Dex-Net 4.0
Dex-Net 4.0 trains an "ambidextrous" grasping policy that learns when to use a parallel jaw and when to use suction, choosing per object to maximize reliability across a heaped bin. It is the headline industrial result of the Dex-Net program at UC Berkeley.
It learns from over five million synthetic grasps across 1,664 objects in simulated heaps, with domain-randomized depth rendering to bridge the sim-to-real gap.
Primary source →- Source
- Mahler et al., Science Robotics 2019
- Year
- 2019
- Scale
- 5M+ synthetic grasps · 1,664 objects in simulated heaps
- Gripper
- Ambidextrous (jaw + suction)
- Modality
- Depth
- Best-known
- HW: 95% reliability · 300 MPPH (ABB YuMi)
- HW: 95% grasp reliability on novel objects (ABB YuMi)
- HW: ~300 mean picks per hour (MPPH)
- Commercialized as Ambidextrous Robotics → Ambi Robotics
SIM = simulation result · HW = physical hardware. Image-wise accuracy is detection quality, not real-robot pick success. Figures cited from Mahler et al., Science Robotics 2019.
Related benchmarks
← Back to the grasping registerGraspNet-1Billion →
De-facto clutter benchmark · AnyGrasp current SOTA (AP)
SuctionNet-1Billion →
HW: 80.65% grasp success · 100% object clearance (their method)
Dex-Net 2.0 →
HW: 93% on adversarial · 99% precision on 40 novel objects (YuMi)
Dex-Net 3.0 →
HW: 98% basic · 82% typical · 58% adversarial