Grasp benchmark · Parallel-jaw (Franka)
ACRONYM
simulationparallel-jawlarge-scaleNVIDIA
ACRONYM is a large-scale, simulation-only grasp dataset from NVIDIA: 17.7 million parallel-jaw grasps across 8,872 objects in 262 categories, with success labels generated by physics simulation in FleX rather than analytic heuristics.
Its significance is the evidence that physics-based grasp labels transfer to hardware better than purely analytic ones — making it the training source for Contact-GraspNet and a reference for synthetic grasp generation.
Primary source →At a glance
- Source
- Eppner et al., ICRA 2021
- Year
- 2021
- Scale
- 17.7M grasps · 8,872 objects · 262 categories · FleX physics
- Gripper
- Parallel-jaw (Franka)
- Modality
- Simulation-only
- Best-known
- SIM: 59.21% of generated grasps succeed (label generation)
Key results
- SIM: 59.21% of the generated candidate grasps succeed in physics
- Physics labels shown to transfer better than analytic labels
- Training source for Contact-GraspNet
SIM = simulation result · HW = physical hardware. Image-wise accuracy is detection quality, not real-robot pick success. Figures cited from Eppner et al., ICRA 2021.