Pith. sign in

REVIEW 3 major objections 6 minor 51 references

GCA-Bench argues that complex robotic grasping must be evaluated from detection to execution, and that current methods top out below 70% task success because execution, not perception, is the bottleneck.

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

2026-08-02 02:21 UTC pith:N2NWBNPQ

load-bearing objection Useful benchmark with a plausible but statistically thin central claim; worth sending to a careful referee, but the authors need to justify the scenario sampling and release the data before I'd trust the headline numbers. the 3 major comments →

arxiv 2607.14341 v1 pith:N2NWBNPQ submitted 2026-07-15 cs.RO cs.AI

Beyond Visual Grasping: Benchmarking Complex Grasping from Detection to Execution

Hanyi Zhang , Khang Nguyen , Charith Munasinghe , Basu Hela , Tianyu Li , Zihong Luo , Hoan Nguyen , Hans Wernher van de Venn
show 5 more authors
Yalin Zheng Ravi Prakash Tung D. Ta Anh Nguyen Baoru Huang
This is my paper
classification cs.RO cs.AI
keywords robotic graspingbenchmarkvision-language-action modelstask success rategrasp executionsimulationsemantic reasoningmanipulation
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

This paper introduces GCA-Bench, a benchmark for robotic grasping that evaluates the full pipeline—detection, grasp, and task-level execution—rather than isolated grasp pose detection. It includes 102 tasks spanning four complexity categories: singulated objects, cluttered scenes, constrained spaces, and semantic or language-constrained tasks. The central claim is that current methods, from classical grasp detection plus motion planning to fine-tuned vision-language-action models, all fall below 70% task success on these complex scenarios. More specifically, the paper reports that even when objects are correctly detected, successful execution remains unlikely, with detection-normalized grasp success consistently below 0.5. A sympathetic reader would care because this reframes the grasping problem: the field's bottleneck is not perception of where to grasp, but reliability of executing the whole action.

Core claim

The paper's core discovery is a measured gap between perception and execution in robotic grasping. By benchmarking a range of methods—AnyGrasp, GraspMAS, with and without cuRobo motion planning, plus zero-shot and fine-tuned VLA models such as GraspVLA, OpenVLA, and the pi-series—on GCA-Bench, the authors show that task success rates remain below 70% across all four categories. The gap is quantified by normalizing grasp success against detection success: the ratio stays below 0.5, meaning a majority of failures happen after the object is correctly found. Failure analysis attributes this to missing closed-loop feedback, inability to handle object slip and tilt, and lack of adaptive replanning

What carries the argument

The central object is GCA-Bench itself: a simulation-and-real benchmark of 102 grasping tasks categorized into singulated, cluttered, constrained-space, and semantic scenarios, with language instructions at three complexity levels. Its diagnostic power comes from staged success metrics—Detection Success Rate, Grasp Success Rate, Task Success Rate—plus efficiency metrics (SPL and execution time). The key analytical tool is the detection-normalized grasp success ratio, which isolates whether failures originate from perception or execution. The benchmark also provides 5,000 simulated and 800 real teleoperated trajectories for fine-tuning VLA baselines.

Load-bearing premise

The headline results rest on 5 hand-picked scenarios per task category—50 trials per category—and if those scenarios don't represent the difficulty of the full 102-task benchmark, the rankings and the below-70% finding could change.

What would settle it

Run the same baselines on a random or stratified sample of all 102 tasks and check whether every method still falls below 70% task success and whether detection-normalized grasp success stays below 0.5; a scenario-selection artifact would show wide variance across draws.

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

If this is right

  • If the paper's central claim is correct, future grasping benchmarks should measure execution-level success, not just grasp pose detection, because detection can succeed while task execution fails.
  • The below-0.5 detection-normalized grasp success implies that improving detectors alone will not raise task success; research should shift toward closed-loop control, recovery behaviors, and adaptive replanning.
  • Fine-tuned VLA models outperform classical detection-plus-planning pipelines on complex grasping, but their drop on semantic tasks suggests their instruction-following is pattern matching rather than robust reasoning.
  • The benchmark's simulation-to-real consistency, if it holds, provides a low-cost testbed for developing and evaluating new grasping policies before physical deployment.
  • The three instruction levels provide a way to isolate semantic reasoning from low-level manipulation, which could guide progress in language-conditioned robotics.

Where Pith is reading between the lines

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

  • The paper's own limitation to parallel-jaw grippers suggests an immediate extension: re-running GCA-Bench with suction or dexterous hands to test whether the execution bottleneck is gripper-specific or general.
  • Because the detection-normalized grasp rate is consistently below 0.5, a testable next step is to add closed-loop feedback or recovery actions to the best VLA baseline and measure whether task success rises—this would directly test the paper's claim that open-loop execution is the bottleneck.
  • The evaluation protocol currently samples only 5 scenarios per category; replacing it with stratified sampling across all 102 tasks would yield per-task difficulty scores and a calibrated difficulty curve, making the benchmark's complexity levels quantitative rather than qualitative.
  • One could extend the semantic tasks to include novel object affordances and unseen instruction phrasings at test time, using GCA-Bench as a measure of whether VLA policies generalize beyond their training distribution.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 6 minor

Summary. The paper introduces GCA-Bench, a simulation benchmark with 102 grasping tasks spanning four categories (Singulated, Cluttered, Constrained Space, Semantic), instructions at three complexity levels, and metrics that separate detection success (DSR), grasp success (GSR), and task success (TSR), plus SPL and execution time. The authors collect 5,000 simulated and 800 real-robot trajectories and evaluate a broad set of baselines: grasp-detection pipelines (AnyGrasp, GraspMAS) with and without cuRobo motion planning, and VLA models (GraspVLA, OpenVLA, OpenVLA-OFT, pi0, pi0.5) after LoRA fine-tuning. The central claim is that existing methods perform well in simple settings but drop sharply on cluttered, constrained, and instruction-driven tasks, with TSR below 70% on complex scenarios and a persistent detection-to-execution gap.

Significance. If the quantitative claims are supported, GCA-Bench would be a valuable contribution: it addresses a real gap in grasping benchmarks by moving beyond isolated grasp-pose detection to full-pipeline evaluation with scene-level and semantic complexity, and it provides a useful failure-stage diagnostic. Strengths include the breadth of baselines, the inclusion of real-world validation, the effort to collect diverse demonstrations, and the proposed multi-level metrics. However, the headline empirical conclusions currently rest on a small, unrepresentative subset of the benchmark and on tables without uncertainty quantification, so the magnitude and generality of the reported gap are not yet established.

major comments (3)
  1. [Section IV.B, Tables II-III] The experimental protocol samples only 5 scenarios per category, i.e., 20 of the 102 benchmark tasks, with 50 trials per category. No selection protocol, difficulty calibration, or inter-rater agreement is reported to show that these scenarios represent the full benchmark. With 50 binary trials, 95% confidence intervals are roughly ±0.14 at p=0.5, so point estimates like 0.42 vs. 0.57 are statistically indistinguishable. Consequently, the claims that 'existing methods perform well in simple settings' and 'success rates drop sharply' in complex tasks are not supported for GCA-Bench as a whole. Please provide per-scenario results, a selection rationale, and confidence intervals or Bayesian credible intervals, or explicitly restrict the conclusions to the evaluated subset.
  2. [Table V] The 'detection-normalized grasp success rate' is undefined. It is not any metric in Eq. (1), and the reported values are nearly identical to the TSR values in Table II (e.g., AnyGrasp Singulated 0.37 vs. 0.37, Constrained 0.14 vs. 0.13), so the reader cannot tell what normalization was applied or whether DSR and GSR were even measured separately. This table is load-bearing for the Q2 conclusion about the gap between detection and execution. Define the ratio, report DSR and GSR separately, and show the uncertainty in the normalized values.
  3. [Section IV.C, Fig. 8] The real-world validation is described as 'four representative tasks for each category' (which is also ambiguous — Fig. 7 appears to show one paired task per category, not four per category), with 20 trials per task and no error bars or statistical comparison. The claim that 'real-world performance closely aligns with simulation results' is therefore not quantitatively supported. Please state the number of real-world tasks, report per-task results with confidence intervals, and provide a sim-to-real agreement measure (e.g., correlation, mean absolute error, or a paired test) rather than visual inspection alone.
minor comments (6)
  1. [Abstract] Typo: 'graspingwith' should be 'grasping with'.
  2. [Section III.D, Eq. (2)] The 'shortest path length' l* is not defined operationally. Clarify whether it is a Euclidean distance, a collision-free path length from a specific planner, or something else; this affects the interpretability of SPL.
  3. [Section III.C] The phrase 'we inherited annotated objects from the MultiGripperGrasp dataset' is odd; 'adopted' or 'used' would be clearer.
  4. [Fig. 3] The example instruction 'Grasp a can from basket and put on right' appears to be missing 'side'.
  5. [Section IV.C] Minor language issue: 'Clutter remain challenging' should be 'Clutter remains challenging' or 'Cluttered scenes remain challenging.'
  6. [Section IV.A] The dataset description reports 5,000 simulation and 800 real trajectories but does not state how many trajectories per task or how many distinct objects/scenes are covered; a breakdown would help assess diversity.

Circularity Check

0 steps flagged

No significant circularity: GCA-Bench is an empirical benchmark whose claims are direct measurements, not derivations from fitted inputs or self-citations.

full rationale

GCA-Bench is presented as a new benchmark, not as a derived theoretical result. The 102 task scenarios, four difficulty categories, and three instruction levels are newly constructed; the evaluation metrics (DSR, GSR, TSR, SPL, ET) are straightforward counting and ratio definitions with no fitted parameters or hidden reductions. The paper's central quantitative claims — that all baselines score below 70% TSR on complex scenarios and that performance degrades from simple to cluttered/constrained/semantic tasks — are direct measurements reported in Tables II–IV, not quantities implied by the benchmark's definitions. Among the baselines, several prior works by the same author group appear (e.g., Grasp-Anything [13], language-driven grasp detection [20,26,29], GraspMAS [48]), but they are used either as examples of the existing pose-detection paradigm or as benchmark contestants; none of them supplies a load-bearing inference that makes the conclusions true by construction. The fine-tuning of VLA models on GCA-Bench data followed by evaluation on GCA-Bench is in-sample benchmarking, but success rates are empirically observed, not analytically forced by the training set, and the real-world validation in Section IV.C provides a separate, external evaluation on physical robots. The only substantive concern — that the headline 'below 70%' result is based on 5 sampled scenarios per category (Section IV.B) — is a representativeness/statistical-validity issue, not a circularity issue. No equation in the paper reduces a predicted quantity to an input or fit, and no uniqueness theorem or ansatz is imported via self-citation. Therefore the paper exhibits no significant circularity.

Axiom & Free-Parameter Ledger

3 free parameters · 5 axioms · 0 invented entities

The benchmark introduces no physical entities. The central claim rests on hand-chosen evaluation design: scenario subset, trial counts, training steps, sim-to-real transfer, LLM instruction validity, and the SPL metric's path-length oracle. None of these are independently verified beyond the paper's own experiments.

free parameters (3)
  • Evaluation scenario sample size = 5 scenarios per category (20 of 102 total)
    Hand-picked subset used for all headline results; no protocol for representativeness; may determine the ranking.
  • Trials per scenario = 10 simulation / 20 real-world
    No power analysis; no confidence intervals; differences among methods may be within noise.
  • VLA fine-tuning steps = 60k LoRA steps on 4 A100s
    Set uniformly across VLA baselines; no ablation showing convergence or fairness across methods.
axioms (5)
  • domain assumption Isaac Lab simulation with MultiGripperGrasp/YCB object assets and Franka gripper approximates real-world complex grasping well enough for benchmark validity.
    Section III-C and IV-C; the sim-to-real transfer claim is central to the benchmark's value and is validated only partially.
  • domain assumption LLM-generated, manually reviewed instructions at three complexity levels are valid natural-language task constraints.
    Section III-B; no human evaluation of instruction difficulty or inter-annotator agreement is reported.
  • domain assumption Manual teleoperation trajectories are appropriate ground truth for training and evaluating policies.
    Section IV-A; no validation that demonstrations are optimal or consistent across operators.
  • ad hoc to paper 5 scenarios per category are representative of the 102-task benchmark.
    Section IV.B; no selection criteria or difficulty calibration is given.
  • domain assumption The SPL shortest-path length l* can be computed for each scene.
    Section III-D; the method for computing l* is unspecified.

pith-pipeline@v1.3.0-alltime-deepseek · 11446 in / 15023 out tokens · 136614 ms · 2026-08-02T02:21:55.984518+00:00 · methodology

0 comments
read the original abstract

Robust robotic grasping remains a fundamental challenge for complex real-world applications. Recent advances in large-scale models demonstrate promising capabilities for reasoning in robotic tasks. However, existing benchmarks for grasping primarily focus on isolated, visual-based grasp pose detection, failing to capture the complexity of grasping tasks that require multi-step reasoning and semantic understanding during execution. To address this gap, we propose GCA-Bench, a benchmark featuring challenging \textit{grasping with complex action} scenarios that involve both scene-level reasoning and semantic constraints. GCA-Bench enables the evaluation of recent large foundation models under the same settings. To demonstrate the effectiveness of our new benchmark, we implement a diverse set of baselines, ranging from traditional grasp detection pipelines to end-to-end learning methods. Empirical studies achieve success rates below 70\% on complex grasping scenarios, underscoring critical limitations. In addition, we propose new evaluation metrics, analyze critical failure models, and provide insights to guide the development of more robust and generalizable grasping strategies.

Figures

Figures reproduced from arXiv: 2607.14341 by Anh Nguyen, Baoru Huang, Basu Hela, Charith Munasinghe, Hans Wernher van de Venn, Hanyi Zhang, Hoan Nguyen, Khang Nguyen, Ravi Prakash, Tianyu Li, Tung D. Ta, Yalin Zheng, Zihong Luo.

Figure 1
Figure 1. Figure 1: We propose GCA-Bench, a benchmark for evaluating complex robotic grasping skills as a multi-stage process. [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overview of scenario design in GCA-Bench. The benchmark comprises four categories of manipulation tasks: ( [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Tasks are designed to evaluate both semantic and scene understanding. Execution is assessed along a pipeline [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Simulation setup in IsaacLab with four cameras: [PITH_FULL_IMAGE:figures/full_fig_p004_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Real-world data collection setup with three robotic [PITH_FULL_IMAGE:figures/full_fig_p005_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Action complexity from our collected data. Complex cluttered scenes cannot be solved by visual detection alone and [PITH_FULL_IMAGE:figures/full_fig_p006_6.png] view at source ↗
Figure 8
Figure 8. Figure 8: Real-world performance with simulation baseline. [PITH_FULL_IMAGE:figures/full_fig_p007_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Failure cases. Constrained tasks benefit from collision-aware behaviors learned in simulation. Semantic tasks achieve the highest success rate, indicating that diverse language instructions in our dataset improve instruction-following in real-world settings. Notably, real-world performance closely aligns with simulation results, suggesting that GCA-Bench provides a re￾liable and practical benchmark suite f… 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

51 extracted references · 14 linked inside Pith

  1. [1]

    Learning-based robotic grasping: A review,

    Z. Xie, X. Liang, and C. Roberto, “Learning-based robotic grasping: A review,”Frontiers in Robotics and AI, 2023

    2023

  2. [2]

    A review of robotic grasp detection technol- ogy,

    M. Dong and J. Zhang, “A review of robotic grasp detection technol- ogy,”Robotica, 2023

    2023

  3. [3]

    Vision-based robotic grasping from object localization, object pose estimation to grasp estimation for parallel grippers: a review,

    G. Du, K. Wang, S. Lian, and K. Zhao, “Vision-based robotic grasping from object localization, object pose estimation to grasp estimation for parallel grippers: a review,”Artificial Intelligence Review, 2021

    2021

  4. [4]

    Fetchbench: A simulation benchmark for robot fetching,

    B. Han, M. Parakh, D. Geng, J. A. Defay, G. Luyang, and J. Deng, “Fetchbench: A simulation benchmark for robot fetching,” arXiv:2406.11793, 2024

    Pith/arXiv arXiv 2024

  5. [5]

    Deep learning approaches to grasp synthesis: A review,

    R. Newbury, M. Gu, L. Chumbley, A. Mousavian, C. Eppner, J. Leit- ner, J. Bohg, A. Morales, T. Asfour, D. Kragic,et al., “Deep learning approaches to grasp synthesis: A review,”IEEE T-RO, 2023

    2023

  6. [6]

    Technological development and optimization of pushing and grasping functions in robot arms: A review,

    A. Efendi, Y .-H. Shao, and C.-Y . Huang, “Technological development and optimization of pushing and grasping functions in robot arms: A review,”Measurement, 2025

    2025

  7. [7]

    Rlbench: The robot learning benchmark & learning environment,

    S. James, Z. Ma, D. R. Arrojo, and A. J. Davison, “Rlbench: The robot learning benchmark & learning environment,”IEEE RA-L, 2020

    2020

  8. [8]

    Robocasa: Large-scale simulation of everyday tasks for generalist robots,

    S. Nasiriany, A. Maddukuri, L. Zhang, A. Parikh, A. Lo, A. Joshi, A. Mandlekar, and Y . Zhu, “Robocasa: Large-scale simulation of everyday tasks for generalist robots,”arXiv:2406.02523, 2024

    Pith/arXiv arXiv 2024

  9. [9]

    Libero: Benchmarking knowledge transfer for lifelong robot learn- ing,

    B. Liu, Y . Zhu, C. Gao, Y . Feng, Q. Liu, Y . Zhu, and P. Stone, “Libero: Benchmarking knowledge transfer for lifelong robot learn- ing,”NeurIPS, 2023

    2023

  10. [10]

    Vlabench: A large-scale benchmark for language- conditioned robotics manipulation with long-horizon reasoning tasks,

    S. Zhang, Z. Xu, P. Liu, X. Yu, Y . Li, Q. Gao, Z. Fei, Z. Yin, Z. Wu, Y .-G. Jiang,et al., “Vlabench: A large-scale benchmark for language- conditioned robotics manipulation with long-horizon reasoning tasks,” arXiv:2412.18194, 2024

    Pith/arXiv arXiv 2024

  11. [11]

    The colosseum: A benchmark for evaluating generalization for robotic manipulation,

    W. Pumacay, I. Singh, J. Duan, R. Krishna, J. Thomason, and D. Fox, “The colosseum: A benchmark for evaluating generalization for robotic manipulation,”arXiv:2402.08191, 2024

    Pith/arXiv arXiv 2024

  12. [12]

    Graspnet-1billion: A large- scale benchmark for general object grasping,

    H.-S. Fang, C. Wang, M. Gou, and C. Lu, “Graspnet-1billion: A large- scale benchmark for general object grasping,” inCVPR, 2020

    2020

  13. [13]

    Grasp-anything: Large-scale grasp dataset from foundation models,

    A. D. Vuong, M. N. Vu, H. Le, B. Huang, H. T. T. Binh, T. V o, A. Kugi, and A. Nguyen, “Grasp-anything: Large-scale grasp dataset from foundation models,” inICRA, 2024

    2024

  14. [14]

    Fmb: a functional manipulation benchmark for generalizable robotic learning,

    J. Luo, C. Xu, F. Liu, L. Tan, Z. Lin, J. Wu, P. Abbeel, and S. Levine, “Fmb: a functional manipulation benchmark for generalizable robotic learning,”IJRR, 2025

    2025

  15. [15]

    Graspvla: a grasping foundation model pre-trained on billion-scale synthetic action data,

    S. Deng, M. Yan, S. Wei, H. Ma,et al., “Graspvla: a grasping foundation model pre-trained on billion-scale synthetic action data,” arXiv:2505.03233, 2025

    Pith/arXiv arXiv 2025

  16. [16]

    Suctionnet-1billion: A large- scale benchmark for suction grasping,

    H. Cao, H.-S. Fang, W. Liu, and C. Lu, “Suctionnet-1billion: A large- scale benchmark for suction grasping,”IEEE Robotics and Automation Letters, 2021

    2021

  17. [17]

    Multi-fingan: Generative coarse-to- fine sampling of multi-finger grasps,

    J. Lundell, E. Corona, T. N. Le, F. Verdoja, P. Weinzaepfel, G. Rogez, F. Moreno-Noguer, and V . Kyrki, “Multi-fingan: Generative coarse-to- fine sampling of multi-finger grasps,” inICRA, 2021

    2021

  18. [18]

    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,” inICRA, 2021

    2021

  19. [19]

    Jacquard: A large scale dataset for robotic grasp detection,

    A. Depierre, E. Dellandr ´ea, and L. Chen, “Jacquard: A large scale dataset for robotic grasp detection,” inIROS, 2018

    2018

  20. [20]

    Language-driven grasp detection,

    A. D. Vuong, M. N. Vu, B. Huang, N. Nguyen, H. Le, T. V o, and A. Nguyen, “Language-driven grasp detection,” inCVPR, 2024

    2024

  21. [21]

    Open-vocabulary affordance detection in 3d point clouds,

    T. Ngyen, M. N. Vu, A. Vuong, D. Nguyen, T. V o, N. Le, and A. Nguyen, “Open-vocabulary affordance detection in 3d point clouds,” inIROS, 2023

    2023

  22. [22]

    Open- world object manipulation using pre-trained vision-language models,

    A. Stone, T. Xiao, Y . Lu, K. Gopalakrishnan, K.-H. Lee, Q. Vuong, P. Wohlhart, S. Kirmani, B. Zitkovich, F. Xia,et al., “Open- world object manipulation using pre-trained vision-language models,” arXiv:2303.00905, 2023

    Pith/arXiv arXiv 2023

  23. [23]

    Review of deep reinforcement learning-based object grasping: Techniques, open challenges, and recommendations,

    M. Q. Mohammed, K. L. Chung, and C. S. Chyi, “Review of deep reinforcement learning-based object grasping: Techniques, open challenges, and recommendations,”IEEE Access, 2020

    2020

  24. [24]

    Learning for a robot: Deep reinforcement learning, imitation learning, transfer learning,

    J. Hua, L. Zeng, G. Li, and Z. Ju, “Learning for a robot: Deep reinforcement learning, imitation learning, transfer learning,”Sensors, 2021

    2021

  25. [25]

    Ocrtoc: A cloud-based competition and benchmark for robotic grasping and manipulation,

    Z. Liu, W. Liu, Y . Qin, F. Xiang,et al., “Ocrtoc: A cloud-based competition and benchmark for robotic grasping and manipulation,” IEEE RA-L, 2021

    2021

  26. [26]

    Lightweight language-driven grasp detection using con- ditional consistency model,

    N. Nguyen, M. N. Vu, B. Huang, A. Vuong, N. Le, T. V o, and A. Nguyen, “Lightweight language-driven grasp detection using con- ditional consistency model,” inIROS, 2024

    2024

  27. [27]

    π 0: A vision-language- action flow model for general robot control

    K. Black, N. Brown, D. Driess, A. Esmail, M. Equi, C. Finn, N. Fusai, L. Groom, K. Hausman, B. Ichter,et al., “π 0: A vision-language- action flow model for general robot control.”arXiv.2410.24164

    Pith/arXiv arXiv

  28. [28]

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

    H.-S. Fang, C. Wang, H. Fang, M. Gou, J. Liu, H. Yan, W. Liu, Y . Xie, and C. Lu, “Anygrasp: Robust and efficient grasp perception in spatial and temporal domains,”IEEE T-RO, 2023

    2023

  29. [29]

    Language-driven 6-dof grasp detection using negative prompt guidance,

    T. Nguyen, M. N. Vu, B. Huang, A. Vuong, Q. Vuong, N. Le, T. V o, and A. Nguyen, “Language-driven 6-dof grasp detection using negative prompt guidance,” inECCV, 2024

    2024

  30. [30]

    Curobo: Parallelized collision-free robot motion generation,

    B. Sundaralingamet al., “Curobo: Parallelized collision-free robot motion generation,” inICRA, 2023

    2023

  31. [31]

    Sage: Bridging semantic and actionable parts for generalizable manipulation of articulated objects,

    H. Geng, S. Wei, C. Deng, B. Shen, H. Wang, and L. Guibas, “Sage: Bridging semantic and actionable parts for generalizable manipulation of articulated objects,”arXiv:2312.01307, 2023

    Pith/arXiv arXiv 2023

  32. [32]

    Efficient end-to-end detection of 6-dof grasps for robotic bin picking,

    Y . Liu, A. Qualmann, Z. Yu, M. Gabriel, P. Schillinger, M. Spies, N. A. Vien, and A. Geiger, “Efficient end-to-end detection of 6-dof grasps for robotic bin picking,” inICRA, 2024

    2024

  33. [33]

    Semgrasp: Semantic grasp generation via language aligned discretization,

    K. Li, J. Wang, L. Yang, C. Lu, and B. Dai, “Semgrasp: Semantic grasp generation via language aligned discretization,” inECCV, 2024

    2024

  34. [34]

    An end-to-end multi- dimensional perception network architecture for robotic grasp detec- tion with target edge collision-aware strategy,

    H. Gao, J. Zhao, Y . Yang, and C. Sun, “An end-to-end multi- dimensional perception network architecture for robotic grasp detec- tion with target edge collision-aware strategy,”IEEE T-ASE, 2025

    2025

  35. [35]

    Simulated and real robotic reach, grasp, and pick-and-place using combined reinforcement learning and traditional controls,

    A. Lobbezoo and H.-J. Kwon, “Simulated and real robotic reach, grasp, and pick-and-place using combined reinforcement learning and traditional controls,”Robotics, 2023

    2023

  36. [36]

    Review of reinforcement learning for robotic grasping: Analysis and recommendations,

    H. Sekkat, O. Moutik, L. Ourabah, B. ElKari, Y . Chaibi, and T. Ait Tchakoucht, “Review of reinforcement learning for robotic grasping: Analysis and recommendations,”SOIC, 2024

    2024

  37. [37]

    A novel robotic grasping method for moving objects based on multi- agent deep reinforcement learning,

    Y . Huang, D. Liu, Z. Liu, K. Wang, Q. Wang, and J. Tan, “A novel robotic grasping method for moving objects based on multi- agent deep reinforcement learning,”Robotics and Computer-Integrated Manufacturing, 2024

    2024

  38. [38]

    Robot grasp planning: A learning from demonstration-based approach,

    K. Wang, Y . Fan, and I. Sakuma, “Robot grasp planning: A learning from demonstration-based approach,”Sensors, 2024

    2024

  39. [39]

    Grasping in the wild: Learning 6dof closed-loop grasping from low-cost demonstrations,

    S. Song, A. Zeng, J. Lee, and T. Funkhouser, “Grasping in the wild: Learning 6dof closed-loop grasping from low-cost demonstrations,” IEEE RA-L, 2020

    2020

  40. [40]

    Dexgraspvla: A vision-language-action framework towards general dexterous grasping,

    Y . Zhong, X. Huang, R. Li, C. Zhang, Y . Liang, Y . Yang, and Y . Chen, “Dexgraspvla: A vision-language-action framework towards general dexterous grasping,”arXiv:2502.20900, 2025

    arXiv 2025

  41. [41]

    Graspgen: A diffusion-based framework for 6-dof grasping with on- generator training,

    A. Murali, B. Sundaralingam, Y .-W. Chao, W. Yuan, J. Yamada, M. Carlson, F. Ramos, S. Birchfield, D. Fox, and C. Eppner, “Graspgen: A diffusion-based framework for 6-dof grasping with on- generator training,”arXiv:2507.13097, 2025

    Pith/arXiv arXiv 2025

  42. [42]

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

    Y . Zheng, X. Chen, Y . Zheng, S. Gu, R. Yang, B. Jin, P. Li, C. Zhong, Z. Wang, L. Liu,et al., “Gaussiangrasper: 3d language gaussian splatting for open-vocabulary robotic grasping,”IEEE R-AL, 2024

    2024

  43. [43]

    Benchmarking in manipulation research: The ycb object and model set and benchmarking protocols,

    B. Calli, A. Walsman, A. Singh, S. Srinivasa, P. Abbeel, and A. M. Dollar, “Benchmarking in manipulation research: The ycb object and model set and benchmarking protocols,”arXiv preprint arXiv:1502.03143, 2015

    Pith/arXiv arXiv 2015

  44. [44]

    Introducing ChatGPT,

    OpenAI, “Introducing ChatGPT,” Software, accessed: July 6th 2023. [Online]. Available: https://openai.com/blog/chatgpt/

    2023

  45. [45]

    Orbit: A unified simulation framework for interactive robot learning environments,

    M. Mittal, C. Yu,et al., “Orbit: A unified simulation framework for interactive robot learning environments,”IEEE RA-L, 2023

    2023

  46. [46]

    Multi- grippergrasp: A dataset for robotic grasping from parallel jaw grippers to dexterous hands,

    L. F. Casas, N. Khargonkar, B. Prabhakaran, and Y . Xiang, “Multi- grippergrasp: A dataset for robotic grasping from parallel jaw grippers to dexterous hands,” inIROS, 2024

    2024

  47. [47]

    Droid: A large-scale in-the-wild robot manipulation dataset,

    A. Khazatsky, K. Pertsch, S. Nair, A. Balakrishna, S. Dasari, S. Karam- cheti, S. Nasiriany, M. K. Srirama, L. Y . Chen, K. Ellis,et al., “Droid: A large-scale in-the-wild robot manipulation dataset,”arXiv preprint arXiv:2403.12945, 2024

    Pith/arXiv arXiv 2024

  48. [48]

    Graspmas: Zero-shot language-driven grasp detection with multi-agent system,

    Q. Nguyen, T. Le, H. Nguyen, T. V o, T. D. Ta, B. Huang, M. N. Vu, and A. Nguyen, “Graspmas: Zero-shot language-driven grasp detection with multi-agent system,” inIROS, 2025

    2025

  49. [49]

    Openvla: An open- source vision-language-action model,

    M. J. Kim, K. Pertsch, S. Karamcheti, T. Xiao, A. Balakrishna, S. Nair, R. Rafailov, E. Foster, G. Lam, P. Sanketi,et al., “Openvla: An open- source vision-language-action model,”arXiv:2406.09246, 2024

    Pith/arXiv arXiv 2024

  50. [50]

    Fine-tuning vision-language- action models: Optimizing speed and success,

    M. J. Kim, C. Finn, and P. Liang, “Fine-tuning vision-language- action models: Optimizing speed and success,”arXiv preprint arXiv:2502.19645, 2025

    Pith/arXiv arXiv 2025

  51. [51]

    π 0.5 : a vision-language-action model with open-world generalization,

    P. Intelligence, K. Black, N. Brown, J. Darpinian, K. Dhabalia, D. Driess, A. Esmail, M. Equi, C. Finn, N. Fusai,et al., “π 0.5 : a vision-language-action model with open-world generalization,”arXiv preprint arXiv:2504.16054, 2025

    Pith/arXiv arXiv 2025