What do you want to do?¶
-
Run a method
A published detector from its released weights: first grasps in minutes.
-
Swap a component
One line swaps an encoder, a grouper or a loss; compose checks it first.
-
Train and evaluate
One evaluator per dataset, collision filter on and off, every number with its record.
-
Your camera
A depth image and its intrinsics; HGGD also reads the colour image.
-
Add a component
Register your own encoder, head or detector without forking.
-
See the evidence
Where each option runs and learns, under which method, and why not elsewhere.
Methods and datasets¶
Seed-based skeleton GraspNet-baseline CVPR 2020 GSNet ICCV 2021 Scale-Balanced-Grasp CoRL 2022 EconomicGrasp ECCV 2024 FineGrasp arXiv 2025
Own family HGGD RA-L 2023
Datasets GraspNet-1Billion GraspClutter6D
How a swap goes¶
Compose
grasppanda compose writes a checked config, or says why the swap cannot work.
Probe
grasppanda probe trains it briefly on real frames and gives one verdict.
Train
grasppanda train and test give its accuracy, with the collision filter on and off.