GraspBench: Execution-Aware Cross-Robot Evaluation of 6-DoF Grasp Generation in MuJoCo

GraspBench: Execution-Aware Cross-Robot Evaluation of 6-DoF Grasp Generation in MuJoCo

Abdul Basit, Bibek Poudel and Muhammad Shafique present GraspBench, a MuJoCo protocol for testing whether generated grasps can be executed by different robots. They argue that a grasp proposal’s score or volume alone cannot show whether a particular robot can carry it out. The protocol evaluates complete configured pipelines after robot-specific inverse-kinematics filtering, motion planning and a fixed approach-close-lift execution. Across six pipelines, four robot embodiments and five diagnostic objects, GraspGen has the highest reported overall success rate at 73.1%, followed by EconomicGrasp at 63.5%; GPG produces the largest native candidate sets but achieves 44.3% success. The authors report substantial variation by robot and object, making the central point that measured performance depends on the generator, robot embodiment, scene and execution protocol together. These results offer a way to compare configured grasping systems under shared scenarios, rather than treating proposal metrics as a stand-in for execution. The authors limit their findings to fixed scenarios and say their intervals do not estimate uncertainty from repeated executions; the results therefore do not establish intrinsic model quality or stochastic repeatability.

Publication: 09908658-d60e-4612-8f01-b4013de81cdc

attribution — pass: The reading correctly credits Abdul Basit, Bibek Poudel and Muhammad Shafique and frames the benchmark results and conclusions as the authors' reported findings and argument.

Authors:Abdul Basit, Bibek Poudel, Muhammad Shafique

summary — pass: The reading leads with the actor and the argument that proposal scores or volume alone do not show robot executability. It gives the numbers and preserves the authors' scope limits. Central terms are kept exactly; tooltip coverage is suggested for MuJoCo, 6-DoF, inverse-kinematics and embodiment. For example, a MuJoCo tooltip could identify it as a physics simulator, so readers do not assume the tests used physical robots. These are tooltip suggestions, not fidelity failures. The question follows from the stated limitation on repeated executions.

proposal quality alone does not determine whether a grasp is executable by a particular robot

date — pass: The originalPublished date of 2026-11-05 matches the arXiv v1 submission date. The later feed timestamp is not used.

[Submitted on 5 Nov 2026]

support — pass: The claims, figures (73.1%, 63.5%, 44.3%), pipeline and scenario counts, and limitations all match the abstract. The statement that the results offer a comparison under shared scenarios reflects the protocol's described purpose and does not overstate real-world reliability. The reading adds no invented benefits or costs.

The results characterize complete configured pipelines on fixed scenarios rather than intrinsic model quality or stochastic repeatability

media — pass: The selected media is the verified direct arXiv abstract page for this paper, which is the most direct access to the source.