ratings of coarse object zones can re-rank geometric grasps toward functionally appropriate regions—handles, not rims—with no training and no loss in physical success.

desk verdict Coarse-zone VLM re-ranking is a sensible idea, but the paper's headline numbers are unverifiable and its main appropriateness metric is circular; treat the technique as promising and the quantitative claims as unproven. read the letter →

arxiv 2607.29374 v1 pith:DEQSMBWL submitted 2026-07-31 cs.RO

classification cs.RO
keywords graspplanningvision-languagemodelssemanticaffordancescoarse-zonedecompositionzero-shotreasoningantipodalgraspsrobotmanipulationpointcloudsegmentation
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

Geometric grasp planners generate stable, collision-free grasps but are functionally blind: a mug by its rim, a knife by its blade, a bottle near its cap. SAGP claims to close that gap with a training-free interface: rather than asking a vision-language model (VLM) to output exact grasp coordinates—the regime where such models hallucinate—it partitions the object point cloud into coarse zones (top, middle, bottom, sides, protrusions) and asks the VLM to rate each zone as good, acceptable, bad, or dangerous. These zone ratings are fused with geometric, reachability, and task-alignment signals to re-rank antipodal grasp candidates. In simulation on 14 everyday objects, the paper reports that the re-ranked planner keeps grasp success above 90%, matching the geometry-only baseline, while substantially raising the share of grasps that land in the VLM-preferred zone, especially on asymmetric, handle-bearing objects. If the claim holds, a general-purpose VLM can inject functional semantics into grasping without fine-tuning, part annotations, or learned segmentation.

What carries the argument

The load-bearing mechanism is the coarse-zone abstraction layer. It maps raw point clouds to eight labeled zones using PCA-based vertical alignment (equal thirds: top, middle, bottom; lateral: left/right/front/back) plus DBSCAN clustering of centroid-outlier points as protrusions. Each grasp candidate is assigned to a zone by its contact midpoint, and a pre-trained VLM's zero-shot ratings of those zones—mapped to numeric values—re-rank the candidates. The layer does the work of translating a VLM's reliable word-level spatial knowledge (handle, rim, blade) into a form a geometric planner can optimize, while deliberately keeping the VLM away from precise coordinate prediction.

What would settle it

Replace the VLM's self-declared preferred zone with human-labeled functional regions on the same 14 objects and recompute preferred-zone accuracy; if the score drops substantially below the reported >60%, the claimed functional gain is an artifact of self-scoring. A complementary test is to run the pipeline on objects rotated far from the VLM's typical view and check whether zone ratings and grasp outcomes diverge from the cached-query setting.

Watch

Extended reading notes

Core claim

The central discovery is that a coarse-zone abstraction is the right interface between VLM semantics and geometric grasp planning. PCA divides the point cloud into vertical thirds and lateral regions, DBSCAN marks protrusions such as handles, and a pre-trained VLM rates each zone on a good/acceptable/bad/dangerous scale. These ratings, fused with geometric, reachability, and task-alignment scores, re-rank antipodal grasps, preserving the geometry-only baseline's high physical success while preferring functionally correct regions.

Load-bearing premise

The load-bearing premise is that a coarse zone computed by PCA vertical thirds and DBSCAN protrusion clustering is a stable proxy for the functional region a task needs, and that the pre-trained VLM's zone ratings are a valid ground truth for that appropriateness; if a tilted object misaligns the PCA or the VLM mislabels a handle as 'bad', the re-ranking will confidently prefer the wrong grasp, and the reported metrics will not reveal it because preferred-zone accuracy and Ta

Editorial extensions

If this is right

  • Functional grasping becomes possible on previously unseen objects: any object a VLM can recognize and whose protrusions survive clustering gets task-appropriate grasp re-ranking without annotations.
  • The semantic layer can be attached to any underlying antipodal grasp generator; only the re-ranking stage changes, so existing planners gain functionality at small engineering cost.
  • On near-symmetric objects where all zones are rated about equally, the method degrades gracefully to the geometry-only baseline rather than hurting physical success.
  • On very small objects, strongly symmetric objects, or objects in atypical orientations, the paper expects zone detection or VLM ratings to degrade; a deployed system would need a fallback or a re-query strategy.

Reading between the lines

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

  • The paper's preferred-zone metric is partly self-confirming: the VLM defines the preferred zone and the experiment then measures how often the grasp lands in it. Testing zone preferences against independent human labels or downstream task outcomes would reveal whether the VLM's choices are truly functional or merely consistent.
  • Because the VLM query is cached per (object, task), the reported comparisons ignore per-instance semantic variation; a stronger stress test would re-query the VLM on every trial, including translated or slightly rotated versions, to measure how stable the zone preferences are under pose noise.
  • The same zone abstraction could be reused beyond grasp planning—for example, to bias exploration during in-hand manipulation or to define finger targets for multi-fingered hands—since the zones carry localizable functional meaning that a planner can consume.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. The paper proposes SAGP, a training-free pipeline that connects VLM semantic reasoning to geometric grasp planning through a coarse-zone abstraction. The method decomposes an object point cloud into zones (top/middle/bottom, protrusions) using PCA alignment and DBSCAN clustering, asks a pretrained VLM (Qwen-VL) to rate each zone's grasp suitability on a four-level scale, and re-ranks antipodal grasp candidates using a weighted sum of geometric, reachability, affordance, avoidance, and task-alignment scores (Eq. 7). The experiments in PyBullet with a Franka Panda on 14 YCB objects are claimed to preserve the geometry-only baseline's grasp success rate above 90% while substantially improving functional appropriateness, e.g. preferred-zone accuracy above 60%, especially on asymmetric handle-bearing objects.

Significance. If the empirical claims were independently grounded, SAGP would be a useful practical contribution: it avoids category-specific part segmentation and VLM direct pose prediction by restricting VLM reasoning to coarse semantic zones, which is a sensible division of labor given VLMs' known spatial-hallucination limitations. The structured JSON prompting, the caching scheme for reproducibility, the comparison against a VLM-Direct baseline, and the explicit discussion of failure modes are strengths. However, the current evaluation does not establish the central claim. Task 2 success and preferred-zone accuracy are defined through the same VLM preferences that SAGP optimizes, so the reported 'functional appropriateness' gain may reflect self-consistency rather than task-relevant correctness. Moreover, Section 5.1 contains no numeric results table, per-object breakdown, error bars, or significance tests, making the magnitude and even direction of the effect impossible to verify. The core idea is original and worth further development, but the evidence presented is not yet commensurate with the strength of the conclusions.

major comments (4)
  1. [§4.3, §4.4, §3.5] Circular evaluation of functional appropriateness. Section 4.3 defines Task 2 success as 'the grasp must occur within the VLM-preferred zone', and Section 4.4 defines preferred-zone accuracy (PZA) as the fraction of successful grasps falling within that same VLM-preferred zone. Section 3.5's scoring function (Eq. 7) uses the same VLM's zone ratings in S_afford (Eq. 8) and the same VLM's preferred direction in S_task, so SAGP is explicitly re-ranking candidates toward the metric on which it is evaluated. The geometry-only baseline is not an independent ground truth; it is merely indifferent to VLM preferences. Under this protocol, a method that blindly implements VLM preferences would score high on Task 2 and PZA regardless of whether those preferences correspond to functionally appropriate grasps. The manuscript therefore 'substantially improving functional appropriateness' is not suppor
  2. [§5.1] The quantitative results are not actually reported. The section makes only qualitative claims: 'grasp success rates above 90%', 'preferred-zone accuracy exceeds 60%', and 'orientation preservation improves correspondingly'. There is no table or figure reporting GSR, PZA, AC, OP, or MPT for SAGP, Geometry-Only, and VLM-Direct, per object or aggregate. There are no error bars, standard deviations, or statistical tests across the 20 trials per (object, task). As a result, the claimed 'substantial improvement' cannot be quantified, checked, or compared across baselines. The paper needs a complete results table (and ideally per-category asymmetric/near-symmetric breakdown) with uncertainties and significance testing. Without these numbers, neither the reader nor the authors can verify the core claim.
  3. [§3.3–§3.5, Eq. (5)] Internal inconsistency in zone definitions and assignment. Section 3.3 declares the zone vocabulary Z = {top, middle, bottom, left, right, front, back, protrusion}. However, Eq. (5) assigns a grasp candidate only to 'protrusion' or, otherwise, to one of top/middle/bottom. The left/right/front/back labels are never assigned to any candidate, so the VLM's ratings of those zones cannot enter S_afford or S_avoid, and any 'preferred zone' named among the lateral labels would make Task 2/PZA impossible to satisfy. The paper's abstract and discussion refer to 'lateral sides' as part of the coarse-zone abstraction, but this component is not implemented as described. The authors should either use the lateral-zone labels in Eq. (5), or remove them from the zone vocabulary and revise the claims about lateral-side reasoning.
  4. [§3.5] The fusion weights α, β, γ, δ, ε are said to be 'selected on a small held-out tuning subset', but no details are given: how large is the subset, which objects/tasks, what selection criterion, and how sensitive are the results to the chosen weights? Since the entire semantic re-ranking effect is mediated by Eq. (7), the usefulness of SAGP depends on this weight vector. Without a sensitivity analysis or a clearly motivated selection procedure, it is hard to know whether the reported advantages are robust or an artifact of tuning. Please provide the tuning protocol and a sensitivity study (e.g., varying each weight or reporting results for a few representative weight sets).
minor comments (5)
  1. [§3.3] Typo and notation: 'The SAGP performs' should be 'SAGP performs'. In Eq. (5), 'midpoint(g)∈protrusion' is ambiguous; better to write 'midpoint(g)∈P_protrusion' where P_protrusion is the set of points labelled as protrusion.
  2. [§4.4] The caching scheme caches VLM responses at the (object, task) level even though each trial uses a random yaw. The authors state they 'verified this assumption by re-querying the VLM under three additional viewpoints for a representative subset' but do not specify the subset, the viewpoints, or the stability criterion. Please give details so this reproducibility claim can be assessed.
  3. [§5.1] The phrase 'With the simulator now correctly configured' reads as if it refers to an earlier revision or debugging process; please remove or replace it with a neutral description of the simulation configuration.
  4. [References] Reference [14] contains a typographical spacing error: 'V oxPoser' should be 'VoxPoser'.
  5. [§5.2] Figure 3 caption says 'three successful grasps (two execution instances each)', which is six images, but the text says 'subfigs 2-7 show three successful grasps'. Consider clarifying whether each grasp is shown twice or whether there are three distinct grasps with two snapshots each.

Circularity Check

1 steps flagged · score 6.0 of 10

The functional-appropriateness metrics are defined by the same VLM that SAGP uses to rank grasps, so the reported improvement may be self-consistent by construction.

  1. self definitional [Sections 3.4-3.5 (Eq. 7-8) and Sections 4.3-4.4 (Task 2, PZA)]
    "The VLM also identifies a preferred-zone set Zpref, an avoid-zone set Zavoid, an approach direction, a grasp style, and orientation constraints. ... The zone-affordance term applies the VLM rating of the candidate's zone: Safford(g) = φ(r_z(g)). ... Task 2: Functional grasp. In addition to Task 1, the grasp must occur within the VLM-preferred zone, and the object orientation must be preserved within 15◦. ... preferred-zone accuracy (PZA), the fraction of successful grasps falling within the VLM-preferred zone."

    The SAGP score in Eq. 7 includes Safford(g) = φ(r_z(g)) from Eq. 8, i.e. the same Qwen-VL zone rating that defines Zpref, plus Stask from the VLM-declared approach/orientation constraints. The paper's headline evidence for improved 'functional appropriateness' is Task 2 success and PZA, both defined as 'within the VLM-preferred zone' by that same VLM. Therefore re-ranking with Eq. 7 directly optimizes the metric used in evaluation; PZA > 60% largely measures self-consistency with the annotator, not agreement with an independent functional ground truth. GSR is an independent physical check, so the circularity is partial, giving a score of 6.

full rationale

The geometric success rate is independently and appropriately measured in PyBullet, so SAGP's claim of preserving >90% GSR is not circular. The circularity concerns the second, central claim of 'substantially improving functional appropriateness.' That claim rests on Task 2 and PZA, which are defined by the VLM-preferred zone. The same VLM's zone ratings are directly inserted into the scoring function (Eq. 8) and its declared task constraints into Stask. Thus the re-ranked policy is evaluated with the very labels it consumes: a grasp is 'functionally appropriate' iff it falls where the VLM said to grasp. The paper offers no human-labeled or physically grounded validation of the VLM ratings (e.g., for mug, drill, scissors), and Section 5.3 only notes symmetric-object collapse, so the asymmetry improvement is not independently confirmed. No other circularity patterns (self-citation chains, imported uniqueness theorems, ansatz-by-citation) appear: the self-citations [18], [19], [25] are not load-bearing. Overall, the physical component is sound and independent, but the semantic-appropriateness evaluation is partially circular.

Assumptions & free parameters 3 free parameters · 6 assumptions · 0 invented entities

SAGP's central claim rests on several domain assumptions: VLM zone ratings are valid task affordances, PCA/DBSCAN zones align with functional regions, viewpoint invariance permits caching, simulation dynamics are faithful, and the antipodal candidate set covers useful grasps. Free parameters include the five fusion weights tuned on a held-out subset and hand-set clustering and tolerance thresholds. No invented physical entities are introduced.

free parameters (3)
  • fusion weights α,β,γ,δ,ε = 0.3, 0.1, 0.3, 0.2, 0.1
    Eq. (7); selected on a small held-out tuning subset rather than derived; the comparison between SAGP and Geometry-Only depends on these values.
  • protrusion detection thresholds = 1.5× median distance, DBSCAN eps=0.01 m, Nmin=10, min cluster=20
    Section 3.3; hand-set thresholds determine which regions are labeled protrusions and therefore which zones the VLM rates.
  • task-alignment and filtering tolerances = approach within 30°, orientation within 15°, gripper width 0.005–0.08 m, IK margin 0.5
    Sections 3.2 and 3.5; hand-chosen thresholds shape both the candidate set and the task-success criteria.
assumptions (6)
  • domain assumption Zero-shot VLM categorical ratings over coarse zones are a valid and sufficiently reliable proxy for task-appropriate grasping.
    The entire re-ranking (Eq. 7–9) rests on Qwen-VL ratings; no human-labeled affordance ground truth or physical task validation is provided.
  • domain assumption PCA-based vertical-axis selection and equal-third partitioning yield zones that align with functional regions across arbitrary object poses.
    Section 3.3 assumes the covariance axis most aligned with gravity identifies the functional vertical axis; this can fail for tilted or symmetric objects.
  • domain assumption VLM zone ratings are robust to viewpoint and pose perturbations so responses can be cached at the (object, task) level.
    Section 4.4 caches only 42 queries; viewpoint stability is verified only on a representative subset.
  • domain assumption Ground-truth segmentation of the target object is available in simulation and real-world SAM substitution will not change results.
    Section 3.1 uses ground-truth segmentation masks; the hardware-transfer claim in Section 5.3 is untested.
  • domain assumption PyBullet dynamics faithfully capture grasp success and orientation preservation for the Franka Panda.
    Section 4.1 evaluates physical success only in simulation, with no real-robot validation.
  • domain assumption Antipodal sampling covers all relevant grasp candidates; semantic re-ranking cannot recover candidates absent from the geometric set.
    The pipeline inherits the geometry planner's coverage; if a good handle grasp is not antipodal or is filtered out, it will never be selected.

how reviews work

0 comments
Cite this review

Pith. "Pith review of SAGP: Semantic Affordance-Guided Grasp Planning via Coarse-Zone VLM Reasoning." pith.science (2026). https://pith.science/paper/DEQSMBWL

@misc{pith2026260729374,
  author       = {Pith},
  title        = {Pith review of: SAGP: Semantic Affordance-Guided Grasp Planning via Coarse-Zone VLM Reasoning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/DEQSMBWL}},
  note         = {Machine review of arXiv:2607.29374}
}
read the original abstract

Geometry-based grasp planners ensure physically valid grasps but ignore functional semantics, often generating grasps that are antipodal and collision-free yet practically inappropriate, for example, gripping a mug by its rim, a knife by the blade, or a bottle near its cap. These inconsistencies cause the downstream task to fail even when traditional grasp metrics are met. Existing vision-language model (VLM) approaches either depend on fine-grained, category-specific part segmentation or attempt to directly infer grasp poses, with the latter prone to spatial hallucinations. As a result, no practical, training-free framework has yet been proposed that robustly links high-level semantic reasoning to geometric grasp planning. We introduce Semantic Affordance-Guided Grasp Planning (SAGP), a training-free pipeline built on a coarse-zone abstraction layer. The method first partitions the object point cloud into spatial regions (top, middle, bottom, lateral sides, and protrusions) by applying PCA-based alignment followed by distance-driven DBSCAN clustering, entirely bypassing learned segmentation. A pre-trained VLM then assesses the grasp quality of each region through a structured zero-shot query, and the resulting zone-wise scores are fused with geometric, reachability, and task-alignment signals to re-rank antipodal grasp candidates. Experiments on YCB objects in PyBullet with a Franka Panda robot show that SAGP preserves the high success rate of geometry-only planning while substantially improving the functional appropriateness of selected grasps, particularly on asymmetric, handle-bearing objects where geometry alone is uninformative. The introduced coarse-zone abstraction offers an effective, training-free bridge between VLM-based reasoning and geometric grasp planning, without the need for fine-grained part segmentation.

Figures

Figures reproduced from arXiv: 2607.29374 by the authors.

Figure 1
Figure 1. SAGP pipeline architecture. Starting from RGB-D perception and segmentation, the pipeline branches into three parallel processes: antipodal grasp sampling, coarse-zone decomposition via PCA and pro￾trusion detection, and VLM-based zone affordance rating. The three streams converge at the semantic grasp scoring stage, where a fused five-component function re-ranks the candidates. The highest-ranked grasp is then exec… view at source ↗
Figure 2
Figure 2. Coarse-zone decomposition for three YCB objects. Vertical thirds (top: red, middle: green, bottom: blue) are combined with protrusion regions (yellow) detected via distance-based clustering. Handles and grips are identified without learned segmentation. cloud into left/right or front/back regions based on each point’s signed displacement from the centroid. To identify functional structures such as handles or grips, … view at source ↗
Figure 3
Figure 3. Qualitative grasp executions of SAGP in PyBullet on the cluttered YCB scene (Task 3). Subfig 1 shows the full workspace; subfigs 2-7 show three successful grasps (two execution instances each): the power drill at its grip (2-3), the banana at its middle (4-5), and the tomato can on its body (6-7). Despite distractors within a 15 cm radius, the gripper contacts the VLM-preferred zone without displacing neighbouring i… view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

27 extracted references · 3 canonical work pages

  1. [1]

    Robotic grasping and contact: a review

    Bicchi A, Kumar V . Robotic grasping and contact: a review. In: Proc. IEEE Int. Conf. on Robotics and Automation (ICRA); 2000. p. 348-353

  2. [2]

    GraspIt!: a versatile simulator for robotic grasping

    Miller AT, Allen PK. GraspIt!: a versatile simulator for robotic grasping. IEEE Robot Autom Mag. 2004;11(4):110-122

  3. [3]

    Grasp pose detection in point clouds

    ten Pas A, Gualtieri M, Saenko K, Platt R. Grasp pose detection in point clouds. Int J Robot Res. 2017;36(13-14):1455-1473

  4. [4]

    Dex-Net 2.0: deep learning to plan robust grasps with synthetic point clouds and analytic grasp metrics

    Mahler J, Liang J, Niyaz S, Laskey M, Doan R, Liu X, et al. Dex-Net 2.0: deep learning to plan robust grasps with synthetic point clouds and analytic grasp metrics. In: Proc. Robotics: Science and Systems (RSS); 2017. M. Ud Din and I. Hussain /

  5. [5]

    Contact-GraspNet: efficient 6-DoF grasp generation in cluttered scenes

    Sundermeyer M, Mousavian A, Triebel R, Fox D. Contact-GraspNet: efficient 6-DoF grasp generation in cluttered scenes. In: Proc. IEEE Int. Conf. on Robotics and Automation (ICRA); 2021. p. 13438-13444

  6. [6]

    AnyGrasp: robust and efficient grasp perception in spatial and temporal domains

    Fang HS, Wang C, Fang H, Gou M, Liu J, Yan H, et al. AnyGrasp: robust and efficient grasp perception in spatial and temporal domains. IEEE Trans Robot. 2023;39(5):3929-3945

  7. [7]

    A human-inspired real-time grasp force selec- tion policy based on load-grip force coupling

    Stachowsky M, Vale J, Abdullah HA, Moussa M. A human-inspired real-time grasp force selec- tion policy based on load-grip force coupling. Int J Mech Eng Robot Res. 2015;4(4):336-342, doi: 10.18178/ijmerr.4.4.336-342

  8. [8]

    GPT-4V(ision) for robotics: multimodal task planning from human demonstration

    Wake N, Kanehira A, Sasabuchi K, Takamatsu J, Ikeuchi K. GPT-4V(ision) for robotics: multimodal task planning from human demonstration. arXiv:2311.12015. 2023

Show all 27 references
  1. [9]

    Same object, different grasps: data and semantic knowledge for task-oriented grasping

    Murali A, Liu W, Marino K, Chernova S, Gupta A. Same object, different grasps: data and semantic knowledge for task-oriented grasping. In: Proc. Conf. on Robot Learning (CoRL); 2020

  2. [10]

    Learning grasp affordance densities

    Detry R, Kraft D, Kroemer O, Bodenhagen L, Peters J, Kr ¨uger N, et al. Learning grasp affordance densities. Paladyn J Behav Robot. 2011;2(1):1-17

  3. [11]

    GraspGPT: leveraging semantic knowledge from a large language model for task-oriented grasping

    Tang C, Huang D, Ge W, Liu W, Zhang H. GraspGPT: leveraging semantic knowledge from a large language model for task-oriented grasping. IEEE Robot Autom Lett. 2023;8(11):7551-7558

  4. [12]

    SemGrasp: semantic grasp generation via language aligned dis- cretization

    Li K, Wang J, Yang L, Lu C, Dai B. SemGrasp: semantic grasp generation via language aligned dis- cretization. In: Proc. European Conf. on Computer Vision (ECCV); 2024

  5. [13]

    Do as I can, not as I say: grounding language in robotic affordances

    Ahn M, et al. Do as I can, not as I say: grounding language in robotic affordances. In: Proc. Conf. on Robot Learning (CoRL); 2022

  6. [14]

    V oxPoser: composable 3D value maps for robotic manipulation with language models

    Huang W, Wang C, Zhang R, Li Y , Wu J, Fei-Fei L. V oxPoser: composable 3D value maps for robotic manipulation with language models. In: Proc. Conf. on Robot Learning (CoRL); 2023

  7. [15]

    CLIPort: what and where pathways for robotic manipulation

    Shridhar M, Manuelli L, Fox D. CLIPort: what and where pathways for robotic manipulation. In: Proc. Conf. on Robot Learning (CoRL), PMLR vol. 164; 2021. p. 894-906

  8. [16]

    RT-2: vision-language-action models transfer web knowledge to robotic control

    Brohan A, et al. RT-2: vision-language-action models transfer web knowledge to robotic control. arXiv:2307.15818. 2023

  9. [17]

    Code as policies: language model programs for embodied control

    Liang J, Huang W, Xia F, Xu P, Hausman K, Ichter B, et al. Code as policies: language model programs for embodied control. In: Proc. IEEE Int. Conf. on Robotics and Automation (ICRA); 2023. p. 9493- 9500

  10. [18]

    Lang2Manip: a tool for LLM-based symbolic-to-geometric planning for manipulation

    Ud Din M, Rosell J, Akram W, Hussain I. Lang2Manip: a tool for LLM-based symbolic-to-geometric planning for manipulation. In: Proc. 12th Int. Conf. on Automation, Robotics and Applications (ICARA); 2026. p. 568-574

  11. [19]

    Planning grasping motions for humanoid robots

    Rosell J, Su ´arez R, Garc ´ıa N, Ud Din M. Planning grasping motions for humanoid robots. Int J Hu- manoid Robot. 2019;16(6):1950041, doi: 10.1142/S0219843619500415

  12. [20]

    LAN-grasp: using large language models for semantic object grasping

    Mirjalili R, Krawez M, Blei Y , Silenzi S, Walter F, Burgard W. LAN-grasp: using large language models for semantic object grasping. arXiv:2310.05239. 2023

  13. [21]

    Open3D: a modern library for 3D data processing

    Zhou QY , Park J, Koltun V . Open3D: a modern library for 3D data processing. arXiv:1801.09847. 2018

  14. [22]

    A density-based algorithm for discovering clusters in large spatial databases with noise

    Ester M, Kriegel HP, Sander J, Xu X. A density-based algorithm for discovering clusters in large spatial databases with noise. In: Proc. 2nd Int. Conf. on Knowledge Discovery and Data Mining (KDD); 1996. p. 226-231

  15. [23]

    Qwen-VL: a versatile vision-language model for understanding, localization, text reading, and beyond

    Bai J, et al. Qwen-VL: a versatile vision-language model for understanding, localization, text reading, and beyond. arXiv:2308.12966. 2023

  16. [24]

    PyBullet, a Python module for physics simulation for games, robotics and machine learning.http://pybullet.org; 2016-2021

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

  17. [25]

    Physics-based motion planning: evaluation criteria and benchmark- ing

    Muhayyuddin, Akbari A, Rosell J. Physics-based motion planning: evaluation criteria and benchmark- ing. In: Robot 2015: Second Iberian Robotics Conf. Advances in Intelligent Systems and Computing, vol. 417. Springer; 2015. p. 43-55, doi: 10.1007/978-3-319-27146-0 4

  18. [26]

    The YCB object and model set: towards common benchmarks for manipulation research

    Calli B, Singh A, Walsman A, Srinivasa S, Abbeel P, Dollar AM. The YCB object and model set: towards common benchmarks for manipulation research. In: Proc. Int. Conf. on Advanced Robotics (ICAR)

  19. [27]

    Segment anything

    Kirillov A, Mintun E, Ravi N, Mao H, Rolland C, Gustafson L, et al. Segment anything. In: Proc. IEEE/CVF Int. Conf. on Computer Vision (ICCV); 2023. p. 4015-4026

Pith tools

Reviewed August 3, 2026 · model on record in the stance chip above.