Jacquard
Jacquard is a large synthetic grasp dataset built in simulation: over 50,000 images of ~11,000 objects annotated with roughly 1.1 million grasp candidates, each verified by a simulated grasp trial.
It was created to overcome the small scale of the Cornell Grasp Dataset, and introduced a simulated-grasp-trial evaluation (SGT) alongside the standard rectangle metric.
Primary source →- Source
- Depierre et al., IROS 2018
- Year
- 2018
- Scale
- 50,000+ images · ~11,000 objects · ~1.1M successful grasps
- Gripper
- Parallel-jaw
- Modality
- RGB-D (synthetic trials)
- Best-known
- ~95% image-wise (GR-ConvNet-class)
- ~95% image-wise detection accuracy for GR-ConvNet-class models
- Introduced the simulated grasp-trial (SGT) evaluation
- Much larger and more diverse than Cornell
SIM = simulation result · HW = physical hardware. Image-wise accuracy is detection quality, not real-robot pick success. Figures cited from Depierre et al., IROS 2018.
Related benchmarks
← Back to the grasping registerGraspNet-1Billion →
De-facto clutter benchmark · AnyGrasp current SOTA (AP)
Dex-Net 2.0 →
HW: 93% on adversarial · 99% precision on 40 novel objects (YuMi)
Grasp-Anything →
Language-driven grasp synthesis · open-vocabulary scenes
Cornell Grasp →
~99% image-wise accuracy — saturated benchmark