Grasp benchmark · Suction
Dex-Net 3.0
suctionseal modelGQ-CNNsynthetic training
Dex-Net 3.0 extended the GQ-CNN approach to suction grasping by introducing a quasi-static analytic model of the suction-cup seal: it predicts whether a seal will form on the local surface and whether that seal can resist the wrench imposed by gravity and dynamics.
Its three-tier object split — basic, typical, adversarial — exposed how sharply suction reliability falls as surfaces become porous, curved, or non-flat.
Primary source →At a glance
- Source
- Mahler et al., ICRA 2018
- Year
- 2018
- Scale
- 2.8M point clouds · 1,500 models · analytic suction-seal labels
- Gripper
- Suction
- Modality
- Depth · point cloud
- Best-known
- HW: 98% basic · 82% typical · 58% adversarial
Key results
- HW: 98% success on basic objects
- HW: 82% on typical objects · 58% on adversarial objects
- Introduced an analytic suction-seal + wrench-resistance model
SIM = simulation result · HW = physical hardware. Image-wise accuracy is detection quality, not real-robot pick success. Figures cited from Mahler et al., ICRA 2018.