pith. sign in

arxiv: 2606.30474 · v1 · pith:FMX763PTnew · submitted 2026-06-29 · 💻 cs.RO

Grasp-Oriented Non-Prehensile Manipulation via Learning a Graspability Field

Licheng Zhong , Gim Hee Lee This is my paper

Pith reviewed 2026-06-30 05:14 UTC · model grok-4.3

classification 💻 cs.RO
keywords non-prehensile manipulationgraspability fieldreinforcement learningrobotic graspingobject reconfigurationclosed-loop policy
0
0 comments X

The pith

Learning a graspability field lets one reinforcement learning policy reconfigure objects for grasping without predefined poses or external planners.

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

The paper reframes non-prehensile manipulation as the task of optimizing an object-centric graspability objective instead of reaching one fixed pose. It builds a graspability field from synthesized grasps that assigns a scalar score to any object configuration based on how ready it is for a successful grasp. This score supplies a dense reward for reinforcement learning and also serves as the signal to stop manipulation and begin grasping. The outcome is a single closed-loop policy that handles both phases without separate planners or hand-tuned termination rules. Real-robot experiments show the policy succeeds at reconfiguration and that the field's predicted distances track actual grasp success rates.

Core claim

A graspability field constructed from synthesized grasps supplies a scalar measure of how suitable any object configuration is for grasp execution. Training a reinforcement learning policy to maximize this measure produces a single controller that manipulates objects into graspable states and automatically switches to grasping once the field value indicates readiness, eliminating the need for target poses, external planners, or manually specified stopping conditions. The same field values correlate with real-world grasp success.

What carries the argument

The graspability field: a scalar function over object configurations that quantifies grasp feasibility, derived from a set of synthesized grasps and used both as reinforcement learning reward and termination criterion.

If this is right

  • A single policy suffices for the full sequence from non-prehensile reconfiguration to grasp execution.
  • Manipulation terminates automatically when the graspability field reaches a suitable threshold.
  • No separate external planner or manually chosen target pose is required.
  • The learned field values serve as a predictor of real-world grasp success.

Where Pith is reading between the lines

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

  • The same field-based objective could be applied to prepare objects for other actions besides grasping.
  • Replacing discrete pose targets with continuous feasibility fields may simplify other multi-stage robotic tasks.
  • Expanding the set of synthesized grasps to cover more object geometries would test how far the current construction generalizes.

Load-bearing premise

The graspability field built from synthesized grasps accurately reflects real-world grasp feasibility and transfers from simulation to physical robots.

What would settle it

Physical-robot trials in which the policy's graspability values show no correlation with measured grasp success rates would falsify the claim that the field captures usable grasp feasibility.

Figures

Figures reproduced from arXiv: 2606.30474 by Gim Hee Lee, Licheng Zhong.

Figure 1
Figure 1. Figure 1: Direct grasps from a fixed approach can fail due to unfavorable object configu￾rations, even when feasible grasps exist. Instead of requiring a predefined target pose, we learn a graspability field over object states to guide manipulation and trigger the manipulation-to-grasp transition. Once the object enters a graspable region, the robot autonomously transitions from manipulation to grasp execution. prob… view at source ↗
Figure 2
Figure 2. Figure 2: Method achitecture. A teacher–student framework that encodes observations, aggregates temporal information with a recurrent module, and predicts manipulation actions and a graspability distance. The teacher is used only during training, and the robot executes a grasp once the predicted graspability indicates a feasible configuration. initial state s 0 o , the manipulation policy aims to produce a trajector… view at source ↗
Figure 3
Figure 3. Figure 3: Real-world experimental setup (a) and perception input (b) [PITH_FULL_IMAGE:figures/full_fig_p010_3.png] view at source ↗
Figure 5
Figure 5. Figure 5: Qualitative results of the full manipulation-to-grasp pipeline on real world objects. Each row shows a temporal sequence where the robot manipulates the object to increase graspability and autonomously transitions to grasp execution once a graspable configuration is reached. configuration reaches any graspable reference within the same tolerances; both therefore share the same geometric criterion, with GOM… view at source ↗
Figure 6
Figure 6. Figure 6: Validation of graspability-based transition in real-world experiments. Graspable configurations correspond to lower predicted distances; a single threshold separates graspable from non-graspable states. only consistently unsuccessful object is Snack, which is highly deformable and slippery. For rigid objects such as PaperBox and DustBin, failures arise from two modes: (1) the object slides during manipulat… view at source ↗
Figure 7
Figure 7. Figure 7: Qualitative results of non-prehensile manipulation in simulation. The policy progressively reconfigures objects through interaction to increase graspability before grasp execution. We provide additional qualitative examples of the learned manipulation be￾havior in simulation as shown in [PITH_FULL_IMAGE:figures/full_fig_p018_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Additional qualitative results of the full manipulation-to-grasp pipeline on real world objects. Each row shows a temporal sequence where the robot manipulates the object to increase graspability and autonomously transitions to grasp execution once a graspable configuration is reached. The graspability is defined related to the home pose of the robot, which is the same for all objects. B Implementation and… view at source ↗
read the original abstract

Non-prehensile manipulation is often used as a preparatory step for robotic grasping, yet existing approaches typically require a predefined target object pose. In practice, however, objects admit multiple graspable configurations and the desired pose is not known in advance. We reformulate non-prehensile manipulation for grasping as optimizing an object centric graspability objective rather than reaching a specific pose. We construct a graspable set from synthesized grasps and define a graspability field that measures how suitable an object configuration is for successful grasp execution. The scalar measure provides a dense learning signal for reinforcement learning and determines when to terminate manipulation. This yields a closed-loop manipulation-to-grasp pipeline driven by a single policy. Experiments in simulation and on a real robot show that the policy reliably reconfigures objects into graspable states and transitions to grasping without external planners or manually specified stopping conditions. The predicted graspability distance correlates with real world grasp success, which indicates that the learned representation captures grasp feasibility of object configurations.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

3 major / 2 minor

Summary. The paper proposes reformulating non-prehensile manipulation as optimization of an object-centric graspability objective rather than a target pose. A graspability field is constructed from synthesized grasps to supply a dense RL reward signal and an automatic termination condition, yielding a single closed-loop policy that reconfigures objects into graspable states and transitions to grasping. Experiments in simulation and on a physical robot are reported to show reliable performance and correlation between the predicted graspability distance and real-world grasp success.

Significance. If the generalization claims hold, the work offers a unified objective for preparatory manipulation and grasping that removes the need for separate pose planners or hand-crafted stopping rules, which could simplify pipelines for robotic manipulation in unstructured settings.

major comments (3)
  1. [Abstract] Abstract: the central claim that the graspability field 'captures grasp feasibility of object configurations' and enables reliable real-robot transfer rests on correlation with grasp success, yet no quantitative metrics (e.g., Pearson coefficient, success-rate tables, or failure-case analysis) are supplied to establish that the correlation is not driven by shared simulation artifacts.
  2. [§3] The construction of the graspable set from synthesized grasps (presumably §3) defines the field via simulation-based synthesis; any unmodeled mismatch in friction, contact compliance, or sensor noise creates a domain gap that directly affects both the RL reward and the termination signal, yet the manuscript provides no ablation on domain-randomization strength or sim-to-real transfer metrics.
  3. [§5] Experiments section: the assertion that the policy 'transitions to grasping without external planners or manually specified stopping conditions' is load-bearing for the closed-loop contribution, but no comparison against baselines that use explicit termination or separate grasp planners is reported, leaving the advantage of the single-policy formulation unquantified.
minor comments (2)
  1. [§3] Notation for the graspability scalar and distance should be introduced with a clear equation reference rather than prose description only.
  2. [§5] Figure captions for real-robot trials should list the number of objects, trials per object, and success criteria to allow reproducibility assessment.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the thoughtful comments, which help clarify the strength of our claims on correlation, domain transfer, and the benefits of the unified policy. We respond to each major comment below.

read point-by-point responses

  1. Referee: [Abstract] Abstract: the central claim that the graspability field 'captures grasp feasibility of object configurations' and enables reliable real-robot transfer rests on correlation with grasp success, yet no quantitative metrics (e.g., Pearson coefficient, success-rate tables, or failure-case analysis) are supplied to establish that the correlation is not driven by shared simulation artifacts.

    Authors: We agree that the abstract would benefit from explicit quantitative support. The manuscript body reports real-robot grasp success rates conditioned on graspability values, but we will revise the abstract to include a Pearson correlation coefficient between predicted graspability distance and observed grasp success, plus a table of success rates across binned graspability thresholds in both simulation and on the physical robot. Failure cases will also be briefly analyzed to address potential simulation artifacts. revision: yes

  2. Referee: [§3] The construction of the graspable set from synthesized grasps (presumably §3) defines the field via simulation-based synthesis; any unmodeled mismatch in friction, contact compliance, or sensor noise creates a domain gap that directly affects both the RL reward and the termination signal, yet the manuscript provides no ablation on domain-randomization strength or sim-to-real transfer metrics.

    Authors: The current manuscript demonstrates successful sim-to-real transfer through physical robot experiments, but we concur that dedicated ablations on randomization strength are absent. We will add an ablation varying domain randomization parameters (friction, compliance, noise) and report the resulting sim-to-real success rates and graspability prediction accuracy to quantify robustness to the domain gap. revision: yes

  3. Referee: [§5] Experiments section: the assertion that the policy 'transitions to grasping without external planners or manually specified stopping conditions' is load-bearing for the closed-loop contribution, but no comparison against baselines that use explicit termination or separate grasp planners is reported, leaving the advantage of the single-policy formulation unquantified.

    Authors: The experiments show reliable closed-loop behavior on the real robot using only the learned graspability signal for both reward and termination. However, we acknowledge that the advantage over pipelines with explicit termination rules remains unquantified without direct baselines. Adding such comparisons would require substantial additional implementation and fair metric alignment; we will therefore expand the discussion section to address this limitation and identify it as future work rather than performing new baseline experiments in the revision. revision: partial

Circularity Check

0 steps flagged

No circularity: graspability field constructed independently from synthesized grasps and validated externally

full rationale

The paper constructs a graspable set from synthesized grasps and defines a graspability field as a scalar measure of configuration suitability for grasping. This field supplies a dense RL reward and termination signal, with the resulting policy evaluated in simulation and on physical robots where predicted distances correlate with real grasp success. No derivation step reduces by construction to fitted parameters from the same data, self-citations, or renamed inputs; the chain relies on external synthesis and physical validation rather than self-definition. This is the most common honest outcome for papers whose central quantities are defined upstream of the learning loop and tested against independent benchmarks.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 1 invented entities

Abstract provides no explicit free parameters, axioms, or invented entities beyond the high-level description of the graspability field.

axioms (1)
  • domain assumption Synthesized grasps can define a graspable set that supports learning a generalizable field for real-world use.
    The method relies on this premise to construct the field from simulated data.
invented entities (1)
  • graspability field no independent evidence
    purpose: Scalar measure of how suitable an object configuration is for grasp execution
    New representation introduced to provide dense learning signal and termination condition.

pith-pipeline@v0.9.1-grok · 5702 in / 1206 out tokens · 32754 ms · 2026-06-30T05:14:58.092231+00:00 · methodology

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 · 12 canonical work pages · 3 internal anchors

  1. [1]

    In: CoRL

    Breyer, M., Chung, J.J., Ott, L., Siegwart, R., Nieto, J.: Volumetric grasping net- work: Real-time 6 dof grasp detection in clutter. In: CoRL. pp. 1602–1611. PMLR (2021)

    2021

  2. [2]

    The International Journal of Robotics Research36(3), 261–268 (2017)

    Calli, B., Singh, A., Bruce, J., Walsman, A., Konolige, K., Srinivasa, S., Abbeel, P., Dollar, A.M.: Yale-cmu-berkeley dataset for robotic manipulation research. The International Journal of Robotics Research36(3), 261–268 (2017)

    2017

  3. [3]

    In: ICRA

    Calli, B., Singh, A., Walsman, A., Srinivasa, S., Abbeel, P., Dollar, A.M.: The ycb object and model set: Towards common benchmarks for manipulation research. In: ICRA. pp. 510–517. IEEE (2015)

    2015

  4. [4]

    IEEE Robotics & Automation Magazine22(3), 36–52 (2015)

    Calli, B., Walsman, A., Singh, A., Srinivasa, S., Abbeel, P., Dollar, A.M.: Bench- marking in manipulation research: Using the yale-cmu-berkeley object and model set. IEEE Robotics & Automation Magazine22(3), 36–52 (2015)

    2015

  5. [5]

    In: ICRA

    Cheng, X., Huang, E., Hou, Y., Mason, M.T.: Contact mode guided motion plan- ning for quasidynamic dexterous manipulation in 3d. In: ICRA. pp. 2730–2736. IEEE (2022)

    2022

  6. [6]

    In: Proceedings of the 2014 conference on empirical methods in natural language processing (EMNLP)

    Cho, K., Van Merriënboer, B., Gulçehre, Ç., Bahdanau, D., Bougares, F., Schwenk, H., Bengio, Y.: Learning phrase representations using rnn encoder–decoder for statistical machine translation. In: Proceedings of the 2014 conference on empirical methods in natural language processing (EMNLP). pp. 1724–1734 (2014)

    2014

  7. [7]

    In: ICLR (2024)

    Cho, Y., Han, J., Cho, Y., Kim, B.: Corn: Contact-based object representation for nonprehensile manipulation of general unseen objects. In: ICLR (2024)

    2024

  8. [8]

    In: CoRL

    Deng, S., Yan, M., Wei, S., Ma, H., Yang, Y., Chen, J., Zhang, Z., Yang, T., Zhang, X., Cui, H., et al.: Graspvla: a grasping foundation model pre-trained on billion-scale synthetic action data. In: CoRL. PMLR (2025)

    2025

  9. [9]

    In: IROS

    Ding, K., Chen, B., Wu, R., Li, Y., Zhang, Z., Gao, H.a., Li, S., Zhou, G., Zhu, Y., Dong, H., et al.: Preafford: Universal affordance-based pre-grasping for diverse objects and environments. In: IROS. pp. 7278–7285. IEEE (2024)

    2024

  10. [10]

    In: IROS

    Dogar, M.R., Srinivasa, S.S.: Push-grasping with dexterous hands: Mechanics and a method. In: IROS. pp. 2123–2130. IEEE (2010)

    2010

  11. [11]

    IEEE TRO (2023)

    Fang, H.S., Wang, C., Fang, H., Gou, M., Liu, J., Yan, H., Liu, W., Xie, Y., Lu, C.: Anygrasp: Robust and efficient grasp perception in spatial and temporal domains. IEEE TRO (2023)

    2023

  12. [12]

    arXiv preprint arXiv:1912.13470 (2019)

    Fang, H.S., Wang, C., Gou, M., Lu, C.: Graspnet: a large-scale clustered and densely annotated dataset for object grasping. arXiv preprint arXiv:1912.13470 (2019)

    work page arXiv 1912

  13. [13]

    In: CVPR

    Fang, H.S., Wang, C., Gou, M., Lu, C.: Graspnet-1billion: A large-scale benchmark for general object grasping. In: CVPR. pp. 11444–11453 (2020)

    2020

  14. [14]

    International Journal of Advanced Robotic Systems14(1), 1729881416682706 (2016) 16 L

    Guo, D., Sun, F., Kong, T., Liu, H.: Deep vision networks for real-time robotic grasp detection. International Journal of Advanced Robotic Systems14(1), 1729881416682706 (2016) 16 L. Zhong and G. H. Lee

    2016

  15. [15]

    In: 2021 30th IEEE International Conference on Robot & Human Interactive Communication (RO-MAN)

    Hager, J., Bauer, R., Toussaint, M., Mainprice, J.: Graspme-grasp manifold estima- tor. In: 2021 30th IEEE International Conference on Robot & Human Interactive Communication (RO-MAN). pp. 626–632. IEEE (2021)

    2021

  16. [16]

    IEEE RAL 10(11), 11180–11187 (2025)

    Hu, B., Tian, H., Wang, D., Huang, H., Zhu, X., Walters, R., Platt, R.: Push-grasp policy learning using equivariant models and grasp score optimization. IEEE RAL 10(11), 11180–11187 (2025)

    2025

  17. [17]

    arXiv preprint arXiv:2502.03072 (2025)

    Huang, Y., Davies, T., Yan, J., Chen, X., Tian, Y., Hu, L.: Robograsp: A universal grasping policy for robust robotic control. arXiv preprint arXiv:2502.03072 (2025)

    work page arXiv 2025

  18. [18]

    good robot!

    Hundt, A., Killeen, B., Greene, N., Wu, H., Kwon, H., Paxton, C., Hager, G.D.: “good robot!”: Efficient reinforcement learning for multi-step visual tasks with sim to real transfer. IEEE RAL5(4), 6724–6731 (2020)

    2020

  19. [19]

    In: RSS (2024)

    Jiang, B., Wu, Y., Zhou, W., Paxton, C., Held, D.: Hacman++: Spatially-grounded motion primitives for manipulation. In: RSS (2024)

    2024

  20. [20]

    arXiv preprint arXiv:2104.01542 (2021)

    Jiang, Z., Zhu, Y., Svetlik, M., Fang, K., Zhu, Y.: Synergies between affordance and geometry: 6-dof grasp detection via implicit representations. arXiv preprint arXiv:2104.01542 (2021)

    work page arXiv 2021

  21. [21]

    arXiv preprint arXiv:2502.18015 (2025)

    Jung, H., Lee, D., Park, H., Kim, J., Kim, B.: Spin: distilling skill-rrt for long-horizon prehensile and non-prehensile manipulation. arXiv preprint arXiv:2502.18015 (2025)

    work page arXiv 2025

  22. [22]

    ACM TOG39(6), 1–14 (2020)

    Laine, S., Hellsten, J., Karras, T., Seol, Y., Lehtinen, J., Aila, T.: Modular primi- tives for high-performance differentiable rendering. ACM TOG39(6), 1–14 (2020)

    2020

  23. [23]

    arXiv preprint arXiv:2412.06931 (2024)

    Lee, H.Y., Zhou, P., Duan, A., Ma, W., Yang, C., Navarro-Alarcon, D.: Non-prehensile tool-object manipulation by integrating llm-based planning and manoeuvrability-driven controls. arXiv preprint arXiv:2412.06931 (2024)

    work page arXiv 2024

  24. [24]

    In: 2024 International Conference on Advanced Robotics and Mechatronics (ICARM)

    Li, D., Zhao, C., Yang, S., Ma, L., Li, Y., Zhang, W.: Learning instruction-guided manipulation affordance via large models for embodied robotic tasks. In: 2024 International Conference on Advanced Robotics and Mechatronics (ICARM). pp. 662–667. IEEE (2024)

    2024

  25. [25]

    In: ICCV (2025)

    Li, G., Tsagkas, N., Song, J., Mon-Williams, R., Vijayakumar, S., Shao, K., Sevilla- Lara, L.: Learning precise affordances from egocentric videos for robotic manipu- lation. In: ICCV (2025)

    2025

  26. [26]

    arXiv preprint arXiv:2511.06240 (2025)

    Lin, T.J., Yeh, J.F., Su, H.T., Lin, C.Y., Chen, Y.T., Hsu, W.H.: Affordance- guided coarse-to-fine exploration for base placement in open-vocabulary mobile manipulation. arXiv preprint arXiv:2511.06240 (2025)

    work page arXiv 2025

  27. [27]

    The international journal of robotics research15(6), 533–556 (1996)

    Lynch, K.M., Mason, M.T.: Stable pushing: Mechanics, controllability, and plan- ning. The international journal of robotics research15(6), 533–556 (1996)

    1996

  28. [28]

    In: ICCV (2025)

    Lyu, J., Li, Z., Shi, X., Xu, C., Wang, Y., Wang, H.: Dywa: Dynamics-adaptive world action model for generalizable non-prehensile manipulation. In: ICCV (2025)

    2025

  29. [29]

    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.,Ojea,J.A.,Goldberg, K.: Dex-net 2.0: Deep learning to plan robust grasps with synthetic point clouds and analytic grasp metrics. arXiv preprint arXiv:1703.09312 (2017)

    work page internal anchor Pith review Pith/arXiv arXiv 2017

  30. [30]

    The International Journal of Robotics Research18(11), 1129–1141 (1999)

    Mason, M.T.: Progress in nonprehensile manipulation. The International Journal of Robotics Research18(11), 1129–1141 (1999)

    1999

  31. [31]

    Isaac Lab: A GPU-Accelerated Simulation Framework for Multi-Modal Robot Learning

    Mittal, M., Roth, P., Tigue, J., Richard, A., Zhang, O., et al.: Isaac Lab - A GPU-Accelerated Simulation Framework for Multi-Modal Robot Learning. arXiv preprint arXiv:2511.04831 (2025)

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  32. [32]

    In: Proceedings of the ACM SIGGRAPH/Eurographics symposium on computer animation

    Mordatch, I., Popović, Z., Todorov, E.: Contact-invariant optimization for hand manipulation. In: Proceedings of the ACM SIGGRAPH/Eurographics symposium on computer animation. pp. 137–144 (2012)

    2012

  33. [33]

    ACM TOG31(4), 1–8 (2012) GOMP 17

    Mordatch, I., Todorov, E., Popović, Z.: Discovery of complex behaviors through contact-invariant optimization. ACM TOG31(4), 1–8 (2012) GOMP 17

    2012

  34. [34]

    NVIDIA: Isaac Sim,https://github.com/isaac-sim/IsaacSim, (accessed June 29, 2026)

    2026

  35. [35]

    In: IROS (2020)

    Pohl, C., Hitzler, K., Grimm, R., Zea, A., Hanebeck, U.D., Asfour, T.: Affordance- based grasping and manipulation in real world applications. In: IROS (2020)

    2020

  36. [36]

    arXiv preprint arXiv:2509.20550 (2025)

    Srinivas, S.K., Shukla, Y., Arnold, A., Chitta, S.: Graspfactory: A large object- centric grasping dataset. arXiv preprint arXiv:2509.20550 (2025)

    work page arXiv 2025

  37. [37]

    Sundermeyer, M., Mousavian, A., Triebel, R., Fox, D.: Contact-graspnet: Efficient 6-dofgraspgenerationinclutteredscenes.In:ICRA.pp.13438–13444.IEEE(2021)

    2021

  38. [38]

    In: ICRA

    Urain, J., Funk, N., Peters, J., Chalvatzaki, G.: Se(3)-diffusionfields: Learning smooth cost functions for joint grasp and motion optimization through diffusion. In: ICRA. IEEE (2023)

    2023

  39. [39]

    In: ICCV

    Wang, C., Fang, H.S., Gou, M., Fang, H., Gao, J., Lu, C.: Graspness discovery in clutters for fast and accurate grasp detection. In: ICCV. pp. 15964–15973 (October 2021)

    2021

  40. [40]

    In: ICRA

    Wang, R., Zhang, J., Chen, J., Xu, Y., Li, P., Liu, T., Wang, H.: Dexgraspnet: A large-scale robotic dexterous grasp dataset for general objects based on simulation. In: ICRA. pp. 11359–11366. IEEE (2023)

    2023

  41. [41]

    arXiv preprint arXiv:2503.23120 (2025)

    Wang, Y., Li, Y., Yang, Y., Chen, Y.: Dexterous non-prehensile manipulation for ungraspable object via extrinsic dexterity. arXiv preprint arXiv:2503.23120 (2025)

    work page arXiv 2025

  42. [42]

    In: ICRA

    Weng, T., Held, D., Meier, F., Mukadam, M.: Neural grasp distance fields for robot manipulation. In: ICRA. pp. 1814–1821. IEEE (2023)

    2023

  43. [43]

    IEEE RAL6(4), 6337–6344 (2021)

    Xu, K., Yu, H., Lai, Q., Wang, Y., Xiong, R.: Efficient learning of goal-oriented push-grasping synergy in clutter. IEEE RAL6(4), 6337–6344 (2021)

    2021

  44. [44]

    IEEE RAL9(12), 10842–10849 (2024)

    Xu, R., Shen, Y., Li, X., Wu, R., Dong, H.: Naturalvlm: Leveraging fine-grained natural language for affordance-guided visual manipulation. IEEE RAL9(12), 10842–10849 (2024)

    2024

  45. [45]

    In: CoRL

    Yuan, W., Duan, J., Blukis, V., Pumacay, W., Krishna, R., Murali, A., Mousavian, A., Fox, D.: Robopoint: A vision-language model for spatial affordance prediction in robotics. In: CoRL. PMLR (2025)

    2025

  46. [46]

    In: IROS

    Zeng, A., Song, S., Welker, S., Lee, J., Rodriguez, A., Funkhouser, T.: Learning synergies between pushing and grasping with self-supervised deep reinforcement learning. In: IROS. pp. 4238–4245. IEEE (2018)

    2018

  47. [47]

    In: CoRL

    Zhang,J.,Liu,H.,Li,D.,Yu,X.,Geng,H.,Ding,Y.,Chen,J.,Wang,H.:Dexgrasp- net 2.0: Learning generative dexterous grasping in large-scale synthetic cluttered scenes. In: CoRL. PMLR (2024)

    2024

  48. [48]

    ActivePusher: Active Learning and Planning with Residual Physics for Nonprehensile Manipulation

    Zhong, Z., Golestaneh, S., Chamzas, C.: Activepusher: Active learning and planning with residual physics for nonprehensile manipulation. arXiv preprint arXiv:2506.04646 (2025)

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  49. [49]

    In: CoRL

    Zhou, W., Held, D.: Learning to grasp the ungraspable with emergent extrinsic dexterity. In: CoRL. pp. 150–160. PMLR (2023)

    2023

  50. [50]

    In: CoRL

    Zhou, W., Jiang, B., Yang, F., Paxton, C., Held, D.: Hacman: Learning hybrid actor-critic maps for 6d non-prehensile manipulation. In: CoRL. pp. 241–265. PMLR (2023)

    2023

  51. [51]

    arXiv preprint arXiv:2511.11052 (2025) 18 L

    Zhu, J., Tie, C., Cao, X., Wang, Y., Guo, J., Chen, Z., Chen, H., Chen, J., Xiao, Y., Wu, R., et al.: Adaptpnp: Integrating prehensile and non-prehensile skills for adaptive robotic manipulation. arXiv preprint arXiv:2511.11052 (2025) 18 L. Zhong and G. H. Lee Appendix A Additional Results A.1 Qualitative Manipulation Simulation Results Fig. 7:Qualitative...

    work page arXiv 2025