Robotics · Grasping · Cornell GraspBenchmark detail
Grasp benchmark · Parallel-jaw

Cornell Grasp

parallel-jawplanar graspclassicimage-wise

The Cornell Grasp Dataset is the classic planar-grasp benchmark: 885 RGB-D images of 240 graspable objects annotated with 8,019 ground-truth grasp rectangles. It defined the 5-parameter planar grasp representation (x, y, θ, width, height).

It is effectively saturated — modern detectors report ~99% image-wise accuracy — which is precisely why it should not be read as a measure of real-robot reliability.

Primary source
At a glance
Source
Lenz et al., IJRR / RSS 2013–15
Year
2011–13
Scale
885 RGB-D images · 240 objects · 8,019 labeled grasp rectangles
Gripper
Parallel-jaw
Modality
RGB-D
Best-known
~99% image-wise accuracy — saturated benchmark
Key results
  • ~99% image-wise accuracy (saturated)
  • Defined the planar grasp-rectangle representation
  • Detection accuracy, not physical pick success

SIM = simulation result · HW = physical hardware. Image-wise accuracy is detection quality, not real-robot pick success. Figures cited from Lenz et al., IJRR / RSS 2013–15.

Related benchmarks

← Back to the grasping register
Parallel-jaw

GraspNet-1Billion

De-facto clutter benchmark · AnyGrasp current SOTA (AP)

Parallel-jaw

Dex-Net 2.0

HW: 93% on adversarial · 99% precision on 40 novel objects (YuMi)

Parallel-jaw

Grasp-Anything

Language-driven grasp synthesis · open-vocabulary scenes

Parallel-jaw

Jacquard

~95% image-wise (GR-ConvNet-class)