Robotics · Grasping · SuctionNet-1BillionBenchmark detail
Grasp benchmark · Suction

SuctionNet-1Billion

suctionclutterseal modelreal RGB-D

SuctionNet-1Billion is the suction counterpart to GraspNet-1Billion, built on the same 190 cluttered scenes and 88 objects. It contributes a physically grounded suction model that analytically evaluates seal formation and wrench resistance at billion scale.

It provides a dedicated metric for suction grasping and a baseline network, filling the gap that most grasp benchmarks ignore suction entirely despite its dominance in real warehouses.

Primary source
At a glance
Source
Cao et al., RA-L 2021
Year
2021
Scale
190 scenes · 88 objects · 97,280 images · ~1.1B suction annotations
Gripper
Suction
Modality
RGB-D
Best-known
HW: 80.65% grasp success · 100% object clearance (their method)
Key results
  • HW: 80.65% grasp success rate (their baseline method)
  • HW: 100% object clearance on the evaluated scenes
  • Analytic seal-formation + wrench-resistance scoring

SIM = simulation result · HW = physical hardware. Image-wise accuracy is detection quality, not real-robot pick success. Figures cited from Cao et al., RA-L 2021.

Related benchmarks

← Back to the grasping register
Suction

Dex-Net 3.0

HW: 98% basic · 82% typical · 58% adversarial